The present invention relates to methods for monitoring activities of users, and for recommending items to users based on such activities. More specifically, the invention relates to methods for providing personalized recommendations of web sites, web pages and/or products that are relevant to a current browsing session of a user.
A recommendation service is a computer-implemented service that recommends items. The recommendations are customized to particular users based on information known about the users. One common application for recommendation services involves recommending products to online customers. For example, online merchants commonly provide services for recommending products (books, compact discs, videos, etc.) to customers based on profiles that have been developed for such customers. Recommendation services are also common for recommending Web sites or pages, articles, and other types of informational content to users.
One technique commonly used by recommendation services is known as content-based filtering. Pure content-based systems operate by attempting to identify items which, based on an analysis of item content, are similar to items that are known to be of interest to the user. For example, a content-based Web site recommendation service may operate by parsing the user's favorite Web pages to generate a profile of commonly-occurring terms, and then using this profile to search for other Web pages that include some or all of these terms.
Content-based systems have several significant limitations. For example, content-based methods generally do not provide any mechanism for evaluating the quality or popularity of an item. In addition, content-based methods require that the items be analyzed, which may be a compute intensive task.
Another common recommendation technique is known as collaborative filtering. In a pure collaborative system, items are recommended to users based on the interests of a community of users, without any analysis of item content. Collaborative systems commonly operate by having the users explicitly rate individual items from a list of popular items. Some systems, such as those described in instead require users to create lists of their favorite items. See U.S. Pat. Nos. 5,583,763 and 5,749,081. Through this explicit rating or list creating process, each user builds a personal profile of his or her preferences. To generate recommendations for a particular user, the user's profile is compared to the profiles of other users to identify one or more “similar users.” Items that were rated highly by these similar users, but which have not yet been rated by the user, are then recommended to the user. An important benefit of collaborative filtering is that it overcomes the above-noted deficiencies of content-based filtering.
As with content-based filtering methods, however, existing collaborative filtering techniques have several problems. One problem is that users frequently do not take the time to explicitly rate items, or create lists of their favorite items. As a result, the operator of a collaborative recommendation system may be able to provide personalized product recommendations to only a small segment of its users.
Further, even if a user takes the time to set up a profile, the recommendations thereafter provided to the user typically will not take into account the user's short term browsing interests. For example, the recommendations may not be helpful to a user who is venturing into an unfamiliar item category.
Another problem with collaborative filtering techniques is that an item in the database normally cannot be recommended until the item has been rated. As a result, the operator of a new collaborative recommendation system is commonly faced with a “cold start” problem in which the service cannot be brought online in a useful form until a threshold quantity of ratings data has been collected. In addition, even after the service has been brought online, it may take months or years before a significant quantity of the database items can be recommended. Further, as new items are added to the catalog (such as descriptions of newly released products), these new items may not recommendable by the system for a period of time.
Another problem with collaborative filtering methods is that the task of comparing user profiles tends to be time consuming, particularly if the number of users is large (e.g., tens or hundreds of thousands). As a result, a tradeoff tends to exist between response time and breadth of analysis. For example, in a recommendation system that generates real-time recommendations in response to requests from users, it may not be feasible to compare the user's ratings profile to those of all other users. A relatively shallow analysis of the available data (leading to poor recommendations) may therefore be performed.
Another problem with both collaborative and content-based systems is that they generally do not reflect the current preferences of the community of users. In the context of a system that recommends products to customers, for example, there is typically no mechanism for favoring items that are currently “hot items.” In addition, existing systems typically do not provide a mechanism for recognizing that the user may be searching for a particular type or category of item.
These and other problems are addressed by providing computer-implemented methods for automatically identifying items that are related to one another based on the activities of a community of users. Item relationships are determined by identifying and analyzing sequences of items viewed or accessed by users. This process may be repeated periodically (e.g., once per day or once per week) to incorporate the latest browsing activities of the community of users. The resulting item relatedness data may be used to provide personalized item recommendations to users (e.g., web site or web page recommendations), and/or to provide users with non-personalized lists of related items (e.g., lists of related web pages or web sites).
In the description that follows, the word “item” will generally be used to refer to things that are viewed by or accessed by users and which can be recommended to users. In the context of this invention, items can be products, web sites, web pages, and/or web addresses. Items can also be other things, for example, where the viewing, use and/or access of those things by users can be tracked.
The present invention provides methods for recommending items to users without requiring the users to explicitly rate items or create lists of their favorite items. The personal recommendations are preferably generated using item relatedness data determined using the above-mentioned methods, but may be generated using other sources or types of item relatedness data (e.g., item relationships determined using a content-based analysis). In one embodiment (described below), the personalized recommendations are based on the web pages or sites viewed by the customer during a current browsing session, and thus tend to be highly relevant to the user's current browsing purpose.
One aspect of the invention thus involves methods for identifying items that are related to one another. In a preferred embodiment, user actions that evidence users' interests in or affinities for particular items are recorded for subsequent analysis. These item-affinity-evidencing actions may include, for example, the viewing of a web page, and/or the searching for a particular item using a search engine. To identify items that are related or “similar” to one another, an off-line table generation component analyzes the histories of item-affinity-evidencing actions of a community of users (preferably on a periodic basis) to identify correlations between items for which such actions were performed. For example, in one embodiment, user-specific browsing histories are analyzed to identify correlations between items (e.g., web pages A and B are similar because a significant number of those who viewed A also viewed B).
In one embodiment, page viewing histories of users are recorded and analyzed to identify items that tend to be viewed in combination (e.g., pages A and B are similar because a significant number of those who viewed A also viewed B during the same browsing session). This may be accomplished, for example, by maintaining user-specific (and preferably session-specific) histories of web pages viewed by the users. An important benefit to using page viewing histories is that the item relationships identified include relationships between items that are pure substitutes for each other.
In one embodiment, a client program executes in conjunction with a web browser on a user's computer to enable the tracking of page viewing histories across multiple web sites. The client program identifies addresses (e.g., URLs) of web pages and/or web sites accessed by the user and transmits the sequence of identifications through the Internet to a server application executing on a recommendation system. Multiple client programs are preferably used by multiple users, therefore, the recommendation system is preferably able to accumulate sequences of web addresses accessed by multiple users during multiple browsing sessions and across multiple web sites. The sequences of web addresses will be referred to herein as browsing histories, click streams or usage trails. During a sequence of proximately visited addresses, users tend to view web pages with similar content. Click streams provide browsing data identifying adjacently or proximately visited addresses based upon which similar web pages or web sites can be effectively identified.
The results of the above processes are preferably stored in a table that maps items to sets of similar items. For instance, for each reference item, the table may store a list of the N items deemed most closely related to the reference item. The table also preferably stores, for each pair of items, a value indicating the predicted degree of relatedness between the two items. The table is preferably generated periodically using a most recent set of click stream data and/or other types of historical browsing data reflecting users' item interests.
Another aspect of the invention involves methods for using predetermined item relatedness data to provide personalized recommendations to users. To generate recommendations for a user, multiple items “known” to be of interest to the user are initially identified (e.g., items currently in the user's shopping cart). For each item of known interest, a pre-generated table that maps items to sets of related items (preferably generated as described above) is accessed to identify a corresponding set of related items. Related items are then selected from the multiple sets of related items to recommend to the user. The process by which a related item is selected to recommend preferably takes into account both (1) whether that item is included in more than one of the related items sets (i.e., is related to more than one of the “items of known interest”), and (2) the degree of relatedness between the item and each such item of known interest. Because the personalized recommendations are generated using preexisting item-to-item similarity mappings, they can be generated rapidly (e.g., in real time) and efficiently without sacrificing breadth of analysis.
In one implementation, the recommendations are generated by monitoring the pages or sites viewed by the user during the current browsing session, and using these as the “items of known interest.” The resulting list of recommended items (web pages or web sites) is presented to the user during the same browsing session. In one embodiment, these session-specific recommendations are displayed on a customized page. From this page, the user can individually de-select the viewed items used as the “items of known interest,” and then initiate generation of a refined list of recommended items. Because the recommendations are based on the items viewed during the current session, they tend to be closely tailored to the user's current browsing interests. Further, because the recommendations are based on items viewed during the session, recommendations may be provided to a user who is unknown or unrecognized (e.g., a new visitor), even if the user has never placed an item in a shopping cart.
The invention also comprises a feature for displaying a hypertextual list of recently viewed pages or other items to the user. For example, in one embodiment, the user can view a list of the pages viewed during the current browsing session, and can use this list to navigate back to such pages. The list may optionally be filtered based on the category of pages currently being viewed by the user. For example, when a user views a page, the page may be supplemented with a list of other recently viewed pages falling within the same category as the viewed page.
