Patent Application
20130173419
The present invention relates to electronically marketing goods, services, content, and other entities through the automated analysis of human behavior. Particularly, the invention relates to the representation of subject and object characteristics for the purpose of efficient generation of recommendations. The system has application in personalization, behavioral targeting, Internet retailing, email segmentation and ad targeting, to name but a few applications.
The consumer faces a profound number of possible choices when selecting most kinds of products, be it movies, music, books, travel, art, dining, employers, and so on, to the extent that the consumer must choose from well-publicized possibilities, such as through advertising, or rely on recommendations of others. In the first case the set of choices is severely limited to those that can be promoted to a broad audience. In the second case the consumer must weigh. the similarity of his or her own tastes to the person making the recommendation, whether it is an acquaintance or media. In addition, the number of possibilities and the cost of acquisition, both in terms of time and money, of assessing possibilities, make it infeasible to sample a large number of possibilities to determine which are of interest to the consumer.
Recommendation systems rely on trying to best match a person's individual preferences to the characteristics of the available items. In general what is known about the subjects and objects is the set of affinities between subjects and objects, where the affinity {Aij} between subject i and object j is determined by explicit feedback from the subject or inferred from the subject's interaction (or non-interaction) with the object The consistency of the affinity scale from subject to subject and object to object derives from the consistency of the goal of the subjects in the given environment, for example to make a purchase in a commerce environment or to read articles in a content environment.
The primary goal of the recommendation system is to predict for a given subject those objects for which the subject will have the greatest affinity. In general the subject characteristics can be represented by a vector S=(S1, S2, . . . , SL) and the object characteristics can be represented by a vector B=(B1, B2, . . . BM), whereby the predicted affinity of the subject to the object is a function P=f(S, B). Various recommendation systems then differ in their representation of subject and object characteristics S and B and the similarity function f.
In a number of contexts, consumers make repeated transactions of the same item. For example, groceries, office supplies, airline flights, and news reports are regularly purchased or consumed in a repeated fashion. Similarly, consumers often make a series of transactions of related products, for example periodically replacing running shoes, getting the latest model of a car, or new offering from a brand of clothing. The frequency and duration between such repeated transactions is in and of itself a distinguishing element of these transactions. This is widely recognized as the premise for a variety of mass marketing efforts targeted at consumers, yet the vast majority of these efforts, by their very nature, are not customized to the purchasing patterns of an individual consumer. Recommendation systems are specifically designed to bridge that gap, but to date have not explicitly leveraged the information available from examining the intervals at which repeated transactions are made.
The invention pertains to recording re-occurring transactions of a product, service, content or other entity, computing the next likely transaction date and time of a specific item by a particular subject, and using that date and time to make recommendations to that subject. The invention also pertains to classifying products by the frequency with which they are re-transacted in order to make product recommendations based on that classification. This invention is sometimes referred to in this patent by its commercial trademarked name, Resonance®.
An object of the invention is to provide a means of recommending objects to subjects based on either explicit or behaviorally interred ratings of that subject or other subjects of those objects and of commonly rated objects.
Another object of the invention is to create object profiles as well as subject profiles, so that objects can be readily indexed by aesthetic or other categories and so that objects can be readily associated across product categories by aesthetic similarity.
Another object of the invention is to create subject and object profiles that can be used to relate the derived aesthetic attributes to other objective measures of subjects, such as personality type or demographics, and objects, such as color, shape, or replacement interval.
Another object of the invention is to combine the recommendations of the system with explicit human merchandising objectives either through “hard” rules that filter results by specified criteria or “soft” rules that bias the results towards a defined business goal.
Another object of the invention is to identify appropriate subjects for the marketing of a particular object.
Another object of the invention is to automatically identify and classify items that are replenishment vs. durable
Another object of the invention is to use the classification of items as replenishment vs. durable to predict a subject's affinity and next transaction date for a particular item.
The present invention is a system and method for predicting subject responses to objects based on that subject and other subjects' repeated responses to that and other objects. The process of matching subject and object profiles produces a predicted response score that can be used to rank recommended content. The scores can be used as is or combined with other business logic to render the final recommendation rank. The invention can be applied to a broad range of applications, including the retailing of single consumption items, such as non-recurring purchases or content views, where the previous purchase or view of an object cannot be used to predict additional purchases or views of the same object. The invention can also be used to predict subject responses to recurring purchases and to recommend new consumables.
The invention considers the interaction of subjects and objects. The subject is an active entity that initiates transactions. The subject consumes or experiences objects and provides feedback on the level of satisfaction with the object. The subject could be a single person or a corporate entity, such as a business. The object is a passive target of interaction by the subject. This could be a physical object, such as a consumer good, for example cars, MP3 player, or ice cream; media, such as music, movies, hooks, art, or plays; or even a person, as in the case of a job search or a matchmaking service. In the case of active entities, it is possible for the subject and object to reverse roles depending on the situation.
