Using machine learning to detect and mitigate transaction fraud

Abstract

A plurality of fraud features is determined for a plurality of merchants. A weight matrix is trained for the fraud features. The fraud features are transformed via the weight matrix. An unsupervised machine learning is applied to the transformed fraud features. Based on the unsupervised machine learning, a plurality of tags is generated that correspond to different types of transaction fraud. The merchants are clustered into different merchant groups based on the tags. For each of the merchant groups, a respective decision tree model is trained. Transactions associated with each of the merchant groups are processed using the decision tree model trained for said merchant group.