Hybridized Oscillating Search Algorithm for Unsupervised Feature Selection

In feature selection, a search problem of finding a subset of features from a given set of measurements has been of interest for a long time. However, unsupervised methods are scarce. An unsupervised criterion, based on SVD-entropy (Singular Value Decomposition), selects a feature according to its contribution to the entropy (CE) calculated on a leave-one-out basis. Based on this criterion, this paper proposes a Hybridized Oscillating Search feature selection method (HOS) which does not follow a pre defined direction of search (forward or backward). It is a randomized search method which begins with a random subset of features. The proposed HOS method makes use of a sequential feature selection method called Simple Ranking based on CE to get the initial feature subset. Repeated modification of the subset is achieved through up and down swings which form the oscillating cycles. The up swing adds good features to the current subset while the down swing removes worst features from the current subset. After each oscillating cycle, the subset is evaluated by comparing its predictive accuracy with known classification. Common indices like Rand Index and Jaccard Coefficient are used for this purpose. If the last oscillating cycle did not find a better subset, then the process ends with the current subset.