Potential Functions for Signals and Symbolic Sequences

• Published in: Braverman Readings in Machine Learning. Key Ideas from Inception to Current State pp. 3 - 31
• Main Authors: Sulimova, Valentina, Mottl, Vadim
• Format: Book Chapter
• Language: English
• Published: Cham Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Featureless approach; Membrane protein prediction; Multi-kernel learning; Potential (Kernel) functions; Probabilistic models; Sequence alignment; SVM
• ISBN: 9783319994918; 3319994913
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-99492-5_1

This paper contains a comprehensive survey of possible ways for potential functions design on sets of signals and symbolic sequences. Significant emphasis is placed on a generalized probabilistic approach to construction of potential functions. This approach covers both vector signals and symbolic sequences at once and leads to a large family of potential functions based on the notion of a random transformation of signals and sequences, which can underlie, in particular, probabilistic models of evolution of biomolecular sequences. We show that some specific choice of the sequence random transformation allows to obtain such important particular cases as Global Alignment Kernel and Local Alignment Kernel. The second part of the paper addresses the multi-kernel situation, which is extremely actual, in particular, due to the necessity to combine information from different sources. A generalized probabilistic featureless SVM-based approach to combining different data sources via supervised selective kernel fusion was proposed in our previous papers. In this paper we demonstrate significant qualitative advantages of the proposed approach over other methods of kernel fusion on example of membrane protein prediction.