Reproducibility of Experimental Results from a Highly Parallelized Classification Algorithm

Abstract

The classification and interpretation of high-dimensional biomedical data is frequently a computationally expensive problem. When analyzing such data it is often advantageous to identify a subset of relevant features that minimize classification errors. During the discovery of such subsets, reproducibility (repeatability) of experimental results is an essential requirement. We present a load balanced parallel algorithm for identifying discriminatory feature subsets. Also presented are various techniques used to ensure the repeatability of the algorithm's experimental results. Experiments conducted on biomedical spectra using a variety of the presented parallelization approaches show some techniques to be significantly more effective than others.