Using RNN to Predict Customer Behavior in High Volume Transactional Data

Abstract

Big data tools and techniques introduce new approaches based on distributed computing methods. When dealing with large data, one of these state-of-art approaches for analysing and predicting in the shortest possible time is the use of deep learning networks that provide real-time, accurate, and comprehensive analysis. This method has provided a new perspective to artificial intelligence with respect to increasing volume of data and complexity of real-world issues. The models used to predict customer behavior have mainly worked with limited features and dimensions. One of the applications of this method is to prevent customer churn, when predicting future behavior of customer transaction on point of sale (POS) devices.

In this paper, transactional data are analyzed using Recurrent Neural Networks (RNNs) that are capable of predicting high-dimensional data or time-series data, by which a model for predicting the behavior of POS holders are proposed. Hence, we used transaction data on POS devices to predict their future behavior using RNN algorithm. In order to validate the generated model, the real data of one of the private banks inside Iran has been used and the predicted results are about 87% accurate. The result is compared with previous methods which outperform other methods discussed later.