Comparison of Clustering Methods for Extraction of Uncorrelated Sparse Sources from Data Mixtures

There is an extensive set of methods to determine sparse sources from mixtures where the mixing coefficients are unknown. Each method involves plotting N sets of mixed data against each other in N-dimensional space. In the approach adopted in this paper, N dimensional normalised vectors are produced by joining data points that are adjacent in time. A novel clustering approach is adopted: the two vectors, not necessarily adjacent in time, which are closest to each other are identified and one of these vectors is taken as the principal direction corresponding to one of the sources. It is shown, using a deflation approach, that it is possible to estimate individual sources to within a multiplicative constant. This novel method is compared with two related methods and the standard FastICA algorithm. This new method has comparable performances to three other methods when applied to examples of purely sparse, semi-sparse and non-sparse sources and also when applied to fetal ECG data.