[Paper Review] Task-Agnostic Graph Explanations

Yaochen Xie, Sumeet Katariya+5 more|arXiv (Cornell University)|Feb 16, 2022

Explainable Artificial Intelligence (XAI)4 citations

TL;DR

This paper proposes Tage, a task-agnostic GNN explainer that learns graph embedding explanations via self-supervised training without access to downstream tasks, enabling efficient, unified explanation of multitask GNN models and individual embedding dimensions. Tage achieves state-of-the-art explanation quality with significantly reduced inference time, especially in multitask settings, outperforming task-specific baselines like GNNExplainer and PGExplainer.

ABSTRACT

Graph Neural Networks (GNNs) have emerged as powerful tools to encode graph-structured data. Due to their broad applications, there is an increasing need to develop tools to explain how GNNs make decisions given graph-structured data. Existing learning-based GNN explanation approaches are task-specific in training and hence suffer from crucial drawbacks. Specifically, they are incapable of producing explanations for a multitask prediction model with a single explainer. They are also unable to provide explanations in cases where the GNN is trained in a self-supervised manner, and the resulting representations are used in future downstream tasks. To address these limitations, we propose a Task-Agnostic GNN Explainer (TAGE) that is independent of downstream models and trained under self-supervision with no knowledge of downstream tasks. TAGE enables the explanation of GNN embedding models with unseen downstream tasks and allows efficient explanation of multitask models. Our extensive experiments show that TAGE can significantly speed up the explanation efficiency by using the same model to explain predictions for multiple downstream tasks while achieving explanation quality as good as or even better than current state-of-the-art GNN explanation approaches. Our code is pubicly available as part of the DIG library at https://github.com/divelab/DIG/tree/main/dig/xgraph/TAGE/.

Motivation & Objective

• To address the inefficiency and inapplicability of existing task-specific GNN explainers in multitask and self-supervised learning settings.
• To enable a single explainer to explain multiple downstream tasks without retraining, improving scalability and deployment efficiency.
• To develop a self-supervised training framework for GNN embedding explainers that does not require downstream task labels or predictions.
• To provide faithful, sparse, and interpretable explanations for individual GNN embedding dimensions without relying on downstream models.
• To reduce the time cost of explanation in real-world graph datasets with hundreds of downstream tasks, where training per-task explainers is infeasible.

Proposed method

• Decompose a GNN model into a GNN embedding model and a downstream prediction head, and train separate explainers for each component.
• Design a task-agnostic embedding explainer (Tage) that is trained using a self-supervised contrastive objective on graph or node embeddings, independent of downstream tasks.
• Use a conditional input mechanism where the explainer takes a one-hot vector as a condition to explain specific embedding dimensions.
• Train the explainer using instance-level contrastive learning between original and perturbed subgraphs, optimizing for fidelity and sparsity in the explanation.
• Leverage the same trained embedding explainer to explain multiple downstream tasks by conditioning on different downstream models.
• Use fidelity scores and edge importance rankings to evaluate explanation quality, with red-highlighted edges indicating top 10% important edges.

Experimental results

Research questions

• RQ1 Can a single GNN explainer be trained without knowledge of downstream tasks and still produce high-quality explanations for multiple downstream tasks?
• RQ2 How does the performance of a task-agnostic explainer compare to task-specific baselines in terms of fidelity and sparsity?
• RQ3 Can the proposed explainer provide meaningful explanations for individual GNN embedding dimensions without relying on downstream predictions?
• RQ4 What is the computational efficiency gain of using a single task-agnostic explainer across multiple downstream tasks compared to training per-task explainers?
• RQ5 Does self-supervised training on embeddings lead to better explanation consistency than training on final predictions, especially when downstream models are non-injective?

Key findings

• Tage achieves higher explanation fidelity and sparsity than task-specific baselines like GNNExplainer and PGExplainer, even with the same number of parameters.
• In multitask settings (T > 1), Tage reduces explanation time cost significantly compared to GNNExplainer and PGExplainer, which require separate training per task.
• Tage produces more coherent explanations, highlighting connected substructures (e.g., functional groups in molecules), unlike scattered edge selections from baselines.
• The explainer successfully generates meaningful explanations for individual embedding dimensions without any downstream task, demonstrating interpretability of learned representations.
• Fidelity scores remain high when explaining embedding dimensions, indicating that the explanations are consistent and relevant to downstream prediction tasks.
• The self-supervised training objective based on embeddings provides stronger supervision than task-based objectives, particularly in non-injective downstream models where similar predictions can arise from dissimilar embeddings.