The present invention also provides a method for recommending pages to a user based on the browse node pages (“browse nodes”) recently visited by the user (e.g., those visited during the current session). In one embodiment, the method comprises selecting pages to recommend to the user based on whether each page is a member of one or more of the recently visited browse nodes. A page that is a member of more than one recently visited browse node may be selected over pages that are members of only a single recently visited browse node. The browse node pages viewed by a user can be tracked using the client program, mentioned above, that executes in conjunction with a web browser on a user computer.
Further, the present invention provides a method for recommending pages to a user based on the searches recently conducted by the user (e.g., those conducted during the current session). In one embodiment, the method comprises selecting pages to recommend to the user based on whether each page is a member of one or more of the results sets of the recently conducted searches. A page that is a member of more than one such search results set may be selected over pages that are members of only a single search results set.
In one embodiment, web page analysis is used to identify products referred to or identified on the web pages reported by the client program. Accordingly, the system can be configured to identify products viewed by users on web pages of multiple web sites. By tracking the viewing of products by multiple users, sequences of products viewed by the users can be accumulated. These sequences of viewed products can be used in accordance with the techniques summarized above to identify products that are related to each other. In addition, a sequence of products viewed by a current user can be used to provide session-specific product recommendations to the current user.
These and other features of the invention will now be described with reference to the drawings summarized below. These drawings and the associated description are provided to illustrate specific embodiments of the invention, and not to limit the scope of the invention.
The various features and methods will now be described in the context of a recommendation service. Sections I through X describe a product recommendation system used to recommend products to users from an online catalog of products. Other features for assisting users in locating products of interest will also be described. Sections XI and XII describe a system for recommending web pages or web sites to users browsing the World Wide Web. Section XIII describes a system for recommending products to users based upon products viewed on web pages.
Throughout the description, the term “product” will be used to refer generally to both (a) something that may be purchased, and (b) its record or description within a database (e.g., a Sony Walkman and its description within a products database.) A more specific meaning may be implied by context.
The more general term “item” will be generally used to refer to things that are viewed by or accessed by users and which can be recommended to users. In the context of this invention, items can be products, web sites, web pages, and/or web addresses. Items can also be other things that can be recommended where the viewing, use and/or access of those things by users can be tracked. Although the items in the embodiments described in Sections I-X and XIII below are products, it will be recognized that the disclosed methods are also applicable to other types of items, such as authors, musical artists, restaurants, chat rooms, and other users. Sections XI and XII relate primarily to embodiments in which the items are web sites and/or web pages.
Throughout the description, reference will be made to various implementation-specific details, including details of implementations on the Amazon.com Web site. These details are provided in order to fully illustrate preferred embodiments of the invention, and not to limit the scope of the invention. The scope of the invention is set forth in the appended claims.
As will be recognized, the various methods set forth herein may be embodied within a wide range of different types of multi-user computer systems, including systems in which information is conveyed to users by synthesized voice or on wireless devices. Further, as described in section X below, the recommendation methods may be used to recommend items to users within a physical store (e.g., upon checking out). Thus, it should be understood that the HTML Web site based implementations described herein illustrate just one type of system in which the inventive methods may be used.
To facilitate an understanding of the specific embodiments described below, an overview will initially be provided of an example merchant Web site in which the various inventive features may be embodied.
As is common in the field of electronic commerce, the merchant Web site includes functionality for allowing users to search, browse, and make purchases from an online catalog of purchasable items or “products,” such as book titles, music titles, video titles, toys, and electronics products. The various product offerings are arranged within a browse tree in which each node represents a category or subcategory of product. Browse nodes at the same level of the tree need not be mutually exclusive.
Detailed information about each product can be obtained by accessing that product's detail page. (As used herein, a “detail page” is a page that predominantly contains information about a particular product or other item.) In a preferred embodiment, each product detail page typically includes a description, picture, and price of the product, customer reviews of the product, lists of related products, and information about the product's availability. The site is preferably arranged such that, in order to access the detail page of a product, a user ordinarily must either select a link associated with that product (e.g., from a browse node page or search results page) or submit a search query uniquely identifying the product. Thus, access by a user to a product's detail page generally represents an affirmative request by the user for information about that product.
Using a shopping cart feature of the site, users can add and remove items to/from a personal shopping cart which is persistent over multiple sessions. (As used herein, a “shopping cart” is a data structure and associated code which keeps track of items that have been selected by a user for possible purchase.) For example, a user can modify the contents of the shopping cart over a period of time, such as one week, and then proceed to a check out area of the site to purchase the shopping cart contents.
The user can also create multiple shopping carts within a single account. For example, a user can set up separate shopping carts for work and home, or can set up separate shopping carts for each member of the user's family. A preferred shopping cart scheme for allowing users to set up and use multiple shopping carts is disclosed in U.S. application Ser. No. 09/104,942, filed Jun. 25, 1998, titled METHOD AND SYSTEM FOR ELECTRONIC COMMERCE USING MULTIPLE ROLES, the disclosure of which is hereby incorporated by reference.
The Web site also implements a variety of different recommendation services for recommending products to users. One such service, known as BookMatcher™, allows users to interactively rate individual books on a scale of 1-5 to create personal item ratings profiles, and applies collaborative filtering techniques to these profiles to generate personal recommendations. The BookMatcher service is described in detail in U.S. Pat. No. 6,064,980, the disclosure of which is hereby incorporated by reference. The site may also include associated services that allow users to rate other types of items, such as CDs and videos. As described below, the ratings data collected by the BookMatcher service and/or similar services is optionally incorporated into the recommendation processes of the present invention.
Another type of service is a recommendation service which operates in accordance with the invention. In one embodiment the service (“Recommendation Service”) used to recommend book titles, music titles, video titles, toys, electronics products, and other types of products to users. The Recommendation Service could also be used in the context of the same Web site to recommend other types of items, including authors, artists, and groups or categories of products. Briefly, given a unary listing of items that are “known” to be of interest to a user (e.g., a list of items purchased, rated, and/or viewed by the user), the Recommendation Service generates a list of additional items (“recommendations”) that are predicted to be of interest to the user. (As used herein, the term “interest” refers generally to a user's liking of or affinity for an item; the term “known” is used to distinguish items for which the user has implicitly or explicitly indicated some level of interest from items predicted by the Recommendation Service to be of interest.)
The recommendations are generated using a table which maps items to lists of related or “similar” items (“similar items lists”), without the need for users to rate any items (although ratings data may optionally be used). For example, if there are three items that are known to be of interest to a particular user (such as three items the user recently purchased), the service may retrieve the similar items lists for these three items from the table, and appropriately combine these lists (as described below) to generate the recommendations.
In accordance with one aspect of the invention, the mappings of items to similar items (“item-to-item mappings”) are generated periodically, such as once per week, from data which reflects the collective interests of the community of users. More specifically, the item-to-item mappings are generated by an off-line process which identifies correlations between known interests of users in particular items. For example, in one embodiment described in detail below, the mappings are generating by analyzing user purchase histories to identify correlations between purchases of particular items (e.g., items A and B are similar because a relatively large portion of the users that purchased item A also bought item B). In another embodiment (described in section IV-B below), the mappings are generated using histories of the items viewed by individual users (e.g., items A and B are related because a significant portion of those who viewed item A also viewed item B). Item relatedness may also be determined based in-whole or in-part on other types of browsing activities of users (e.g., items A and B are related because a significant portion of those who put item A in their shopping carts also put item B in their shopping carts). Further, the item-to-item mappings could reflect other types of similarities, including content-based similarities extracted by analyzing item descriptions or content.
An important aspect of the Recommendation Service is that the relatively computation-intensive task of correlating item interests is performed off-line, and the results of this task (item-to-item mappings) are stored in a mapping structure for subsequent look-up. This enables the personal recommendations to be generated rapidly and efficiently (such as in real-time in response to a request by the user), without sacrificing breadth of analysis.
In accordance with another aspect of the invention, the similar items lists read from the table are appropriately weighted (prior to being combined) based on indicia of the user's affinity for or current interest in the corresponding items of known interest. For example, in one embodiment described below, if the item of known interest was previously rated by the user (such as through use of the BookMatcher service), the rating is used to weight the corresponding similar items list. Similarly, the similar items list for a book that was purchased in the last week may be weighted more heavily than the similar items list for a book that was purchased four months ago.
Another feature of the invention involves using the current and/or recent contents of the user's shopping cart as inputs to the Recommendation Service. For example, if the user currently has three items in his or her shopping cart, these three items can be treated as the items of known interest for purposes of generating recommendations, in which case the recommendations may be generated and displayed automatically when the user views the shopping cart contents. If the user has multiple shopping carts, the recommendations are preferably generated based on the contents of the shopping cart implicitly or explicitly designated by the user, such as the shopping cart currently being viewed. This method of generating recommendations can also be used within other types of recommendation systems, including content-based systems and systems that do not use item-to-item mappings.
