ExplainabilityTransparencyAccountabilitycs.LG
Higher-Order Explanations of Graph Neural Networks via Relevant Walks
Thomas Schnake, Oliver Eberle, Jonas Lederer, Shinichi Nakajima, Kristof T. Schütt, Klaus-Robert Müller, Grégoire Montavon
June 5, 2020
IEEE Transactions on Pattern Analysis and Machine Intelligence
319 citations
21 influential
Key Takeaways
-
GNNs can now be explained through interpretable patterns: The paper demonstrates that Graph Neural Networks—previously considered "black boxes"—can be made transparent by identifying groups of connected edges (called "walks") that jointly influence predictions, making their decision-making process more understandable to users.
-
A practical method exists to extract these explanations: The researchers developed GNN-LRP, a technique that breaks down how GNNs make decisions by applying existing explanation methods layer-by-layer, producing concrete, traceable paths through the input data that explain why a specific prediction was made.
-
The approach works across diverse real-world applications: The method was successfully tested on three different domains—sentiment analysis of text, predicting chemical properties in quantum chemistry, and image classification—suggesting it could be broadly applicable to many GNN-based systems used in practice.
-
This addresses a critical gap in AI transparency: Since GNNs tightly integrate input data into their structure in ways that "common explainable AI approaches are not applicable," this work fills an important gap in making an increasingly popular class of AI systems more accountable and trustworthy.
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.