Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

Technical Summary: Overcoming Shortcut Learning in Graph Neural Networks through Active Explanation Guidance

Problem Statement
Graph Neural Networks (GNNs) frequently suffer from "shortcut learning," where models exploit spurious correlations (confounders) rather than causal patterns to achieve high in-distribution accuracy. These shortcuts—such as specific node features, metadata, or simple input statistics—allow models to perform well on training data but fail to generalize when the data distribution shifts (out-of-distribution). Existing solutions often require access to multi-domain data or extensive observational data, which are not always available. Furthermore, while GNN explanations can reveal reliance on shortcuts, obtaining ground-truth explanations for an entire dataset is prohibitively expensive and time-consuming due to the need for expert annotation.

Methodology: XIGL
The authors propose XIGL (eXplanatory Interactive Graph shortcut unLearning), an architecture-agnostic, human-in-the-loop strategy designed to deconfound GNNs by iteratively removing shortcut dependencies. The framework operates on two key insights: (1) GNN explanations can expose reliance on shortcuts, and (2) expert users can provide corrective feedback on these explanations to guide the model toward causal reasoning.

The core methodology involves three components:

1. Explanation-Guided Loss Function:
   XIGL utilizes input gradients as a proxy for node importance. To penalize reliance on irrelevant (spurious) nodes ($V_{ir}$), the authors introduce a loss function $L = L_{ce} + \lambda L_e$.

   • $L_{ce}$ is the standard cross-entropy loss.
   • $L_e$ minimizes the ratio of input gradients of irrelevant nodes to the total input gradients.
   • A normalization term is included in the denominator to prevent the trivial solution where all gradients collapse to zero.
   • The ground-truth explanation vector $e$ (where $e_j=1$ for relevant nodes and $0$ otherwise) is used to weight this penalty.

2. Active Learning Pipeline:
   Recognizing that annotating explanations for all data is impractical, XIGL employs an active learning (AL) strategy with a fixed annotation budget $B$.

   • Initialization: The model is trained on a small initial set of graphs with expert-corrected explanations.
   • Iterative Selection: In each iteration, an active learning query strategy selects a batch of the most informative graphs from the unlabeled pool.
   • Correction & Fine-tuning: An expert annotates the selected graphs (indicating which nodes the model should not rely on), and the model is fine-tuned using the explanation-guided loss.

3. Query Strategies:
   The paper proposes and evaluates two novel explanation-based query strategies alongside standard prediction-based baselines:

   • Maximum Explanation Entropy (MaEE): Selects graphs where the model's explanation (input gradients) is most uncertain (uniform across nodes).
   • Minimum Explanation Entropy (MiEE): Selects graphs where the model is most confident in its explanation. This strategy leverages the "simplicity bias" of neural networks, hypothesizing that highly confident explanations are often overfitted to spurious shortcuts and thus most in need of correction.
   • Baselines: Maximum Classification Entropy (MaCE) and Random Sampling.

Key Contributions

• XIGL Framework: An interactive pipeline that integrates human feedback on GNN explanations to actively remove shortcut dependencies.
• Novel Query Strategies: The introduction of explanation-based active learning strategies (MaEE and MiEE) that prioritize instances based on the quality of the model's reasoning rather than just prediction uncertainty.
• Efficiency: Demonstrating that high-performance deconfounding can be achieved with a fraction of the annotation budget required for passive supervision (full dataset annotation).

Results
Experiments were conducted on two datasets: ER-color (synthetic) and CPatchMNIST (real-world extension), across four GNN architectures (GCN, GIN, GraphSAGE, GAT).

• Passive Supervision: When provided with explanations for the entire dataset, XIGL significantly improves generalization on unconfounded test sets compared to models trained without explanation supervision. For instance, on ER-color, models without supervision performed near chance level, while those with passive supervision showed substantial accuracy gains.
• Active Supervision: Active learning strategies generally outperformed random sampling.
  • On CPatchMNIST, explanation-based strategies (MaEE and MiEE) consistently outperformed prediction-based strategies (MaCE) across all architectures.
  • MiEE emerged as the most consistently high-performing strategy, suggesting that targeting instances where the model is overly confident in its (potentially spurious) reasoning is highly effective.
  • While active learning achieved slightly lower accuracy than passive learning (due to the smaller annotation budget), the gap was relatively small, with active learning using less than 20% of the training data.

Significance and Claims
The paper claims that XIGL offers a practical route to improving GNN reliability in out-of-distribution scenarios without requiring multi-domain data. By leveraging the fact that explanations naturally expose shortcut reliance, the method allows expert users to efficiently "unlearn" spurious patterns. The authors emphasize that their approach reduces both annotation and cognitive costs by focusing expert effort only on instances where feedback is most beneficial.

The work concludes that explanation-based active learning is a promising direction for deconfounding, particularly strategies that target the model's confidence in its own reasoning (MiEE). The authors note that future work will explore self-explainable GNNs, the impact of explanation faithfulness on deconfounding success, and the application of these methods to a broader range of real-world datasets and shortcut types.