For many years, the main goal of the Netflix personalized recommendation system has been to get the right titles in front of our members at the right time. But the job of recommendation does not end there. The homepage should be able to convey to the member enough evidence of why a title may be good for her, especially for shows that the member has never heard of. One way to address this challenge is to personalize the way we portray the titles on our service. An important aspect of how to portray titles is through the artwork or imagery we display to visually represent each title. The artwork may highlight an actor that you recognize, capture an exciting moment like a car chase, or contain a dramatic scene that conveys the essence of a movie or show. It is important to select good artwork because it may be the first time a member becomes aware of a title (and sometimes the only time), so it must speak to them in a meaningful way. In this talk, we will present an approach for personalizing the artwork we use on the Netflix homepage. The system selects an image for each member and video to give better visual evidence for why the title might be appealing to that particular member.
There are many challenges involved in getting artwork personalization to succeed. One challenge is that we can only select a single piece of artwork to represent each title. In contrast, typical recommendation engines present multiple items (in some order) to a member allowing us to subsequently learn about preferences between items through the specific item a member selects from the presented assortment. In contrast, we only collect feedback from the one image that was presented to each member for each title. This leads to a training paradigm based on incomplete logged bandit feedback . Moreover, since the artwork selection process happens on top of a recommendation system, collecting data directly from the production experience (observational data) makes it hard to detangle whether a play was due to the recommendation or from the incremental effect of personalized evidence. Another challenge is understanding the impact of changing the artwork between sessions and if that is beneficial or confusing to the user. We also need to consider how diverse artworks perform in relation to one another. Finally, given that the popularity and audiences for titles can change or drop quickly after launch, the system needs to quickly learn how to personalize images for a new item.
All these considerations naturally lead us to frame the problem as online learning with contextual multi-arm bandits. Briefly, contextual bandits are a class of online learning algorithms that balance the cost of gathering randomized training data (which is required for learning an unbiased model on an ongoing basis) with the benefits of applying the learned model to each member context (to maximize user engagement). This is known as the explore-exploit trade-off. In this setting, for a given title the set of actions is the set of available images for the title. We aim to discover the underlying unknown reward, based on probability of play, for each image given a member, a title, and some context. The context could be based on profile attributes (geo-localization, previous plays, etc), the device, time, and other factors that might affect what is the optimal image to choose in each session.
With a large member base, many titles in the catalog, and multiple images per title, Netflix's product is an ideal platform to test ideas for personalization of artwork. At peak, over 20 million personalized image requests per second need to be handled with low latency. To train our model, we leveraged existing logged data from a previous system that chose images in an unpersonalized manner. We will present results comparing the contextual bandit personalization algorithms using offline policy evaluation metrics , such as inverse propensity scoring and doubly robust estimators . We will conclude with a discussion of opportunities to expand and improve our approach. This includes developing algorithms to handle cold-start by quickly personalizing new images and new titles. We also discuss extending this personalization approach across other types of artwork we use and other evidence that describe our titles such as synopses, metadata, and trailers. Finally, we discuss potentially closing the loop by looking at how we can help artists and designers figure out what new imagery they should create to make a title even more compelling and personalizable.