Disentangled and Self-Explainable Node Representation Learning

TL;DR

DiSeNE introduces a disentangled, self-explainable framework for unsupervised node embeddings in graphs, where each embedding dimension is tied to a distinct topological substructure. It combines a random-walk connectivity objective with a dimension-wise disentanglement mechanism and an entropy-based regularizer, producing per-dimension explanation subgraphs via edge-attribution masks. The authors propose five novel metrics (Topological Alignment, Sparsity, Overlap Consistency, Positional Coherence, Plausibility) to quantify interpretability alongside representation quality, and demonstrate that DiSeNE achieves state-of-the-art interpretability across real and synthetic graphs while remaining competitive on downstream tasks. The work shifts explainability from post-hoc analysis of embeddings to inherently interpretable representations, enabling human-centric insights and potential downstream benefits in transparent graph learning applications.

Abstract

Node representations, or embeddings, are low-dimensional vectors that capture node properties, typically learned through unsupervised structural similarity objectives or supervised tasks. While recent efforts have focused on explaining graph model decisions, the interpretability of unsupervised node embeddings remains underexplored. To bridge this gap, we introduce DiSeNE (Disentangled and Self-Explainable Node Embedding), a framework that generates self-explainable embeddings in an unsupervised manner. Our method employs disentangled representation learning to produce dimension-wise interpretable embeddings, where each dimension is aligned with distinct topological structure of the graph. We formalize novel desiderata for disentangled and interpretable embeddings, which drive our new objective functions, optimizing simultaneously for both interpretability and disentanglement. Additionally, we propose several new metrics to evaluate representation quality and human interpretability. Extensive experiments across multiple benchmark datasets demonstrate the effectiveness of our approach.

Figures (19)

• Figure 1: DiSeNE generates dimension-wise disentangled representations in which each embedding dimension is mapped to a mesoscale substructure in the input graph. The vector represents the embedding for the node marked in blue and the bars depict feature values.
• Figure 2: The overlap in dimension explanations aligns with the correlation between the node feature values for those dimensions. The dimension referenced by the blue subgraph shows a stronger correlation with the red dimensions and a lower correlation with the green dimension.
• Figure 3: The node feature value indicates its proximity to the explanation substructure mapped to the corresponding dimension. The black node has a higher value for the dimension corresponding to the green subgraph (since it is 1 hop away) than for the dimension corresponding to the red subgraph (3 hops away).
• Figure 4: Sketch of Plausibility metric computation. High plausibility scores indicate that the dimensions deemed more comprehensible also received higher importance scores from the post-hoc attribution technique.
• Figure A1: Downstream tasks results on real-world datasets (link prediction on the top panel, multi-label node classification on the bottom panel) with varying feature dimensions size.