Exception tolerant Ensemble with a novel combination method and its application

Abstract

This thesis intends to create a tool to identify the Ayurvedic medicinal plants considering the root samples. The main job of the tool would be classification. The act of classification should be efficient and at the same time, optimized. Deep learning, variants of Support Vector Machines, Boosted learning etc. have quite successfully established themselves as state-of-the-art solutions to decision making / classification problems. However, none of them would be a smart choice when the training data is limited as these well performing methods require quite an elaborate training. Hence, in terms of cost effectiveness these machine learning techniques are not the most impressive. Again, limited training restricts them from optimally performing. Time complexity of these algorithms are comparatively higher too. This thesis, in a way discourages the \survival of the fittestquot theory for classifiers and establishes the fact that co-existence of strong and weak classifiers could actually be beneficial in certain aspects. In relevance to that, the decision tree/ rule-learning algorithms, the weak learners are worked with in this thesis. A few decision learners/rule learners(namely PRISM, RISE,C4.5 and CN2)are employed to learn from these microscopic properties. PRISM, RISE, C4.5 and CN2 are rule generating classifiers and are also considered as weak learners. The first phase of the thesis deals in individual betterment of the algorithms. While, processing the data from several data sets from the UCI Machine Learning repository it was observed that the data sets not only contained noise but also exceptions which when dealt with a strategy can improve the performance of the classifiers. ExIORISE is proposed as an enhancement of the RISE algorithm. Later enhanced versions of PRISM, RISE and C4.5 named as OE2PRISM, OE2RISE and OE2C4.5 respectively are proposed. When statistically analyzed, ExIORISE is found to be significantly outperforming all the conventional versions.