Graph Convolutional Networks using Heat Kernel for Recommendation

Graph convolutional networks (GCNs) have great potential in recommendation systems by aggregating signals from high-order neighbors to generate user and item embeddings. A key component of GNN-based recommendation systems is the message propagation module, which disseminates information across multiple layers of the interaction graph to capture high-order collaborative signals. However, GCN-based models often suffer from the over-smoothing problem, where node embeddings become indistinguishable after passing through several layers, undermining recommendation effectiveness. This issue arises from the lack of differentiation between user and item nodes during propagation, causing neighboring nodes to be assigned equal importance. To address this, we propose a novel heat kernel-based adaptive convolutional network (HKACN), which adapts the significance of neighboring nodes using a heat transfer model. Additionally, a self-attention mechanism is applied to optimize layer weights, improving the GCN’s effectiveness in recommendation tasks. Experimental results on three benchmark datasets show that HKACN outperforms state-of-the-art GCN-based recommendation models.