Classification Performance Measure Using Missing Insurance Data: A Comparison Between Supervised Learning Models

摘要

The aim of this paper is to measure and compare the classification performance of five supervised learning models in the presence of missing data. These models are the support vector machines, naive Bayes, k-NN, Ripper and the Iogistic discriminant analysis algorithm. Five different proportions of “missingness” are artificially simulated on a test data set obtained from the UCI repository. For each simulated proportion, aIJ model are tested against the data to access the impact on accuracy of the classification. The classification involves determining whether a group of customers are likely to have an insurance cover or not. This is done by scrutinizing the attributes associated with the customers. The models are built using completely observable data. The results show that with a small proportion of missing data, all models perform well with high accuracies. However, as the number of missing values increase, the performance drops with some models experiencing a greater impact on classification accuracy than others. The logistic discriminant analysis algorithm performs better overall, but the naive Bayes model shows more resilience when there is an increase in proportions of missing data in the test data set.