NICE: Scale-Stable Perturbations for Graph Neural Network Explanations via Noise Corruption

In the world of artificial intelligence, there is a growing class of systems known as Graph Neural Networks. These are not the standard image-recognition tools found in your phone; instead, they are designed to understand data that exists as a web of connections, like molecules in a chemical compound, social networks, or traffic routes. To make a decision, these systems pass information along the links between nodes, much like a rumor spreading through a crowd. Because these systems are incredibly complex and their internal logic is often a tangled black box, scientists have developed "explainers." These are tools designed to look at a finished decision and point out which specific connections or pieces of information were most important. The standard way these tools work is by playing a game of "what if." They take a specific connection, remove it or silence it, and then ask the AI: "Does your answer change?" If the answer changes drastically, the tool assumes that connection was vital. This method has been the industry standard for years, but it relies on a fragile assumption: that the act of removing a connection doesn't accidentally break the entire system in a way that has nothing to do with the connection itself.

A team of researchers from Central South University, Hong Kong Polytechnic University, and Griffith University has discovered that this standard method is fundamentally flawed. They found that the way these explainers currently silence connections introduces a hidden distortion that makes the AI's answers unreliable. When an explainer tries to "turn off" a link in the network, it typically multiplies the information flowing through that link by a number close to zero. While this successfully stops the specific message, it also shrinks the overall volume of information moving through the entire network. The researchers call this phenomenon "Scale Drift." Imagine a conversation where you ask a person to whisper a secret to a friend, but in doing so, you accidentally force everyone in the room to speak at a fraction of their normal volume. The secret might be gone, but the entire conversation has also become faint and distorted. The AI, receiving this shrunken signal, might change its mind not because the secret was important, but simply because the whole room got too quiet to hear anything clearly. The researchers demonstrated that this drift accumulates with every step the information takes through the network, meaning the deeper the network goes, the more the AI's internal state drifts away from reality, rendering the explanation untrustworthy.

To solve this, the researchers proposed a new way to test the AI that avoids shrinking the signal. Instead of simply silencing a connection, they introduced a method called "Noise Corruption." Rather than turning the volume down to zero, this method replaces the clear message with a random, scrambled version that has the exact same overall strength or "loudness." It is like taking a clear sentence and replacing it with a burst of static that is just as loud as the original voice. The information is corrupted, but the energy flowing through the network remains stable. By keeping the volume constant, the researchers ensured that any change in the AI's answer was due to the loss of specific meaning, not a collapse in the system's overall power. They built a new framework around this idea, which they named NICE. This system first learns a boundary of how much information needs to be restored to get the AI back to its original decision, and then it calculates exactly which connections contributed most to that restoration.

The results of this new approach were significant. When tested across eight different benchmarks involving chemical molecules and other complex structures, the new method outperformed the best existing tools. It was able to identify the correct critical connections more accurately, improving its ability to find the right answers by an average of nearly nine percent compared to the previous best methods. More importantly, the researchers confirmed that their new technique successfully eliminated the "Scale Drift" problem. When they compared the old method of silencing connections against their new method of adding noise, the old method consistently pushed the AI's internal state far away from its normal operating range, while the new method kept the AI's internal state much closer to how it behaved before any changes were made. This suggests that for years, many explanations generated by AI might have been misleading, not because the AI was wrong, but because the tools used to explain it were inadvertently breaking the very system they were trying to understand. By fixing the way we perturb the data, the researchers have provided a more faithful way to see inside the mind of these complex networks.