Improving Sequential Service Recommendation via a Novel Neighborhood-Augmented Graph Collaborative Attention Network

As an effective tool to alleviate information overload on the service platform, sequential recommender systems aim to predict the service in which users are interested by analyzing their historical behaviors. To leverage transition patterns among services, some solutions apply the graph attention networks for service representation learning via neighborhood information aggregation. However, existing solutions struggle to sufficiently leverage the graph structure due to two significant challenges. Firstly, some high-correlation services may not appear in adjacent positions and thus have no connections, which makes it difficult to aggregate comprehensive information in graph learning. Secondly, the attention network parameters are randomly initialized, which makes the information propagation unstable in the early training phase. To tackle the two challenges, we propose a novel neighborhood-augmented graph collaborative attention network (NA-GCAN). For the former challenge, we augment the graph structure by screening potential neighbors with high correlation for each service node based on the attention network, to ensure effective aggregation of global-wise information. For the latter challenge, we exploit the co-occurrence information to pre-train service embeddings and the transition information to guide the information propagation. In addition, we devise a novel two-stage learning strategy to enable a warm start for model training and make full use of the augmented graph structure. Extensive experiments have demonstrated the superiority of our proposed NA-GCAN.