A machine learning approach to customer prediction

Abstract

Direct promotion like mailings or phone calls to a company’s potential customers is an effective way to market a product or a service. Being able to make accurate potential customer predictions, and to promote the product or service to them timely is certainly valuable to companies. Recently, machine learning approaches has been widely employed in the subject of customer prediction. In this project, machine learning techniques are used to solve a real-world problem in predicting potential customers of a Dutch insurance company. Two of the machine learning models, namely ANN (Artificial Neural Network) and Bayesian Learning are the focus of study. Prototypes from both machine learning paradigms are built and their performance evaluated. In addition, comparisons are made on the prediction capabilities of the two approaches. Since the design parameters (e.g. learning rates, m-estimates) have great impacts on the performance (like training time, prediction capability, etc.) of the learning models, the available design choices are discussed and their values of selection are justified. The effects of some of the common practical issues like overfitting and feature selection are also examined. From the result of the experiment, one can tell the importance of such issues which are often ignored by machine learning practitioners.