Concurrency in feature analysis

Abstract

Classifier-independent feature analysis is a classic problem whose solution exhibits exponential growth. In previous research, we developed a new approach to classifier-independent feature analysis based on relative feature importance (rfi), a metric for the relative usefulness of the feature in the optimal subset. Because finding the optimal subset requires exhaustive search, we have also developed an estimator for rfi. The estimator uses adaptive techniques to reduce the computational load. The implementation of both algorithms, direct calculation of rfi and the estimator, on a Connection Machine (CM-5) in CM Fortran is described in this paper. Direct calculation of rfi lends itself naturally to implementation in CM Fortran because the computationally intensive components of the algorithm involve manipulation of large arrays. The adaptive nature of the estimator, however, makes implementing it on the CM-5 more challenging and less efficient.

This research partially supported by a grant from the NRL/ARPA Connection Machine Facility, an effort to rapidly prototype Massively Parallel Processing.