Ordered-unordered Network For Session-based Recommendation

With a broad spectrum of application,Recommender Systems play an important role in modern commercial online systems.Specifically,session-based recommendation has become an increasingly hot topic recently.With the progress in the deep learning community and the availability of large-scale datasets for session-based recommendation,deep neural networks started to show their effectiveness in this research filed.In this thesis,we focus on improving the effectiveness of existing session-based recommendation approaches.The task of session-based recommendation is to predict users' next actions based on anonymous sessions.Though sorted by time,items in a session may not necessarily have sequential relation due to users' random choices.However,such unordered information is ignored by previous methods because they only treat each session as a pure item sequence.In fact,unorderedness is an important property of sessions and it is useful for the session-based recommendation tasks.Based on this observation,we propose to treat each session as an unordered set and introduce an order-invariant module to explicitly exploit unordered information of the session.An ideal model for session-based recommendation tasks should be able to capture both ordered and unordered information of sessions.To this end,we present a novel neural network by integrating the proposed order-invariant module with an RNN-based method.The proposed network,named Ordered-Unordered Network(OUN),can learn session representations by combining both ordered and unordered information in an end-to-end fashion.Experiments demonstrate that our OUN achieves state-of-the-art performances on real-world benchmark datasets.