The present disclosure relates to personalized, generalized content recommendation systems and methods.
The ubiquitous presence of the Internet in the modern era has placed the world of information and entertainment at our finger tips. Users can access vast resources of infotainment at any time with nearly minimum effort so that almost any need for information can be instantly answered. In view of the vast resources available and the busy life styles, the users' attention is highly fragmented between disparate information sources so that it is difficult for any one information source to hold the user's attention for a considerable time period.
This disclosure provides for personalized, generalized recommendation systems and methodologies that facilitate determining relevance and recommending relevant content selected from different public and private data. In one embodiment a method of providing content recommendations is disclosed. This method includes receiving, at a processor, a plurality of content items obtained from content streams of different content types to be forwarded to at least one user. In an embodiment, the content streams of different content types are received from multiple content sources. Attributes of the received content items are obtained and content items to be forwarded to the user are obtained based on the attributes. In an embodiment, the content items determined to be relevant based on the attributes are forwarded to a user model associated with the user. The content items forwarded to the user model are scored and content items determined to be relevant based on the score are added to the lists of unseen content items. At least a first subset of the scored content items are added to respective lists of unseen content items for the user based at least on the type of content in each content item of the subset. When a request is received from the user for relevant unseen content items, the first subset of content items are rescored based on a current context of the user. Top scoring content items are selected from the first subset of content items and forwarded to the user. In an embodiment, a user selection of a content item from the selected top scoring content items is received by the processor that recalculates the relevance of the user selected content item to user models of other users on receiving the user selection.
In another embodiment, a computing device, comprising a processor and a storage medium for tangibly storing thereon programming logic associated with the personalized, generalized, recommendation system, for execution by the processor, is disclosed. The programming logic comprises content item receiving logic, executed by the processor, for receiving a plurality of content items obtained from content streams of different content types to be forwarded to at least one user. Scoring logic is executed by the processor, for scoring at least a subset of the plurality of content items. Adding logic, executed by the processor, adds at least a first subset of the scored content items to respective lists of unseen content items for the user based at least on the type of content in each content item of the subset. User request receiving logic, executed by the processor, configures the processor to receive a request from the user for relevant unseen content items. The first subset of content items are rescored based on a current context of the user upon receiving the user request by rescoring logic, executed by the processor. Selecting logic, executed by the processor, selects top scoring content items from the first subset of content items and the selected top scoring content items are forwarded to the user in response to the user request for the relevant items by forwarding logic, executed by the processor.
A computer readable storage medium, having stored thereon, instructions for execution by a processor is disclosed in accordance with another embodiment. The instructions cause the processor to receive a plurality of content items obtained from content streams of different content types to be forwarded to at least one user and score at least a subset of the received content items. The instructions further cause the processor to add at least a first subset of the scored content items to respective lists of unseen content items for the user based at least on the type of content in each content item of the subset. When a user request for relevant unseen content items is received, the instructions further cause the processor to rescore the first subset of content items based on a current context of the user. In addition, the instructions also comprise instructions that configure the processor to select top scoring content items from the first subset of content items and forward the selected top scoring content items to the user in response to the user request.
In one embodiment, a method of providing content is disclosed. The method comprises receiving, at a processor, a request from a user for a predetermined number of unseen content items relevant to the current user context and transmitting, by the processor, the user request to a recommendation server. A list of seen content items is obtained by the processor in response to the user request, and displayed on a display device. The list of seen content items comprises unseen content items relevant to the current user context. In one embodiment, a combination of the list of seen content items and at least one content item displayed prior to receiving the list of seen content items can be displayed.
In an embodiment, a method of generating a user model in terms of categories of interest to the user and providing recommendations based on such a user model is disclosed. The method commences with a processor receiving, a user request for content recommendations along with information regarding the user which can comprise at least a subset of content sources from which the user desires content recommendations. In an embodiment, the information regarding the user can also comprise information associated with prior user activity, such as, previous user searches, selections of content or prior user feedback. Based on such user information, a plurality of categories are provided for selection to the user. The user selections of categories are received and the respective category vectors of the user selected categories are aggregated. A user model representing the user's interests which is employed in making relevant content item recommendations is generated from the aggregated category vectors in combination with vector representations of the user information. In an embodiment, keywords, entities, content item features and other user information can be represented as vectors which are combined with the aggregated category vectors to generate the user model. Based on a determination of relevance by the user model, content recommendations are forwarded to the user. In one embodiment the recommended content items are selected from the subset of content sources wherein each content source provides a respective content type different from other sources.
In one embodiment, the user model is updated based on a user selection of a recommended content item. This can trigger discovery of new content and consequently new content recommendations. In an embodiment, such updates to user model in response to user selections and recommendations of new content can occur in real time. Hence, in response to a user selection of a particular content item, updated recommendations of content items can be transmitted to the user. Such updated recommendations of content items can comprise recommendations for new content items identified by the processor as being relevant to the user in accordance with the updated user model. In an embodiment, updated recommendations of content items can be based on updates to other user models that are similar to the generated user model. In an embodiment, updated recommendations of content items can be based on updates to the category vectors included in the user model.
In an embodiment the user model can be updated on a periodic basis, for example, on a daily basis. User vectors associated with the user model over the preceding ‘N’ days are obtained, N being a natural number, for example, thirty. In a further embodiment, the user vectors of the preceding thirty days are obtained and are weighed by a function of the number of days back and aggregated to generate an aggregated user vector. Current category vectors of the respective categories are combined with the aggregated user vector to generate an updated user vector for the day.
In one embodiment, the user model parameters such as relevance threshold are adjusted based on user behavior and/or system response. In one embodiment, the number of times that a user requests new content is recorded. If the user requests new content frequently, a relevance threshold associated with determination of relevance for providing the recommendations is lowered, such that more content items can be recommended. Conversely, if the user requests new content less frequently, the relevance threshold is increased such that fewer relevant content items can be recommended. In one embodiment, types of content requested by the user can be stored and the user model can be updated such that a greater number of category vectors are aggregated within the user model if the user is requesting greater variety of content or content of different content types.
In one embodiment, a computing device for generating a user model in terms categories of interest to the user and providing recommendations based on such a user model is disclosed. The computing device comprises a processor and a storage medium for tangibly storing thereon program logic for execution by the processor. The program logic when executed by the processor, causes the processor to receive a user request for content recommendations in addition to receiving information regarding the user comprising at least a subset of content sources from which the user desires content recommendations. The processor also executes logic for providing categories for selection by the user and logic for receiving category selections made by the user. Aggregating logic, executed by the processor, aggregates category vectors of respective categories selected by the user. A user model representing the user's interests is generated by the generating logic executed by the processor which generates the user model by combining the aggregated category vectors with vector representations of the user information. Recommendations providing logic is executed by the processor, for providing recommendations of content items to the user based on the user model. In one embodiment, the recommendations providing logic further comprises, logic for selecting the recommended content items from the subset of content sources, wherein each content source provides a respective content type different from other sources. The processor also executes logic for receiving user selection of at least one recommended content item and updating logic, for updating the user model based on the user selection. In addition, the processor also executes logic for updating the recommendations of content items based on the updated user model and transmitting the updated recommendations of content items to the user in real-time.
In one embodiment, the processor also executes logic for providing, updated recommendations of content items based on updates to other user models that are similar to the generated user model. In an embodiment, the processor executes logic for providing updated recommendations of content items based on updates to the category vectors included in the user model. In an embodiment, the processor executes logic for updating the user model on a periodic basis by obtaining, user vectors associated with the user model over prior ‘N’ days, N being a natural number, aggregating, the user vectors of the preceding days weighed by a function of the number of days back to generate an aggregated user vector and by combining, current category vectors of the respective categories with the aggregated user vector to generate an updated user vector.
In an embodiment, the processor executes storing logic, for storing frequency of user requests for new content so that the processor can lower a relevance threshold associated with determination of relevance for providing the recommendations such that more content items can be recommended if the user request new content frequently. Conversely, the processor can increase the relevance threshold such that fewer relevant content items can be recommended if the user requests new content infrequently. In an embodiment, the processor executes logic for storing types of content requested by the user and updating, the user model such that a greater number of category vectors are aggregated within the user model.
