Exception-Tolerant Decision Tree / Rule Based Classifiers

A plenty of classifiers had already been proposed but none of them deals with exceptions. Theories are formed by observing examples, the phenomenon being called ‘Superarticulacy’ [1] and there would certainly be instances not abiding by the theories. These had not been dealt with in machine learning till date. The challenge was to differentiate between an exception and a noise. The improvement in the performance of the classifier when exceptions are considered, however, totally depends on the data sets. But in a real life scenario, exceptions are better not to be ignored. The proposed methodology also includes features like n-fold crossvalidation, exclusion of inefficient rules, ordering the rules by weighted voting and inclusion of a default rule. This methodology has been applied on three popular classifiers C4.5, PRISM and RISE. These are old algorithms but can perform pretty well with small training sets whereas a few state-of-the-art classifiers like Random Forest [2], Xgboost [3] etc. fail to perform well without extensive training. C4.5 [4, 5] was proposed by J. Ross Quinlan which, from the training set, forms a decision tree. Decision trees come up with amazing outputs but are not easy to understand, work with, or to be manipulated. PRISM was proposed by Jazdia Cendrowska [6] which produces rule sets which are more comprehensible than a tree. The RISE algorithm was proposed by Pedro Domingos [7] which keeps on generalizing instances until a are in rule is obtained. The rule sets formed and their representation the most intelligible form amongst the three algorithms dealt with.