The invention provides a novel solution to the problem of how to identify objects, for example products, that will appeal to a particular subject, for example a person, where the large number of possible objects, including less desirable objects that are descriptively similar but aesthetically different or where some objects may appeal highly to a limited population of subjects while being, undesirable to the broader population, makes it difficult for the subject to notice the objects that the subject wants simply by browsing the entire set of objects. This provides a breakthrough for target marketing and retail applications because it allows the consumer, solely by behavior, to “self-market” or “pull” those products which are of interest, rather than requiring that retailers “push” potentially unwanted products through advertising or other inefficient means,
The invention also addresses the issue of consumer privacy because it does not profile the consumer using personal demographics information, which consumers find both invasive and tedious to enter. Thus Resonance improves retailers' ability to target customers, while simultaneously making it easier for consumers to participate.
The invention works by forming profiles of subjects, for example consumers, and objects, such as goods or media, based on aesthetic evaluations of objects by subjects. The invention does not require a priori information about either subjects, such as demographics or psychographics, or objects, such as classifications or genres. Rather, it automatically generates representations of subjects and objects solely from the subjects' interaction with the objects. Because it creates its own abstract representation of subjects, it allows retailers to transparently target the subject without compromising, subject privacy through the collection and modeling of sensitive personal information. The profiles can also be extended across catalogs, product or content domains, or across websites or stores.
Note that the identification of subjects and objects is not necessarily a physical one and may change depending on the application. For example, in a consumer movie recommendation application, the person requesting recommendations is the subject and the movie is the object, in a dating, service application, a person would be considered a subject when searching for matches and an object when being searched by others. Similarly, in the case of employer/employee matching, companies and persons would alternate between the roles of subject and object. Note that in cases where an entity can assume different roles, a different profile would he created for each role.
Because the profiles are symmetric (both subjects and objects are profiled to the same representation), subjects can be matched to other subjects or objects, and objects can be matched to other objects or subjects. For example subject-subject matching could be used on a social networking site to connect people of like interests or on an online store to order product reviews according to the similarity of the reviewer to the reader. Similarly, object-object matching can be used to match keywords to products or content, advertisements to news articles, or promotional banners to referring affiliate sites.
Because of the large. number of weights or adapted parameters of the system, which scales as the number of subjects and objects, a key aspect of the method is that the weights for each subject or object are decoupled from other subjects and objects when updated separately. This allows individual subjects and objects to be trained by different processing units, which allows the method to scale up to large numbers of subjects and objects, which may ultimately total millions or tens of millions or more.
A key improvement of the invention over collaborative filters is that it creates not just profiles of the subjects, but profiles of the objects as well. This provides several advantages, including rapid and scalable prediction of subject to object affinities; straightforward cross marketing across product categories; and sorting of objects by aesthetic categories for purposes of browsing and selecting, items for consumption or association, such as selecting, musical recordings to go with a movie production.
Another key improvement of the invention over previous recommendation systems is that it exploits the duration between transactions, not just the quantity or date of transactions to predict subject to object affinities. This provides the advantage that the system can accurately predict when a subject will transact in addition to how likely they are to transact a particular item. The additional prediction of when a transaction will occur allows the system to recommend products when they are most likely to be appealing to a subject, maximizing the relevancy of the recommendations.
A subject interacts with the user interface. The user interface makes a request to the recommendation system, which returns personalized content based on the context of the request, where the context of the request includes the identity of the subject, the specific type of system, such as desktop computer or mobile phone, the physical location of the subject, the specific page of the application being viewed, or any other known attributes that may be relevant to determining the interest or intent of the subject. In addition to requesting and displaying recommended content, the user interface submits information on the activity of the subject, including whether the subject completes a desired or targeted outcome, such as making a purchase, booking a hotel, completing a survey, accepting an offer, or an other conversion event for the site. The recommendation system stores all of the recommendation requests and subject outcomes, which are used to build subject profiles in accordance with the present invention.
Additionally, a profiling engine retrieves affinity and other data from an application database and uses the data to generate the segmentation models, which are then stored back into the application database. The predictive segmentation models, content metadata, and any additional business rules logic are also cached on the web servers for .faster match generation during live operation. In order to process an arbitrarily large number of visitors the web servers are multiplexed using a load balancer, which makes a collection of web servers appear to the Internet as a single server. Also, when a web server becomes unavailable or out of service for any reason, the load balancer automatically transfers traffic to another server. This provides a high level of fault tolerance for the system. In order to provide additional service availability the database and web servers can be replicated to other data centers, through geographic load balancing.
Note that in this embodiment the service has been distributed over multiple servers. In an alternative embodiment all of the functions of the service could be put onto a single or smaller set of servers without a substantive change in the overall functionality of the service. This embodiment also supports multiple Service Customers making simultaneous requests to the web services by allocating different requests to different subsets of servers in the server farm and by creating a separate database for each Service Customer.
