Enhancing Interpretability of Graph Convolutional Networks for Multi-view Learning

The explosion of multimedia data, collected from heterogeneous sources and represented in multiple formats, has led to the formation of comprehensive multi-view or multi-modal datasets, facilitating deeper analysis and insight. These datasets cover various physical features captured by different sensors, which makes the quality distribution of data among views uneven. In recent years, Graph Convolutional Networks (GCNs) have attracted substantial attention from the academic community, leading to their widespread adoption in a variety of application domains. However, GCNs are often considered black-box models due to their complex internal operations, thereby posing significant challenges to understanding and interpreting their decision-making processes. The lack of interpretability in GCNs undermines confidence in their predictions and makes it difficult to identify and address potential biases. To tackle these issues, we propose a generic multi-view graph convolutional network, which is applied to semi-supervised classification tasks. By maximizing subspace independence and restricting network transmission weights, we aim to find interpretability for the constructed network framework from both the spatial and transmission domains. The main contributions are summarized as follows: First, we propose a general and efficient graph learning framework for multi-view representation, which simplifies both feature fusion and downstream classification tasks. Second, we design two learning strategies, focusing on constraining the weights of forward propagation and maximizing the independence of subspaces, respectively, to effectively capture the inherent characteristics between multiple views and stably propagate label information. Finally, we develop a joint network based on the proposed framework that integrates both constrained weights and learned embeddings to emphasize the most informative features from each view. We conduct extensive experiments on eight benchmark datasets, where our proposed method consistently outperforms ten state-of-the-art approaches, demonstrating its superior effectiveness across diverse multi-view learning tasks.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.