Cross Product and Attention Based Deep Neural Collaborative Filtering

Matrix factorization and its subsequent models have been widely used in recommendation systems due to their simple and efficient performance. However, this simple and intuitive operation has its natural limitations, which limit performance improvements. The reason is that it assumes that the embedded dimensions of the user and item are independent and identically distributed (IID), and that each dimension contributes equally to the predicted score. To overcome these limitations, we propose a new deep neural network architecture CADNCF to perform ranking recommendation. The idea is to use the cross product to build a two-dimensional correlation matrix, which can not only explicitly model the pairwise correlations between the dimensions of the embedded space, but also be more expressive and semantic. We also designed an attentionmechanism learning module to extract these useful information from the correlation matrix for the final prediction and eliminate noise. Then, we adopt the MLP to learn their interactions function and make predictions. We conducted extensive experiments on three benchmark data sets, and the results show that our proposed CADNCF model is superior to some baseline methods and other sate-of-the-art methods.