From Abductive Explanations to Global Logical Rules for Node Classification in SGCs
This paper proposes a logic-based framework for Simple Graph Convolution (SGC) networks that extracts compact, high-fidelity global logical rules by converting minimal abductive explanations of node predictions into decision trees, thereby overcoming the redundancy and limited generalizability of previous subgraph-based explanation methods.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the modern digital world, vast networks of information connect everything from scientific papers to social media profiles. To make sense of these tangled webs, computer scientists use a powerful type of artificial intelligence called a graph neural network. These systems act like expert navigators, learning to predict the nature of individual items within a network by looking at their neighbors and the connections between them. While these tools have become incredibly accurate at tasks like sorting documents or identifying entities, they often operate as black boxes. They produce a correct answer, but they rarely explain how they arrived at it. This lack of transparency is a growing problem; as these systems begin to influence real-world decisions, from scientific research to operational logistics, understanding their reasoning has become just as important as the predictions themselves. Researchers are now racing to build methods that can translate the complex, internal logic of these networks into clear, human-readable rules.
A team of researchers from the Federal Institute of Ceará in Brazil has taken a significant step forward in this quest by developing a new framework called AXSGC. Their work focuses on a specific, streamlined version of graph neural networks known as Simple Graph Convolution. Unlike more complex models that layer multiple non-linear transformations, this simpler version moves information through the network in a direct, linear fashion. The researchers realized that this simplicity offered a unique opportunity: because the math behind the model is so straightforward, they could trace exactly which pieces of information were absolutely necessary to make a specific prediction. They began by examining individual nodes, or points, within the network and asking a precise question: what is the smallest, most essential set of features and connections that guarantees this specific outcome?
To answer this, the team employed a technique called abductive reasoning, which seeks the minimal set of facts required to explain an observation. For every single node in their test networks, they stripped away all unnecessary data, leaving behind only the critical "node-feature pairs"—specific combinations of a location in the network and a piece of information attached to it—that were sufficient to lock in the predicted class. Imagine a network where a document is classified as "biology" not because of every word it contains, but because of a specific handful of terms appearing in its immediate vicinity and the documents linked to it. The researchers found that for many nodes, they could reduce the explanation down to a surprisingly small number of these critical pairs, discarding the rest as redundant.
Once they had these minimal, essential explanations for thousands of individual nodes, the researchers faced a new challenge: how to turn these thousands of tiny, specific stories into a single, global set of rules that could describe the entire network's behavior. They achieved this by abstracting the specific identities of the neighbors. Instead of saying "the document from node 42," they described the pattern as "a feature appearing at a distance of one hop." By converting these thousands of individual explanations into a standardized format based on distance and feature type, they created a clean, organized dataset. They then fed this data into a decision tree, a simple algorithm that learns to make decisions by asking a series of yes-or-no questions. The paths through this tree revealed the global logical rules that govern the network's predictions.
The results of this approach were striking when compared to existing methods. In tests across four different benchmark datasets, ranging from small synthetic graphs to large citation networks of scientific papers, the new framework produced rules that were both more accurate and far more concise. On one large dataset involving nearly twenty thousand nodes, the new method achieved a fidelity of 99.9 percent, meaning its rules matched the original model's predictions almost perfectly, while using only 31 rules. In contrast, a leading competing method achieved a fidelity of only 69.7 percent and required 153 rules to do so. The researchers found that their approach could explain up to 83.8 percent fewer rules than previous techniques while maintaining a higher level of trustworthiness. This efficiency suggests that by focusing on the minimal, sufficient causes for a prediction, the system avoids the noise and redundancy that often clutter other explanations.
The study also revealed that the size of these minimal explanations varied depending on the complexity of the data. In simpler networks, the essential explanations were often just a few items long, while in denser networks with more features, they grew larger but still remained manageable. Crucially, the researchers demonstrated that these compact rules were not just mathematical artifacts; they carried clear structural meaning. For instance, in a synthetic network designed with specific shapes, the extracted rules correctly identified the presence of high-degree connections and the absence of specific local patterns, effectively rediscovering the underlying design of the network. This ability to distill complex, high-dimensional behavior into a few short, logical statements offers a promising path toward making artificial intelligence more transparent and accountable. The researchers plan to extend this work to more complex network architectures, hoping that the principles of finding the minimal sufficient cause can be applied to a wider range of intelligent systems.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.