Study on Financial Fraud Account Detection Based on Imbalanced Datasets

For the detection of bank accounts involved in fraud, this paper proposes a framework, iForest-SMOTE, which is applicable to the imbalanced financial datasets. Based on the dynamic transaction features of the accounts, the transaction behavior features are extracted from the dimensions of statistical information, sequential order information and supervision information. Then a datasets equalization strategy for data pre-processing is proposed to address the problem of cross-region sample synthesis, which is faced by the oversampling technology, ADASYN, on the financial account datasets. The strategy uses the iForest algorithm for mixed sampling of the data to remove the majority of noisy data and reduce the difficulty of the classifier learning from the minor classes. On this basis, a random forest classifier is designed to implement the detection of the accounts involved in financial fraud. The experimental results on the datasets of actual financial account transactions show that iForest-SMOTE has a clear advantage in the recall rate and accuracy over ADASYN, SMOTE and other sampling techniques. Its F-value is at least 2.13 percentage points higher than that of the other algorithms.