In an embodiment, the processor executes explicit training logic, for receiving terms from the user for explicitly adding to or deleting from the user model and item weighing logic, for respectively weighing favorably and unfavorably content items comprising the received terms when providing the recommendations. Additionally, the processor also executes training logic, for suggesting terms from the content items to the user for adding to the user model and/or for content searches that may be issued by the user. In an embodiment, the processor executes user model comparison logic, for comparing the user model to a disparate user model of a second user. Contact suggestion logic, executed by the processor, suggests the second user as a contact to the user based on a similarity of the user model to the disparate user model of the second user as determined by the user model comparison logic. For example, the contact suggestion logic can be activated by the processor, based on the similarity between the two user models crossing a predetermined threshold value.
In one embodiment, a computer readable storage medium, having stored thereon, instructions which when executed by a processor, cause the processor to provide relevant content recommendations is disclosed. These include instructions that cause the processor to receive, a user request for content recommendations in addition to information regarding the user comprising at least a subset of content sources from which the user desires content recommendations. The instructions cause the processor to provide categories for selection by the user and receive, category selections made by the user. The instructions further include those that cause the processor to aggregate category vectors of respective categories selected by the user and generate a user model representing the user's interests. In an embodiment, the user model is generated from the aggregated category vectors in combination with vector representations of the user information and recommendations of content items to the user are based on the user model. In an embodiment, the user information can include information associated with user activity prior to generation of the user model. In an embodiment, the computer readable medium also includes instructions to select the recommended content items from the subset of content sources, each content source providing a respective content type different from other sources of the subset of content sources. In one embodiment, the computer readable storage medium, further comprises instructions to store the recommended content items in a data storage and generate an index on the stored content items that facilitates later user retrieval of the stored content items from the data storage. Relevant content items are retrieved from the stored content items in response to a user query and ranked based on prior user actions associated therewith. The ranked content items are transmitted in response to the user query.
A method for providing content recommendations is disclosed in accordance with one embodiment. The method comprises, receiving, at a processor, a user request for content recommendations. A plurality of content categories for user selection are provided, by the processor. At least a subset of the categories selected by the user are received and based at least on such user selected categories a subset of avatars are generated for further selection by the user. Upon receiving a user selection of an avatar from the subset of avatars, the user selected avatar is associated with a respective user model of the user. The user model can be customized by exchanging communication with the user via the avatar. In an embodiment, the user model can be customized by receiving search terms for whitelisting from the user via exchange of messages between the user and the avatar and generating, standing search queries for each of the whitelisted terms. User feedback to content retrieved via the standing search queries can be monitored and contextual messages can be provided to the user via the avatar based on the user feedback. The standing search queries are updated based on user response to the contextual messages. In an embodiment, a health of the avatar is indicated based at least on interaction of the user with the user model wherein the health of the avatar is indicated by the processor via a color of the avatar, via an expression of the avatar or some other visual representation. In an embodiment, a sentiment of a content item is indicated via an expression of the avatar. In an embodiment, an indication of personalized content can be indicated via a rendering of the avatar.
A computing device comprising a processor and a storage medium for tangibly storing thereon program logic for execution by the processor are disclosed in one embodiment. The program logic comprises request receiving logic, executed by a processor, for receiving a user request for content recommendations and providing logic, executed by the processor, for providing a plurality of content categories for user selection in response to the user request. Category selection receiving logic is executed by the processor for receiving at least a subset of the categories selected by the user. Based at least on user selected categories, avatar generating logic, executed by the processor, generates a subset of avatars for selection by the user. Avatar selection receiving logic, executed by the processor, receives a user selection of an avatar from the subset of avatars and associating logic, executed by the processor, associates the avatar to a respective user model of the user. Additionally, customizing logic is executed by the processor, for customizing the user model by exchanging communication with the user via the avatar.
A method of providing content recommendations is disclosed in accordance with one embodiment. The method commences with a processor transmitting a user request for content recommendations. The processor receives a plurality of content categories for user selection wherein at least a subset of the categories are selected by the user. A subset of avatars for selection by the user are displayed wherein the subset of avatars displayed to the user are indicative of the user selected categories. A user selection of an avatar from the subset of avatars is received and communication from between user model and the user is facilitated via the avatar, wherein the user model is associated with the avatar for example, by displaying messages from the user model to the user.
A computing device comprising a processor and a storage medium for tangibly storing thereon program logic for execution by the processor is disclosed in accordance with another embodiment. The storage medium comprises transmitting logic, executed by a processor, for transmitting a user request for content recommendations. Content category receiving logic is executed by the processor, for receiving a plurality of content categories for user selection and user category selection receiving logic, executed by the processor, for receiving at least a subset of the categories selected by the user. Avatars set display logic is executed by the processor, for displaying a subset of avatars for selection by the user wherein, the subset of avatars are indicative of the user selected categories. Avatar selection receiving logic, executed by the processor receives a user selection of an avatar from the subset of avatars so that communication display logic, executed by the processor, can facilitate communication from a user model to the user via the avatar, wherein the user model is associated with the avatar.
These and other embodiments and embodiments will be apparent to those of ordinary skill in the art by reference to the following detailed description and the accompanying drawings.
In the drawing figures, which are not to scale, and where like reference numerals indicate like elements throughout the several views:
a shows an avatar informing the user about how it can become healthy and the various levels of health in accordance with an embodiment of the present disclosure;
b shows various facial expressions associated with an avatar in accordance with embodiments of the present disclosure;
Subject matter will now be described more fully hereinafter with reference to the accompanying drawings, which form a part hereof, and which show, by way of illustration, specific example embodiments. Subject matter may, however, be embodied in a variety of different forms and, therefore, covered or claimed subject matter is intended to be construed as not being limited to any example embodiments set forth herein; example embodiments are provided merely to be illustrative. Likewise, a reasonably broad scope for claimed or covered subject matter is intended. Among other things, for example, subject matter may be embodied as methods, devices, components, or systems. Accordingly, embodiments may, for example, take the form of hardware, software, firmware or any combination thereof (other than software per se). The following detailed description is, therefore, not intended to be taken in a limiting sense.
In the accompanying drawings, some features may be exaggerated to show details of particular components (and any size, material and similar details shown in the figures are intended to be illustrative and not restrictive). Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting, but merely as a representative basis for teaching one skilled in the art to variously employ the disclosed embodiments.
The present invention is described below with reference to block diagrams and operational illustrations of methods and devices to select and present media related to a specific topic. It is understood that each block of the block diagrams or operational illustrations, and combinations of blocks in the block diagrams or operational illustrations, can be implemented by means of analog or digital hardware and computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, ASIC, or other programmable data processing apparatus, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, implements the functions/acts specified in the block diagrams or operational block or blocks.
In some alternate implementations, the functions/acts noted in the blocks can occur out of the order noted in the operational illustrations. For example, two blocks shown in succession can in fact be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality/acts involved. Furthermore, the embodiments of methods presented and described as flowcharts in this disclosure are provided by way of example in order to provide a more complete understanding of the technology. The disclosed methods are not limited to the operations and logical flow presented herein. Alternative embodiments are contemplated in which the order of the various operations is altered and in which sub-operations described as being part of a larger operation are performed independently.
For the purposes of this disclosure the term “server” should be understood to refer to a service point which provides processing, database, and communication facilities. By way of example, and not limitation, the term “server” can refer to a single, physical processor with associated communications and data storage and database facilities, or it can refer to a networked or clustered complex of processors and associated network and storage devices, as well as operating software and one or more database systems and applications software which support the services provided by the server.
For the purposes of this disclosure a “network” should be understood to refer to a network that may couple devices so that communications may be exchanged, such as between a server and a client device or other types of devices, including between wireless devices coupled via a wireless network, for example. A network may also include mass storage, such as network attached storage (NAS), a storage area network (SAN), or other forms of computer or machine readable media, for example. A network may include the Internet, one or more local area networks (LANs), one or more wide area networks (WANs), wire-line type connections, wireless type connections, cellular or any combination thereof. Likewise, sub-networks, such as may employ differing architectures or may be compliant or compatible with differing protocols, may interoperate within a larger network. Various types of devices may, for example, be made available to provide an interoperable capability for differing architectures or protocols. As one illustrative example, a router may provide a link between otherwise separate and independent LANs.
