Multi-view Feature Fusion Recommendation Algorithm based on Representation Learning

Abstract

With the rapid development of representation learning, more and more external side-information like users ‘comment on item is introduced into the recommendation system to alleviate the problem of data sparseness. Recommendation algorithm based on multi-view learning considers those external side-information as independent views feature which is used to deal with data sparsity. However, these views are interdependent. By assuming that the view feature is independent, reducing the computational complexity, but resulting in poor recommendation performance. To solve this problem,this paper proposes a multi-view feature fusion recommendation algorithm based on representation learning,namely MVF. First, the algorithm uses an automatic encoder to extract the features of each view, and constructs second-order and third-order interactive features based on those features. Then, the singular value decomposition algorithm is used to compress the second-order interaction feature to extract the main interaction features, and the Tucker tensor decomposition algorithm is used to compress the third-order interaction feature to extract the main interaction features. After getting the main interaction feature,using the attention mechanism to fuse those features to get the representation of item. Considering that users have different preferences for different items, the attention mechanism is used to fuse user's items to obtain the user's preference model. Finally, Extensive experiments on real data sets from Amaza and compared with multiple baseline algorithms to verify the effectiveness of the proposed algorithm.