TL;DR — Instead of learning black-box node vectors and employing post-hoc explainers, DiSeNE learns unsupervised embeddings whose dimensions are interpretable by design: each dimension maps to a concrete mesoscale substructure (e.g., an anchor/community) in the graph. You get dimension-wise explanations, competitive utility, and a clean way to tease apart structure from node features.

What is the problem with “explain the GNN after the fact”

End-to-end GNNs mix structure and features during message passing: many different combinations of neighborhoods and attributes can yield the same prediction, so post-hoc attributions lead to multiple plausible explanations. DiSeNE sidesteps this by separating concerns: it learns structural factors first—one per dimension—so you can study what the graph alone contributes before adding metadata.

Key Idea

DiSeNE jointly optimizes three goals: connectivity preservation (a random-walk/skip-gram objective), dimensional interpretability (each dimension “explains” edges via a subgraph), and structural disentanglement (dimensions capture distinct structures). A light entropy regularizer prevents empty/degenerate dimensions.

Why this matters. Each coordinate now has a reason (“close to subgraph Sₖ”), not just a number that happens to work in a downstream task. That makes the embedding space auditable and easy to discuss with domain experts

Measuring interpretability (not just accuracy)

The paper proposes new task-agnostic metrics for self-explainable embeddings, including Topological Alignment (do per-dimension explanation masks align with human-interpretable structures like communities?), Overlap Consistency (do overlaps between explanation subgraphs match correlations between dimensions?), and Positional Coherence (do higher coordinate values mean closer to the subgraph?).