I. Introduction
Recent years have witnessed the success of graph convolution network (GCN)-based recommender models, such as PinSAGE [1] and LightGCN [2], which performs node representation learning over the interaction graph and demonstrates promising performance [3]. The core of them is neighborhood aggregation that enhances a node’s representation with the information from its neighbors. In this way, the graph structure can be explicitly integrated into the embedding space, improving the representations of users and items. In practical usage, a recommender system needs to be periodically (e.g., daily) retrained to keep the model fresh with the new interaction data. In this work, we study the problem of GCN model retraining for the recommendation, which has received relatively little scrutiny.