xAI-Drop: Don't Use What You Cannot Explain

TL;DR

It is argued that explainability should be a key indicator of a model's quality throughout its training phase and introduced xAI-Drop, a novel topological-level dropping regularizer that leverages explainability to pinpoint noisy network elements to be excluded from the GNN propagation mechanism.

Abstract

Graph Neural Networks (GNNs) have emerged as the predominant paradigm for learning from graph-structured data, offering a wide range of applications from social network analysis to bioinformatics. Despite their versatility, GNNs face challenges such as lack of generalization and poor interpretability, which hinder their wider adoption and reliability in critical applications. Dropping has emerged as an effective paradigm for improving the generalization capabilities of GNNs. However, existing approaches often rely on random or heuristic-based selection criteria, lacking a principled method to identify and exclude nodes that contribute to noise and over-complexity in the model. In this work, we argue that explainability should be a key indicator of a model's quality throughout its training phase. To this end, we introduce xAI-Drop, a novel topological-level dropping regularizer that leverages explainability to pinpoint noisy network elements to be excluded from the GNN propagation mechanism. An empirical evaluation on diverse real-world datasets demonstrates that our method outperforms current state-of-the-art dropping approaches in accuracy, and improves explanation quality.

Figures (11)

• Figure 1: Illustration of the rationale behind xAI-Drop. Panel (a) shows a Barabási-Albert network with house-shaped motifs randomly attached. The task here is to classify nodes as either the top of a house (label 1) or otherwise (label 0). It is easy to see that a triangle is an approximate pattern for the positive class. The figure highlights three prototypical nodes (A, B, C) which are parts of a triangle, where only two (A, B) are also the top of a house (triangle and houses highlighted for readability). Panel (b) reports the explanation of a GNN trained on the network, for the three highlighted nodes. Node A has a high confidence, because it has both the correct (the house) and spurious (the two triangles) patterns. However its explanation is mostly based on the (simpler) spurious triangle, which is insufficient to explain its confidence (as shown by the lower confidence of nodes B and C). Removing A (Panel c) prevents the network from focusing on the spurious pattern, so that the correct pattern is eventually learned (Panel d, with C omitted as no longer predicted as label 1).
• Figure 2: A graphical representation of the node dropping strategy (xAI-DropNode) employed by the xAI-Drop algorithm in node classification tasks. Panel a) illustrates the confidence-based selection process, where nodes are selected if the model’s confidence is equal or greater than a specified threshold $\theta$. Panel b) presents the computation of fidelity sufficiency scores and dropping probabilities for the nodes selected in panel a). Lastly, panel c) illustrates the computation of the node dropping mask by Bernoulli sampling, and the resulting graph after dropping nodes and their associated edges.
• Figure 3: Test accuracy (left axis) and training time (right axis) when using different explainers for xAI-DropNode applied on Cora for node classification with GCN architecture. The dotted line represents the accuracy achieved when using the baseline DropNode random strategy.
• Figure 4: The histogram of the time in seconds required for training GCN on Cora, Citeseer, and PubMed with each regularization method used for node classification.
• Figure 5: The log-scale histogram of the parameters used for training GCN for node classification task with Cora, Citeseer, and Pubmed datasets.