Discovering Invariant Rationales for Graph Neural Networks (2201.12872v1)

Abstract: Intrinsic interpretability of graph neural networks (GNNs) is to find a small subset of the input graph's features -- rationale -- which guides the model prediction. Unfortunately, the leading rationalization models often rely on data biases, especially shortcut features, to compose rationales and make predictions without probing the critical and causal patterns. Moreover, such data biases easily change outside the training distribution. As a result, these models suffer from a huge drop in interpretability and predictive performance on out-of-distribution data. In this work, we propose a new strategy of discovering invariant rationale (DIR) to construct intrinsically interpretable GNNs. It conducts interventions on the training distribution to create multiple interventional distributions. Then it approaches the causal rationales that are invariant across different distributions while filtering out the spurious patterns that are unstable. Experiments on both synthetic and real-world datasets validate the superiority of our DIR in terms of interpretability and generalization ability on graph classification over the leading baselines. Code and datasets are available at https://github.com/Wuyxin/DIR-GNN.

Overview of "Discovering Invariant Rationales for Graph Neural Networks"

The paper "Discovering Invariant Rationales for Graph Neural Networks" introduces a novel approach to enhance the interpretability and generalization of Graph Neural Networks (GNNs) by identifying invariant rationales. GNNs have demonstrated substantial prowess in various applications, yet their intrinsic interpretability remains a challenging task. Interpretability in GNNs involves isolating subsets of input features that fundamentally guide the prediction process. However, many existing methods overly depend on spurious correlations or biases within training data, thereby potentially causing significant performance degradation when these models encounter out-of-distribution (OOD) data.

Methodology and Contributions

The authors propose a strategy called Discovering Invariant Rationale (DIR), which aims to construct intrinsically interpretable GNNs by creating multiple interventional distributions through causal interventions on the training data. This process helps in distinguishing critical causal rationales that remain stable across different distributions from spurious patterns that may fluctuate. The paper makes several notable contributions:

1. DIR Framework: This encompasses a rationale generator, a distribution intervener, a feature encoder, and two classifiers. The rationale generator separates the input graph into causal and non-causal subgraphs. Through causal interventions, the distribution intervener generates perturbed distributions, allowing the inference of invariant causal features.
2. Invariant Risk Objective: The DIR method minimizes the variance across interventional distributions, promoting the learning of causal features and discarding unstable spurious correlations.
3. Empirical Results: Experiments conducted on synthetic and real-world datasets demonstrate that DIR outperforms state-of-the-art methods concerning interpretability and generalization, particularly in OOD scenarios. For instance, on Spurious-Motif datasets, DIR achieved an improvement in precision for identifying causal features, significantly surpassing traditional methods like graph attention and pooling approaches.

Strong Numerical Results and Implications

The findings underscore DIR's superior capacity for retaining interpretative power and enhancing prediction accuracy beyond the training data's underlying distribution biases. Notably, the method's success in reducing variance in the interventional risk shows promise for improving model robustness, thus signifying a critical step in model reliability for real-world applications where datasets often present biases.

Theoretical and Practical Implications

Theoretically, the paper contributes a causal learning perspective to GNN interpretability, addressing how invariant causal relationships can be utilized to bolster model performance. Practically, this approach can have far-reaching implications, particularly in scientific fields like bioinformatics and chemistry, where understanding the causal interactions within data is crucial. Such advancements further suggest the potential for broader applications, where robust and interpretable models are essential.

Future Directions

As an extension to this work, research might explore the incorporation of more complex causal models and alternative methods of generating interventional distributions. The domain could also benefit from studies on the scalability of DIR across larger and more complex graph datasets, alongside exploring its adaptability in different GNN-based architectures. Furthermore, integrating DIR with other interpretability frameworks or extending it to semi-supervised learning domains could yield substantial advancements in GNN research.

Overall, the methodology presented in this paper marks a promising advancement toward more interpretable and reliable graph-based models, offering insights that could lead to the development of next-generation AI systems with enhanced transparency and robustness.