Presentation Information
[B-11-15]A Preliminary Study on Incorporating Concept Layers into Graph Neural Networks
〇Naoki Shibao1, Yuto Tamura1, Sho Tsugawa1 (1. Tsukuba University)
Keywords:
Graph Neural Network,Information Diffusion,explainability
Predicting the final size of information diffusion from its early-stage dynamics is an important task in the study of information diffusion on social networks. In recent years, prediction methods using graph neural networks (GNNs), which can take graph structures into account, have been proposed.However, the intermediate representations learned by such models are often difficult for humans to interpret. To address this issue, Concept Bottleneck Models (CBMs), which introduce concept layers into deep learning models and use predefined concepts as intermediate representations, have been proposed.In this study, we investigate the effect of introducing a concept layer consisting of network statistics and related features into a GNN for cascade popularity prediction. The results show that the proposed GCN-CBM maintains prediction accuracy comparable to that of a standard GCN without a concept layer.These findings suggest that introducing a concept layer into GNNs can incorporate intermediate representations that may contribute to improved interpretability without substantially impairing prediction performance.