Learning by Knowledge Assimilation A New Machine Learning Paradigm Applied to Elicit Improved Decision Trees

Abstract

There are various paradigms of machine learning which have been widely researched, newlinestudied, and applied to practical and industrial tasks. In this thesis, we have proposed a newlinenew machine learning paradigm - learning by knowledge assimilation, which is a common newlinehuman cognitive and applied learning faculty. For example, learning a topic from multiple newlinetextbooks or from multiple teachers is very common. In this thesis, we have formalized and newlineapplied learning by knowledge assimilation to elicit improved form of decision trees from newlinethe concept (in the form of sets of rules) generated by multiple classifiers from the same newlinetraining set. Decision trees are mathematical representations of concepts that mimic human newlinethought processes in many real-life applications, such as in the medical diagnosis, control, newlineand strategic planning, etc. We have envisaged three steps for obtaining a decision tree from newlinemultiple sets of rules: (C)ombine, (R)estucture, and (I)nclude new observations. The first of newlinethese three steps is compulsory and the last two are optional but may be repetitive. Broadly, newlinewe have used different heuristics to follow the Occam Razor Principle while implementing newlinethe first two stages (C and R). For the last stage and for the general decision tree, a change in newlineconventional decision trees structure has been suggested in the form of identifying and storing newlineexceptions (False-Positive instances) from the training set at the leaves of the decision trees newline(human beings are also known to store exceptions). For the cost-sensitive decision trees, we newlinehave applied an existing pruning method as a form of (R)estructuring such decision trees as newlinethe last step, instead of storing exceptions at the leaves. Empirical evaluations are carried newlineout to validate our methodology with real-world data sets drawn from UCI machine learning newlinedata repository. We have proposed and used a discretization algorithm to preprocess the newlinedata sets that are used in the experiments. The discretization algorithm combines one local newlineand one global heuristics to res