Generating Automatic Projections by Means of Genetic Programming

The aim of inductive machine learning (ML) is to generate models that can make predictions from analysis of data sets. These data sets consist of a number of instances or examples, each example described by a set of attributes. It is known that the quality or relevance of the attributes of a data set is a key issue when trying to obtain models with a satisfactory level of generalization. There are many techniques of feature extraction, construction, and selection [1] that try to improve the representation of data sets, thus increasing the prediction capabilities of traditional ML algorithms. These techniques work by filtering nonrelevant attributes or by recombining the original attributes into higher-quality ones. Some of these techniques were created in an automatic way by means of genetic programming (GP). GP is an evolutionary technique for evolving symbolic programs [2]. Most research has focused on evolving functional expressions, but the use of loops and recursion has also been considered [3]. Evolving circuits are also among the successes of GP [4]. In this work we present a method for attribute generation based on GP called the GPPE (genetic programming projection engine). Our aim is to evolve symbolic mathematical expressions that are able to transform data sets by representing data on a new space, with a new set of attributes created by GP. The goal of the transformation is to be able to obtain higher accuracy in the target space than in the original space. The dimensions of the new data space can be equal to, larger, or smaller than those of the original. Thus, we also intend that GPPE be used as a dimension reduction technique as