All Times ET
Keywords: discrete empirical interpolation method, class identification, subset selection
The discrete empirical interpolation method (DEIM) has demonstrated viability in detecting representatives of different classes present within a data matrix. As is, however, this index selection method is not able to detect more classes than the data matrix rank; while this is not an issue in many data sets, it may be prohibitive when the number of expected classes represented in the data set exceeds the number of features or time samples per observation. We present an extension of DEIM (E-DEIM) that overcomes this limitation in a deterministic manner. To evaluate the newly developed algorithm in practice, we perform subset selection on matrices formed from the MIT-BIH Arrhythmia Database and the Letter Recognition Data Set. The E-DEIM results are compared to those generated from applying the more commonly known leverage score subset selection technique as well as k-medoids clustering. This comparison suggests that E-DEIM is able to perform comparably to or better than the other methods considered, further supporting the use of DEIM-related index selection schemes for the purposes of class identification.