Augmenting Netflix Search with In-Session Adapted Recommendations

“Machine Learning Ranker: We focus on building a machine learning model that can jointly optimize for the member’s long-term and short-term viewing preferences. To do so we took the following approaches:

(1) We develop a personalized ranking model capable of generating relevant recommendations for a member, given her historical engagement with the platform, as well as her real-time interaction signals. We consider a few different deep-learning architectures and objectives to jointly learn different types of positive interactions for a given candidate title in a given context. For the in-session adaptation part of the model we use different member interactions observed in the current session to engineer features that can capture information about the member’s in-the-moment intent. For example, we can add signals such as the members’ genres and titles they interacted positively with during the current session.

(2) We next improve upon the model developed above by taking into account the member’s temporal interaction sequence. Specifically, we can leverage the member’s raw interaction sequence during the current and past few sessions (cross-session adaptation). Although the above approach successfully captures the member’s short-term as well as long-term intents, we hypothesize that the raw sequence could enable the model to learn feature representations and signals that the carefully engineered features might have missed. For this stage we experiment with multiple different sequential models, including simple RNN, LSTM, bi-directional LSTM and transformer architecture to leverage a member’s interaction sequence.

Offline Results: Our offline analysis shows that an in-session adaptation model significantly outperforms (6% relative increase in offline ranking metrics) the current production model for the pre-query. Fig. 2 show an example of recommendations generated from the model for pre-query where the member’s pre-query page within the current session updates to capture her most recent interest based on her in-session browsing signals.

4 CHALLENGES AND FUTUREWORK

The member signals leveraged for the approaches developed capture the member’s current interest state and enable the model to identify what the member wants to watch now. However, we still balance it with information about the member’s historical preference, since lack of long-term preference information can result in a limited level of personalization, where a small set of popular items may be recommended to many members. Moreover, we hypothesize that such repeated exposure of a subset of the titles through only a short-term intent prediction model can lead to a concentration effect over time [2]. However, a model that jointly optimizes for both the short-term and long-term intent can avoid such myopic recommendations while still catering to the member’s current needs. This is especially important for sessions where there is no prior browsing before landing on search. Additionally, such an adaptive model can improve user cold starting. We faced several challenges in developing an end-to-end in-session adapted recommender systems:

(1) To generate fresh recommendations that use interaction signals from the current session we cannot precompute and cache recommendations on the server or on the client devices. This may significantly increase the number of calls to the recommender thus increasing timeouts and an increase in infrastructural cost, putting severe constraints on runtime model complexity. Moreover, the increased number of server calls, unreliable or slow networks can degrade the member experience.

(2) In-session adaptation can make the recommendations too dynamic in some scenarios. Members may not always prefer such fast-changing rankings. For developers it becomes harder to reproduce and debug issues to understand why certain recommendations are showing up.

(3) Browsing signals from the ongoing sessions can yield extremely sparse features, which needs to be taken into account when developing the model.”