Effective Customer Churn Prediction on Large Scale Data using Metaheuristic Approach

Abstract

Objectives: Customer retention is one of the major requirements of any organization to gain competitive advantage. Accurately predicting the customer’s status can help organizations reduce and prevent churns. Methods/Analysis: This paper presents an analysis of churn data and issues related to churn data in terms of data size, attribute density, data sparsity and abstraction contained in the data. It discusses the advantages of utilizing metaheuristic techniques for churn prediction and in specific analyses ACO for churn prediction and performs a comparison with other metaheuristic algorithms and emphasizes the importance of using ACO. Findings: Experiments were conducted by implementing ACO and applying it on Orange Dataset. It was observed from the ROC curve that the points plotted falls to the top left of the graph, hence indicating good efficiency and a fluctuation from low to moderate false positive rates were observed. It could be observed from the PR curve that the ACO algorithm exhibits high recall rates and moderate precision rates. ROC and the PR plots indicate that there is still scope for enhancement in terms of reduction in false positive rates and increase in precision levels. It was identified that though ACO exhibits effective performance, the size of the dataset acted as a huge downside increasing the time taken. Due to the huge size of the data, memory requirements are very high, but due to the skewed nature of the data most of them contain null values. Applications/Improvement: Findings exhibited scope for improvement, hence research directions namely data structure identification to reduce memory requirements, graph based churn prediction and fuzziness incorporation in the prediction process were proposed.