Pair-Wise Convolution Network with Transformers for Sequential Recommendation

Abstract

Sequential recommendations seek to employ the sequence of interactions between users and commodities to predict their next behavior based on the behavior they have recently made. Previously, some recommendation systems have been built on Markov chains and recurrent neural networks (among others). However, these methods have many limitations that they emphasize too much sequence change to fully emphasize the correlation between adjacent items; Besides, they generally ignore the influence of contextual information. To solve the shortcomings of the existing sequential recommendations, we try to model the relationship between items, get an effective representation of sequential features, and capture complex sequence correlations. Specifically, we propose a pair-wise convolution network with transformers for the sequential recommendation. The two-dimensional convolution networks encodes the sequence into a three-dimensional tensor and learns the relationships of features between the sequences. We adopt a residual connection to prevent the gradient from disappearing and solve the loss of feature information. The experimental results show that our method is superior to various advanced sequential models on sparse and dense data sets and different evaluation indicators.