GraphLIME: Local Interpretable Model Explanations for Graph Neural Networks (2001.06216v2)

Abstract: Graph structured data has wide applicability in various domains such as physics, chemistry, biology, computer vision, and social networks, to name a few. Recently, graph neural networks (GNN) were shown to be successful in effectively representing graph structured data because of their good performance and generalization ability. GNN is a deep learning based method that learns a node representation by combining specific nodes and the structural/topological information of a graph. However, like other deep models, explaining the effectiveness of GNN models is a challenging task because of the complex nonlinear transformations made over the iterations. In this paper, we propose GraphLIME, a local interpretable model explanation for graphs using the Hilbert-Schmidt Independence Criterion (HSIC) Lasso, which is a nonlinear feature selection method. GraphLIME is a generic GNN-model explanation framework that learns a nonlinear interpretable model locally in the subgraph of the node being explained. More specifically, to explain a node, we generate a nonlinear interpretable model from its $N$-hop neighborhood and then compute the K most representative features as the explanations of its prediction using HSIC Lasso. Through experiments on two real-world datasets, the explanations of GraphLIME are found to be of extraordinary degree and more descriptive in comparison to the existing explanation methods.

• The paper introduces GraphLIME, a novel method to generate local, interpretable explanations for GNN predictions using HSIC Lasso.
• It employs an N-hop sampling strategy that captures feature and structural dependencies to ensure contextually faithful explanations.
• Experimental results on datasets like Cora and Pubmed show GraphLIME reduces noisy features and outperforms methods like GNNExplainer.

Overview of GraphLIME for Explaining Graph Neural Networks

The paper introduces GraphLIME, a methodology designed to provide local, interpretable model explanations specifically for Graph Neural Networks (GNNs), leveraging the Hilbert-Schmidt Independence Criterion (HSIC) Lasso. Given the rising prominence of GNNs in representing graph-structured data from various domains such as social networks, chemistry, and biology, the lack of interpretability in their predictions poses notable challenges. The need for models to be trusted beyond mere accuracy is emphasized, especially in critical applications like medical diagnosis, where understanding the decision process is paramount.

Key Contributions and Methodology

1. Problem Formulation: The paper articulates the need for interpretable models for GNNs, which inherently deal with complex structured data. Traditional deep models often function as black boxes, providing limited insights into model decisions. GraphLIME addresses this by providing explanations of node predictions derived from their N-hop neighborhood within the graph, ensuring local fidelity in interpretations.
2. Utilization of HSIC Lasso: The authors employ HSIC Lasso, a nonlinear feature selection method, as the backbone of their explanation model. This choice allows the model to capture non-linear dependencies between the input features and the GNN's outputs, thereby identifying the most representative features that lead to specific predictions.
3. Local Sampling Strategy: To ascertain explanations, GraphLIME considers the N-hop neighborhood of the node, which ensures that explanations are contextually localized. This sampling captures both feature and graph structural dependencies, allowing for more comprehensive insights into the model's behavior in local subgraphs.
4. Experimental Validation: The framework's efficacy is substantiated through experiments on real-world datasets (Cora and Pubmed). The results underscore GraphLIME's superior ability to filter out noisy features and offer clearer explanations compared to existing methods like GNNExplainer and LIME.
5. Comparison with Other Methodologies: The results demonstrate that GraphLIME consistently selects fewer noisy features and provides more reliable explanations, helping users determine the trustworthiness of model predictions. Moreover, the proposed method's explanations facilitate better model selection by highlighting the less spurious classifier consistently.

Implications and Future Directions

The proposed GraphLIME method significantly improves the interpretability of GNN models, ensuring that predictions are not just accurate but also understandable and trustworthy. The implications of such a framework are manifold:

• Enhanced Trust: Providing explanations aligned with human reasoning enhances trust in ML systems, especially in safety-critical applications.
• Model Transparency: It offers insights into the model's decision pathways, promoting transparency and aiding in debugging and model improvement.
• Guidance for Feature Engineering: By highlighting informative and influential features, GraphLIME can guide feature engineering efforts for better model performance.

For future research, the authors suggest extending GraphLIME to explain graph structural patterns and provide group-level explanations across sets of nodes. This would not only enhance individual node interpretability but also offer insights into community or cluster behaviors within graphs.

In conclusion, GraphLIME represents a valuable addition to the toolkit for GNN interpretability, offering a robust approach to elucidate complex model predictions in graph structures. As AI continues to advance, methodologies such as GraphLIME become instrumental in reinforcing the bridge between machine learning models and human interpretability.