A Promising Direction towards Automatic Construction of Relevance Measures

Abstract

A relevance measure is a measure over the space of features of a learning problem that quantifies the degree of relatedness of a single feature or a subset of features to a target variable. The measure can be used to both detect relevant features (when the target variable is the response variable) and detect redundant features (when the target variable is another input feature). Measuring relevance and redundancy is a central concept in feature selection. In this chapter, the authors show that there is a lack of generality in the features selected based on heuristic relevance measures. Through some counter-examples, the authors show that regardless of the type of heuristic measure and search strategy, heuristic methods cannot optimise the performance of all learning algorithms. They show how different measures may have different notions of relevance between features and how this could lead to not detecting important features in certain situations. The authors then propose a hyper-heuristic method that through an evolutionary process automatically generates an appropriate relevance measure for a given problem. The new approach can detect relevant features in difficult scenarios.