MetaCost: a general method for making classifiers cost-sensitive
KDD '99 Proceedings of the fifth ACM SIGKDD international conference on Knowledge discovery and data mining
Expert Systems with Applications: An International Journal
Journal of Artificial Intelligence Research
The foundations of cost-sensitive learning
IJCAI'01 Proceedings of the 17th international joint conference on Artificial intelligence - Volume 2
Hybrid approaches for classification under information acquisition cost constraint
Decision Support Systems
Determination of optimal inspection sequence within misclassification error bound
Expert Systems with Applications: An International Journal
A genetic algorithm-based approach to cost-sensitive bankruptcy prediction
Expert Systems with Applications: An International Journal
A maximum-margin genetic algorithm for misclassification cost minimizing feature selection problem
Expert Systems with Applications: An International Journal
A cost-sensitive decision tree approach for fraud detection
Expert Systems with Applications: An International Journal
Hi-index | 12.05 |
Linear discriminant analysis models to minimize misclassification cost have recently gained popularity. It is well known that the misclassification cost minimizing linear discriminant analysis problem is an -complete problem that is difficult to solve to optimality for large scale datasets. As a result, heuristic techniques have gained popularity but it is difficult to assess how well these heuristic techniques perform. One way to aid assessment of the performance of heuristic techniques is to establish a lower-bound on the optimal value of misclassification cost. In this paper, we propose and use a hybrid particle swarm optimization (PSO) and Lagrangian relaxation (LR) based heuristic to establish a misclassification cost lower bound (MCLB) for two-group linear classifiers. We use the subgradient optimization procedure to tighten the MCLB. Using simulated and real-world datasets, we test a misclassification cost minimizing linear genetic algorithm classifier and two commercial non-linear classifiers (C5.0 and C&RT) to compare their performances with the MCLB. Our holdout sample tests indicate that the proposed MCLB works well for both linear and non-linear classifiers when class data distributions are normal. Additionally, as misclassification cost asymmetry increases, the proposed MCLB appears to provide better results.