Recommendation System over Multi-layer Complex Networks: A Wide and Deep Graph Convolutional Neural Network Approach

In recent years, the design of recommendation systems has emerged as a trending, yet challenging topic. Concurrently, there has been a significant increase in graph-structured data, including social networks. Recommendation systems can be built over multi-layer graph-based data (complex networks). A layer represents a certain information domain, and the layers can be interconnected based on specific relations. Such complex networks can convey additional information by considering inter- or intra-layer relations. Due to their complexity, processing multi-layer complex networks demands efficient algorithms. This paper suggests a deep-learning-based recommendation system that, on input a multi-layer network, analyzes both inter-layer and intra-layer connections, to achieve better recommendations. The proposed method is a variant of wide and deep neural networks utilized over multi-layer complex networks. The wide part uses hand-crafted features, while the deep part extracts deep features from the graph convolutional network. The proposed method was evaluated on several real datasets. Compared with recent matrix completion methods, our method is more accurate than the previous approaches, while outperforming them or achieving essentially similar performance on the same hardware.