Learning to Model Graph Structural Information on MLPs via Graph Structure Self-Contrasting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Representation Learning

Background:

• Graph Neural Networks (GNNs) excel at graph tasks but rely on message passing, which can be sensitive to noise and perturbations.
• Explicitly coupling node features with structural information via message passing in GNNs can lead to error propagation and reduced robustness.

Purpose of the Study:

• To propose a novel framework, Graph Structure Self-Contrasting (GSSC), that learns graph structural information without relying on message passing.
• To enhance the robustness and generalization capabilities of graph-based machine learning models.

Main Methods:

• The GSSC framework utilizes multilayer perceptrons (MLPs) and implicitly incorporates structural information as prior knowledge.
• It employs structural sparsification (STR-Sparse) to remove noisy edges and structural self-contrasting (STR-Contrast) for learning robust node representations.
• STR-Sparse and STR-Contrast are formulated as a bilevel optimization problem within a unified framework.

Main Results:

• GSSC demonstrates superior performance compared to leading competitors in graph-related tasks.
• Experiments show that the GSSC framework achieves better generalization and robustness.
• Qualitative and quantitative results validate the effectiveness of the proposed approach.

Conclusions:

• The GSSC framework offers an effective alternative to traditional message-passing GNNs for learning from graph data.
• By avoiding explicit message passing, GSSC mitigates issues related to noise and perturbations, leading to more reliable models.
• The proposed method provides a promising direction for developing more robust graph representation learning techniques.