KASGCN integrates Kolmogorov-Arnold Network (KAN) layers into Signed Graph Convolutional Networks (SGCNs) for representation learning on signed graphs. Where a standard SGCN uses MLP-style transforms with fixed activations to aggregate node information along balanced and unbalanced paths, KASGCN substitutes KAN's learnable univariate (spline) functions on edges, aiming for more flexible, interpretable embeddings with fewer parameters. Evaluated on signed community detection and link sign prediction, it performs competitively or comparably to standard SGCNs, with benefits that depend on the characteristics of the specific signed graph and the chosen hyperparameters.
Shebaro and Tesic investigate whether Kolmogorov-Arnold Networks (KANs) improve signed graph neural networks. Signed graphs have edges labeled positive or negative (e.g. friend/foe, trust/distrust), …