Long short-term enhanced memory for sequential recommendation

Abstract Sequential recommendation is a stream of studies on recommender systems, which focuses predicting the next item user interacts with by modeling dynamic sequence user-item interactions. Since being born to explore tendency variable-length temporal sequence, Recurrent Neural Networks (RNNs) have been paid much attention in this area. However, inherent defects caused network structure RNNs limited their applications sequential recommendation, are mainly shown two factors: tend make point-wise predictions and ignore collective dependencies because relationships between items change monotonically; likely forget essential information during processing long sequences. To solve these problems, researchers done work enhance memory mechanism RNNs. although previous RNN-based methods achieved promising performance taking advantage external knowledge other advanced techniques, improvement intrinsic property existing has not explored, still challenging. Therefore, work, we propose novel architecture based Long Short-Term Memories (LSTMs), broadly-used variant RNNs, specific for called enhanced Memory (LSTeM) , boosts original LSTMs ways. Firstly, design new gates introducing “Q-K-V” triplet, accurately properly model correlation current user’s historical behaviors at each time step. Secondly, “recover gate” remedy inadequacy forgetting mechanism, works global embedding. Extensive experiments demonstrated that LSTeM achieves comparable state-of-the-art challenging datasets recommendation.