A communication link or channel may include, for example, analog telephone lines, such as a twisted wire pair, a coaxial cable, full or fractional digital lines including T1, T2, T3, or T4 type lines, Integrated Services Digital Networks (ISDNs), Digital Subscriber Lines (DSLs), wireless links including satellite links, or other communication links or channels, such as may be known to those skilled in the art. Furthermore, a computing device or other related electronic devices may be remotely coupled to a network, such as via a telephone line or link, for example.
A computing device may be capable of sending or receiving signals, such as via a wired or wireless network, or may be capable of processing or storing signals, such as in memory as physical memory states, and may, therefore, operate as a server. Thus, devices capable of operating as a server may include, as examples, dedicated rack-mounted servers, desktop computers, laptop computers, set top boxes, integrated devices combining various features, such as two or more features of the foregoing devices, or the like. Servers may vary widely in configuration or capabilities, but generally a server may include one or more central processing units and memory. A server may also include one or more mass storage devices, one or more power supplies, one or more wired or wireless network interfaces, one or more input/output interfaces, or one or more operating systems, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, or the like.
Throughout the specification and claims, terms may have nuanced meanings suggested or implied in context beyond an explicitly stated meaning Likewise, the phrase “in one embodiment” as used herein does not necessarily refer to the same embodiment and the phrase “in another embodiment” as used herein does not necessarily refer to a different embodiment. It is intended, for example, that claimed subject matter include combinations of example embodiments in whole or in part. In general, terminology may be understood at least in part from usage in context. For example, terms, such as “and”, “or”, or “and/or,” as used herein may include a variety of meanings that may depend at least in part upon the context in which such terms are used. Typically, “or” if used to associate a list, such as A, B or C, is intended to mean A, B, and C, here used in the inclusive sense, as well as A, B or C, here used in the exclusive sense. In addition, the term “one or more” as used herein, depending at least in part upon context, may be used to describe any feature, structure, or characteristic in a singular sense or may be used to describe combinations of features, structures or characteristics in a plural sense. Similarly, terms, such as “a,” “an,” or “the,” again, may be understood to convey a singular usage or to convey a plural usage, depending at least in part upon context. In addition, the term “based on” may be understood as not necessarily intended to convey an exclusive set of factors and may, instead, allow for existence of additional factors not necessarily expressly described, again, depending at least in part on context.
The information age has made available large amounts of data for users to access via different modalities. This data includes both public or private data which is available for access via a computer, telephone, television or any other systems which can include information processors. As more and more personal data goes online and into the cloud, content consumption and other online user activity takes place in a large number of different places and via various devices. A user may employ dozens of different web sites and applications during the course of a day, with all of them competing for the user's attention. Therefore, while a user needs information from a variety of content sources, the sheer volume of information coming from these sources can be overwhelming for any one person to reasonably process, especially within a short time span as the case may be for many customers. Various embodiments described herein provide for recommendation systems and methods that provide recommendations for content from various private and public sources personalized to a user's context that includes the users current needs and preferences. A shortlist of information items can be created that the user might need to know here and now. Alternately, the embodiments provide for a “mother of all feeds” that aggregates and forwards, in real time, relevant, actionable and/or engaging content that matters most to the user.
Turning now to the figures,
In one embodiment, a query-less search procedure can be implemented based on an understanding of a user's needs in real-time, adapting to user's interests changes, location changes and changes in the time of the day by employing a filtering mechanism that is adaptive and personalized on a per-user basis. In one embodiment, a search based on a user query for a particular type of content from a particular content stream or for content items having user-specified attributes from various content streams can be aggregated for presentation to the user.
In order to build the personalized content recommendation system 100 wherein the content is constantly changing, hyper-personally relevant, the following are non-limiting examples of preferred considerations to be addressed:
A. Cold-start: Ensuring that the recommended items are relevant the first time a user begins to use the recommendation system 100 in order to mitigate the users from abandoning usage of the recommendation system 100.
B. Implicit model: Automatically building a robust representation of the users' interests with as little explicit customization as possible since users typically will only invest a small amount of cognitive energy on customizing the recommendation system 100.
C. Learning & forgetting: Adjusting the user model based on user behavior (or absence of behavior).
D. Discovery: Ensuring that the user model does not get “trapped” in a local optimum, and allows the user to discover content on topics they may not have previously seen, including trending topics.
E. Transparency: Representing the user model in such a way as to aid interpretation, as well as supporting a user-friendly view (for example, to let the user know what the model thinks are relevant topics for him/her).
F. De-duping: Detecting duplicate or near-duplicate content, and being smart about when to recommend it and when to not recommend it.
A recommendation system or methodology may address at least partially the aforementioned challenges by combining the following features:
1. An Information Retrieval vector-space model to represent users, articles, and categories, which allows for a simple similarity criterion used for both recommendation and duplicate detection.
2. Leveraging lightweight explicit customization to prime the cold-start model, while also providing a mechanism to support discovery of new topics as well as trending topics.
3. Exploiting simple “Rocchio” style feedback in real-time to update the model based on user behavior.
4. Adapting parameters of the model in real-time to target desired user and system behavior, for example, based on click-through rate of the user or recommendation rate of the recommendation system.
The personalized, generalized content recommendation module 210 receives the content items 252 from the feed manager 212 and processes them in order to determine those content items that are relevant to the user. In an embodiment, the relevance of the received content items 252 can be calculated and a subset of the content items that are determined to be relevant can be stored in unseen items list (not shown) for example, in the recommendations database 214 prior to being forwarded to the user. The relevance of the received content items 252 can be determined by employing a user model that incorporates the user attributes and preferences. The user attributes/preferences can be collected explicitly via user input, for example, when the user initially signs up with the recommendation module 210 and/or implicitly via automated learning mechanisms incorporated into the recommendation system 100 as will be further detailed infra.
In one embodiment, a request 254 for unseen content items that are relevant to the current user context can be received from a user employing the user device 204. The user's current context can be determined based on the information associated with the user request 254, such as, the time at which the request was received or the location information. In an embodiment, the location information associated with the user request 254 can be determined from the user's current GPS coordinates as obtained from the device 204. In addition, the current user context can be also be determined based on social context of the user and its combinations with other context information available from different public/private sources such as, for example, the user's calendar. Based on the current user context information gathered by the recommendation module 210, a relevance score of the items stored in the unseen items lists for the user of the device 204 is determined. As shown in
In an embodiment, the received content items 252 can be stored in a database 214 without indexing as their relevance to the different users is calculated in real time even as they are received by the recommendations module 210 from the feed manager 212. In one embodiment, the recommendations module 210 can further implement a feedback mechanism in real time. The recommendations module 210 receives a user's selections of the items in the seen items list 256 and employs the information associated with such selections in filtering content items to be forward to that user. Thus, for example, based on the user selections, the seen items list 256 can be refreshed in real time. Additionally, it can also affect, in real-time, the content items forwarded to other users related to the user such as those that are connected to the user in an online social network or those that the recommendation module 210 has determined have similar profiles as the user making the selections.
In an embodiment, the recommendations module 210 can function as a centralized storage for different types of personalized content consumed by a user or it can store the content items forwarded to the user. The stored contents can be indexed in order to facilitate search and retrieval of specific content items. In an embodiment, the recommendations module 210 can also store not only content recommended to the user, it can also store and index content that is under consideration for recommendation regardless of whether or not it gets recommended. In this context, the recommendations module 210 can provide an uber-index of all of the personally relevant content for a user. In an embodiment, the recommendations module 210 returns search results from such stored content which are ranked based on past user actions. When issuing a keyword search against the items indexed by the recommendations module 210, the recommendations module 210 takes into consideration the user's past actions/interactions with the items when determining their rankings so that, for example, items that have been read or otherwise provided positive feedback for example, will be ranked higher.
In an embodiment, some content that is particularly urgent for a user can be recognized and forwarded to the user so that the recommendations module 210 acts as a notification platform. The content that is pushed can comprise for example, alerts associated with weather, stock prices, traffic.