Using the current and/or recent shopping cart contents as inputs tends to produce recommendations that are highly correlated to the current short-term interests of the user—even if these short term interests are not reflected by the user's purchase history. For example, if the user is currently searching for a father's day gift and has selected several books for prospective purchase, this method will have a tendency to identify other books that are well suited for the gift recipient.
Another feature of the invention involves generating recommendations that are specific to a particular shopping cart. This allows a user who has created multiple shopping carts to conveniently obtain recommendations that are specific to the role or purpose to the particular cart. For example, a user who has created a personal shopping cart for buying books for her children can designate this shopping cart to obtain recommendations of children's books. In one embodiment of this feature, the recommendations are generated based solely upon the current contents of the shopping cart selected for display. In another embodiment, the user may designate one or more shopping carts to be used to generate the recommendations, and the service then uses the items that were purchased from these shopping carts as the items of known interest.
As will be recognized by those skilled in the art, the above-described techniques for using shopping cart contents to generate recommendations can also be incorporated into other types of recommendation systems, including pure content-based systems.
Another feature, which is described in section V-C below, involves displaying session-specific personal recommendations that are based on the particular items viewed by the user during the current browsing session. For example, once the user has viewed products A, B and C, these three products may be used as the “items of known interest” for purposes of generating the session-specific recommendations. The recommendations are preferably displayed on a special Web page that can selectively be viewed by the user. From this Web page, the user can individually de-select the viewed items to cause the system to refine the list of recommended items. The session recommendations may also or alternatively be incorporated into any other type of page, such as the home page or a shopping cart page.
The Web site 30 also includes a “user profiles” database 38 which stores account-specific information about users of the site. Because a group of individuals can share an account, a given “user” from the perspective of the Web site may include multiple actual users. As illustrated by
As depicted by
The external components 40 also include a shopping cart process (not shown) which adds and removes items from the users' personal shopping carts based on the actions of the respective users. (The term “process” is used herein to refer generally to one or more code modules that are executed by a computer system to perform a particular task or set of related tasks.) In one embodiment, the shopping cart process periodically “prunes” the personal shopping cart listings of items that are deemed to be dormant, such as items that have not been purchased or viewed by the particular user for a predetermined period of time (e.g. Two weeks). The shopping cart process also preferably generates and maintains the user-specific listings of recent shopping cart contents.
The external components 40 also include recommendation service components 44 that are used to implement the site's various recommendation services. Recommendations generated by the recommendation services are returned to the Web server 32, which incorporates the recommendations into personalized Web pages transmitted to users.
The recommendation service components 44 include a BookMatcher application 50 which implements the above-described BookMatcher service. Users of the BookMatcher service are provided the opportunity to rate individual book titles from a list of popular titles. The book titles are rated according to the following scale:
The recommendation services components 44 also include a recommendation process 52, a similar items table 60, and an off-line table generation process 66, which collectively implement the Recommendation Service. As depicted by the arrows in
In the embodiments described in detail below, the items of known interest are identified based on information stored in the user's profile, such as by selecting all items purchased by the user, the items recently viewed by the user, or all items in the user's shopping cart. In other embodiments of the invention, other types of methods or sources of information could be used to identify the items of known interest. For example, in a service used to recommend Web sites, the items (Web sites) known to be of interest to a user could be identified by parsing a Web server access log and/or by extracting URLs from the “favorite places” list of the user's Web browser. In a service used to recommend restaurants, the items (restaurants) of known interest could be identified by parsing the user's credit card records to identify restaurants that were visited more than once.
The various processes 50, 52, 66 of the recommendation services may run, for example, on one or more Unix or NT based workstations or physical servers (not shown) of the Web site 30. The similar items table 60 is preferably stored as a B-tree data structure to permit efficient look-up, and may be replicated across multiple machines (together with the associated code of the recommendation process 52) to accommodate heavy loads.
The general form and content of the similar items table 60 will now be described with reference to
As indicated above, the similar items table 60 maps items to lists of similar items based at least upon the collective interests of the community of users. The similar items table 60 is preferably generated periodically (e.g., once per week) by the off-line table generation process 66. The table generation process 66 generates the table 60 from data that reflects the collective interests of the community of users. In the initial embodiment described in detail herein, the similar items table is generated exclusively from the purchase histories of the community of users (as depicted in
Further, in other embodiments, the table 60 may additionally or alternatively be generated from other indicia of user-item interests, including indicia based on users viewing activities, shopping cart activities, and item rating profiles. For example, the table 60 could be built exclusively from the present and/or recent shopping cart contents of users (e.g., products A and B are similar because a significant portion of those who put A in their shopping carts also put B in their shopping carts). The similar items table 60 could also reflect non-collaborative type item similarities, including content-based similarities derived by comparing item contents or descriptions.
Each entry in the similar items table 60 is preferably in the form of a mapping of a popular item 62 to a corresponding list 64 of similar items (“similar items lists”). As used herein, a “popular” item is an item which satisfies some pre-specified popularity criteria. For example, in the embodiment described herein, an item is treated as popular of it has been purchased by more than 30 customers during the life of the Web site. Using this criteria produces a set of popular items (and thus a recommendation service) which grows over time. The similar items list 64 for a given popular item 62 may include other popular items.
In other embodiments involving sales of products, the table 60 may include entries for most or all of the products of the online merchant, rather than just the popular items. In the embodiments described herein, several different types of items (books, CDs, videos, etc.) are reflected within the same table 60, although separate tables could alternatively be generated for each type of item.
Each similar items list 64 consists of the N (e.g., 20) items which, based on correlations between purchases of items, are deemed to be the most closely related to the respective popular item 62. Each item in the similar items list 64 is stored together with a commonality index (“CI”) value which indicates the relatedness of that item to the popular item 62, based on sales of the respective items. A relatively high commonality index for a pair of items ITEM A and ITEM B indicates that a relatively large percentage of users who bought ITEM A also bought ITEM B (and vice versa). A relatively low commonality index for ITEM A and ITEM B indicates that a relatively small percentage of the users who bought ITEM A also bought ITEM B (and vice versa). As described below, the similar items lists are generated, for each popular item, by selecting the N other items that have the highest commonality index values. Using this method, ITEM A may be included in ITEM B's similar items list even though ITEM B in not present in ITEM A's similar items list.
In the embodiment depicted by
Although the recommendable items in the described system are in the form of book titles, music titles and videos titles, and other types of products, it will be appreciated that the underlying methods and data structures can be used to recommend a wide range of other types of items.
The general sequence of steps that are performed by the recommendation process 52 to generate a set of personal recommendations will now be described with reference to
The
Any of a variety of other methods can be used to initiate the recommendations generation process and to display or otherwise convey the recommendations to the user. For example, the recommendations can automatically be generated periodically and sent to the user by e-mail, in which case the e-mail listing may contain hyperlinks to the product information pages of the recommended items. Further, the personal recommendations could be generated in advance of any request or action by the user, and cached by the Web site 30 until requested.
As illustrated by
In the embodiment depicted in
For each item of known interest, the service retrieves the corresponding similar items list 64 from the similar items table 60 (step 82), if such a list exists. If no entries exist in the table 60 for any of the items of known interest, the process 52 may be terminated; alternatively, the process could attempt to identify additional items of interest, such as by accessing other sources of interest information.
In step 84, the similar items lists 64 are optionally weighted based on information about the user's affinity for the corresponding items of known interest. For example, a similar items list 64 may be weighted heavily if the user gave the corresponding popular item a rating of “5” on a scale of 1-5, or if the user purchased multiple copies of the item. Weighting a similar items list 64 heavily has the effect of increasing the likelihood that the items in that list will be included in the recommendations ultimately presented to the user. In one implementation described below, the user is presumed to have a greater affinity for recently purchased items over earlier purchased items. Similarly, where viewing histories are used to identify items of interest, items viewed recently may be weighted more heavily than earlier viewed items.
The similar items lists 64 are preferably weighted by multiplying the commonality index values of the list by a weighting value. The commonality index values as weighted by any applicable weighting value are referred to herein as “scores.” In some embodiments, the recommendations may be generated without weighting the similar items lists 64 (as in the Shopping Cart recommendations implementation described below).
If multiple similar items lists 64 are retrieved in step 82, the lists are appropriately combined (step 86), preferably by merging the lists while summing or otherwise combining the scores of like items. The resulting list is then sorted (step 88) in order of highest-to-lowest score. By combining scores of like items, the process takes into consideration whether an item is similar to more than one of the items of known interest. For example, an item that is related to two or more of the items of known interest will generally be ranked more highly than (and thus recommended over) an item that is related to only one of the items of known interest. In another embodiment, the similar items lists are combined by taking their intersection, so that only those items that are similar to all of the items of known interest are retained for potential recommendation to the user.
