MKGCN: a novel multi-channel KAN-GCN model for scalable link prediction

Abstract Link Prediction (LP) is a classic topic in network science, aiming to predict missing and unknown links using existing network data. With the expansion of network scale and the sparsity of network structure, accurately predicting links among a vast number of node pairs has become a severe challenge. Traditional Graph Convolutional Network (GCN) emphasizes local features more during the message passing process. Increasing the number of network layers can collect more global and abstract features, but it also brings the risk of over-smoothing. Blindly increasing the size of hidden layers can also lead to an increase in model complexity and the danger of overfitting. To address this issue, we propose a Multi-channel KAN-GCN Model (MKGCN). It combines KAN and GCN and constructs of three GCN channels to capture feature information at different scales. Meanwhile, the model also introduces a new neural network based on the Kolmogorov-Arnold representation theorem - Kolmogorov-Arnold Network (KAN). Unlike the fixed activation functions used in Multi-layer perceptrons (MLPs), KAN adopts learnable activation functions, providing greater flexibility in capturing complex relationships in high-dimensional data. The MKGCN model maps and fuses features from different channels through KAN. Experimental results demonstrate that the proposed architecture achieves competitive results on eleven real-world network.