The document discusses recommendations based on graph neural networks. It provides an overview of general recommendation approaches categorized by: graph construction, neighbor aggregation, information update, final node representation. It notes that graph construction approaches deal with insufficient or excessive neighbor information, and that neighbor aggregation looks at mean pooling, degree normalization, and attention mechanisms. Information update approaches consider replacing the central node, updating while retaining central node information, or concatenating. Final node representations may use the central node alone or combine it with related nodes. The document also discusses enhancing recommendations using social networks and knowledge graphs.