The output of the recommendation engine is fed through the content management module to get the display attributes of the recommended content and then displayed on the website. The content management module not only generates content for the website, it also feeds content information (metadata) to the catalog management module, which stores both information necessary to display content and descriptive attributes that may be applied to filter targeted content according to website-defined business rules or objectives. When a request is made to the recommendation engine for targeted content, the recommendation engine combines the subject profiles with the catalog metadata to find the optimal content subject to any business rules, or other restrictions, put on the results.
In a sample embodiment, the shopper history is stored in a computer database memory to facilitate the computation of the replenishment interval as described below.
The system uses Ri,j to compute the following averages. The system computes the Average Replenishment Rate
If
α·Ri,j+(1−α)·
where α is a continuous value between 0 and 1, inclusive, that determines the weighting or balance between the two rates. The system calculates α as
where σi,j is the sample variance of Ri,j and Nmin is the minimum number of transactions, for example 5. The advantage of this method of calculation is that as the variance of Ri,j decreases, i.e. the individual replenishment rate becomes more and more precise, the system relies more on it. Alternatively, when it is not reasonable to assume a homogenous set of shoppers or the statistical variance of Ni,j across different shoppers of pj is large, the system uses {circumflex over (R)}j, if available, to predict the next likely transaction date and time by adding a convex combination of Ri,j and {circumflex over (R)}j seconds to the last date and time pj was purchased by si.
The advantage of using a convex combination of the Individual Replenishment rate with either the Average Replenishment Rate or the Weighted Average Replenishment Rate is the ability to make a tradeoff between the statistical evidence supporting the Individual Replenishment Rate and the statistical evidence supporting the Average Replenishment Rate. If a subject has a large number of transactions for a particular product, there is little need to rely on the Average Replenishment Rate or the Weighted Replenishment Rate to inform the system on when their next transaction will be for that product. On the other hand, if a subject has purchased a product only once, the system must rely on the Average Replenishment Rate or the Weighted Replenishment Rate to determine the next likely purchase date.
An alternate embodiment of the system uses the above calculations, but instead of computing replenishments for individual products, the system computes replenishment rates for groups of products. That is, Ni,j is the number of times any product in a group j was purchased by subject si, and Ri,j is the corresponding replenishment rate. Groups can be based on an attribute of the product, such as category, subcategory, model, or brand. Alternatively, groups can be based on a measure of object similarity such as object vector matching as described in U.S. patent application Ser. No. 12/415,758, US Patent Application Publication No. 20094-248599 A1, filed on Mar. 31, 2009. An advantage of this embodiment is to handle the case of products that are purchased only once, but replaced with similar products from the same manufacturer, supplier, or creator of such products. For example, a subject might purchase multiple pairs of shoes from the same producer over their lifetime, each with a distinct name or design. Another advantage of this embodiment is that the system can recommend (see
Another alternate embodiment decreases the score as a function, of 1 divided by how far Di,j is from the present time. The advantage of this alternate embodiment is when the system has a high degree of confidence in Di,j, it will decrease the score more severely when Di,j is farther from the present. An alternate embodiment of the system can use other decreasing functions of how far Di,j is from the present to alter the score a product gets, or combine the above methods. The system ranks all products by their score and then selects the best scoring products to recommend.
An alternate embodiment of the system ranks all products by their score, then adjusts the ranking by preferentially boosting the score of those products that have average or weighted average replenishment rates within a day of the individual replenishment rates of the ten highest ranking products and last purchase dates within a day of the ten highest ranking products. An advantage of this alternate embodiment is that products the subject has not transacted with can be recommended based on the fact that they are commonly transacted at a similar frequency and in coordination with high ranking products. In another embodiment the replenishment scores can be combined with predicted affinities from behavior modeling or satisfaction ratings, and then re-ranked by the combine score. In another embodiment the score of low variance items (hence high likelihood of transaction repetition) can be given an additional score and re-ranked.
When the system receives a request for recommendations, the context in which the request is made is evaluated to determine if the shopper is considering a durable or consumable product. If the shopper is considering a consumable product, the system might, for example, recommend both durable and consumable products. On the other hand, if the shopper is considering durable products, the system might recommend just consumable products, depending on how the service customer decides to merchandise to their end-user customers. Note that the behavioral classification of products into consumables and durables does not rely on any expert knowledge of the products themselves or any explicit classification supplied by the supplier of the products. Rather, the threshold can be set experimentally by comparing the response of the subject population to different thresholds. Also note that by exploiting a behavioral classification of products, the system makes more relevant product recommendations than if it leveraged just the individual shopper's behavior.
An alternate embodiment of the system classifies products by their propensity to be transacted multiple times. In this embodiment, the propensity of a product to be transacted multiple times is defined as the fraction of all transactions for that product that do not include transactions by subjects with a single transaction with that product. If the propensity to be transacted multiple times is greater than a specified threshold, for example 50%, then the product is classified as a consumable; otherwise it is classified as a durable product. The advantage of this embodiment is that it is unaffected by the natural variability in replenishment rates.
| Number | Date | Country | |
|---|---|---|---|
| 61581696 | Dec 2011 | US |
