Inducing Hierarchical Multi-label Classification rules with Genetic Algorithms

Abstract

Hierarchical Multi-Label Classification is a challenging classification task where the classes are hierarchically structured, with superclass and subclass relationships. It is a very common task, for instance, in Protein Function Prediction, where a protein can simultaneously perform multiple functions. In these tasks it is very difficult to achieve a high predictive performance, since hundreds or even thousands of classes with imbalanced data distributions have to be considered. In addition, the models should ideally be easily interpretable to allow the validation of the knowledge extracted from the data. This work proposes and investigates the use of Genetic Algorithms to induce rules that are both hierarchical and multi-label. Several experiments with different fitness functions and genetic operators are preformed to obtain different Hierarchical Multi-Label Classification rules. The different proposed configurations of Genetic Algorithms are evaluated together with state-of-the-art methods for HMC rule induction based on Ant Colony Optimization and Predictive Clustering Trees, using many datasets related to the Protein Function Prediction task. The experimental results show that it is possible to recommend the best configuration in terms of predictive performance and model interpretability.