Evaluating and Constructing Features For Identification of Tau Leptons
Ricardo Vilalta · 2003
In this paper we show the importance of choosing the right feature representation in attempting to improve the quality of a predictive model. We explain how to evaluate and construct new features using information-theoretic measures (information gain, gain ratio) and statistical tests (e.g., 2, G statistic). Our experiments use Monte-Carlo simulated data containing both lepton signals and background events. Results show how our evaluation process can identify a small set of relevant features that bear correlation with the class ( signals). We also show how to construct new features by exploring the space of logical feature combinations using genetic algorithms; the set of newly constructed features can eectively improve the quality of the feature representation. 1.