Convolutional context encoding for hybrid session-based recommendation

Abstract

Recently, session-based recommendation and context-aware recommendation have attracted great attention from the recommender systems community. The marriage of the two topics has activated a new interesting research direction: context-aware session-based recommendation. However, since previous context-aware session models mainly focused on improving short-term modeling power using short-term contexts, the effect of long-term user interests on user session behaviors has been largely ignored for context-aware session models. To fill this gap, in this thesis, a CNN-based context-aware session model called CCE(Convolutional Context Encoding) is proposed. CCE is a hybrid context-aware session model that extends traditional RNN-based session models by incorporating long-term user interests which are extracted from various user context information using CNN. Furthermore, CCE supports any types of contexts and has linear scalability to the context size, thereby being suitable for commercial environments where the numbers and types of contexts can be extremely large. The experiments on two real-world datasets verified that, using the user context information, CCE outperforms the pure session models that ignore long-term user interests.