Self-attention Based Collaborative Neural Network for Recommendation

With the rapid development of e-commerce, various types of recommendation systems have emerged in an endless stream. Collaborative filtering based recommendation methods are either based on user similarity or item similarity. Neural network as another choice of recommendation method is also based on item similarity. In this paper, we propose a new model named Self Attention based Collaborative Neural Network (SATCoNN) to combine both user similarity and item similarity. SATCoNN is an extension of Recurrent Neural Network (RNN). SATCoNN model uses self-attention mechanism to compute the weight of products in multi aspects from user purchase history which form a user purchase history vector. Borrowing the idea of image style transfer, we model the users' shopping style by gram matrix. We exploit the max-pooling technique to extract users style as a style vector in gram matrix. The experimental results show that our model has better performance by comparison with other recommendation algorithms.