IBCS: Learning Information Bottleneck-Constrained Denoised Causal Subgraph for Graph Classification.

Area of Science:

• Graph Learning
• Causal Inference
• Machine Learning

Background:

• Extracting precise causal subgraphs is crucial for interpreting and improving graph learning predictions.
• Existing methods often fail because they only address spurious or noisy data, not both simultaneously.
• Real-world scenarios frequently involve a coexistence of causal, spurious, and noisy subgraphs.

Purpose of the Study:

• To propose a more realistic problem formulation for causal subgraph extraction where spurious and noisy data coexist.
• To develop a novel model capable of simultaneously identifying and excluding spurious and noisy components from causal subgraphs.
• To accurately extract the true causal substructure for reliable interpretation and prediction.

Main Methods:

• Introduced a novel problem formulation hypothesizing graphs as a mixture of causal, spurious, and noisy subgraphs.
• Developed an Information Bottleneck-constrained denoised Causal Subgraph (IBCS) learning model.
• Designed a causal learning objective incorporating intervention on spurious features and an information bottleneck constraint to filter noise.

Main Results:

• Theoretically proved that the causal subgraph extracted by IBCS can approximate the ground-truth.
• Empirically demonstrated the superiority of IBCS over state-of-the-art baselines on nine benchmark datasets.
• IBCS effectively excludes both spurious and noisy parts, enabling accurate causal subgraph extraction.

Conclusions:

• The proposed problem formulation and IBCS model offer a more practical approach to causal subgraph learning.
• IBCS successfully addresses the challenge of simultaneous spurious and noisy subgraph exclusion.
• The findings highlight the effectiveness of IBCS in extracting reliable causal structures for enhanced graph learning applications.