Active Learning and Sample Imbalance in Graph Data for Risk Control

The talk begins with an overview of graph data applications in risk control, highlighting how user transaction networks can be modeled as graphs for fraud detection, community detection, and user risk analysis.

Two major challenges are identified: difficulty in obtaining reliable labels for rare malicious users and severe class imbalance that degrades model robustness.

To address these issues, a semantic‑aware active learning framework is introduced, which selects informative samples by combining model uncertainty, graph structural properties (e.g., node degree, centrality), and semantic influence measures, thereby focusing labeling effort on high‑impact nodes.

The presentation also examines node labeling on imbalanced graphs, discussing strategies such as oversampling minority nodes, loss re‑weighting, and advanced techniques like GraphSMOTE that synthesize node features and edges while preserving graph topology.

A “dual‑channel information alignment” mechanism is proposed, leveraging pretrained GNN embeddings for both classification confidence and clustering proximity to select reliable nodes for pseudo‑labeling, thus mitigating both label scarcity and imbalance.

Experimental results on public datasets (e.g., Cora, Citeseer) and Huawei’s financial transaction data demonstrate that the proposed methods outperform existing SOTA baselines, achieving notable gains with limited labeled samples.

The conclusion summarizes the effectiveness of integrating semantic information, prototype‑based diversity, and graph‑aware sampling to solve node classification under severe imbalance in risk‑control graphs.