Using Navigation to Improve Recommendations in Real-Time

Implicit feedback is a key source of information for many recommendation and personalization approaches. However, using it typically requires multiple episodes of interaction and roundtrips to a recommendation engine. This adds latency and neglects the opportunity of immediate personalization for a user while the user is navigating recommendations. We propose a novel strategy to address the above problem in a principled manner. The key insight is that as we observe a user’s interactions, it reveals much more information about her desires. We exploit this by inferring the within-session user intent on-the-fly based on navigation interactions, since they offer valuable clues into a user’s current state of mind. Using navigation patterns and adapting recommendations in real-time creates an opportunity to provide more accurate recommendations. By prefetching a larger amount of content, this can be carried out entirely in the client (such as a browser) without added latency. We define a new Bayesian model with an efficient inference algorithm. We demonstrate significant improvements with this novel approach on a realworld, large-scale dataset from Netflix on the problem of adapting the recommendations on a user’s homepage.

Introduction. Recommender systems are particularly important for services with large numbers of items. They rely on the promise that through effective algorithmic use of collected data, they can help the user discover novel content by presenting personalized, relevant items to the user [18]. Recommendation is fundamental to many places, such as news readers [5, 8], blogs [10], homepages [4], search engines [12], and streaming video [11]. Click models are often used to model the viewing behaviors and estimate intrinsic relevance [12, 5]. A key challenge in such approaches is that a user’s intention can vary significantly between sessions in ways that cannot be captured by prior knowledge. For instance, movie consumption preferences are context-dependent: whether a viewer watches alone, with friends, with a romantic love interest, or with children. Likewise, they are situation dependent (dedicated viewing vs. background entertainment during a chore), depend on the available time (short episode vs. full movie) and can even depend on mood.

Discussion / Conclusion. AND FUTURE WORK We proposed inferring user intent through in-session navigation information and then updating pages of recommendations based on that intent. To accomplish this, we defined a probabilistic model capable of incorporating current session navigational information as well as logged navigation and consumption data. The model infers interest in candidate rows of recommended items based on homepage navigation to a certain point on the page, and can populate the remaining rows on the page with more relevant content that reflects the current likely intent as predicted by the model. This is the first work of its kind that performs online updating of recommendations based on user navigation within a single page. Thus, it is the intention of this paper to demonstrate feasibility of such a method, which we do on real-world dataset from Netflix.