K-Partition Ensemble Multi-label Classifier for Insurance Purchase Prediction

Abstract

Individual insurance purchase prediction can help effectively and accurately advertise to maximize sales. Most current insurance purchase prediction model only considers whether a customer will buy certain insurance. In this way, the prediction will be time-consuming with the rapid growth of the number of insurance products. In order to provide more effective and efficient marketing methods for insurance companies, this work proposes a K-Partition Ensemble Multi-Label classification model to predict the customer’s possible future insurance purchase. First, by transforming the insurance purchase prediction problem into a multi-label classification problem, the balance between features and labels of data division in the insurance purchasing dataset is explored. Second, a k-partition ensemble multi-label classification model is introduced, where each distinct label constitutes in the training set as a new category of a single-label classification task, and the random forest is used for multi-class classification. The empirical test is carried out using the Insurance Company Case data from CoIL Challenge 2000. We find prediction classifiers perform the best when the number of labels is around 20. Empirical evidence indicates that our model manages to improve substantially over other 3 classical multi-label classification algorithms with relatively little time, especially in domains with a large number of labels. The research results also provide a new idea and useful reference for the application in specific fields construction of data models based on the multi-label evaluation.