An Experimental Evaluation of Data Mining Algorithms Using Hyperparameter Optimization

Abstract

The challenge to choose the best algorithm and its best parameters for a given problem is known as Combined Algorithm Selection and Hyperparameter Optimization Problem. Among all the classification algorithms available are those based on human comprehensible representations, such as decision trees and classification rule induction. These algorithms are usually chosen by the clarity of the results obtained and the interpretability of its models. In this paper, we evaluated the six most used algorithms based on human comprehension. We conducted experiments with 28 datasets often used in the literature in different ways: using default parameters, using ExpDB parameters and using a tool based in genetic algorithm to find the best parameter combination. The results obtained have shown the strategy of combining the data from ExpDB via GA is effective in finding classification models with good accuracy.