Higher-Order Explanations of Graph Neural Networks via Relevant Walks (2006.03589v3)

Abstract: Graph Neural Networks (GNNs) are a popular approach for predicting graph structured data. As GNNs tightly entangle the input graph into the neural network structure, common explainable AI approaches are not applicable. To a large extent, GNNs have remained black-boxes for the user so far. In this paper, we show that GNNs can in fact be naturally explained using higher-order expansions, i.e. by identifying groups of edges that jointly contribute to the prediction. Practically, we find that such explanations can be extracted using a nested attribution scheme, where existing techniques such as layer-wise relevance propagation (LRP) can be applied at each step. The output is a collection of walks into the input graph that are relevant for the prediction. Our novel explanation method, which we denote by GNN-LRP, is applicable to a broad range of graph neural networks and lets us extract practically relevant insights on sentiment analysis of text data, structure-property relationships in quantum chemistry, and image classification.

• The paper introduces GNN-LRP, a novel method employing higher-order Taylor expansions to reveal the contribution of edge sequences in GNN predictions.
• It leverages deep Taylor decomposition to iteratively propagate relevance scores across layers, outperforming traditional node-level attribution methods.
• Experimental validation on synthetic, sentiment, quantum chemistry, and image classification tasks confirms its effectiveness in enhancing model transparency.

Higher-Order Explanations of Graph Neural Networks via Relevant Walks

This paper, authored by Schnake et al., addresses the explainability of Graph Neural Networks (GNNs), a potent class of models for processing graph-structured data. GNNs are extensively applied across various domains such as quantum chemistry, sentiment analysis, and image classification. Despite their utility, GNNs primarily function as black-box models, presenting challenges in deriving interpretability from their predictions. To address this, the authors propose a method for generating higher-order explanations for GNNs—explanations that elucidate the contribution of walks or edge sequences to the model's predictions.

Conceptual Framework and Methodology

The central innovation presented in the paper is the GNN-LRP (Layer-wise Relevance Propagation) method. The authors leverage the polynomial nature of GNNs to establish a theoretically grounded method for explaining GNN predictions through higher-order Taylor expansions. Traditional attribution methods often focus on node or edge-level contributions. However, the GNN-LRP framework provides a nested attribution scheme, attributing relevance through sequences of connected edges—termed as "walks" in the input graph—thereby offering a more nuanced and comprehensive understanding of model predictions.

This relevance propagation is theoretically justified using insights from deep Taylor decomposition, enabling propagation of relevance scores from the output layer back to the input graph layers. GNN-LRP is applied iteratively across multiple layers of a GNN to identify sequences of nodes and edges that are pivotal to the model's output, thereby transforming the ways GNN explanations are constructed.

Experimental Validation

The paper presents a thorough validation of the proposed method using several benchmark and real-world tasks. The authors demonstrate the application of GNN-LRP on synthetic datasets, sentiment analysis tasks, molecular property prediction, and image classification.

• Synthetic Data Tests: In synthetic graphs, GNN-LRP effectively identifies distinct topological features critical to classification decisions, outperforming conventional node-based methods.
• Sentiment Analysis: By analyzing textual input structured as tree graphs, the method reveals the sentiment conveyed by specific paths through conjunctions of words—showcasing its capability in NLP contexts.
• Quantum Chemistry: The method highlights chemical structures contributing to molecular properties, aligning with domain-specific knowledge, such as bond stability influenced by bond order.
• Image Classification: Adapting classic CNNs as GNNs, the method offers insights into feature assimilation from successive neural layers, enhancing the interpretative depth of traditional pixel-attribution methods.

Implications and Future Directions

The implications of GNN-LRP are manifold. It provides a robust framework for unpacking GNN decisions, facilitating transparency and trust in AI systems. This is essential in domains where understanding the model's reasoning is as crucial as the prediction itself—such as healthcare, finance, and legal domains.

Moreover, GNN-LRP could pave the way for advancements in model refinement and debugging by identifying and remedying erroneous decision paths. As GNN applications expand, GNN-LRP stands as a critical tool for model interpretability, potentially leading to more responsible and ethical AI deployment, aligning with ongoing movements in AI transparency and accountability.

Future research avenues include refining the computational efficiency of GNN-LRP, exploring its applications in dynamic and time-evolving graphs, and integrating it with other interpretative techniques for a holistic understanding of model behavior. Continued exploration of GNN-LRP's integration into complex multi-modal architectures also presents an exciting frontier for AI research.

In conclusion, this paper contributes a significant methodological advancement in GNN interpretability, redefining how practitioners can probe into the decision-making processes of sophisticated graph-based models. GNN-LRP not only fills a critical gap in current XAI techniques but sets the stage for future explorations into the mechanics of complex AI systems.