Feature selection with applications to text classification

摘要

Application of a feature selection algorithm to a textual data set can improve the performance of some classifiers. Due to the characteristics, specifically the size, of textual data sets researchers have traditionally relied on a family of simple heuristics to perform feature selection. These heuristics, which in practice are quite effective, use functions of individual feature statistics, that we refer to as feature ranking functions, to order the feature set.;We are interested in identifying the most effective feature ranking functions. To do this we begin by defining a feature set evaluation methodology. Traditionally the performance of feature selection algorithms has been measured by comparing the performance of classification algorithms before and after feature selection. Instead, we measure various criteria of the selected feature set itself, including measures of separation, noise, size, and robustness. We demonstrate that many of these criteria are competing, and show how the tools of multicriteria optimization can be employed to rank the performance of feature selection algorithms.;Using this methodology we evaluate the performance of a large set of feature ranking functions, including a function that measures the rareness of a feature assuming that relevant and irrelevant documents are generated by two independent stochastic processes. Motivated by the results, we identify the defining characteristics of the functions that are most successful, noting that many of these can be written as ratios of measures of separation to measures of noise.;Next we introduce a set of axioms which we believe that feature ranking functions should satisfy, and study the set of these functions that can be represented as a linear combination of some finite set of basis functions. We demonstrate that many of the functions or approximations to the functions that we studied are members of this set. Next consider the set of coefficient vectors of this set and show that it is convex, bounded, and not empty. We conclude by investigating the performance of other approaches to feature selection including greedy and ensemble algorithms that use feature ranking functions.