Publication

Abstract :

Churn prediction is required by most of the service companies to improve their business. The machine learning approaches concentrate on the selection of algorithms or features to improve the accuracy of churn prediction. Algorithms to understand what went wrong and why a prediction is not accurate are needed to improve the system. This paper gives special attention to the error analysis of those approaches and the overall analysis of the dataset. This paper analyses the working of various machine learning approaches for customer retention prediction based on bank customer’s transaction data. It also gives a detailed error analysis using distance and similarity metrics like Mahalanobis distance, Hamming distance, and Jaccard Similarity Score. It provides a ranking for the features in the dataset based on error analysis and also lists their importance in a quantified manner by removing highly ranked features.