Classifying non-uniformly sampled vector-valued curves

Abstract

We propose a novel means of classifying vector-valued curves using functional principal components. This uses cross-validation to select curve components, the degree of smoothing and scores associated with the best classification performance. Our approach is well suited to data generated by sensors of different modalities and time varying sampling frequencies. Experimental comparisons show this approach has significant advantages over conventional techniques using non-uniformly sampled data. Our approach also generates novel forms of derivative analysis, a widely used technique for classifying spectral functions from contaminated data.