Customer Analysis using Machine Learning with Feature Selection Approaches: A Comparative Study

In every enterprise, the customer is the focal point of the business, and it is essential to comprehend how customer satisfaction with the enterprise's product or service is evaluated. To understand how the customer is satisfied with the enterprise's business the process of carrying out customer analysis is inevitable. Moreover, due to furtherance of the Internet, CRM and technology, generation of customer data is enormous. Using these customer data efficient understanding of the customer patterns can be carried. In this analysis, ML methods are implemented to perform customer analysis. However, due to existence of irrelevant, inappropriate, & correlated data, the ML algorithm cannot help to get quality performance results for the customer analysis, and also computational time is increased with these datasets. To address these issues, the feature selection (FS) mechanism is applied. Goal of FS is to improve accuracy of the classifier and also reducing the amount of computation that the algorithm requires. This is accomplished by removing correlated, inappropriate, and irrelevant variables and focusing on selecting relevant attributes. In this study, four distinct filter FS methods are implemented: Symmetrical Uncertainty, Information Gain, Chi-square, and ReliefF. Further feature subsets captured from the FS approaches are individually applied with an ML classifier to build the model. Here, two distinct supervised ML approaches are utilized: Nave Bayes (NB) and the other one is (K-NN). The empirical methodology is conducted using a bank data collected from UCI and experiment is done in two distinct perspectives: one using full features with ML and the other one using selected features obtained from FS approaches. Empirical results observed are presented & evaluated using diverse validity scores. The comparison of the two sets of results demonstrates quite clearly that ML with feature selection performs significantly better than ML without feature selection.