Incremental Feature Subsetting useful for Big Feature Space Problems

Dimensionality Reduction process is a means to overcome curse of dimensionality in general. When all features are available together, it is a way to extract knowledge from a population in a big feature space. On the contrary, dimensionality reduction is intriguing when update to feature space is streaming and the question arises whether one could reduce the feature space as and when the features become available instead of waiting for all the features to arrive .This could not only enable the creation of knowledge that can incrementally align with the incremental access to feature space, but would also facilitate decision making locally at every incremental stage. While facilitating the local decision making parameters, it would eventually generate the most optimal reduced feature space. Moving in this direction, the possibility of implementing feature subsetting in an incremental framework is explored. The incremental streaming could be due to the temporal arrival of features or due to collection of features arriving from distributed sources. In this paper, the adoption of incremental dimensionality reduction model is also explored to observe the complexity reduction of working with a big feature space. The speciality of the proposed incremental framework is that the dimensionality reduction is performed to obtain a cumulative reduced feature space at every stage without having to look back at the earlier features. General Terms Pattern Recognition, Dimensionality Reduction, Feature Subsetting, Incremental Learning