In an embodiment, an advertisement server 220 can serve context sensitive advertisements to be displayed on web pages or mobile applications associated with the recommendations module 210 based at least in part upon one or more terms associated with the requesting user as will be detailed further infra. Although the advertisement server 220 is shown in this embodiment as located on the same server computer 206 as the recommendations module 210, it can be appreciated that this is not necessary. The advertisement server 220 can also be located with the feed manager 212 or it can also be located independently on an external machine that is disparate from both the server computer 206 and the feed manager 212. In an embodiment, an “ad server” can comprise a server that stores online advertisements for presentation to users. “Ad serving” refers to methods used to place online advertisements on websites, in applications, or other places where users are more likely to see them, such as during an online session or during computing platform use, for example. Advertising may be beneficial to users, advertisers or web portals if displayed advertisements are relevant to interests of one or more users. In an embodiment, advertisements can be presented to users in a targeted audience based at least in part upon predicted user behavior(s) or user profile information.
Each time a new piece or item of content is received by the recommendation module 210, a representation of the item or an item processing element 312 of the content item is generated by the item processing module 310. As described supra, the item processing element 312 can comprise data 3120 and code 3122 associated with events that an item processing element 312 is configured to process. The data 3120 included in the item processing element 312 can comprise the keywords or features associated with the content item represented by the item processing element 312. In addition, the data 3120 can comprise attributes of the content item which can include, for example, the title, abstract, source, author, or location associated with or referred to in the content item, in addition to other characterizing and/or statistical data. By the way of illustration and not limitation, the statistical data can include data that determines the popularity of the content item which can further comprise the number of users to which the content item was determined to be relevant, the number of users who actually clicked on the content item, the number of users who liked/disliked the content item or the number of users who performed other action such as storing, deleting or forwarding the content item to social contacts. In one embodiment, the item processing element 312 can also store the statistical data associated with each user action, like the number of times the content item was saved/deleted/forwarded so that the relevance of the content item represented by the item processing element 312 can be better determined relative toss the users of the recommendations module 210.
In an embodiment, users of the recommendations module 210 are represented by the user processing elements 332 in a manner similar to the content items. Thus, each active user of the recommendations module 210 is represented by a respective user processing element. When a new user registers with the recommendation module 210, information related to the user is received and processed by the user processing module 330 which generates a representation of the user as the user processing element 332. As described herein, the user processing element 332 comprises data 3320 and code 3322 associated with the user it represents. The data 3320 can include an initial user model 3324 that is generated from the information provided explicitly by the user, for example, while signing up to use the recommendation module 210 and implicit information gathered by observing the user's interaction as will be detailed further infra. The user model 3324 thus generated is further constantly updated based on user interaction with the recommendations module 210. This is achieved by configuring the user processing element 332 to listen for actions taken by the user it represents, e.g., clicking on a given item or forwarding items to contacts, voting (“like” or “dig”) for the items or other user action that is recorded and can be employed to update the user model 3324.
In addition to the aforementioned modules, the recommendation module 210 also includes a feature processing module 320 that maintains a correspondence between users 254 and the received items of content 252. When an item of content is received, features such as keywords and/or metadata as described herein regarding the item are extracted, for example, via natural language processing or other techniques. A feature can generally include any piece of data derived from an item or a user and used as input to the filtering decision. Features can include explicit data of content items such as words, or entities. By the way of illustration and not limitation, entities can comprise information associated with the content item such as a source of the content item, author of the content item, a location, a category or a sentiment associated with the content item. They can also include implicit data such as inherent attributes of a content item, for example, an encoding format of a video item. For each feature that is extracted, a feature processing element 322 comprising data 3220 and code 3226 is generated by the feature processing module 320. For each of the recognized features, the feature processing module 320 maintains, for example, in the data 3220 of the corresponding feature processing element 322, a list of content items 3222 associated with a given feature e.g., a word or phrase and a list of users 3224 who express interest in the given feature. For example, for a feature, such as a word “microhybrids”, the feature processing module 320 maintains a list of a subset of the received content items 252 which include or are otherwise associated with the word “microhybrids” and a list of active users of the recommendation module 210 who have either implicitly or explicitly expressed interest in the content items that include the term “microhybrids”. The code 3224 can further comprise modules for communicating with the item processing module 310 in order to update the list of items 3222 as new items are received by the recommendations module 210. The code 3226 can also comprise modules for communicating with the user processing module 330 in order to update the list of users 3224 so that users can be added or removed from the users list 3224. In an embodiment, only public items such as news items are analyzed by the feature processing module 320, whereas a user's private items such as, emails, are forwarded directly to the user PE 332 for scoring.
When a new item is received by the recommendations module 210, it is analyzed by the item processing module 310 in order to generate the corresponding item processing element 312 and to extract various features from the newly received item that will be forwarded for further analysis to the feature processing module 320. In different aspects, the feature processing elements can be generated from both public and private items based on, for example, user permissions. For each feature obtained from the received content items 252, the feature processing module 320 determines if a feature processing element exists that corresponds to each of the features thus obtained. For example, if the received feature is associated with the feature processing element 322, the newly received item is added to the list of items 3222 comprised within that feature processing element 322 and the newly received item is forwarded to all the users in the users list 3224 comprised within the feature processing element 322 for further scoring. If a feature processing element does not exist for a received feature, it is generated by the feature processing module 320 and the users and items lists are generated and they can be updated as new data is received by the recommendations module 210. In one embodiment, a feature processing element can also be generated from user input to the recommendations module 210. In this case, it is assumed that the feature processing element does not exist within the recommendations module 210 because if there was a feature processing element for the user input feature, then the user would have been added to the users list of the corresponding feature processing element. However, in this embodiment, a feature processing element is generated from user input and its list of items and users is expanded as the feature gains popularity among the users at large and is used more often in published/public content. The feature processing elements, therefore, act as gates that sort the various content items that need to be distributed to the different users based on their attributes. The feature processing elements act as go-betweens executing a firstpass matching of items with users. They can act as a dynamic, always-up-to-date index and thus mitigate the need for scoring each content item by the user model 3324. In one embodiment, only public data such as news items or public blog posts are processed by the feature processing elements whereas private data such as emails intended for a particular user bypass the feature processing elements and are transmitted directly to the user processing element 332 for scoring.
In accordance with an embodiment the different processing elements are generated by modeling the users, categories/features and items of content within the recommendations module 210 as vectors. The feature space for user, item, and category vectors consists of terms (single words or bi-grams) that have been stripped of some special and non-ASCII characters, and stopped (common, non-meaningful words are ignored, and bigrams with one or more stopwords are also ignored). Depending on what field of the item they are taken from, they may be lowercased as well. All vectors V are kept normalized by setting Vi=1 for all Vi>1, any time a new vector is computed (this ensures the highest weighted feature has a value of 1). Furthermore, the number of non-zero entries in each vector is limited to some number (e.g., 1000 for user and item vectors, 10,000 for category vectors). Vectors are also normalized by setting their lowest-value entries to 0 when this maximum is exceeded.
In an embodiment, every item or article is preferably represented by a vector: Dg, c, d where the item has a guid (global unique identifier) g, is possibly labeled with a category c, and was published/received on day d. A modified version of tf-idf (term frequency—inverse document frequency) weighting is used to populate the item vector. Specifically, term frequency is the sum of occurrences of the term in title or snippet (with unigrams in the title counting twice), and the document frequency term is:
Additionally, when an item has an author or source, that string is added as a feature with a weight of 1. All new users are asked to select from a list of preferred categories, for example, from a list of eight possible categories. The initial user model for user x who has selected a list of categories C, is simply a term vector that is taken from a set of cold-start vectors: Vc. These cold-start vectors are computed on an ongoing basis. By the way of illustration and not limitation, at ten minute intervals, the vectors for the hundred most recent items for a category are summed and added to Vc, which is then decayed using a constant that equates to a 0.8 daily multiplier.