In step 90, the sorted list is preferably filtered to remove unwanted items. The items removed during the filtering process may include, for example, items that have already been purchased or rated by the user, and items that fall outside any product group (such as music or books), product category (such as non-fiction), or content rating (such as PG or adult) designated by the user. The filtering step could alternatively be performed at a different stage of the process, such as during the retrieval of the similar items lists from the table 60. The result of step 90 is a list (“recommendations list”) of other items to be recommended to the user.
In step 92, one or more additional items are optionally added to the recommendations list. In one embodiment, the items added in step 92 are selected from the set of items (if any) in the user's “recent shopping cart contents” list. As an important benefit of this step, the recommendations include one or more items that the user previously considered purchasing but did not purchase. The items added in step 92 may additionally or alternatively be selected using another recommendations method, such as a content-based method.
Finally, in step 94, a list of the top M (e.g., 15) items of the recommendations list are returned to the Web server 32 (
The table-generation process 66 is preferably executed periodically (e.g., once a week) to generate a similar items table 60 that reflects the most recent purchase history data (
IV-A. Use of Purchase Histories to Identify Related Items (
As depicted by
The product IDs may be converted to title IDs during this process, or when the table 60 is later used to generate recommendations, so that different versions of an item (e.g., hardcover and paperback) are represented as a single item. This may be accomplished, for example, by using a separate database which maps product IDs to title IDs. To generate a similar items table that strongly reflects the current tastes of the community, the purchase histories retrieved in step 100 can be limited to a specific time period, such as the last six months.
In steps 102 and 104, the process generates two temporary tables 102A and 104A. The first table 102A maps individual customers to the items they purchased. The second table 104A maps items to the customers that purchased such items. To avoid the effects of “ballot stuffing,” multiple copies of the same item purchased by a single customer are represented with a single table entry. For example, even if a single customer purchased 4000 copies of one book, the customer will be treated as having purchased only a single copy. In addition, items that were sold to an insignificant number (e.g., <15) of customers are preferably omitted or deleted from the tables 102A, 104B.
In step 106, the process identifies the items that constitute “popular” items. This may be accomplished, for example, by selecting from the item-to-customers table 104A those items that were purchased by more than a threshold number (e.g., 30) of customers. In the context of a merchant Web site such as that of Amazon.com, Inc., the resulting set of popular items may contain hundreds of thousands or millions of items.
In step 108, the process counts, for each (popular_item, other_item) pair, the number of customers that are in common. A pseudocode sequence for performing this step is listed in Table 1. The result of step 108 is a table that indicates, for each (popular_item, other_item) pair, the number of customers the two have in common. For example, in the hypothetical table 108A of
In step 110, the process generates the commonality indexes for each (popular_item, other_item) pair in the table 108A. As indicated above, the commonality index (CI) values are measures of the similarity between two items, with larger CI values indicating greater degrees of similarity. The commonality indexes are preferably generated such that, for a given popular_item, the respective commonality indexes of the corresponding other_items take into consideration both (a) the number of customers that are common to both items, and (b) the total number of customers of the other_item. A preferred method for generating the commonality index values is set forth in equation (1) below, where Ncommon is the number of users who purchased both A and B, sqrt is a square-root operation, NA is the number of users who purchased A, and NB is the number of users who purchased B.
CI(item—A, item—B)=Ncommon/sqrt(NA×NB) Equation (1)
CI(item—P, item—X)=20/sqrt(300×300))=0.0667
CI(item—P, item—Y)=25/sqrt(300×30,000))=0.0083
Thus, even though items P and Y have more customers in common than items P and X, items P and X are treated as being more similar than items P and Y. This result desirably reflects the fact that the percentage of item_X customers that bought item_P (6.7%) is much greater than the percentage of item_Y customers that bought item_P (0.08%).
Because this equation is symmetrical (i.e., CI(item_A, item_B)=CI(item_B, item_A)), it is not necessary to separately calculate the CI value for every location in the table 108A. In other embodiments, an asymmetrical method may be used to generate the CI values. For example, the CI value for a (popular_item, other_item) pair could be generated as (customers of popular_item and other_item)/(customers of other_item).
Following step 110 of
In step 114, the sorted other_items lists are filtered by deleting all list entries that have fewer than 3 customers in common. For example, in the other_items list for POPULAR_A in table 108A, ITEM_A would be deleted since POPULAR_A and ITEM_A have only two customers in common. Deleting such entries tends to reduce statistically poor correlations between item sales. In step 116, the sorted other_items lists are truncated to length N to generate the similar items lists, and the similar items lists are stored in a B-tree table structure for efficient look-up.
IV-B. Use of Product Viewing Histories to Identify Related Items (
One limitation with the process of
Another limitation is that the purchase-history based method is generally incapable of identifying relationships between items that are substitutes for (purchased in place of) each other. Rather, the identified relationships tend to be exclusively between items that are complements (i.e., one is purchased in addition to the other).
In accordance with one aspect of the invention, these limitations are overcome by incorporating user-specific (and preferably session-specific) product viewing histories into the process of determining product relatedness. Specifically, the Web site system is designed to store user click stream or query log data reflecting the products viewed by each user during ordinary browsing of the online catalog. This may be accomplished, for example, by recording the product detail pages viewed by each user. Products viewed on other areas of the site, such as on search results pages and browse node pages, may also be incorporated into the users' product viewing histories.
During generation of the similar items table 60, the user-specific viewing histories are analyzed, preferably using a similar process to that used to analyze purchase history data (
An important benefit to incorporating item viewing histories into the item-to-item mapping process is that relationships can be determined between items for which little or no purchase history data exists (e.g., an obscure product or a newly released product). As a result, relationships can typically be identified between a far greater range of items than is possible with a pure purchase-based approach.
Another important benefit to using viewing histories is that the item relationships identified include relationships between items that are pure substitutes. For example, the purchase-based item-to-item similarity mappings ordinarily would not map one large-screen TV to another large-screen TV, since it is rare that a single customer would purchase more than one large-screen TV. On the other hand, a mapping that reflects viewing histories would likely link two large-screen TVs together since it is common for a customer to visit the detail pages of multiple large-screen TVs during the same browsing session.
The query log data used to implement this feature may optionally incorporate browsing activities over multiple Web sites (e.g., the Web sites of multiple, affiliated merchants). Such multi-site query log data may be obtained using any of a variety of methods. One known method is to have the operator of Web site A incorporate into a Web page of Web site A an object served by Web site B (e.g., a small graphic). With this method, any time a user accesses this Web page (causing the object to be requested from Web site B), Web site B can record the browsing event. Another known method for collecting multi-site query log data is to have users download a browser plug-in, such as the plug-in provided by Alexa Internet Inc., that reports browsing activities of users to a central server. The central server then stores the reported browsing activities as query log data records. Further, the entity responsible for generating the similar items table could obtain user query log data through contracts with ISPs, merchants, or other third party entities that provide Web sites for user browsing.
Although the term “viewing” is used herein to refer to the act of accessing product information, it should be understood that the user does not necessarily have to view the information about the product. Specifically, some merchants support the ability for users to browse their electronic catalogs by voice. For example, in some systems, users can access voiceXML versions of the site's Web pages using a telephone connection to a voice recognition and synthesis system. In such systems, a user request for voice-based information about a product may be treated as a product viewing event.
As depicted by
Each query log record is preferably in the general form of a browsing session identification together with a list of the identifiers of the items viewed in that browsing session. The item IDs may be converted to title IDs during this process, or when the table 60 is later used to generate recommendations, so that different versions of an item are represented as a single item. Each query log record may alternatively list some or all of the pages viewed during the session, in which case a look up table may be used to convert page IDs to item or product IDs.
In steps 302 and 304, the process builds two temporary tables 302A and 304A. The first table 302A maps browsing sessions to the items viewed in the sessions. A table of the type shown in
In step 306, the process identifies the items that constitute “popular” items. This may be accomplished, for example, by selecting from table 304A those items that were viewed within more than a threshold number (e.g., 30) of sessions. In the context of a Web site of a typical online merchant that sells many thousands or millions of different items, the number of popular items in this embodiment will desirably be far greater than in the purchase-history-based embodiment of
In step 308, the process counts, for each (popular_item, other_item) pair, the number of sessions that are in common. A pseudocode sequence for performing this step is listed in Table 2. The result of step 308 is a table that indicates, for each (popular_item, other_item) pair, the number of sessions the two have in common. For example, in the hypothetical table 308A of
In step 310, the process generates the commonality indexes for each (popular_item, other_item) pair in the table 308A. The commonality index (CI) values are measures of the similarity or relatedness between two items, with larger CI values indicating greater degrees of similarity. The commonality indexes are preferably generated such that, for a given popular_item, the respective commonality indexes of the corresponding other_items take into consideration the following (a) the number of sessions that are common to both items (i.e, sessions in which both items were viewed), (b) the total number of sessions in which the other_item was viewed, and (c) the number of sessions in which the popular_item was viewed. Equation (1), discussed above, may be used for this purpose, but with the variables redefined as follows: Ncommon is the number of sessions in which both A and B were viewed, NA is the number of sessions in which A was viewed, and NB is the number of sessions in which B was viewed. Other calculations that reflect the frequency with which A and B co-occur within the product viewing histories may alternatively be used.
