Adaptive Boosting Graph Convolution Network for Recommendation

Graph Convolution Networks (GCNs) have been widely applied to recommender systems during the last few years. Existing GCNs propagate the collaborative signals within the user-item interacted graph to find the similar nodes. However, this kind of network only uses implicit features that limited by the sparsity of the graph, and many nodes cannot be well represented. Here implicit feature refers to the interaction of user-item. In this paper, we propose an adaptive boosting graph convolution network algorithm (BoostGCN) for recommender, which combines multiple homogeneous component graph-based models linearly. In the framework, we choose the Normalized Discounted Cumulative Gain (NDCG) as the parameter of re-weight function for each component models. At last, we conduct experiments on two public datasets and our proposed method outperforms the state-of-the-art approaches on common metrics and convergence time.