Active PU Learning for Cash‑Out Fraud Detection in Alipay’s AlphaRisk Engine

The Ant Financial Risk and Decision Center team introduces AlphaRisk, the fifth‑generation risk engine powered by AI, which processes hundreds of models in real‑time to detect fraud, account theft, and cash‑out risks during Alipay transactions.

Cash‑out fraud is particularly challenging because it lacks explicit negative labels; users rarely report cash‑out activities, making supervised learning difficult and prompting reliance on unsupervised methods such as anomaly detection and graph algorithms, which have high feature and computational requirements.

To address the labeling bottleneck, the authors propose a hybrid approach that integrates Active Learning with a two‑step Positive‑Unlabeled (PU) learning method, termed Active PU Learning. This method reduces manual annotation effort while improving model performance.

Algorithm Workflow

Algorithm: Active PU Learning
1. Generate sample pool: select required samples and assign positive labels using transferred knowledge.
2. while stopping condition not met do
   3.  Sample: select samples for labeling based on a specific sampling strategy.
   4.  Label: manually annotate the selected samples.
   5.  Update samples: refresh the sample repository.
   6.  Update model: apply two‑step PU Learning to retrain the model.
7. end while

The sampling strategy combines uncertainty and diversity: new data are scored by the current model, the most uncertain samples are clustered via K‑Means, and a diverse subset from each cluster is chosen for labeling.

During labeling, only samples with high confidence are marked as positive to ensure the reliability of the P (positive) set. Unlabeled or low‑confidence samples are treated as part of the U (unlabeled) set, and negative samples are generated through a spy‑based mechanism within the two‑step PU process.

Model updates use GBRT (Gradient Boosting Regression Tree) as the base classifier for active learning, and the two‑step PU procedure iteratively refines scores for P and U sets via EM (Expectation‑Maximization) iterations.

Experimental Results

Three experiments demonstrate the effectiveness of the proposed method: (1) two‑step PU Learning outperforms both Isolation Forest (unsupervised) and standard GBRT (supervised) with a 70% vs. 60% accuracy at the top 95‑100 percentile; (2) Active Learning improves model performance, raising accuracy from 91% (Isolation Forest) to 94% (Active Learning‑enhanced Random Forest); (3) The combined Active PU Learning approach consistently matches or exceeds the accuracy of both baseline models across multiple percentile ranges, confirming its superiority.

In the cash‑out detection scenario, the Active PU Learning model identifies three times more fraudulent cash‑out transactions than the Isolation Forest baseline at comparable accuracy, while requiring significantly fewer manual labels.

The authors conclude that Active PU Learning offers a cost‑effective way to incorporate external information and leverage limited labeled data, though it demands high‑quality annotations and incurs higher training overhead compared to conventional GBRT pipelines.