Partial imputation to improve predictive modelling in insurance risk classification using a hybrid positive selection algorithm and correlation-based feature selection

Abstract

We propose a hybrid missing data imputation technique using positive selection and correlation-based feature selection for insurance data. The hybrid is used to help supervised learning methods improve their classification accuracy and resilience in the presence of increasing missing data. The positive selection algorithm searches for potential candidates for imputation and the correlation-based feature selection method searches for attributes have a significant effect on the target outcome. The imputation is performed only on those attributes that have an impact on the target outcome. The results show that the classification accuracy and resilience of supervised learning methods improve significantly when applied with the imputation strategy under these assumptions.