Semantic‑Aware Active Learning on Graph Data for Risk Control: Tackling Sample Imbalance

This talk explores active learning and sample‑imbalance issues on graph data, with a focus on risk‑control applications.

Graph data is ubiquitous: astronomical bodies, molecular structures, social networks, and, in risk control, user transaction networks can all be modeled as graphs for tasks such as fraud detection, community detection, and node‑level risk classification.

The two main challenges are (1) the difficulty of obtaining reliable labels because malicious users are rare, and (2) severe class imbalance that harms model training and robustness.

To address label scarcity, the presentation introduces semantic‑aware active learning. Traditional active learning selects samples with high uncertainty, but on graphs we also consider structural cues such as node degree and centrality. By combining uncertainty with graph‑specific importance, we can select representative, informative nodes.

Furthermore, a prototype‑based diversity strategy is proposed: after obtaining node embeddings, we compute class prototypes, measure each candidate’s distance to its prototype, and prioritize samples that are far from the prototype (high information) while also having high uncertainty.

For sample‑imbalance, several remedies are discussed: oversampling the minority class, undersampling the majority class, loss re‑weighting, and graph‑specific synthesis such as GraphSMOTE, which generates synthetic node attributes and predicts edges to integrate new nodes into the graph.

Recent works like Renode and TAM also exploit graph topology to guide sampling. Beyond synthesis, the talk proposes a double‑channel information alignment mechanism. A lightweight GNN is pretrained to obtain node embeddings, then two tasks are performed: (1) node classification to obtain confidence scores, and (2) clustering to obtain prototype proximity. Nodes with both high confidence and close proximity to their cluster center are selected as reliable candidates for labeling.

Experiments on public citation graphs (Cora, Citeseer) and Huawei’s financial transaction data demonstrate that the proposed method consistently outperforms state‑of‑the‑art baselines (random selection, degree‑based, uncertainty‑entropy) across various imbalance ratios, achieving higher accuracy with far fewer labeled samples.

In conclusion, semantic‑aware active learning combined with prototype‑based diversity and double‑channel alignment effectively mitigates label scarcity and class imbalance in graph node classification, offering a practical solution for risk‑control scenarios.