Similarity Search using Cluster based Ensemble Classification

摘要

An automatic retrieval system for similarity search using cluster based ensemble classifier is proposed in this work to help the doctors. The proposed system has designed with the modules (i) metric database construction (ii) cluster based ensemble classifier and (iii) similar image/test report retrieval. Metric database is constructed by first and second order features of lung tomography images available in the internet, heart, liver and diabetes datasets collected from UCI machine learning repository. Dataset is further reduced by association rule mining as a feature vector. The cluster based ensemble classifier technique includes two phases namely training and testing phase. The phase clusters the reduced feature vector into groups using clustering algorithms (k-means and fuzzy c-means). The defined cluster id is passed along with the metric data as an attribute value. These cluster vectors are further classified using ensemble classifier with the decision tree, SVM and bayesian classifiers. Accuracy obtained by cluster based ensemble classifier is 95.33% for lung images, 83.14% for diabetes and 86.97% for heart dataset and 93.32% for liver dataset which is higher than the accuracy obtained by clustering or by ensemble classifier (mixture of decision tree, SVM and Bayesian classifier) or by individual classifiers namely decision tree classifier, SVM or Bayesian classifier. Hence cluster based ensemble classifier is better for similarity search than other methods.