The Interpretable and Effective Graph Neural Additive Networks

TL;DR

The paper tackles the need for transparency in graph-based learning by introducing Graph Neural Additive Networks (GNAN), an interpretable-by-design GNN that extends Generalized Additive Models to graphs. GNAN learns a distance-weighted aggregation using a global distance function $ρ$ and univariate feature shape functions $f_k$, yielding node representations $[\mathbf{h}_i]_k$ that support exact, human-understandable visualizations and explanations. Empirically, GNAN achieves competitive accuracy on standard node and graph prediction benchmarks, with particular strength on long-range dependency tasks, while providing both global and local explanations and debugging capabilities. The work demonstrates that interpretability need not come at substantial accuracy cost and outlines future directions for smoother shape functions, per-feature weighting, and readout enhancements.

Abstract

Graph Neural Networks (GNNs) have emerged as the predominant approach for learning over graph-structured data. However, most GNNs operate as black-box models and require post-hoc explanations, which may not suffice in high-stakes scenarios where transparency is crucial. In this paper, we present a GNN that is interpretable by design. Our model, Graph Neural Additive Network (GNAN), is a novel extension of the interpretable class of Generalized Additive Models, and can be visualized and fully understood by humans. GNAN is designed to be fully interpretable, offering both global and local explanations at the feature and graph levels through direct visualization of the model. These visualizations describe exactly how the model uses the relationships between the target variable, the features, and the graph. We demonstrate the intelligibility of GNANs in a series of examples on different tasks and datasets. In addition, we show that the accuracy of GNAN is on par with black-box GNNs, making it suitable for critical applications where transparency is essential, alongside high accuracy.

Figures (13)

• Figure 1: Visualization of the distance and feature functions, learned on Mutagenicity. As the features are binary, the feature functions are evaluated only on the value $1$. These plots provide an exact description of the functions' signal processing and a global explanation of how the model uses the distances and features.
• Figure 2: Visualization of products of the outputs of the distance function and the feature functions, trained on Mutagenicity. Each cell shows the exact contribution, positive or negative, of features at a certain distance to the prediction. Positive values (green) contribute to classifying a molecule as mutagenic, and negative values (red) contribute to classifying a molecule as non-mutagenic.
• Figure 3: Visualization of the distance shape function learned on the PubMed dataset. As the output of the function is of dimension three, we plot it as three shape functions, one for each class. We plot them on the same figure to compare them. The shape functions show that the model uses only the local neighborhood of each node. It also shows a difference between the classes; while for type 2 diabetes, the longer the distance, the less their information is used (converges to 0), for type 1 and gestational diabetes, nodes of long-distance have a negative effect.
• Figure 4: Visualization of nine features' shape functions, learned over the PubMed dataset.
• Figure 5: Visualization of the products between the outputs of the 'children' feature function over the input range $[0, 1]$ and the outputs of the distance function, learned over the PubMed dataset.