Cross-level graph contrastive learning for community value prediction.

Area of Science:

• Social Commerce
• Graph Machine Learning
• Artificial Intelligence

Background:

• Community Value Prediction (CVP) is crucial in social commerce but challenging due to complex community and individual structures.
• Existing graph machine learning methods struggle to adequately address CVP tasks.
• A novel approach is needed to effectively model multi-level social connections for CVP.

Purpose of the Study:

• To introduce a novel cross-level graph contrastive learning method, Cross-level Community Contrastive Learning (CCCL), for subgraph-level tasks like CVP.
• To enhance the prediction of community values by leveraging both node-level and community-level graph information.
• To establish a new benchmark in graph contrastive learning for social commerce applications.

Main Methods:

• Developed CCCL, a cross-level graph contrastive learning framework.
• Generated two distinct graph views: an augmented node-level graph and a community-level graph via graph coarsening.
• Employed a cross-view contrastive loss to capture mutual information between the node and community views.

Main Results:

• CCCL effectively learns embeddings that utilize multi-level community and node information.
• The proposed method significantly outperforms both end-to-end and self-supervised baselines on the CVP task.
• The CCCL model demonstrates robust resistance to edge perturbation attacks, indicating its stability.

Conclusions:

• CCCL is the first graph contrastive learning method specifically designed for the CVP problem.
• The theoretical analysis shows CCCL maximizes a lower bound of mutual information between different graph representations.
• CCCL offers a highly effective and robust solution for Community Value Prediction in social commerce.