RULE-BASED RECOMMENDATION MODEL FOR SELECTING CLASSIFIERS IN WEKA BASED ON USER SPECIFICATIONS

Abstract

Machine learning field has many different algorithms; selecting an algorithm for non-expert user and aiming for maximizing empirical performance could be a tough task. This thesis considers the problem of selecting an algorithm based on the user specifications, for instance, training speed, memory, and interpretation. Specifically, considers the classification problem in the range of 20 models (2 meta-methods, 18 based classifiers) arising from 10 different families (Bayesian, Decision trees, Rule-based methods, Nearest neighbor methods, Logistic, multinomial regression, Neural networks, Support vector machines, Boosting, Bagging, and other ensembles), all implemented in WEKA. Different characteristics have been gathered for each model such as training speed and the available memory, then a set of rules have been defined based on these characteristics by using a tree architecture in order to choose one or for given user requirements. Finally, the model evaluated on 10 datasets from the UCI repository, the classification results show a better than or close to a previous work that addressed the similar problem.