Meta-learning based framework for helping non-expert miners to choice a suitable classification algorithm: An application for the educational field

Abstract

One of the most challenging tasks in the knowledge discovery process is the selection of the best classification algorithm for a data set at hand. Thus, tools which help practitioners to choose the best classifier along with its parameter setting are highly demanded. These will not only be useful for trainees but also for the automation of the data mining process. Our approach is based on meta-learning, which relies on the application of learning algorithms on meta-data extracted from data mining experiments in order to better understand how these algorithms can become flexible in solving different kinds of learning problems. This paper presents a framework which allows novices to create and feed their own experiment database and later, analyse and select the best technique for their target data set. As case study, we evaluate different sets of meta-features on educational data sets and discuss which ones are more suitable for predicting student performance.