Optimised search strategies to improve structural pattern recognition techniques

Abstract

Data retrieval methods are based on three essential aspects: feature extraction (to reduce signal dimensionality), the classification system (to index objects according to some criteria) and similarity measures (to compare how similar two objects are); but there is not a single solution to handle these key elements. This paper provides a new solution to the localisation and extraction of similar patterns in time-series data. Alternative searches are proposed to objectively increase the recognition of similar patterns so as to achieve better results on the data retrieval. These search strategies have been studied with excellent results in the detection of long subpatterns. Long subpatterns are not very easy to identify since even a single mismatch in one character can compromise similarity between two patterns. Identifying long patterns in a fast, fault-tolerant and intelligent way is the aim of the analysed strategies, which are formally based on statistical criteria and some aspects of probability theory.