A subspace projection based feature fusion: An application to EEG clustering

Abstract

Goal: In most decision models, feature biasing (FB) is the major concern that greatly impacts the performance management burden. The objective of this framework is to present a multi-view feature fusion strategy using canonical correlation analysis (CCA) that can effectively classify various classes of Electroencephalogram (EEG) patterns. Method: To make the best use of inherent class information, we first created multi-view vectors (MVVs) registering templates (i.e., signals) associated with study specific class groups through a given strategy, followed by projection to extract compact views, which are then fused via parallel fusion and then, applied to classification. Results: On EEG data, the learned patterns effectively represent underlying information they were trained, with significant performance in terms of markers. Further, its comparison with state-of-thearts manifests the efficacy of adopted model. Conclusion: The methodology effectively classify various EEG patterns. Significance: The significant reduction of complexity and dimensionality with enhanced information space envisages the possible extension of this work to alleviate the onus of clinician of large volume data and also expedite the large scale research.