An Algorithm for Sample and Data Dimensionality Reduction Using Fast Simulated Annealing

摘要

This paper deals with dimensionality and sample length reduction applied to the tasks of exploratory data analysis. Proposed technique relies on distance preserving linear transformation of given dataset to the lower dimensionality feature space. Coefficients of feature transformation matrix are found using Fast Simulated Annealing - an algorithm inspired by physical annealing of solids. Furthermore the elimination or weighting of data elements which, as an effect of above mentioned transformation, were moved significantly from the rest of the dataset can be performed. Presented method was positively verified in routines of clustering, classification and outlier detection. It ensures proper efficiency of those procedures in compact feature space and with reduced data sample length at the same time.