Graph Contrastive Learning with Adaptive Augmentation

Graph Contrastive Learning with Adaptive Augmentation

April 19–23, 2021, Ljubljana, Slovenia | Yanqiao Zhu, Yichen Xu, Feng Yu, Qiang Liu, Shu Wu, and Liang Wang
This paper introduces a novel graph contrastive representation learning method called Graph Contrastive Learning with Adaptive Augmentation (GCA). GCA aims to improve the performance of graph representation learning by incorporating adaptive data augmentation schemes that preserve intrinsic structures and attributes of graphs. The key contributions of GCA are: 1. **Adaptive Data Augmentation**: GCA proposes a joint, adaptive data augmentation scheme at both topology and node attribute levels. On the topology level, edges are randomly removed with higher probabilities assigned to unimportant edges to highlight important connective structures. On the node attribute level, node features are corrupted by adding more noise to unimportant feature dimensions to enforce the model to recognize underlying semantic information. 2. **Theoretical Justification**: The paper provides theoretical justifications for the proposed method. It shows that the objective function of GCA is a lower bound of Mutual Information (MI) between input node features and learned node representations in two graph views. Additionally, it draws connections between the optimization problem and the triplet loss, highlighting the importance of appropriate data augmentation schemes. 3. **Experimental Results**: Extensive experiments on five real-world datasets for node classification demonstrate that GCA consistently outperforms existing state-of-the-art methods and even surpasses some supervised counterparts. The results validate the effectiveness of the proposed contrastive framework with adaptive augmentation. The paper also includes a detailed discussion on the broader impact of the proposed method, emphasizing its potential to alleviate label scarcity issues in real-world applications and its practical benefits in machine learning models.This paper introduces a novel graph contrastive representation learning method called Graph Contrastive Learning with Adaptive Augmentation (GCA). GCA aims to improve the performance of graph representation learning by incorporating adaptive data augmentation schemes that preserve intrinsic structures and attributes of graphs. The key contributions of GCA are: 1. **Adaptive Data Augmentation**: GCA proposes a joint, adaptive data augmentation scheme at both topology and node attribute levels. On the topology level, edges are randomly removed with higher probabilities assigned to unimportant edges to highlight important connective structures. On the node attribute level, node features are corrupted by adding more noise to unimportant feature dimensions to enforce the model to recognize underlying semantic information. 2. **Theoretical Justification**: The paper provides theoretical justifications for the proposed method. It shows that the objective function of GCA is a lower bound of Mutual Information (MI) between input node features and learned node representations in two graph views. Additionally, it draws connections between the optimization problem and the triplet loss, highlighting the importance of appropriate data augmentation schemes. 3. **Experimental Results**: Extensive experiments on five real-world datasets for node classification demonstrate that GCA consistently outperforms existing state-of-the-art methods and even surpasses some supervised counterparts. The results validate the effectiveness of the proposed contrastive framework with adaptive augmentation. The paper also includes a detailed discussion on the broader impact of the proposed method, emphasizing its potential to alleviate label scarcity issues in real-world applications and its practical benefits in machine learning models.
Reach us at info@study.space
Understanding Graph Contrastive Learning with Adaptive Augmentation