GraphNNK -- Graph Classification and Interpretability

TL;DR

The paper addresses interpretability in graph classification by substituting the parametric softmax with a non-parametric NNK interpolator applied to GNN embeddings. It demonstrates that NNK provides explicit, example-based explanations and can achieve competitive accuracy on NCI1 when the embedding space is well-structured. Predictions are derived from a sparse set of active training neighbors found via nearest-neighbor retrieval and solved through a constrained optimization in kernel space. This approach offers transparent decision-making for graph analytics and points to future work on adaptive non-parametric classifiers.

Abstract

Graph Neural Networks (GNNs) have become a standard approach for learning from graph-structured data. However, their reliance on parametric classifiers (most often linear softmax layers) limits interpretability and sometimes hinders generalization. Recent work on interpolation-based methods, particularly Non-Negative Kernel regression (NNK), has demonstrated that predictions can be expressed as convex combinations of similar training examples in the embedding space, yielding both theoretical results and interpretable explanations.

Figures (2)

• Figure 1: Proposed architecture for graph classification.
• Figure 2: Supervised vs. NNK comparison: best validation checkpoint (top) and last training snapshot (bottom). Each pair shows accuracy (left) and F1 score (right).