CI(item—P, item—X)=20/sqrt(300×300))=0.0667
CI(item—P, item—Y)=25/sqrt(300×30,000))=0.0083
Thus, even though items P and Y have more sessions in common than items P and X, items P and X are treated as being more similar than items P and Y. This result desirably reflects the fact that the percentage of item_X sessions in which item_P was viewed (6.7%) is much greater than the percentage of item_Y sessions in which item_P was viewed (0.08%).
Because this equation is symmetrical (i.e., CI(item_A, item_B)=CI(item_B, item_A)), it is not necessary to separately calculate the CI value for every location in the table 308A. As indicated above, an asymmetrical method may alternatively be used to generate the CI values.
Following step 310 of
In one embodiment, the items in the other_items list are weighted to favor some items over others. For example, items that are new releases may be weighted more heavily than older items. For items in the other_items list of a popular item, their CI values are preferably multiplied by the corresponding weights. Therefore, the more heavily weighted items (such as new releases) are more likely to be considered related and more likely to be recommended to users.
In step 316, the sorted other_items lists are truncated to length N (e.g., 20) to generate the similar items lists, and the similar items lists are stored in a B-tree table structure for efficient look-up.
One variation of the method shown in
Another variation is to use the “distance” between two product viewing events as an additional indicator of product relatedness. For example, if a user views product A and then immediately views product B, this may be treated as a stronger indication that A and B are related than if the user merely viewed A and B during the same session. The distance may be measured using any appropriate parameter that can be recorded within a session record, such as time between product viewing events, number of page accesses between product viewing events, and/or number of other products viewed between product viewing events. Distance may also be incorporated into the purchase based method of
As with generation of the purchase-history-based similar items table, the viewing-history-based similar items table is preferably generated periodically, such as once per day or once per week, using an off-line process. Each time the table 60 is regenerated, query log data recorded since the table was last generated is incorporated into the process—either alone or in combination with previously-recorded query log data. For example, the temporary tables 302A and 304A of
IV-C. Determination of Item Relatedness Using Other Types of User Activities
The process flows shown in
With the above and other types of item-affinity-evidencing actions, equation (1) above may be used to generate the CI values, with the variables of equation (1) generalized as follows:
As indicated above, any of a variety non-user-action-based methods for evaluating similarities between items could be incorporated into the table generation process 66. For example, the table generation process could compare item contents and/or use previously-assigned product categorizations as additional or alternative indicators of item relatedness. An important benefit of the user-action-based methods (e.g., of
Another important benefit of the Recommendation Service is that the bulk of the processing (the generation of the similar items table 60) is performed by an off-line process. Once this table has been generated, personalized recommendations can be generated rapidly and efficiently, without sacrificing breadth of analysis.
Three specific implementations of the Recommendation Service, referred to herein as Instant Recommendations, Shopping Basket Recommendations, and Session Recommendations, will now be described in detail. These three implementations differ in that each uses a different source of information to identify the “items of known interest” of the user whose recommendations are being generated. In all three implementations, the recommendations are preferably generated and displayed substantially in real time in response to an action by the user.
Any of the methods described above may be used to generate the similar items tables 60 used in these three service implementations. Further, all three (and other) implementations may be used within the same Web site or other system, and may share the same similar items table 60.
V-A Instant Recommendations Service (
A specific implementation of the Recommendation Service, referred to herein as the Instant Recommendations service, will now be described with reference to
As indicated above, the Instant Recommendations service is invoked by the user by selecting a corresponding hyperlink from a Web page. For example, the user may select an “Instant Book Recommendations” or similar hyperlink to obtain a listing of recommended book titles, or may select a “Instant Music Recommendations” or “Instant Video Recommendations” hyperlink to obtain a listing of recommended music or video titles. As described below, the user can also request that the recommendations be limited to a particular item category, such as “non-fiction,” “jazz” or “comedies.” The “items of known interest” of the user are identified exclusively from the purchase history and any item ratings profile of the particular user. The service becomes available to the user (i.e., the appropriate hyperlink is presented to the user) once the user has purchased and/or rated a threshold number (e.g. three) of popular items within the corresponding product group. If the user has established multiple shopping carts, the user may also be presented the option of designating a particular shopping cart to be used in generating the recommendations.
In step 184, the process 52 weights each similar items list based on the duration since the associated popular item was purchased by the user (with recently-purchased items weighted more heavily), or if the popular item was not purchased, the rating given to the popular item by the user. The formula used to generate the weight values to apply to each similar items list is listed in C in Table 2. In this formula, “is_purchased” is a boolean variable which indicates whether the popular item was purchased, “rating” is the rating value (1-5), if any, assigned to the popular item by the user, “order_date” is the date/time (measured in seconds since 1970) the popular item was purchased, “now” is the current date/time (measured in seconds since 1970), and “6 months” is six months in seconds.
In line 1 of the formula, if the popular item was purchased, the value “5” (the maximum possible rating value) is selected; otherwise, the user's rating of the item is selected. The selected value (which may range from 1-5) is then multiplied by 2, and 5 is subtracted from the result. The value calculated in line 1 thus ranges from a minimum of −3 (if the item was rated a “1”) to a maximum of 5 (if the item was purchased or was rated a “5”).
The value calculated in line 1 is multiplied by the value calculated in lines 2 and 3, which can range from a minimum of 1 (if the item was either not purchased or was purchased at least six months ago) to a maximum of 2 (if order_date=now). Thus, the weight can range from a minimum of −6 to a maximum of 10. Weights of zero and below indicate that the user rated the item a “2” or below. Weights higher than 5 indicate that the user actually purchased the item (although a weight of 5 or less is possible even if the item was purchased), with higher values indicating more recent purchases.
The similar items lists 64 are weighted in step 184 by multiplying the CI values of the list by the corresponding weight value. For example, if the weight value for a given popular item is ten, and the similar items list 64 for the popular item is
(productid_A, 0.10), (productid_B, 0.09), (productid_C, 0.08), . . .
the weighted similar items list would be:
(productid_A, 1.0), (productid_B, 0.9), (productid_C, 0.8), . . .
The numerical values in the weighted similar items lists are referred to as “scores.”
In step 186, the weighted similar items lists are merged (if multiple lists exist) to form a single list. During this step, the scores of like items are summed. For example, if a given other_item appears in three different similar items lists 64, the three scores (including any negative scores) are summed to produce a composite score.
In step 188, the resulting list is sorted from highest-to-lowest score. The effect of the sorting operation is to place the most relevant items at the top of the list. In step 190, the list is filtered by deleting any items that (1) have already been purchased or rated by the user, (2) have a negative score, or (3) do not fall within the designated product group (e.g., books) or category (e.g., “science fiction,” or “jazz”).
In step 192 one or more items are optionally selected from the recent shopping cart contents list (if such a list exists) for the user, excluding items that have been rated by the user or which fall outside the designated product group or category. The selected items, if any, are inserted at randomly-selected locations within the top M (e.g., 15) positions in the recommendations list. Finally, in step 194, the top M items from the recommendations list are returned to the Web server 32, which incorporates these recommendations into one or more Web pages.
The general form of such a Web page is shown in
The user can also select a specific category such as “non-fiction” or “romance” from a drop-down menu 202 to request category-specific recommendations. Designating a specific category causes items in all other categories to be filtered out in step 190 (
V-B Shopping Cart Based Recommendations (
Another specific implementation of the Recommendation Service, referred to herein as Shopping Cart recommendations, will now be described with reference to
The Shopping Cart recommendations service is preferably invoked automatically when the user displays the contents of a shopping cart that contains more than a threshold number (e.g., 1) of popular items. The service generates the recommendations based exclusively on the current contents of the shopping cart (i.e., only the shopping cart contents are used as the “items of known interest”). As a result, the recommendations tend to be highly correlated to the user's current shopping interests. In other implementations, the recommendations may also be based on other items that are deemed to be of current interest to the user, such as items in the recent shopping cart contents of the user and/or items recently viewed by the user. Further, other indications of the user's current shopping interests could be incorporated into the process. For example, any search terms typed into the site's search engine during the user's browsing session could be captured and used to perform content-based filtering of the recommended items list.
In step 286, these similar items lists are merged while summing the commonality index (CI) values of like items. In step 288, the resulting list is sorted from highest-to-lowest score. In step 290, the list is filtered to remove any items that exist in the shopping cart or have been purchased or rated by the user. Finally, in step 294, the top M (e.g., 5) items of the list are returned as recommendations. The recommendations are preferably presented to the user on the same Web page (not shown) as the shopping cart contents. An important characteristic of this process is that the recommended products tend to be products that are similar to more than one of the products in the shopping cart (since the CI values of like items are combined). Thus, if the items in the shopping cart share some common theme or characteristic, the items recommended to the user will tend to have this same theme or characteristic.
