Design of efficient classifier integration and performance evaluation in machine learning

Characteristics of any classifier heavily depend upon the nature of data set taken for training and verification. Area of app lications like health care suffered from having the large and suitable dataset. Classifier designed for health care should show a better generalization and robustness characteristics so that end results presented by classifier can consider with high reliability and confidence. In this paper consistency problem associated with classifier has presented, which is a big issue from practical point of view. Defining committee of experts is one of natural way to increase the reliability in classifier design but at the same time, way of integration rules the end performance. To overcome problem of generalization and consistency of classifier, two methods for developing the mixture of classifier namely TMQD and MVFD are presented. Estimation of quality associated with a classifier is very challenging task for researcher, because there is no single parameter which could alone represents the absolute performance .To measure the quality of classifier rather than having the conventional parameters like sensitivity and specificity, receiver operating characteristics is always a better choice. But in practical environment of health care use of ROC hardly has seen. In this paper detail understanding of ROC and estimation of area under curve has also presented. Selection of threshold value is one of the most important factor to determine the performance of classifier. Dependency of threshold value with population and geographical area making difficult to decide a optimal value. A graphical approach has presented to select the best threshold value as according to environment and need. Index Terms – Data Mining, Classifier, Classifier integration, ROC, Area under ROC, Sensitivity, Specifity, Heart Diseases, Neural Networks, .