HEAD-DT: Automatic Design of Decision-Tree Algorithms

Abstract

AbstractAs presented in Chap. 2, for the past 40 years researchers have attempted to improve decision-tree induction algorithms, either by proposing new splitting criteria for internal nodes, by investigating pruning strategies for avoiding overfitting, by testing new approaches for dealing with missing values, or even by searching for alternatives to the top-down greedy induction. Each new decision-tree induction algorithm presents some (or many) of these strategies, which are chosen in order to maximize performance in empirical analyses. Nevertheless, the number of different strategies for the several components of a decision-tree algorithm is so vast after these 40 years of research that it would be impracticable for a human being to test all possibilities with the purpose of achieving the best performance in a given data set (or in a set of data sets). Hence, we pose two questions for researchers in the area: “is it possible to automate the design of decision-tree induction algorithms?”, and, if so, “how can we automate the design of a decision-tree induction algorithm?” The answer for these questions arose with the pioneering work of Pappa and Freitas [30], which proposed the automatic design of rule induction algorithms through an evolutionary algorithm. The authors proposed the use of a grammar-based GP algorithm for building and evolving individuals which are, in fact, rule induction algorithms. That approach successfully employs EAs to evolve a generic rule induction algorithm, which can then be applied to solve many different classification problems, instead of evolving a specific set of rules tailored to a particular data set. As presented in Chap. 3, in the area of optimisation this type of approach is named hyper-heuristics (HHs) [5, 6]. HHs are search methods for automatically selecting and combining simpler heuristics, resulting in a generic heuristic that is used to solve any instance of a given optimisation problem. For instance, a HH can generate a generic heuristic for solving any instance of the timetabling problem (i.e., allocation of any number of resources subject to any set of constraints in any schedule configuration) whilst a conventional EA would just evolve a solution to one particular instance of the timetabling problem (i.e., a predefined set of resources and constraints in a given schedule configuration). In this chapter, we present a hyper-heuristic strategy for automatically designing decision-tree induction algorithms, namely HEAD-DT (Hyper-Heuristic Evolutionary Algorithm for Automatically Designing Decision-Tree Algorithms). Section 4.1 introduces HEAD-DT and its evolutionary scheme. Section 4.2 presents the individual representation adopted by HEAD-DT to evolve decision-tree algorithms, as well as information regarding each individual’s gene. Section 4.3 shows the evolutionary cycle of HEAD-DT, detailing its genetic operators. Section 4.4 depicts the fitness evaluation process in HEAD-DT, and introduces two possible frameworks for executing HEAD-DT. Section 4.5 computes the total size of the search space that HEAD-DT is capable of traversing, whereas Sect. 4.6 discusses related work.KeywordsAutomatic designHyper-heuristic decision-tree inductionHEAD-DT