Improvement of accuracy of under-performing classifier in decision making using discrete memoryless channel model and Particle Swarm Optimization

Abstract

Listen

Translate article

In-spite of availability of wide range of algorithms for constructing multi-class classifier, there are applications like Decision support system, Game prediction in the sports, where the multi-class classifier performs relatively poor in terms of achieving the reasonable classification accuracy. In this paper, poorly performing multi-class classifier (constructed using the classical methods like Artificial Neural Network, cascade of Support Vector Machine, etc.) is treated as the discrete memoryless channel model with known transition probabilities (channel matrix) and the unknown priors. It is further used to construct M-ary Mini-Max technique based randomized decision rule to improve the performance of the multi-class classifier in terms of the classification accuracy. The prior probabilities and the probabilities associated with the M-ary randomized decision rule are further solved using the Particle Swarm Optimization. The experimental results based on the Monte-Carlo simulation using the synthetic data set and the real data set reveal the consistent improvement in the performance of the poorly performing classifier using the proposed technique.