Fraud detection using time-series transaction data

Abstract

Disclosed are various embodiments for leveraging deep learning-based recurrent neural networks (RNNs) using time-series data to evaluate fraud risk for an incoming transaction associated with a user account. Time-series attributes can be extracted from historical transaction data and the incoming transaction data. The time-series attributes can be defined as an array of sequential events that are inputted into an RNN-based machine-learning framework to predict whether an incoming or otherwise pending transaction is fraudulent given the spending sequence. An RNN-based time-series prediction model can be trained to understand and predict patterns associated with a user's spending history according to the inputted time-series data in order to predict whether the transaction is fraudulent.