Affiliations: Irisa, Université de Rennes 1, Campus de
Beaulieu, 35042 Rennes cedex, France. [email protected] | Department of Computer Science, University of Wales,
Aberystwyth, Aberystwyth, SY23 3DB, Wales, UK. [email protected]
Note: [] Address for correspondence: Irisa, Université de Rennes
1, Campus de Beaulieu, 35042 Rennes cedex, France
Abstract: Many application domains make use of specific data structures such
as sequences and graphs to represent knowledge. These data structures are
ill-fitted to the standard representations used in machine learning and
data-mining algorithms: propositional representations are not expressive
enough, and first order ones are not efficient enough. In order to efficiently
represent and reason on these data structures, and the complex patterns that
are related to them, we use domain-specific logics. We show these logics can be
built by the composition of logical components that model elementary data
structures. The standard strategies of top-down and bottom-up search are
ill-suited to some of these logics, and lack flexibility. We therefore
introduce a dichotomic search strategy, that is analogous to a dichotomic
search in an ordered array. We prove this provides more flexibility in the
search, while retaining completeness and non-redundancy. We present a novel
algorithm for learning using domain specific logics and dichotomic search, and
analyse its complexity. We also describe two applications which illustrates the
search for motifs in sequences; where these motifs have arbitrary length and
length-constrained gaps. In the first application sequences represent the
trains of the East-West challenge; in the second application they represent the
secondary structure of Yeast proteins for the discrimination of their
biological functions.