In an embodiment, category vectors Vc are therefore, used as both the initial user vector, as well as a component of the evolving user vector. This provides a direct way to give the users reasonable content recommendations for the outset thereby addressing the cold start challenge, with an implicit model that requires little input from the user. It also directly helps set user expectations by making explicit to them what types of content they can expect to receive as recommendations. As the category vectors Vc are computed continuously, they always represent the recent content in any category most strongly, and thus, as part of the user vector, they help make sure users are able to discover new, trending content thereby ensuring that a user model does not get “trapped” in a local optimum. Furthermore, since different users may use the recommendations module 210 in different ways. Some users may want hyper-relevance while others may expect more discovery. An explicit parameter αx as detailed further infra to trade-off how much of the category vectors are used in the user vector and thereby will allow the recommendations module 210 (or the user) to adjust this property.
The data 3320 also includes user preferences 420 that are collected from the user, for example, via a settings screen. These can include, for example, the privacy setting of the user, the themes to be used for a user interface, the information to be displayed for various screens of the user interface and other user preferences. In addition to simply providing the stream of recommended content, one embodiment of the recommendations module 210 also provides for “lenses” which are filters based on different content attributes. The user can therefore, “drill down” on specific attributes of recommended content, so they see only trending content, content from a particular publisher or author, content from a specified time period in the past or future, content associated with a given location, content containing specific keywords, content based on sentiment (positive, negative, happy, sad, shocking, etc.) and content based on any other meta-data. In an embodiment, such lenses can be associated with specific content sources under the user preferences 420.
Again, it may be appreciated that the user settings are listed herein by the way of illustration and not limitation and that other user settings/attributes obtained explicitly or implicitly can be included in the data 3320 as user settings. In one embodiment, the user settings can also include the user's avatar selection. As will be described further infra, the avatar of a user is employed to represent the user's user model and a user's selection of an avatar can be a feature of the user model 3324 which is employed in making personalized, generalized recommendations to the user.
The user model 3224, also comprised within the user PE 332, generally includes all of those parameters of the recommendations module 210 that are used on a per-user basis to make the filtering decision for an item. It is employed to score and recommend items to a user that it represents. As described supra, the user model 3324 is built using the Vector Space Model which is a list of terms (words and phrases) or other features, with associated weights and modified dot-product for measuring similarity. Representing the user in a manner similar to which the content items are represented, which is in terms of words and phrases, makes users and items somewhat interchangeable leading to many uses as will be discussed further herein. Since word-based features are used to represent the user, parts of the user models can be exposed to the users themselves which allows direct user manipulation or otherwise provides insights which can be useful in improving recommendations made to the users. In fact, modeling the user on word-based features can also assist the user with other search activities. For example, when the user is issuing a keyword search (either within or outside of the recommendations module 210), terms from the respective user model can be used as search suggestions or as input to the ranking functions so that content items including terms from the user model 3324 are ranked higher in the result set.
In one embodiment, user model 3324 comprises, among other data, the following parts:
It may be appreciated that other user attributes or item features that are used to determine relevance of content items to a particular user can also be part of the user model 3324 in accordance with other embodiments. As discussed supra, the initial user model for a user x who has selected a list of categories C, is a term vector that is taken from a set of cold-start vectors Vc. In an embodiment, the category vectors included in a user model 3324 are based on the high-level categories, for example, News categories, which are specified as being of particular interest to a user when the user signs up for the recommendation module 210. The vector representation of each category is maintained by a real time process based on the content items published in that category. A combination of these category vectors based on user preferences is used as the users' initial user model.
Another source for cold-start models are the user's activities on the accounts they have linked to the recommendations module 210. For example, if a user links their FACEBOOK account or other social network account(s) in order to receive recommendations from their FACEBOOK stream, the recommendations module 210 (depending on what gets exposed by FACEBOOK) can gather more important features (like terms) from that account, and add them to the user's model. As an example, words from the user's posts on FACEBOOK or from the posts of entities they follow, from their profile, from what they have “liked”, can be retrieved in order to be included as part of the vectors used to build up the user model 3324. In general, any activity on any service the user has linked can provide features (not limited to terms) that may be added or removed from their initial model. In an embodiment, the user model 3324 can be configured to include content the user has viewed prior to registering with the recommendations module 210. This content can be obtained for example, via the user's search history and the content, such as search terms for the user's search history is also aggregated to be included in the user model 3324. Thus, in the case of new users for whom a user model has never been generated, an initial model is built from initial explicit user input and other explicit or implicit information available via various sources so that the possibility of providing poor recommendations and thereby discouraging the users from using the recommendations module 210 is mitigated. This is the so-called “cold-start” challenge, and it also encompasses the ongoing challenge of making sure that the user model does not get “trapped” in always recommending the same type of content.
In a typical information retrieval model, the range of values for the elements in a vector are taken from the interval [0 . . . infinity]. In one embodiment, the user vector U is a special vector called a “Max One” vector, which differs from a typical representation. A list of terms 430 along with their respective weights that can be used to determine the user vector U is shown in accordance with one embodiment. The weights of the user vector (which is a “Max One” vector) can be associated with a specific lower-bound, a parameter called MIN_WEIGHT. Only real-valued weights between MIN_WEIGHT and 1 are allowed for the “Max One” vector, with any operations that cause the weight to go above 1 will instead reset it to 1. For example, the vector values can be limited to the range [0 . . . 1]. Any operations that cause the weight to go below MIN_WEIGHT instead cause the term to be removed from the list. A fixed most important terms to represent a user's interests and which, if some operation causes to grow longer, will automatically remove the lowest weighted terms so as to stay within the maximum length. This version of a vector space model has the effect of naturally supporting two of the abilities required by the recommendation module 210. By limiting the maximum weight to 1, it is ensured that the recommendations module 210 never overweighs a given term for a user, which can result in over-emphasizing content about a specific topic. In addition, by automatically having terms “fall off the end” of the vector, the user model 3324 naturally “forgets” terms that do not otherwise get reinforced by the user so that the possibility of entries growing unbounded as a result of feedback is mitigated. For example, a user may have read an article about soccer at one time, which does not necessarily indicate a continued, deeper interest in soccer. In fact, if this interest is not reinforced by further user selections of content items related to soccer, it is likely that the user was only briefly interested in that particular content item. Therefore, assigning a minimum weight to a term and removing terms that go below the MIN_WEIGHT, automatically keeps finite, the list of terms associated with the user so that only highly rated terms are maintained in the user model 3324 and a user's fleeting or less enduring interest in a particular category/term is forgotten and not maintained within the user model 3324. This mitigates the possibility that content items associated with such passing interests are discarded as less relevant and are not forwarded to the user.
The next parameter in the user model 3324 is the threshold parameter θx, which is used to determine if an item is relevant enough to be shown in the user's item list. In addition to the daily updates and feedback updates to the user vector portion of the model, some parameters are also adjusted dynamically. Specifically, the threshold parameter θx is adjusted on a periodic basis to regulate the observed article recommendation rate or the number of articles added to the user's item list per unit time. In one embodiment, the threshold parameter θx is adjusted to achieve a substantially optimal item type mixture weights wherein content items from various types of content streams like news items, social networking feeds, emails or other content types are selected for presentation to the user based on feedback received from the user. This is done by pre-defining a target rate, and then periodically adjusting θx up or down using small increments depending on whether the recommendation rate is above or below this target. Thus, the threshold parameter θx for a given user is not a fixed number, but is dynamically adjusted based on feedback received from the user. In an embodiment, the recommendations module 210 can track the number times the user requests new content and adjusts θx up or down accordingly. Thus, the threshold parameter θx is lowered if the user is requesting new content frequently so the more content is determined to be relevant to the user thereby fulfilling the frequent user requests. Conversely, the threshold parameter θx is raised for a user who does not request content frequently so that the recommendations module 210 maintains a more stringent standard of relevance for recommending items to such a user. In addition, the next parameter of the user model, the explore/exploit parameter αx is also adjusted in a similar manner, depending on observed click through rate as will be described herein. In one embodiment, a mechanism can be included to let users directly manipulate both θx and αx to be able to control the flow of recommendations by themselves.
