An improved multiple classifier combination scheme for pattern classification

Combining multiple classifiers are considered as a new direction in the pattern recognition to improve classification performance. The main problem of multiple classifier combination is that there is no standard guideline for constructing an accurate and diverse classifier ensemble. This is due to the difficulty in identifying the number of homogeneous classifiers and how to combine the classifier outputs. The most commonly used ensemble method is the random strategy while the majority voting technique is used as the combiner. However, the random strategy cannot determine the number of classifiers and the majority voting technique does not consider the strength of each classifier, thus resulting in low classification accuracy. In this study, an improved multiple classifier combination scheme is proposed. The ant system (AS) algorithm is used to partition feature set in developing feature subsets which represent the number of classifiers. A compactness measure is introduced as a parameter in constructing an accurate and diverse classifier ensemble. A weighted voting technique is used to combine the classifier outputs by considering the strength of the classifiers prior to voting. Experiments were performed using four base classifiers, which are Nearest Mean Classifier (NMC), Naive Bayes Classifier (NBC), k-Nearest Neighbour (k-NN) and Linear Discriminant Analysis (LDA) on benchmark datasets, to test the credibility of the proposed multiple classifier combination scheme. The average classification accuracy of the homogeneous NMC, NBC, k-NN and LDA ensembles are 97.91%, 98.06%, 98.09% and 98.12% respectively. The accuracies are higher than those obtained through the use of other approaches in developing multiple classifier combination. The proposed multiple classifier combination scheme will help to develop other multiple classifier combination for pattern recognition and classification.