UTM Lab meeting - "Tree-based Models to Predict Rankings"

"This project addresses the recommendation of Machine Learning/Data Mining (DM) algorithms. This problem is relevant because many different algorithms are available today for DM tasks such as classification and regression, and the computational cost of executing all off them on a given dataset is very high. Therefore, the user must decide which algorithm(s) should be tested. The Metalearning approach to this problem consists of using information about the past performance of a set of learning algorithms on a set of datasets to induce a mapping between characteristics of those datasets to the relative performance of those algorithms. We propose a new approach to the problem of recommending algorithms based on a combination of simple regression trees. We present a model that provides individual ranks and other model that attempts to use information from other ranks to make a prediction. We obtain surprisingly results when compared this models experimentally with the ranking trees algorithm. "