Node Classification Based on Kolmogorov-Arnold Networks

Most graph deep learning methods extract feature information from graph data by using learnable weights and specific activation functions. The selection of different activations has a significant impact on performance. Aiming at the above problems, this paper proposes a fully connected neural model G-KAN based on the Kolmogorov-Arnold network (KAN), which does not require specific activations and explicit node message passing modules. It dynamically learns activations through KAN and introduces a contrastive loss guided by node similarity to implicitly extract the original graph information. Firstly, G-KAN maps the graph data to the feature space through a linear layer. Secondly, the KAN layers extract the latent features from the input data. Thirdly, G-KAN maps the output to a probability distribution of labels using a linear layer and a Softmax function. Finally, the contrastive loss is used to optimize the output of the KAN layer, encouraging high-similarity nodes to be close to each other and low-similarity nodes to be far away. In the node classification task, G-KAN outperforms the currently advanced baselines and raises Micro-F1 and Macro-F1 by 50.42 and 52.84 percentage points respectively compared with graph convolutional networks (GCN) on BlogCatalog. In activation comparison experiments, the method introduced by KAN outperforms the variants that use different activation functions and demonstrates stronger generalization across various datasets. The results show that the learnable activation strategy used by G-KAN can enhance the representation capacity of fully connected neural networks and generate low-dimensional node representations with higher discriminability.