If the user has defined multiple shopping carts, the recommendations generated by the
The various uses of shopping cart contents to generate recommendations as described above can be applied to other types of recommendation systems, including content-based systems. For example, the current and/or past contents of a shopping cart can be used to generate recommendations in a system in which mappings of items to lists of similar items are generated from a computer-based comparison of item contents. Methods for performing content-based similarity analyses of items are well known in the art, and are therefore not described herein.
V-C Session Recommendations (
One limitation in the above-described service implementations is that they generally require users to purchase or rate products (Instant Recommendations embodiment), or place products into a shopping cart (Shopping Cart Recommendations embodiment), before personal recommendations can be generated. As a result, the recommendation service may fail to provide personal recommendations to a new visitor to the site, even though the visitor has viewed many different items. Another limitation, particularly with the Shopping Cart Recommendations embodiment, is that the service may fail to identify the session-specific interests of a user who fails to place items into his or her shopping cart.
In accordance with another aspect of the invention, these limitations are overcome by providing a Session Recommendations service that stores a history or “click stream” of the products viewed by a user during the current browsing session, and uses some or all of these products as the user's “items of known interest” for purposes of recommending products to the user during that browsing session. Preferably, the recommended products are displayed on a personalized Web page (
The click-stream data used to implement this service may optionally incorporate product browsing activities over multiple Web sites. For example, when a user visits one merchant Web site followed by another, the two visits may be treated as a single “session” for purposes of generating personal recommendations.
As illustrated, the system includes an HTTP/XML application 37 that monitors clicks (page requests) of users, and records information about certain types of events within a click stream table 39. The click stream table is preferably stored in a cache memory 39 (volatile RAM) of a physical server computer, and can therefore be rapidly and efficiently accessed by the Session Recommendations application 52 and other real time personalization components. All accesses to the click stream table 39 are preferably made through the HTTP/XML application, as shown. The HTTP/XML application 37 may run on the same physical server machine(s) (not shown) as the Web server 32, or on a “service” layer of machines sitting behind the Web server machines. An important benefit of this architecture is that it is highly scalable, allowing the click stream histories of many thousands or millions of users to be maintained simultaneously.
In operation, each time a user views a product detail page, the Web server 32 notifies the HTTP/XML application 37, causing the HTTP/XML application to record the event in real time in a session-specific record of the click stream table. The HTTP/XML application may also be configured to record other click stream events. For example, when the user runs a search for a product, the HTTP/XML application may record the search query, and/or some or all of the items displayed on the resulting search results page (e.g., the top X products listed). Similarly, when the user views a browse node page (a page corresponding to a node of a browse tree in which the items are arranged by category), the HTTP/XML application may record an identifier of the page or a list of products displayed on that page.
A user access to a search results page or a browse node page may, but is preferably not, treated as a viewing event with respect to products displayed on such pages. As discussed in sections VIII and XI below, the session-specific histories of browse node accesses and searches may be used as independent or additional data sources for providing personalized recommendations.
In one embodiment, once the user has viewed a threshold number of product detail pages (e.g., 1, 2 or 3) during the current session, the user is presented with a link to a custom page of the type shown in
The process for generating the Session Recommendations is preferably the same as or similar to the process shown in
As depicted by the dashed arrow in
The browsing session ID can be any identifier that uniquely identifies a browsing session. In one embodiment, the browsing session ID includes a number representing the date and time at which a browsing session started. A “session” may be defined within the system based on times between consecutive page accesses, whether the user viewed another Web site, whether the user checked out, and/or other criteria reflecting whether the user discontinued browsing.
Each page ID uniquely identifies a Web page, and may be in the form of a URL or an internal identification. For a product detail page (a page that predominantly displays information about one particular product), the product's unique identifier may be used as the page identification. The detail page list may therefore be in the form of the IDs of the products whose detail pages were viewed during the session. Where voiceXML pages are used to permit browsing by telephone, a user access to a voiceXML version of a product detail page may be treated as a product “viewing” event.
The search query list includes the terms and/or phrases submitted by the user to a search engine of the Web site 30. The captured search terms/phrases may be used for a variety of purposes, such as filtering or ranking the personal recommendations returned by the
In one embodiment, the process of converting page IDs to corresponding product IDs is handled by the Web server 32, which passes a session_ID/product_ID pair to the HTTP/XML application 37 in response to the click stream event. This conversion task may alternatively be handled by the HTTP/XML application 37 each time a click stream event is recorded, or may be performed by the Session Recommendations component 52 when personal recommendations are generated.
As illustrated in
In another embodiment, the Web page of
The “page I made” Web page may also include other types of personalized content. For instance, in the example shown in
In embodiments that support browsing by voice, the customized Web page may be in the form of a voiceXML page, or a page according to another voice interface standard, that is adapted to be accessed by voice. In such embodiments, the various lists of items 402, 404, 406 may be output to the customer using synthesized and/or pre-recorded voice.
An important aspect of the Session Recommendations service is that it provides personalized recommendations that are based on the activities performed by the user during the current session. As a result, the recommendations tend to strongly reflect the user's session-specific interests. Another benefit is that the recommendations may be generated and provided to users falling within one or both of the following categories: (a) users who have never made a purchase, rated an item, or placed an item in a shopping cart while browsing the site, and (b) users who are unknown to or unrecognized by the site (e.g., a new visitor to the site). Another benefit is that the user can efficiently refine the session data used to generate the recommendations.
The Session Recommendations may additionally or alternatively be displayed on other pages of the Web site 30. For example, the Session Recommendations could be displayed when the user returns to the home page, or when the user views the shopping cart. Further, the Session Recommendations may be presented as implicit recommendations, without any indication of how they were generated.
As described above with reference to
In one embodiment, each hyperlink within the list 402 is to a product detail page visited during the current browsing session. This list is generated by reading the user's session record in the click stream table 39, as described above. In other embodiments, the list of recently viewed items may include detail pages viewed during prior sessions (e.g., all sessions over last three days), and may include links to recently accessed browse node pages and/or recently used search queries.
Further, a filtered version of a user's product viewing history may be displayed in certain circumstances. For example, when a user views a product detail page of an item in a particular product category, this detail page may be supplemented with a list of (or a link to a list of) other products recently viewed by the user that fall within the same product category. For instance, the detail page for an MP3 player may include a list of any other MP3 players, or of any other electronics products, the user has recently viewed.
An important benefit of this feature is that it allows users to more easily comparison shop.
In addition to using the similar items table 60 to generate personal recommendations, the table 60 may be used to display “canned” lists of related items on product detail pages of the “popular” items (i.e., items for which a similar items list 64 exists).
An important benefit to using a similar items table 60 that reflects viewing-history-based similarities, as opposed to a table based purely on purchase histories, is that the number of product viewing events will typically far exceed the number of product purchase events. As a result, related items lists can be displayed for a wider selection of products—including products for which little or no sales data exists. In addition, for the reasons set forth above, the related items displayed are likely to include items that are substitutes for the displayed item.
As indicated above and shown in
For example, in one embodiment, the Session Recommendations process 52 identifies items that fall within one or more browse nodes viewed by the user during the current session, and recommends some or all of these items to the user (implicitly or explicitly) during the same session. If the user has viewed multiple browse nodes, greater weight may be given to an item that falls within more than one of these browse nodes, increasing the item's likelihood of selection. For example, if the user views the browse node pages of two music categories at the same level of the browse tree, a music title falling within both of these nodes/categories would be selected to recommend over a music title falling in only one.
As with the session recommendations based on recently viewed products, the session recommendations based on recently viewed browse nodes may be displayed on a customized page that allows the user to individually deselect the browse nodes and then update the page. The customized page may be the same page used to display the product viewing history based recommendations (
A hybrid of this method and the product viewing history based method may also be used to generate personalized recommendations.
Each user's history of recent searches, as reflected within the session record, may be used to generate recommendations in an analogous manner to that described in section VIII. The results of each search (i.e., the list of matching items) may be retained in cache memory to facilitate this task.
In one embodiment, the Session Recommendations component 52 identifies items that fall within one or more results lists of searches conducted by the user during the current session, and recommends some or all of these items to the user (implicitly or explicitly) during the same session. If the user has conducted multiple searches, greater weight may be given to an item falling within more than one of these search results lists, increasing the item's likelihood of selection. For example, if the user conducts two searches, a music title falling within both sets of search results would be selected to recommend over a music title falling in only one.
As with the session recommendations based on recently viewed products, the session recommendations based on recently conducted searches may be displayed on a customized page that allows the user to individually deselect the search queries and then update the page. The customized page may be the same page used to display the product viewing history based recommendations (
Any appropriate hybrid of this method, the product viewing history based method (section V-C), and the browse node based method (section VIII), may be used to generate personalized recommendations.
