Explain Graph Neural Networks to Understand Weighted Graph Features in Node Classification

Abstract

Real data collected from different applications that have additional topological structures and connection information are amenable to be represented as a weighted graph. Considering the node labeling problem, Graph Neural Networks (GNNs) is a powerful tool, which can mimic experts' decision on node labeling. GNNs combine node features, connection patterns, and graph structure by using a neural network to embed node information and pass it through edges in the graph. We want to identify the patterns in the input data used by the GNN model to make a decision and examine if the model works as we desire. However, due to the complex data representation and non-linear transformations, explaining decisions made by GNNs is challenging. In this work, we propose new graph features' explanation methods to identify the informative components and important node features. Besides, we propose a pipeline to identify the key factors used for node classification. We use four datasets (two synthetic and two real) to validate our methods. Our results demonstrate that our explanation approach can mimic data patterns used for node classification by human interpretation and disentangle different features in the graphs. Furthermore, our explanation methods can be used for understanding data, debugging GNN models, and examine model decisions.

Graph Decipher: A transparent dual‐attention graph neural network to understand the message‐passing mechanism for the node classification

“…Since only the local dominant and representative features are needed in this step, the computational burden can be further reduced by using masks that record the position of local representative features, allowing the unnecessary zeros to be ignored in the computation. The new nonlinear node v's feature is applied by the activation function φ in Equation (6).…”

“…Node classification is a primary graph task for a wide range of applications. 5,6,10 Many researches 34,35 have demonstrated that neural networks are more inclined to learn features from categories with larger amounts of data, which results in relatively lower accuracy of minor categories.…”

“…Graph neural networks (GNNs) offer effective graph-based techniques applied to solve abundant real-world problems in diverse fields, such as social science, 1,2 remote sensing, 3 protein-protein interaction networks, 4 brain neuroscience, 5 knowledge graphs, 6,7 image processing, 8,9 physical systems, 10,11 edge computing, 12,13 information safety, 14 big data, [15][16][17] and so forth. The power of current GNNs 18,19 is largely due to the message-passing mechanism, which recursively aggregates information along edges and updates these newly incorporated features on the center node.…”

Graph Decipher: A transparent dual-attention graph neural network to understand the message-passing mechanism for the node classification

“…Node classification is a primary graph task for a wide range of applications [1,2,3,4,5,6], as it determines the category of the central node by comparing neighbors based on the messagepassing mechanism which learns the node attributes in the multiclass graph dataset. The distribution of multi categories in the graph dataset is directly tied to the network's overall performance.…”

“…Graph neural networks (GNNs) offer effective graph-based techniques applied to solve abundant real-world problems in diverse fields, such as social science [1], physical systems [2,3], protein-protein interaction networks [4], brain neuroscience [5], knowledge graphs [6], etc. The power of current GNNs [1,7,8,9,10] is largely due to their message-passing mechanisms.…”