Self-Attentive Sequential Recommendation

The paper targets sequential dynamics in recommender systems, where two approaches have dominated: Markov Chains, which predict a user's next action from their last few actions and perform best in extremely sparse datasets where parsimony is critical, and Recurrent Neural Networks, which capture longer-term semantics and perform better in denser datasets that can afford higher complexity. To balance these goals, the authors propose SASRec, a self-attention based sequential model that captures long-term semantics like an RNN but, via an attention mechanism, makes predictions based on relatively few actions like a Markov Chain. At each time step, SASRec identifies which items from a user's action history are relevant and uses them to predict the next item.

Extensive empirical studies show that SASRec outperforms various state-of-the-art sequential models, including MC-, CNN-, and RNN-based approaches, on both sparse and dense datasets, while being an order of magnitude more efficient than comparable CNN/RNN-based models. Visualizations of the attention weights show how the model adaptively handles datasets of varying density and uncovers meaningful patterns in activity sequences. This mattered by delivering an accurate, efficient, and interpretable attention-based recommender.