A high performance fraud detection strategy prediction model

Fraud is one of the most important problems facing the financial sector. It's very expensive. Because the model based on machine learning can scan huge transaction data sets, detect abnormal activities, and identify all cases that may be vulnerable to fraud. Therefore, we choose to build a model based on LightGBM to predict whether the transaction is fraudulent. For the transaction data set conducted by European cardholders within two days in September 2013, the factor of unequal data labels is very important. We use random undersampling to balance the sample. Then the sample is segmented and trained. The results show that our model has an accuracy of more than 99.5%. Compared with popular classification models, our model has higher performance. The solution can detect the default rate or fraud tendency of each potential customer and transaction, and provide key alerts and insights for financial institutions.