Research And Application Of Small Sample Anomaly Detection In Customer Finance

Financial anomaly detection plays an important role in consumer finance,and its core part is anomaly detection of customer transactions,aiming to prevent fraud and reduce credit risks,thereby reducing the company’s asset losses.Currently,anomaly detection in consumer finance faces some problems and challenges:abnormal behaviors such as fraud and bad credit are rare in consumer finance scenarios,resulting in a serious imbalance between abnormal data and normal data.Models trained directly on unbalanced data are prone to overfitting;In the new channel consumer finance scenario,the lack of user data results in a lack of necessary training samples for the model,making it difficult for the model to converge;There are differences in the distribution of customer groups in different consumer finance scenarios,and existing models have poor ability to detect anomalies in new data.To solve the problem of mall sample size and low proportion of anomaly data when new channel consumer finance products are launched,this thesis proposes a method for anomaly detection of small sample consumer finance.Firstly,this thesis proposes a dynamic data augmentation consumer finance anomaly detection model based on OSS(One Side Selection)Clustering CTGAN(Condition Tabular Generative Adversarial Networks),which expands a small number of samples to balance the distribution of normal and anomaly samples,and improve the model’s learning ability on anomaly samples in small samples.Secondly,to address the problem of distribution differences between different consumer finance scenarios,this thesis proposes a consumer finance abnormal detection model based on instance weight transfer learning with confidence probability.The model uses self-owned consumer finance data and specific new channel consumer finance data for transfer learning.Through differential sample instance transfer learning based on confidence probability,an anomaly detection model suitable for small sample new channel consumer finance products is trained to quickly achieve anomaly detection capability.This thesis evaluates the effectiveness of the model using indicators such as AUC(Area Under Curve),Recall,and KS(Kolmogorov-Smirnov),experimental results demonstrate that the small sample consumer finance anomaly detection solution proposed in this thesis is feasible in practice and has significant performance improvements in anomaly detection compared to traditional models.