Lambda pruning: an approximation of the string subsequence kernel for practical SVM classification and redundancy clustering

Abstract

The support vector machine (SVM) is a powerful learning algorithm, e.g., for classification and clustering tasks, that works even for complex data structures such as strings, trees, lists and general graphs. It is based on the usage of a kernel function for measuring scalar products between data units. For analyzing string data Lodhi et al. (J Mach Learn Res 2:419–444, 2002) have introduced a String Subsequence kernel (SSK). In this paper we propose an approximation to SSK based on dropping higher orders terms (i.e., subsequences which are spread out more than a certain threshold) that reduces the computational burden of SSK. As we are also concerned with practical application of complex kernels with high computational complexity and memory consumption, we provide an empirical model to predict runtime and memory of the approximation as well as the original SSK, based on easily measurable properties of input data. We provide extensive results on the properties of the proposed approximation, SSK-LP, with respect to prediction accuracy, runtime and memory consumption. Using some real-life datasets of text mining tasks, we show that models based on SSK and SSK-LP perform similarly for a set of real-life learning tasks, and that the empirical runtime model is also useful in roughly determining total learning time for a SVM using either kernel.