The recommendation methods described above can also be used to provide personalized recommendations within physical stores. For example, each time a customer checks out at a grocery or other physical store, a list of the purchased items may be stored. These purchase lists may then be used to periodically generate a similar items table 60 using the process of
Once a similar items table has been generated, a process of the type shown in
As mentioned in section IV-B above, a browser plug-in can be used to report browsing activities of users to a central server.
A recommendation system 1400 preferably uses a client program or browser plug-in 1402 that executes in conjunction with a web browser 1404 on a user computer 34 to monitor web addresses (e.g. URLs) of web pages viewed by a user of the computer. The web pages can be hosted by any number of different web sites 1406. By monitoring a user's browsing actions through a client program rather than through a web server, a user's browsing actions can be tracked as the user moves from site to site.
In
As the client program identifies each web address, it transmits the address to a server application 1408, which can be similar in functionality to the HTTP/XML application 37 discussed with reference to
Each web address that is accumulated in the click stream table 1410 for a user's browsing session is preferably stored in a click stream database 1412, which can be similar to the query log database 42 discussed with reference to
In one embodiment, a browsing session can include a set of web addresses that are accessed by a user within a certain time period. The time period of a browsing session can be defined as a certain length of time, such as 15 minutes or 1 day. Alternatively, the time period can be variable, in which case it can be based upon a maximum interval between clicks (page visits). For example, a browsing session can be defined as a sequence of clicks where each click occurs within 2 minutes of the last click.
In order to create a set of recommendations, the system 1400 preferably relies upon both the current user's click stream, which is stored in the click-stream table 1410, as well as click-streams of other users that have been accumulated in the click-stream database 1412. The click-streams of multiple users are preferably processed by a table generation process 1414 to generate a similar items table 1416, which identifies similar or related web pages, web sites and/or addresses. Generation of the similar items table 1416 is preferably performed off-line, in advance of the gathering of the current user's click stream.
In one embodiment, the table generation process 1414 generates the similar items table 1416 substantially in accordance with the method described above with reference to
As depicted by the arrows in
The session recommendation process 1418 preferably generates the personal recommendations substantially in accordance with the method described above with reference to
The personal recommendations are preferably incorporated into a web page 1420, which can be hosted and served by a web server 1422. The web page 1420 preferably includes hypertext links to the web addresses of the web pages being recommended. In one embodiment, each link can be labeled with the title of the web page being recommended. In one embodiment, the client program 1402 can be configured to display an icon or link on the user computer 34 that the user can select in order to drive the web browser 1404 to the web page 1420 that displays the set of personal recommendations. The client program 1402 can alternatively be configured to display the recommendations in a separate window that can be maintained and even updated as the user continues browsing.
In accordance with this embodiment, the click stream data accumulated for each user is preferably used in two ways. In one aspect, the click stream data for a current user is used, in conjunction with the similar items table 1416, to create a set of personal recommendations for the current user. In another aspect, the click stream data for a current user is accumulated and used in conjunction with other click stream data to create the similar items table 1416 for subsequent users.
In the case that web pages are being recommended, as described above, the table generation process 1414 and the session recommendation process 1418 are preferably based upon the web addresses in the click stream data. As mentioned above however, web sites and/or web addresses can be recommended similarly. In the case that web sites are being recommended in addition to web pages, the web sites visited during each click stream of web pages can be derived from the web addresses (of web pages) stored in the click stream table 1410 and click-stream database 1412. The web sites derived from the click stream data can then be used by the table generation process 1414 and session recommendations process 1418 to generate a set of web site recommendations. In the case that only web sites are being recommended, the web addresses stored in the click stream table 1410 and click-stream database 1412 can be addresses of web site home pages or domain names. As discussed above, the session recommendations process 1418 preferably provides the web addresses of recommended web pages. Accordingly, in one embodiment, these web addresses can be included on the recommendation web page 1420 to recommend web addresses in addition to or instead of the corresponding web pages or web sites.
In one embodiment, web addresses, such as URLs, are used to identify web pages and/or web sites. Alternatively, other identifiers can be used to identify web pages and/or web sites. For example, each web address can be truncated or modified to remove any session ID information or other session-specific information. In addition, multiple addresses that map to the same web page or site can be translated into a common identifier, such as one of the addresses that map to the page or site. Web sites can be identified, for example, through their domain names or through the addresses of their home pages. In alternative embodiments, any identifier, such as a name or a number, can be used by the client program and/or system 1400 to identify web sites and/or web pages.
Other methods or processes for identifying similar items or creating similar items tables 1416 can alternatively be used, including methods that do not use browsing histories of users. For instance, web site relatedness can be determined by performing a content-based analysis of site content and identifying sites that use the same or similar characterizing terms and phrases. In certain embodiments, the results of multiple methods of identifying similar items can be combined. In one embodiment, the table generation process 1416 generates the similar items table 1416 using a minimum sensitivity calculation as described in the next section.
In accordance with one embodiment, the relatedness (similarity) of two web sites A and B can be determined using a sensitivity calculation that takes into consideration the number of transitions (user clicks) between A and B, the number of transitions between A and other web sites, and/or the number of transitions between B and other web sites within a set of browsing history data including user click streams. This process for determining relatedness of web sites presumes that web sites accessed by the user during a browsing session, and/or within some threshold number of web site transitions from one another, tend to be related.
In accordance with one embodiment, this minimum sensitivity calculation is used to create the similar items table 1416 based upon click stream data stored in the click-stream database 1412. The calculation is preferably based upon data collected from many user browsing sessions and from many users.
The description that follows will be presented in the context of identifying similar web sites, which can be identified through the web addresses of their home pages. This method can also be applied to web pages and/or web addresses in a similar manner.
For any two web sites A and B, a transition between site A and site B in a click stream (also referred to herein more generally as a “usage trail”) can be either an accessing of site A followed by an accessing of site B, or an accessing of site B followed by an accessing of site A. In one embodiment, the only type of transition recognized between web sites A and B is a 1-step transition, meaning that site B is the first site browsed immediately after site A, or vice versa. In an alternative embodiment, the transition between web sites A and B can be an n-step transition, meaning that site B is the n-th site browsed after site A, or vice versa. In still other embodiments, the transition between web sites A and B can be an m to n step transition, meaning that B is at least the m-th site and at most the n-th site browsed after site A, or vice versa.
In accordance with one embodiment, the sensitivity calculation is preferably a minimum sensitivity calculation. The minimum sensitivity between A and B can be defined as follows:
where T(A,B) is defined as the number of transitions between A and B, MAX(x,y) is a function that yields the greater of x and y, and all_sites denotes all web sites within the data set. The minimum sensitivity, as defined here, has a range of 0 to 1 inclusive. A minimum sensitivity of 0 indicates that no transitions occur between web sites A and B in the sample set of usage trail data. A minimum sensitivity of 1 indicates that any transitions involving A or B are always between A and B.
The above calculation of minimum sensitivity can also be described by the following process: divide the number of transitions between web sites A and B by the greater of (i) the number of transitions between A and all web sites and (ii) the number of transitions between B and all web sites. In this embodiment, minimum sensitivity is used as a measure of the relatedness of two web sites.
An example calculation of the minimum sensitivity between web sites A and B follows:
100 transitions between A and B;
100 transitions between A and all web sites; and
100 transitions between B and all web sites.
In this example, the since there are 100 transitions between A and all web sites, there are 100 transitions between B and all web sites, and there are 100 transitions between A and B, then all the transitions involving A and B were between A and B. Therefore, the sensitivity between A and B is 1.
In performing the table generating process 1414, minimum sensitivity is preferably determined based upon a set of transitions included in the click stream database 1412. Preferably all, but possibly only some of the transitions recorded in the database 1412 are used in the calculation. Each transition is preferably a transition between two sites or pages visited in a single session. As mentioned above, the sites can be visited one after another, or alternatively the sites can be visited after some number of intervening sites have been visited. Other than for the purpose of identifying transitions, browsing sessions need not be used in determining minimum sensitivity.
The table generation process in this embodiment is preferably accomplished by applying sorting, matching, cataloguing, and/or categorizing functions to the usage trail data gathered by the server application 1408. Depending upon the objectives of the implementation and the desired accuracy of the sensitivity measure, approximation measures, rounding, and other methods that will be apparent to one skilled in the art can be used to gain efficiencies in the determinations of minimum sensitivity.
Note that the aforementioned minimum sensitivity calculation is symmetric, MS (A, C)=MS (C, A), since the transitions do not take direction into account. The minimum sensitivity calculation, however, is not symmetric when directional transitions are used as will be discussed below.
In the preferred embodiment, web sites are identified by the domain name portions of their URLs. Personal home pages and their associated pages are preferably also considered web sites, but are identified, in addition, by their addresses (relative or absolute pathnames) on their host systems. A table of web site aliases may also be used to identify different domain names that refer to the same web site.
