Multi-axis Gating MLP with Bidirectional Feature Extraction for Sequence Recommendation

The sequence recommendation task uses neural network models such as GRU and Self-Attention to extract potential preference characteristics from the user's behavior data, This makes the recommendation more accurate. However, they use the sequence information of items to build embedded features, which may destroy the representation of interaction between items. Therefore, based on the traditional RNN network, we use multi-axis gating unit to capture the user behavior characteristics inside and across channels. In our experiments, we found that multi-axis gating unit can significantly improve the performance of sequence recommendation model. Compared with Self-Attention model, MLP architecture has less parameters and lower time complexity. Specifically, we propose a spatial gating unit MLP model for sequential recommendation tasks, which uses multi axis gating units to capture the behavior characteristics of multiple users in the spatial dimension. A large number of experiments on multiple real data sets show that our proposed method is superior to RNN and Self-Attention based methods.