As described supra, the initial user model for user x who has selected a list of categories C, is simply a term vector that is taken from a set of cold-start vectors: Vc which are computed on an ongoing basis. In an embodiment, on a daily basis, the user vector is updated or recomputed as:
and αx is a user-specific parameter which controls the amount of user-specific features versus cold-start features; Nx is the number of days of history available for user x, and C is the set of category preferences for user x. In an embodiment, this update is done nightly by combining the user model 3324 of the preceding thirty days and the current cold start categories of vectors included in the user model 3324 or when the user changes their category preferences. The term wi is the weight that will bias more recent user models over older user models and 1/Nx is used because in the initial thirty days that the user registers with the recommendation module 210, thirty snapshots of the user model are not available of updating the user vector. Thus, at any given point in time the user vector is a sum of two components: a weighted sum of the user vector across time (emphasizing recent vectors)—this is the ‘exploit’ component and a weighted sum of category vectors across time (also emphasizing recent vectors)—which is the ‘explore component’. Therefore, if the user associated whose user vector is determined in accordance with Expression (1) above, requests content more frequently, the user vector is adjusted by lowering αx on the assumption that the user likes the recommended content and therefore the recommendations module 210 is serving the relevant content. On the other hand if the user does not request content frequently, the user vector is adjusted to be weighed more towards the exploitation part of Expression (1). In real-time the user vector is then adjusted using simple “Rocchio” algorithm as will be described further infra.
In an embodiment, when a user x indicates that item D is relevant (for example, by clicking on it, or “liking” it), the user vector is updated as:
U
↓(χ,C)(γU↓(χ,C)±δD Ex. (3)
where γ and δ are two global parameters, which, in an embodiment are set as γ=0.99, δ=0.85. In an embodiment, δ takes positive values for all positive forms of feedback (item clicks, shares, saves, thumb-ups, and maximizes) and a negative value for negative feedback (deletes and thumb-downs). However, the user model 3324 can be extended to use different δ or different types of feedback, and in accordance with other embodiments it may be different on a per-user basis or may be adjusted dynamically.
In an embodiment, the relevance score of a particular item to a particular user is normalized, squashed dot product given below in expression 4:
where F is a “fudge factor” that depends on the content type of the item, when the item was published, as well as the number of positive and negative feedback signals the item has received from the overall user population. The tan h is used to ensure the score stays in the [0 . . . 1] range, as well as to shape the score and spread out the distribution. In one embodiment, mail items have a slightly different relevance score as shown below in expression 5:
It can be appreciated that all constants have been manually determined by examining data, and are not necessarily optimal and may vary in different embodiments.
Various parameters detailed herein thus aid in optimizing the user model 3324. These operational parameters can be adjusted based on observed behavior of the user and the recommendations module 210. Other recommender systems use traditional machine learning techniques to build classifiers targeting a prediction of probability of click with the hope of maximizing click through rate. The recommendations module 210 can monitor key system metrics on a per-user basis (e.g., click rate, recommendation rate, “gimme” rate, etc.) and then adjust some key parameters (namely, the threshold used to determine recommendation θx and the parameter used to mix cold-start with historical models αx) based on the user's behavior in real-time. Thus, for example, if the user is requesting new items often, their threshold θx is lowered so they get more items, or if the users are clicking on a lot of items, the cold-start part of their model (the term including 1−αx) is weighed more heavily in order to give them a more diverse set of items.
The user PE 332 also includes various code 3322 modules that enable it to handle different tasks related to updating the user model 3324 and thereby providing relevant content to the user. One of the various functions that a user PE 332 can execute includes determining a cold-start model for the user. As discussed previously, when a user initially registers to use the recommendations module 210, providing accurate recommendations to the user can be a challenging task based only on the minimum user preferences that are normally collected by recommender systems. Hence, the recommendations module 210 collects information regarding specific content categories that a user is interested in and defines the initial user model based on the user selected categories so that the user receives substantially relevant recommendations even during the initial stages of interaction with the recommendations module 210. Subsequently, the initial user model is updated and fine tuned based on explicit and implicit feedback from the user as described in accordance with different embodiments herein.
Another function that the user PE 332 is configured to handle is de-duping, scoring and filtering incoming items. When the recommendations module 210 receives a new content item, it is forwarded to the user PE 332 based on an initial determination of relevance provided by the feature PEs associated with the new item. Upon receiving the new item or a vector representation of the new item, the user PE 332 initially determines whether the newly received item is a duplicate of a previously received content item. This can be determined, for example, using cosine similarity of the received item to items in seen, unseen and deleted caches. If it is determined that the new item is not a duplicate, the user PE 332 proceeds with scoring and filtering the item in accordance with various relevance criteria associated therewith.
The user PE 332 is also configured to expose certain features of the user model 3324 to the user for explicit training. By the way of illustration and not limitation, features for explicit training can include selected uni/bigrams that are in the item, biasing towards uppercased, author/source, words in title. As the recommendation module 210 models users in the feature space of words and phrases, parts of the user model 3324 can be exposed to the user for fine-tuning so that more accurate recommendations can be obtained. It may be appreciated that explicit training of the user model by the user 3324 can be optional and that the user model 3324 can implement machine learning techniques to automatically fine-tune the user model 3324 based on implicit feedback obtained from different sources including the user actions and/or actions of other users who may share similar interests and hence who have similar user models.
The user PE 332 is designed to service client requests for content items. The user PE 332 can make a determination of relevance each time it obtains a new article and additionally determine relevance between various content types upon receiving user request. Based on the predetermined item type weights associated with each content type, the user PE assembles a list of seen items to be forwarded to the user in response to the user request.
A user PE 332 forwards a list of seen items to a user and records user actions on the content items in the received list. Events generated by user actions such as, clicking on a particular content item, are received by the user PE 332 and employed to update the user model in real time In addition, the user PE 332 also functions to update the user model on a periodic basis in accordance with various embodiments as detailed further infra.
In an embodiment, the user PE 332 can be used to locate context sensitive, targeted advertisements for presentation from the user. For example, information from the user model 3324 can be employed in identifying targeted advertisements for presentation to the user along with the content recommendations. In an embodiment, different types of advertisements which are relevant to information in each of the content items in the aggregated content stream 130.
Moreover, the updated category vector, for example, the updated health category vector described above is also folded into the user models of the users who indicated their interest in the health category. Thus, the user models of all the users are automatically updated with the new features from the health category 530 in real-time and/or on a periodic basis.
At 614, the item of content is forwarded to the user PEs associated with the users in the users list comprised within each of the feature PEs. The users list associated with a feature PE is assembled based on explicit or implicit user input. For example, the user may have explicitly added a particular keyword to the list of terms 430 in the user PE 322 or the user may have provided an implicit input via selecting or clicking on articles associated with the keyword corresponding to the feature PE. Therefore, the user has been included in the users list of the feature PE corresponding to the keyword. As discussed supra, the feature PE is configured for a first-pass estimation of relevance of the received content items to the users so that the user PEs of each individual user is not overwhelmed with content whose relevance is to be determined. The received content of item is again scored at the user PEs to determine relevance prior to being forwarded to the users as will be detailed further infra.
At 804, the unseen items cache is accessed. The cache of unseen items comprises of those content items that are determined to be relevant to the user upon scoring by the user PE 332. In an embodiment, the cache comprises separate list of unseen items for each content type. Accordingly, the lists can comprise by the way of illustration and not limitation, an unseen emails list that comprises, for example, the ten latest, unseen emails that are determined to be relevant to the user, an unseen news items list which can comprise latest, relevant news items which were not seen by the user, a list of latest, unseen, FACEBOOK posts, and a list of latest, unseen tweets from TWITTER. At 806, the content items are rescored in order to determine their relevance to the user based on, for example, current user context in accordance with embodiments as described herein. At 808, the relevance score of the content items obtained at step 806 is compared with the relevance threshold and a predetermined number of content items that still exceed the relevance threshold are selected at 810. The selected content items that exceed the relevance threshold are sorted by the temporal metadata associated therewith at 812. In one embodiment, the temporal metadata can be a time at which the particular content item was received. In one embodiment, the temporal metadata can be a time at which a particular content item was published. The selected content items can either be sorted from the latest to the oldest content item or conversely from the oldest to the latest content items based on, for example, user preferences associated with different types of content items. The recommendations module 210 can permit the user to select the sort order so that the user may select different sort orders for different content types. For example, for private data such as emails, the user can choose to receive the oldest unseen relevant email first whereas for public data such as news item types, the user can choose to receive the latest unseen news item first.
