Balanced detection of anomalous transctions using hybrid awred on real data under strong class imbalance

Abstract

The article considered the problem of detecting anomalous transactions in strongly imbalanced financial data. These data included a very small amount of the rare class compared with the number of normal operations. Such formulation was difficult not only because of the small number of anomalous examples. It was also caused by the fact that part of suspicious transactions did not form a clearly separated group in the feature space, therefore the usual reconstruction error of the autoencoder did not always provide sufficient ranking quality and reliable threshold selection. To solve this problem, the seventh version of the HYBRID AWRED method was used. It applies a hybrid approach in which the autoencoder reconstruction basis is combined with a stable profile of calibrated residuals.
All experiments were carried out on the Credit Card Fraud Detection dataset, which included 284807 transactions, 30 features and 492 anomalous records. The share of the anomalous class was only 0.1727%. The training sample was formed only from normal transactions. The validation sample was used for threshold selection and parameter tuning of HYBRID AWRED v7, and the test sample was used only for the final quality evaluation. This approach reduced the risk of information leakage from the test data into the threshold tuning procedure.
In the HYBRID AWRED v7 method, fixed calibration of reconstruction residuals was used after the initial training of the autoencoder. For each feature, a calibrated residual was calculated, after which its positive part was taken and then soft logarithmic compression ln (1 + rj +) was applied. This reduced the influence of too large values in the right tail of the anomaly score. The final score was formed as a normalized linear combination of the calibrated residual profile with non-negative weights. The AWRED component was trained using pseudo-anomalies and a ranking loss function. Such approach made it possible not only to reduce reconstruction errors, but also to improve the ordering of normal and anomalous transactions by risk level.
According to the results of five independent runs, HYBRID AWRED v7 achieved AUC-ROC 0.9432+/-0.0117, AUC-PR 0.6639+/-0.0634, F1 0.7153+/-0.0402, MCC 0.7180+/-0.0399, Precision 0.6608+/-0.0565 and Recall 0.7831+/-0.0501. Compared with such competitors as AE, AE-finetuned, DAE, SVDD and DAGMM, the proposed approach showed the best quality by AUC-PR, F1, MCC and Recall. But it should be taken into account that this result should not be interpreted as a complete separation of anomalous transactions from normal ones. Part of weak suspicious transactions remains close to the normal profile. However, HYBRID AWRED v7 noticeably improved the practically important threshold quality in the task with extreme class imbalance.

Keywords: anomalous transactions, anomaly detection, autoencoder, HYBRID AWRED v7, AUC-PR, class imbalance, calibrated residuals, threshold decision.