- Conference Article
3
- 10.1109/icmla51294.2020.00038
GCN with Clustering Coefficients and Attention Module
- Dec 01, 2020
- Rakesh Kumar Yadav + 3 more +3
Graph convolutional networks (GCN) exploit graph connectivity through their adjacency matrix. However, the assignment of equal importance to every one-hop neighbor and incognizance of intra-neighbor connectivity restricts its performance. Graph attention networks (GAT) address the problem of treating all neighbors equally by employing a self-attention mechanism, which encodes the node of interest using the weighted features of its one-hop neighbors. This imparts appropriate similarity based weight between connected nodes. The learning of weights of nodes takes into consideration the intra-neighbor connectivity between nodes common to a neighborhood. However, this ignores the nature of neighborhood of the node of interest. In this work, we propose to encode the nature of the neighborhood of a node onto itself by appropriately weighing the node's self-loop. The local clustering coefficient of a graph explores to what extent which nodes in a graph tend to bunch together. A vertex with a low clustering coefficient and a large degree shows connectivity between different classes. Similarly, a vertex with a high clustering coefficient and low degree represents tightly coupled similar vertices belonging to the same class. Thus, the node self-loop is initially weighed with the local clustering coefficient. Experiment results on the benchmark graph dataset, including Cora, Citeseer, and Pubmed, shows that even with sparse two layers, the local clustering coefficient based self-loop outperforms the baseline GCN by ≈ 3%. The Ablation study and learning curve prove that the proposed model remains stable and learns the optimal feature representation in comparatively less number of epochs.
Read more