Weight learning from cost matrix in weighted least squares model based on genetic algorithm

Abstract

In real life, it is a common phenomenon that different misclassification causes different cost. Given a misclassification cost matrix (MCM), cost-sensitive learning is aiming at decreasing the overall misclassification cost rather than simply reducing the misclassification rate. Weighted least squares (WLS) model is acknowledged as an effective way of cost sensitive learning. However, the weights in WLS model are generally unknown and finding these weights is usually difficult. In this paper, we put forward a new approach to learning these weights of WLS model from a given MCM based on a genetic algorithm. A comparative study shows that our proposed approach has an overall cost of misclassification significantly smaller than the existing cost-sensitive learning methods.