The application of machine learning techniques to time-series data

Abstract

"Knowledge discovery" is one of the most recent and fastest growing fields of research in computer science. It combines techniques from machine learning and database technology to find and extract meaningful knowledge from large, real world databases. Much real world data is temporal in nature, for example stock prices, dairy cow milk production figures or meteorological data. Most current knowledge discovery systems utilise similarity-based machine learning methods "learning from examples" which are not in general well suited to this type of data. Time-series analysis techniques are used extensively in signal processing and sequence identification applications such as speech recognition, but have not often been considered for knowledge discovery tasks. This report documents new methods for discovering knowledge in real world time-series data. Two complementary approaches were investigated: 1) manipulation of the original dataset into a form that is usable by conventional similarity-based learners; and 2) using sequence identification techniques to learn the concepts embedded in the database. Experimental results obtained from applying both techniques to a large agricultural database are presented and analysed.