Explainable Recommendations via Attentive Multi-Persona Collaborative Filtering

“Collaborative Filtering (CF) methods model users’ taste and inclinations based on user-item interactions. Often, a user is represented by a single latent vector that encodes all of the user’s different interests. For example, suppose the user likes both horror movies as well as comedies, the latent user representation should represent a positive inclination towards movies of both genres. At prediction time, both comedies and horror movies will receive high scores, but there is no easy way to distinguish between items from each genre, or explain the reason for each item recommendation. Moreover, in some cases, the more dominant genre may overtake the entire recommendation list. To mitigate this, additional diversity enhancing mechanisms are often added on top of the original CF model in order to ensure that both types of movies are properly represented within the list of recommendations [26].

In the example above, we have used known movie genres (Comedy and Horror) to indicate two groups of user tastes or “personas”. In the general case, a user can have multiple latent personas that could not always be labeled as distinct genres. However, we assume that user interests are not monolithic and should be represented by several distinct personas, each corresponding to a different taste. This concept is somewhat related to a well-known dilemma in recommendations literature known as the accuracy-vs-diversity tradeoff [33]. However, “diversity” implies that the items in the recommendation lists should be different from each other (diversified). In this work, on the other hand, we wish to identify and discern the users’ different tastes and inclinations. Our goal is to learn the correct distribution of tastes per user and the contribution or importance of each persona when different items are considered. Then, when generating a recommendation list for the user, each recommendation is associated with the user persona that best explains it.

We present Attentive Multi-Persona Collaborative Filtering (AMP-CF). AMP-CF is an explainable recommender system that models a user via an attentive mixture of personas that discerns and captures the user’s different inclinations, and explain each recommendation in the final recommendation list. Using a novel attention mechanism, the distinct user personas are dynamically weighted and combined to generate a single attentive user representation that depends on the candidate item (the item under consideration). When considering a specific item, some user personas may be indifferent to the item, while other personas may have more distinct positive or negative affinities. The attention mechanism in AMP-CF determines the “importance” of each persona w.r.t. the specific item and the attentive user representation dynamically updates accordingly. Thus, the AMP-CF model facilitates a type of contextual user representation: while the user personas are static at prediction time, the resulting user representation dynamically changes based on the item under consideration. Arguably, this approach better emulates the way human beings actually assess items - different characteristic components of our taste profile contribute differently in reaction to different items under consideration. Moreover, the dominant characteristics are the ones that best explain our final choices.

It is true that “traditional” single-persona user representations have the ability to encode multiple tastes through the use of orthogonal latent dimensions. However, there is no easy way to distinguish and identify the distinct tastes. In this work, we separate the traditional user representation into several latent personas, each representing a different inclination. As an additional contribution, we introduce a new evaluation procedure named Taste Distribution Distance (TDD). TDD compares the distribution of “tastes” in a recommendation list with that of the user’s interests based on her historical items. Importantly, TDD is different than diversity. While diversity concerns with how different the recommended items are from one another, TDD measures the ability of the recommendation list to proportionally match the user’s full range of interests. We present extensive experimental result.”