At 814, the item mixture weights for different content types are determined. As described herein, the recommendations module 210 takes a wide variety of content types as input and determines not only how to score them but also how to ensure that there is a reasonable amount of diversity in the recommendations or content lists forwarded to the user. In an embodiment, the fudge factor F described supra with respect to expression (4) is employed for scaling the scoring function based on content types. In addition, each major content type (e.g., News, FACEBOOK, TWITTER, Email) has a target percentage associated with it, which varies over time as the user interacts with the recommendations module 210. By the way of illustration and not limitation, the content types that can be included in the recommendations module 210 can comprise: Email (from YAHOO! or any other email provider), FACEBOOK Newsfeed (personal), GOOGLE+ feeds (personal), Horoscope, LINKEDIN News, status updates, Local News, Local deals and ads, Local events, Local points of interest, Local traffic reports, Local weather and weather alerts, RSS(Really Simple Syndication) feeds (user selected), RSS feeds (curated), Stock market News (from personal portfolio), Stock market alerts (from personal portfolio), TWITTER accounts (curated), TWITTER stream (personal), YAHOO Answers (personal question and answers), YAHOO Fresh (web content from trends on Twitter), YAHOO News (by category), YAHOO! News For You, YAHOO! News friend activity (aka Social Chrome), YAHOO! content sites (Shine, OMG, Green, etc.), YAHOO Groups (personal), photos from photo-sharing sites. When selecting which items to recommend, the user PE 322 walks through each content type (starting with the one with the largest target), and tries to recommend enough items from that type that pass the relevance threshold so that the ratio is preserved. If it cannot recommend the requisite amount of a particular content type, it then readjusts the target percentages of the remaining content types so as to maintain their relative ratios. Furthermore, as the user provides positive or negative feedback for an item of a given content type, its percentage is adjusted upwards or downwards by a small amount, so that for example, interacting positively with a given type will increase its prevalence in the items recommended. There are both minimum and maximum limits on the percentages, so no one type comes to dominate the stream of recommended items.
At 816, a target number of items for each of the item types can be selected to forward to the user based on the item type weights and the number of new content items requested by the user as described herein. Based on the item weight associated with each content type, a ratio of the number of content items of each type to be included in the mixture of item types forwarded to the user is determined. Again, as described herein various content items including public and private data such as emails, new items, alerts, advertisements, social networking feeds and input from other selected sources can be included in the content sent to the user. Thus, if the target ratio to include a particular type of content in the resultant items list to be forwarded to the user is less, there will be less number of that particular type of content items in the resultant items list forwarded to the user. Conversely, if the target ratio to include a particular type of content in the resultant items list to be forwarded to the user is higher, there will be more of that particular type of content items in the resultant items list forwarded to the user.
The resultant items list with the selected number of each item type is forwarded or transmitted to the user as a seen items list as shown at 818. In an embodiment, it can happen that the number of new content items ‘X’ determined to be relevant to the user upon rescoring at 806 is less than the number of new content items ‘N’ requested by the user. In this case, either only the relevant new content items ‘X’ as requested by the user are displayed to the user on a user device or a combination of the ‘X’ relevant new content items and top ‘N-X’ content items previously viewed by the user can be displayed to the user.
If the user, upon receiving the content item D 502 in the seen queue 960, clicks on it, the click event 962 is received by the user PE 332 and used as a feedback to update item mixture weights in real time as shown at 964. In one embodiment, as the item D 502 was a news type item, the News Weight associated with the news content type is increased so that subsequent seen item lists forwarded to the user will be biased to include a greater number of news type items. Conversely, if the user had not clicked on the content item D 502, such information is also received by the user PE 332 and employed to update the mixture weights 964 so that less number of content items of the content type that remained un-selected by the user are included in the subsequent seen lists that are forwarded to the user.
Additionally, the feedback regarding the user selection of the content item D 502 is employed in updating the user vector in real time as shown at 966. Thus, actions that the user takes on items translate into real-time feedback using a standard technique called Rocchio feedback, wherein the item's vector, in this instance vector 540 is added or subtracted from the user vector 944 after multiplying it by some weight. If the user clicks on the content item D 502, the user vector 944 is combined with the representation of the content item D 502, the information regarding the click event, for example, the temporal metadata associated with the click event, and the user preferences in order to recalculate the user vector. Therefore, each time a user selects an item of content, the event generated by such selection affects the user model 3324 in real time. In an embodiment, the selection information of the user can also affect other user models which are similar to the user model 3324 so that the content items forwarded to the other users are also affected in real-time. In an embodiment, in addition to being updated in real-time, the user vector can also be updated on a periodic basis, for example every twenty four hours, as will be detailed further herein.
In addition to updating the user model 3324, the feedback regarding the user click events can also be employed for content discovery. For example, any new terms that get added to a user's model 3324 as a result of feedback immediately (in real-time) cause their user PEs to re-evaluate any item in the recommendations module 210 that contains one of the new terms. This has the effect of making the recommendations module 210 extremely responsive to the user's current interests, by recommending related items to what the user has just read or otherwise interacted with. Furthermore, actions taken by individual users on publicly available items are noted by the Item PE, which resends itself to all interested users for rescoring, for example, via the users list associated with the various feature PEs which include that particular item in their list of items. And because the aggregate actions of other users of the recommendations module 210 are also used when scoring an item, this can result in, for example, more popular items being more likely to get recommended as they become more popular. Also, whenever the user model changes, items that have previously passed the threshold for recommendation are re-scored to ensure they still pass.
In addition to being updated with the cold start category vectors on a periodic basis, the user vector 944 is additionally aggregated with weighted user vectors over the previous thirty days as shown at 1006. Snapshots of the user model are obtained on a periodic basis (for example, daily), and kept for some time into the past. When folding the cold-start model in, an aggregate of the user's model over the past (weighted exponentially for example, to emphasize recent user models) is also folded in. In this way, long-term interests are kept in the user's model, so that even if they haven't interacted with an item with a particular term recently, it won't completely drop out of their model from lack of reinforcement. Accordingly, the user vectors of the user PE 332 of the last thirty days weighted by days back are obtained and combined with the current version of cold start category vectors in order to obtain the updated user vector 944 as shown at 1008 and as given by Expression (1) supra. The recommendation module 210 records what the user is interested in and exploits this information so that more content that is aligned to the user's interests is recommended. Saving the user vectors on a daily basis, and then rolling them back into the next day's user vector, allows for incorporation of short-term interests while also not forgetting long-term interests. This also ensures that short-term interests (caused by spiking news events) do not overwhelm a user vector. Therefore, the updated user vector 944 is a combination of the cold start category vectors of those categories selected for inclusion into the user PE 322 and the user vectors of the previous thirty days as shown at 1008. Thus, user model for the next day is:
In addition to being automatically updated in real time and on periodic basis, the recommendations module 210 allows a user to manipulate the user vector 944 to include keywords directly as shown at 1010 so that the user vector 944 can be configured to better represent the user. Users can explicitly add terms to their respective user models, effectively “whitelisting” these terms. Such terms are added to the user vector with a weight of 1. Items that contain these terms are also upweighted, so that terms on the whitelist are effectively standing searches and content items comprising such terms will be weighed more favorably in comparison to other results that do not comprise such terms. Additionally, users have a “blacklist” of terms that they can edit, effectively removing these terms from their model, thus while presenting recommendations to the user, the content items comprising the blacklisted terms will be weighed unfavorably thereby greatly reducing the probability that items with these terms will be recommended. In an embodiment, UI (user interface) of the recommendations module 210 can also treat whitelist terms specially when rendering items for recommendation. For example, by highlighting these terms as ones the user has told the system are important to them. This makes it easy for the user to see items about terms they are particularly interested in. The UI also has a mode in which terms from an item are recommended to the user, who then has the option of adding them to their whitelist or blacklist. The recommended terms are selected from the corresponding item, and reflect the most salient terms from that item, including for example, a source of the content item, a location associated with the content item, or the author's name. Thus, the recommendations module 210 can operate in a “training mode” suggesting material to the user that can be used to fine-tune the user model 3324.