In one embodiment, the table generation process is based upon 1-step transitions determined from the sample set of usage trail data. In addition, transitions through certain types of web sites, such as web portals and search engines may by filtered out of a usage trail or not considered in identifying a transition. For example, a user may transition from a search engine site to a first site of interest. Next, the user may transition back to the search engine and then to a second site of interest. By filtering out the transition to the search engine between the first and second web sites, the possibility that the first and second web sites are related is captured in the usage trail data.
In alternative embodiments, an n-step transition or an m-n step transition can be used. In still other embodiments, 1-step, n-step, and m-n step transitions can be combined in order to modify the characteristics of the resulting sensitivity calculation. For example, the various types of transitions can be combined by weighting each type of transition. In a more specific example, the number of 1-step transitions and the number of 2-step transitions between A and B could each be weighted by 0.5. The weighted numbers could be added to yield a combined number of transitions that takes into account both 1-step and 2-step transitions. The combined number of transitions could then be used to perform the sensitivity calculation. As another alternative, a sensitivity can be determined for each of two or more types of transitions, and the resulting sensitivities can be combined by weighting. For example, a 1-step sensitivity and a 2-step sensitivity can each be calculated between A and B. The two sensitivities can then be combined, for example, by weighting each by a factor, such as 0.5, and adding the weighted sensitivities.
In some embodiments, the sensitivity need not be a minimum sensitivity. In one embodiment, for example, the taking of the maximum in the denominator of the minimum sensitivity calculation can be replaced with another function. The calculated sensitivity could be the number of transitions between web sites A and B divided by the number of transitions between A and all web sites. In another embodiment, the calculated sensitivity could be the number of transitions between web sites A and B divided by the number of transitions between all web sites and B. In still another embodiment the number of transitions between A and B could be divided by the sum of (i) the number of transitions between A and all web sites and (ii) the number of transitions between B and all web sites.
In additional embodiments, equivalent metrics to numbers of transitions could be used in the sensitivity calculation, such as, for example, frequencies of transitions. As another example, the number of transitions between A and B could be excepted from the number of transitions between A and all sites, or the number of transitions between B and all sites, respectively.
The table generation process 1414 is preferably repeated to calculate a sensitivity for all pairs of web sites between which transitions exist in the sample set of usage trail data. In addition, the sensitivity calculation may be modified to incorporate other types of information that may also be captured in conjunction with the usage trail data. For example, page request timestamps may be used to determine how long it took a user to navigate from web site A to web site B, and this time interval may be used to appropriately weight or exclude from consideration the transition from A to B. In addition, a transition between A and B could be given greater weight if a direct link exists between web sites A and B as may be determined using an automated web site crawling and parsing routine.
The table generation process 1414, can also be applied in determining the relatedness of web pages in addition to or instead of web sites. In this case, for any two web pages A and B, a transition between A and B in a usage trail can be either an accessing of page A followed by an accessing of page B, or an accessing of page B followed by an accessing of page A. Like a transition between web sites, a transition between web pages A and B can be a 1-step transition, an n-step transition, or an m-n step transition, where a step involves the following of a link from one page to a next.
Additional factors can also be used to determine how much to weight a particular directional transition. For example, a transition may be given an increased weight if it is detected that a user makes a purchase, performs a search, or performs some other type of transaction at a web site following the transition.
The table generation process 1414 can also be adapted to determine the relatedness of a web site A to a web site B (as opposed to the relatedness between web sites A and B) based upon directional transitions. A transition from a web site A to a web site B in a usage trail is an accessing of site A followed by an accessing of site B. A transition from a web site A to a web site B is a subset of a transition between A and B in that it includes a transition in only a single direction.
The determination of minimum sensitivity based upon directional transitions can be described as follows: divide the number of transitions from web site A to web site B by the greater of (i) the number of transitions from A to all web sites and (ii) the number of transitions from all web sites to B. 1-step, n-step, and m-n step directional transitions can be used to determine a minimum sensitivity from a web site A to a web site B. In this embodiment, the minimum sensitivity has a range of 0 to 1 inclusive. A minimum sensitivity of 0 indicates that no transitions occur from web site A to web site B in the sample set of usage trail data. A minimum sensitivity of 1 indicates that all transitions from web site A are to web site B. Sensitivity based upon directional transitions can also be used as a measure of the relatedness of a web site A to a web site B.
At a first step 1502, a sample set of usage trail data is gathered from users over a period of time by the server application 1408. The server application 1408 receives identifications of web pages or web sites from the client programs 1402 executing in conjunction with users' web browsers 1404. In one embodiment, the server application 1408 gathers usage trail data over a period of approximately four weeks from the users of the system 1404. The time period may be varied substantially to account for the actual number of users and other considerations.
At step 1504, for each subject web site (the web site for which similar sites are to be identified) the table generation process 1414 calculates the sensitivities between a subject web site and other web sites preferably using a minimum sensitivity calculation. The subject web site may be any web site for which related sites are to be identified and for which there is at least one transition within the usage trail data. The other web sites are preferably all web sites having at least one transition in common with the subject web site within the usage trail data. Web sites that are not identified in at least one transition can be effectively dropped from consideration as potential related sites as their sensitivities would be zero.
At step 1506 the process 1414 identifies the other sites with the highest sensitivities as related sites for the subject web site. The related sites are preferably identified by their domain names, or in the case of web pages, by their URLs. In one embodiment, approximately eight related sites are identified for each subject site. In alternative embodiments, however, any number of related links could be identified.
The process 1414 preferably performs steps 1504 and 1506 for each subject web site for which there is at least one transition in the usage trail data. The process 1414 preferably stores the resulting lists of related sites in the similar items table 1416 for subsequent retrieval and use in creating personal recommendations. The sequence of steps 1502-1506 involved in identifying related sites is preferably repeated periodically, such as every four weeks.
The process illustrated in flowchart 1500 can also or alternatively be adapted to provide related web pages, in addition to or in place of related web sites. The process 1414 can also be configured to provide related sites or pages for subject web pages in addition to or instead of subject web sites. Alternative and additional embodiments by which relatedness of web sites can be determined are described in U.S. application Ser. No. 09/470,844, filed Dec. 23, 1999, which is assigned to the assignee of the present application and which is hereby incorporated herein by reference in its entirety.
In one embodiment, the web addresses reported by the client program 1402, discussed in Section XI above, can be used to (1) identify products that are related to each other, and/or (2) provide session-specific product recommendations to users. More generally, this embodiment can be adapted to recommend any item that can be identified through the World Wide Web.
The recommendation system 1400 can be configured to fetch each web page identified by each client program 1402 and perform an analysis of the fetched page in order to identify products that may be identified on the page. The analysis can be a content-based analysis that may include searching the page for product names, manufacturer names, part numbers, and/or catalog numbers. Alternatively or additionally, a structure-based analysis can be used as described in U.S. patent application Ser. No. 09/794,952 filed Feb. 27, 2001 and titled “RULE-BASED IDENTIFICATION OF ITEMS REPRESENTED ON WEB PAGES,” which is incorporated herein by reference. In one embodiment, once a web page is analyzed to identify any products on the web page, the products are associated with the web page in a database so that the analysis need not be performed again the next time the web page is identified by a client program 1402.
U.S. patent application Ser. No. 09/820,207 filed Mar. 28, 2001 and titled “SUPPLEMENTATION OF WEB PAGES WITH PRODUCT-RELATED INFORMATION,” which is incorporated herein by reference, describes a system that associates products with web pages based upon the input of users browsing the pages. Such a system can be used to identify products displayed on web pages without having to separately fetch and analyze each web page provided by each client program. This system can be used in addition to or instead of fetching and analyzing web pages.
By tracking and analyzing sequences of web pages viewed by users, sequences of products viewed by users on those web pages can be accumulated in a database. These sequences of viewed products can be used to generate a similar items table 60 (
Although this invention has been described in terms of certain preferred embodiments, other embodiments that are apparent to those of ordinary skill in the art, including embodiments that do not provide all of the features and benefits set forth herein, are also within the scope of this invention. Accordingly, the scope of the present invention is intended to be defined only by reference to the appended claims.
In the claims which follow, reference characters used to denote process steps are provided for convenience of description only, and not to imply a particular order for performing the steps.
This application is a continuation of U.S. application Ser. No. 10/050,579, filed Jan. 15, 2002, which claims the benefit of Provisional Application 60/343,797 filed Oct. 24, 2001. The disclosures of the aforesaid applications, and of U.S. application Ser. No. 09/821,826, filed Mar. 29, 2001, are hereby incorporated herein by reference.
Number | Date | Country | |
---|---|---|---|
60343797 | Oct 2001 | US |
Number | Date | Country | |
---|---|---|---|
Parent | 10050579 | Jan 2002 | US |
Child | 12041498 | US |