Fast Kernel Methods for SVM Sequence Classifiers

In this work we study string kernel methods for sequence analysis and focus on the problem of species-level identification based on short DNA fragments known as barcodes. We introduce efficient sorting-based algorithms for exact string k-mer kernels and then describe a divide-and-conquer technique for kernels with mismatches. Our algorithm for the mismatch kernel matrix computation improves currently known time bounds for these computations. We then address the fast kernel vector evaluation problem during SVM testing. Finally, we introduce the mismatch kernel problem with feature selection, and present efficient algorithms for it. Our experimental results show that, for string kernels with mismatches, kernel matrices can be computed 100-200 times faster. Evaluations of kernel vectors on new sequences using new techniques also require far less time than the standard approaches. In experiments with several DNA barcode datasets, k-mer string kernels also considerably improve identification accuracy compared to prior results. String kernels with feature selection demonstrate enhanced or similar classification performance with substantially fewer computations than full feature kernels.