CSF: convolution sequential fusion model for session recommender system

Abstract

Capturing users’ specific preferences is a fundamental problem in a large-scale recommender system. Currently, Sequence-based session recommendation has become a hot research direction in the industry. However, the current approach is somewhat inadequate in capturing user’s long-term preferences. We propose a session model based on convolution sequential fusion (CSF) to capture users' specific preferences by fusing users' long-term behaviors and short-term sessions. Compared with existing sequential recommendations, we address two inherent weaknesses of session recommendations: 1) The interaction of long-term interest features is not sufficiently extracted. 2) The combination of long-term interest and short-term interest is too simplistic. The long-term interests of users are varied and complex. simultaneously, users' long-term interests are closely related to their short-term interests, so they should be better integrated. We propose to encode the behavior sequence with two corresponding components: the convolutional network for interactive extraction of users' long-term interests and the long-short term gated fusion module for better combination of long-short preferences Our entire model has been test on multiple real-world data sets, and the results demonstrate that our model is more effective than other recent models on multiple evaluation benchmarks.