On Consistency in Graph Neural Network Interpretation

Abstract

Uncovering rationales behind predictions of graph neural networks (GNNs) has received increasing attention over recent years. Instance-level GNN explanation aims to discover critical input elements, like nodes or edges, that the target GNN relies upon for making predictions. These identified sub-structures can provide interpretations of GNN's behavior. Though various algorithms are proposed, most of them formalize this task by searching the minimal subgraph which can preserve original predictions. An inductive bias is deep-rooted in this framework: the same output cannot guarantee that two inputs are processed under the same rationale. Consequently, they have the danger of providing spurious explanations and fail to provide consistent explanations. Applying them to explain weaklyperformed GNNs would further amplify these issues. To address the issues, we propose to obtain more faithful and consistent explanations of GNNs. After a close examination on predictions of GNNs from the causality perspective, we attribute spurious explanations to two typical reasons: confounding effect of latent variables like distribution shift, and causal factors distinct from the original input. Motivated by the observation that both confounding effects and diverse causal rationales are encoded in internal representations, we propose a simple yet effective countermeasure by aligning embeddings. This new objective can be incorporated into existing GNN explanation algorithms with no effort. We implement both a simplified version based on absolute distance and a distribution-aware version based on anchors. Experiments on 5 datasets validate its effectiveness, and theoretical analysis shows that it is in effect optimizing a more faithful explanation objective in design, which further justifies the proposed approach.Preprint. Under review.

Disambiguated Node Classification with Graph Neural Networks

“…Additionally, works like Factorizable Graphs [46,60] and Decoupled Graphs [55] propose methods to unveil latent groups of nodes or edges and convey messages on disentangled graphs. Recent investigations have also delved into the trustworthiness [12,38,39,52] and interpretability [11,62,64] of GNNs. More recent progresses of GNN can be found in these surveys [25,71].…”

Faithful and Consistent Graph Neural Network Explanations with Rationale Alignment

“…Second, it increases the models' transparency to enable trusted applications in decision-critical fields sensitive to fairness, privacy and safety challenges, such as healthcare and drug discovery [11]. Thus, studying the explainability of GNNs is attracting increasing attention and many efforts have been taken [12], [13], [14].…”

T-SaS: Toward Shift-aware Dynamic Adaptation for Streaming Data

“…This occurs when the statistical characteristics of newly incoming data differ from those observed by the model in a dynamically changing environment [3,13,17,21,32]. Generally, the shift can happen without any precursor, be unknown to users, cause dramatic personal injury for systems like self-driving [8], robotics [24,31], sleep tracking [22] and irreparable economic damage on financial trading algorithms [16,20,22,27]. At any moment, the model is expected to 1) provide early warnings when the data distribution changes and 2) make accurate predictions by adapting to new data.…”