In an embodiment, user PEs of different users can send messages to each other. A direct comparison of two user models using a similarity measure describe herein can result in a recommendation that users connect to each other, either on one of the client services (TWITTER, FACEBOOK etc.) or within the recommendations module 210. If the users already have established a connection within the recommendations module 210, user A could “follow” user B, with the result being that any publicly available recommendations that user B's User PE recommends are also sent to user A's User PE for evaluation and possible recommendation.
In the context of the recommendations system 100, an avatar is a graphical image of a person or other animated or anthropomorphized object that represents the user model 3324. In an embodiment it acts as a bridge between the user and their otherwise opaque model 3324. In an embodiment, the avatar is selected by the user while initially configuring the recommendations module 210 and a selection of the avatar can be factored into building the user model 3324.
As mentioned above, a “training mode” recommends terms for the user to whitelist or blacklist—the avatar is featured in this mode in order to make it clear that the user's actions are directly affecting their model.
An avatar can have different levels of “health” (which are rendered in the avatar's image, as well as elsewhere) depending on the user's engagement level with the recommendations module 210, as well as their level of customization of the recommendations module 210.
In another embodiment, the avatar 1802 health can be indicated via only a single color but the intensity of the color can indicate the health of the avatar 1802. For example, a healthy avatar will be vividly colored indicating a user who interacts frequently with the recommendations module 210 and is actively involved in developing the respective user model. Another user who does not interact as frequently with the recommendations module 210 and seldom provides feedback may have an avatar of the same color as the frequent user, but the occasional user's avatar can be associated with a dull shade which is less intense than that of the frequent user.
b shows various facial expressions associated with the avatar 1850 by the way of illustration and not limitation. In an embodiment, the different facial expressions of the avatar can be associated with the various levels of health associated with the corresponding user model. The sad expressions 1860 and 1870 of the avatar indicate that it is not healthy. In one embodiment, the health level/facial expression can indicate that the corresponding user is an occasional user of the recommendations module 210 who infrequently provides feedback or who does not have enough number of accounts/content streams associated with the recommendations module 210 and thereby is not providing enough information about himself/herself to the user model. In one embodiment, this can be the expression of an avatar at the initial stages of the user interaction with the recommendations module 210. But eventually as the user interacts more frequently with the recommendations module 210 and gets more involved with customizing the user model 3324 by providing explicit input or implicit feedback via various actions, the expression of the avatar can brighten up to display a more cheerful countenance as shown at 1880 and finally at 1890 the expression indicates a very healthy avatar associated with or personifying a user model whose user is involved with the recommendations module 210 and provides plenty of information for customizing the user model.
In one embodiment, the avatar can likewise reflect the emotional sentiment of the content. In this case, the avatar's facial expression, or other way of rendering the avatar such as, for example, different colors, or other visual treatment can indicate sentiment.
As a representative of the user's recommendation model 3324, the avatar could travel with the user across the web. In this context, a representation of the avatar can communicate to the user where and when content is being personalized or not.
As shown in the example of
Memory 2104 interfaces with computer bus 2102 so as to provide information stored in memory 2104 to CPU 2112 during execution of software programs such as an operating system, application programs, device drivers, and software modules that comprise program code, and/or computer-executable process steps, incorporating functionality described herein, e.g., one or more of process flows described herein. CPU 2112 first loads computer-executable process steps from storage, e.g., memory 2104, storage medium/media 2106, removable media drive, and/or other storage device. CPU 2112 can then execute the stored process steps in order to execute the loaded computer-executable process steps. Stored data, e.g., data stored by a storage device, can be accessed by CPU 2112 during the execution of computer-executable process steps.
Persistent storage medium/media 2106 is a computer readable storage medium(s) that can be used to store software and data, e.g., an operating system and one or more application programs. Persistent storage medium/media 2106 can also be used to store device drivers, such as one or more of a digital camera driver, monitor driver, printer driver, scanner driver, or other device drivers, web pages, content files, playlists and other files. Persistent storage medium/media 2106 can further include program modules and data files used to implement one or more embodiments of the present disclosure.
A client device may vary in terms of capabilities or features. The client device can include standard components such as a CPU 2202, power supply 2228, a memory 2218, ROM 2220, BIOS 2222, network interface(s) 2230, audio interface 2232, display 2234, keypad 2236, illuminator 2238, I/O interface 2240. Claimed subject matter is intended to cover a wide range of potential variations. For example, the keypad 2236 of a cell phone may include a numeric keypad or a display 2234 of limited functionality, such as a monochrome liquid crystal display (LCD) for displaying text. In contrast, however, as another example, a web-enabled client device 2200 may include one or more physical or virtual keyboards 2236, mass storage, one or more accelerometers, one or more gyroscopes, global positioning system (GPS) 2224 or other location identifying type capability, Haptic interface 2242, or a display with a high degree of functionality, such as a touch-sensitive color 2D or 3D display, for example. The memory 2218 can include Random Access Memory 2204 including an area for data storage 2208.
A client device may include or may execute a variety of operating systems 2206, including a personal computer operating system, such as a Windows, iOS or Linux, or a mobile operating system, such as iOS, Android, or Windows Mobile, or the like. A client device 2200 may include or may execute a variety of possible applications 2210, such as a client software application 2214 enabling communication with other devices, such as communicating one or more messages such as via email, short message service (SMS), or multimedia message service (MMS), including via a network, such as a social network, including, for example, Facebook, LinkedIn, Twitter, Flickr, or Google+, to provide only a few possible examples. A client device 2200 may also include or execute an application to communicate content, such as, for example, textual content, multimedia content, or the like. A client device 2200 may also include or execute an application 2212 to perform a variety of possible tasks, such as browsing, searching, playing various forms of content, including locally stored or streamed video, or games (such as fantasy sports leagues). The foregoing is provided to illustrate that claimed subject matter is intended to include a wide range of possible features or capabilities.
For the purposes of this disclosure a computer readable medium stores computer data, which data can include computer program code that is executable by a computer, in machine readable form. By way of example, and not limitation, a computer readable medium may comprise computer readable storage media, for tangible or fixed storage of data, or communication media for transient interpretation of code-containing signals. Computer readable storage media, as used herein, refers to physical or tangible storage (as opposed to signals) and includes without limitation volatile and non-volatile, removable and non-removable media implemented in any method or technology for the tangible storage of information such as computer-readable instructions, data structures, program modules or other data. Computer readable storage media includes, but is not limited to, RAM, ROM, EPROM, EEPROM, flash memory or other solid state memory technology, CD-ROM, DVD, or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other physical or material medium which can be used to tangibly store the desired information or data or instructions and which can be accessed by a computer or processor.
For the purposes of this disclosure a module is a software, hardware, or firmware (or combinations thereof) system, process or functionality, or component thereof, that performs or facilitates the processes, features, and/or functions described herein (with or without human interaction or augmentation). A module can include sub-modules. Software components of a module may be stored on a computer readable medium. Modules may be integral to one or more servers, or be loaded and executed by one or more servers. One or more modules may be grouped into an engine or an application.
Those skilled in the art will recognize that the methods and systems of the present disclosure may be implemented in many manners and as such are not to be limited by the foregoing exemplary embodiments and examples. In other words, functional elements being performed by single or multiple components, in various combinations of hardware and software or firmware, and individual functions, may be distributed among software applications at either the client or server or both. In this regard, any number of the features of the different embodiments described herein may be combined into single or multiple embodiments, and alternate embodiments having fewer than, or more than, all of the features described herein are possible. Functionality may also be, in whole or in part, distributed among multiple components, in manners now known or to become known. Thus, myriad software/hardware/firmware combinations are possible in achieving the functions, features, interfaces and preferences described herein. Moreover, the scope of the present disclosure covers conventionally known manners for carrying out the described features and functions and interfaces, as well as those variations and modifications that may be made to the hardware or software or firmware components described herein as would be understood by those skilled in the art now and hereafter.
While the system and method have been described in terms of one or more embodiments, it is to be understood that the disclosure need not be limited to the disclosed embodiments. It is intended to cover various modifications and similar arrangements included within the spirit and scope of the claims, the scope of which should be accorded the broadest interpretation so as to encompass all such modifications and similar structures. The present disclosure includes any and all embodiments of the following claims.