Predicting the Silent Majority on Graphs: Knowledge Transferable Graph Neural Network

Wendong Bi, Bingbing Xu, Xiaoqian Sun, Easton Li Xu, Huawei Shen, Xueqi Cheng

Abstract

Graph Neural Networks (GNNs) typically assume that all nodes possess complete features and labels. However, real-world graphs often feature a "vocal minority" with rich data and a "silent majority" with missing attributes and labels, creating a distribution shift that standard GNNs fail to address. This paper introduces the Vocal-Silent Graph (VS-Graph) framework and proposes the Knowledge Transferable Graph Neural Network (KTGNN). KTGNN employs domain-adapted feature completion, message passing, and classifier transfer to bridge the gap between vocal and silent nodes. Evaluated on financial risk and political tendency datasets, KTGNN significantly outperforms state-of-the-art baselines, demonstrating robust performance in data-scarce graph scenarios.

Introduction: The Challenge of the Silent Majority

In the ideal world of Graph Neural Networks (GNNs), every node comes with a complete set of features and plenty of labels for training. In reality, graphs are often "data-hungry." Consider a social network: politicians (the vocal minority) have detailed profiles and clear political leanings, while ordinary users (the silent majority) often leave their profiles blank and rarely disclose their affiliations. Similarly, in finance, listed companies publish exhaustive reports, while unlisted companies provide very little.

Overview of the Vocal-Silent (VS) Graph concept and real-world applications in political and financial networks.

Standard GNNs struggle in these scenarios because they assume all nodes come from the same distribution (the I.I.D. hypothesis). However, the "vocal" and "silent" groups often exhibit a distribution shift—listed companies are fundamentally different in scale and behavior from unlisted ones. Standard models either fail to handle missing features or produce poor predictions because they cannot bridge the gap between these two distinct domains.

Methodology: Knowledge Transferable Graph Neural Network (KTGNN)

To address the distribution shift and data scarcity in VS-Graphs, the authors propose KTGNN, which consists of three specialized modules:

1. Domain-Adapted Feature Complementor (DAFC)

Instead of filling missing features with zeros, DAFC iteratively completes the features of silent nodes. It uses a calibration mechanism to account for the domain difference between a vocal neighbor and a silent target, effectively translating knowledge across the gap. This process works like a breadth-first search, starting from silent nodes with vocal neighbors and gradually spreading completed feature knowledge through the graph.

2. Domain-Adapted Message Passing (DAMP)

Standard message passing mixes information indiscriminately. DAMP introduces a "distribution-shift variable" that projects features from the source domain (vocal) into the target domain (silent) during the aggregation process, preserving the unique characteristics of each group.

3. Domain Transferable Classifier (DTC)

Rather than forcing the model to learn a single "one-size-fits-all" classifier, DTC trains two: one for the vocal domain and one for the silent. It then uses KL-divergence minimization to transfer the logic of the well-trained vocal classifier to the label-scarce silent classifier.

Experimental Evaluation

Datasets

The model was evaluated on two real-world VS-Graphs:

  • Company Equity Graph: 10,641 companies in China. Vocal nodes are listed companies (with financial statements); silent nodes are unlisted companies (missing statements). The goal is to predict financial risk.
  • Twitter Political Network: Vocal nodes are famous politicians (with descriptions and known parties); silent nodes are ordinary civilians. The goal is to predict political tendency.

Key Results

KTGNN consistently outperformed nine representative baselines, including GCN, GAT, and OODGAT. In financial risk assessment, KTGNN achieved an AUC of 67.11%, a nearly 6% improvement over the best-performing baseline. On the Twitter network, it reached an AUC of 95.08%, significantly higher than standard GNNs which hovered around 86-92%.

Exploratory analysis using t-SNE and box plots confirmed that vocal and silent nodes live in different statistical distributions. Treating them as the same leads to noise that degrades model performance. KTGNN's success lies in its ability to preserve these differences in representations while transferring the underlying logic of the classifier.

Conclusion and Future Work

KTGNN provides a robust framework for predicting traits in the "silent majority" of graphs by explicitly modeling domain shifts. While the model is highly effective, its performance depends on the iteration hyperparameter for feature completion. Future work may involve automating parameter selection based on graph topology and exploring VS-Graphs in other domains, such as citation networks or recommendation systems where active and inactive users present a similar dichotomy.

Summary updated

BibTeX
@misc{bi2023predicting,
  author = {Wendong Bi and Bingbing Xu and Xiaoqian Sun and Easton Li Xu and Huawei Shen and Xueqi Cheng},
  title = {{Predicting the Silent Majority on Graphs: Knowledge Transferable Graph Neural Network}},
  year = {2023},
  eprint = {2302.00873},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2302.00873},
}
Figure