[Paper Review] Graph-MVP: Multi-View Prototypical Contrastive Learning for Multiplex Graphs

Baoyu Jing, Yuejia Xiang+3 more|arXiv (Cornell University)|Sep 8, 2021

Advanced Graph Neural Networks63 references10 citations

TL;DR

This paper proposes Graph-MVP, a multi-view prototypical contrastive learning framework for multiplex graphs that addresses semantic errors in contrastive learning and leverages multiple relational views. By combining graph prototypical contrastive learning with view alignment via hard and soft clustering strategies, Graph-MVP improves node representation learning, achieving state-of-the-art performance on downstream tasks across real-world datasets.

ABSTRACT

Contrastive Learning (CL) is one of the most popular self-supervised learning frameworks for graph representation learning, which trains a Graph Neural Network (GNN) by discriminating positive and negative node pairs. However, there are two challenges for CL on graphs. On the one hand, traditional CL methods will unavoidably introduce semantic errors since they will treat some semantically similar nodes as negative pairs. On the other hand, most of the existing CL methods ignore the multiplexity nature of the real-world graphs, where nodes are connected by various relations and each relation represents a view of the graph. To address these challenges, we propose a novel Graph Multi-View Prototypical (Graph-MVP) framework to extract node embeddings on multiplex graphs. Firstly, we introduce a Graph Prototypical Contrastive Learning (Graph-PCL) framework to capture both node-level and semantic-level information for each view of multiplex graphs. Graph-PCL captures the node-level information by a simple yet effective data transformation technique. It captures the semantic-level information by an Expectation-Maximization (EM) algorithm, which alternatively performs clustering over node embeddings and parameter updating for GNN. Next, we introduce Graph-MVP based on Graph-PCL to jointly model different views of the multiplex graphs. Our key insight behind Graph-MVP is that different view-specific embeddings of the same node should have similar underlying semantic, based on which we propose two versions of Graph-MVP: Graph-MVP_hard and Graph-MVP_soft to align embeddings across views. Finally, we evaluate the proposed Graph-PCL and Graph-MVP on a variety of real-world datasets and downstream tasks. The experimental results demonstrate the effectiveness of the proposed Graph-PCL and Graph-MVP frameworks.

Motivation & Objective

• To address semantic errors in contrastive learning where semantically similar nodes are incorrectly treated as negative pairs.
• To exploit the multiplex structure of real-world graphs, where nodes are connected via multiple relations, by modeling each relation as a distinct view.
• To improve node representation learning by jointly modeling multiple views using prototypical clustering and cross-view alignment.
• To develop a unified framework that captures both node-level and semantic-level information across views in a self-supervised manner.

Proposed method

• Introduces Graph-PCL, a prototypical contrastive learning framework that uses data augmentation for node-level contrast and an Expectation-Maximization (EM) algorithm to capture semantic-level structure.
• Employs an EM-based clustering procedure that alternates between clustering node embeddings and updating GNN parameters to refine semantic representations.
• Applies a multi-view learning strategy in Graph-MVP to jointly model different relational views of the same node in a multiplex graph.
• Proposes two variants—Graph-MVP_hard and Graph-MVP_soft—using hard and soft clustering to align view-specific embeddings based on shared semantic content.
• Uses instance-level and prototype-level contrastive learning to enhance both local and global structure awareness in node representations.
• Aligns embeddings across views by encouraging consistent clustering assignments and minimizing divergence between view-specific representations.

Experimental results

Research questions

• RQ1 Can prototypical contrastive learning reduce semantic errors in graph contrastive learning by capturing underlying node semantics?
• RQ2 How can multiple relational views in multiplex graphs be effectively leveraged to improve node representation learning?
• RQ3 What is the impact of cross-view alignment on the quality of learned node embeddings in multi-relational graphs?
• RQ4 How do hard versus soft clustering strategies affect the performance of multi-view graph representation learning?

Key findings

• Graph-MVP achieves state-of-the-art performance on multiple downstream node classification and clustering tasks across diverse real-world multiplex graph datasets.
• The Graph-PCL component significantly improves representation quality by reducing semantic errors through EM-based clustering and prototype learning.
• Graph-MVP_hard and Graph-MVP_soft both outperform baseline contrastive methods, with the soft alignment variant showing stronger performance on datasets with complex relational structures.
• The ablation study confirms that both view-level contrastive learning and cross-view alignment are essential for optimal performance.