Analysis of Cost-Sensitive Algorithms for Degree of Imbalancing

Abstract

Class imbalance, along with other characteristics of the data, such as feature set and class separability, could affect the performance of most machine learning algorithms. This can be attributed to the algorithm’s primary assumption about the data being class-balanced and indifferent weightage among different misclassification errors. Class imbalance can be handled using both data and algorithmic-level methods. In this paper, we work on cost-sensitive learning, applied at the algorithmic level. We use the Cost-Sensitive Logistic Regression (CSLR) algorithm as a reference. We propose a methodology to empirically evaluate the performance of cost-sensitive algorithms over varying degrees of imbalanced data. Cost-sensitive learning induces a cost matrix consisting of the weighting scheme of different misclassification errors into the algorithm’s training process. This, in turn, forces the model to penalize the misclassification errors according to the skewness of the data to reduce the learning bias of the model towards the majority class. we present empirical evaluations of the reference model over four popular datasets and analyse its behaviour using MAE and Kappa values.