Introduction

Machine learning has generally ignored time in supervised classification. While there are many tools for learning static information, such as classifying different flowers according to their attributes, or determining people's credit worthiness, these are not the only type of real world classification problem.

Most real domains changes over time. What is interesting and useful to learn is not just to recognise when our classification is obsolete (changes in the economic environment, for example, may affect the accuracy of a credit-worthiness classifier); but also to use the patterns of change over time itself directly as a means of classification.

Consider a typical real-world temporal domain: speech. Our vocal chords generate amplitude and frequency values that vary over time. These variations denote a higher level concept, such as a word. Looking at the amplitude or frequency at one point in time is unlikely to help recognise words; it is only by looking at how the amplitude and frequency vary that classification becomes possible. Other examples include:

• Recognising action sequences or gestures. These arise in areas such as human-computer interaction, handwriting recognition and robot imitation.

• Medical applications. A patient's body rhythms are frequently recorded and used for diagnosis, for example: electrocardiographs, electroencephalographs, levels of various chemicals in the body and so on.

• Observing sequences of events and extracting higher level meaning from them. For example, looking at network logs and trying to identify causes of congestion problems.

• Economic and financial time series, where the user may be interested in event patterns indicating a particular phenomena or preceding particular phenomena. For example, the user might be interested in the events preceding a large market crash.

• Industrial/scientific data often includes temporal data; increasingly, today's production facilities have many embedded sensors which produce measurements at regular time intervals. The task of interest may, for example, be detecting when catastrophic events will occur, based on measurements of volume, pressure, temperature or thickness.

This thesis explores the design and implementation of a general classification tool for such temporal domains that tries to balance the specific properties of each domain against the general issues that arise in temporal classification. To do so, it proceeds in the following way:

• A brief explanation of the terminology and key problems are presented together with a theoretical foundation and formalisation.

• Existing techniques for dealing with temporal classification are discussed. These include hidden Markov models, dynamic time warping and recurrent neural networks. Current research in the artificial intelligence community is also explored.

• Metafeatures, the core of the TClass system, are presented.

• A general system for classification of multivariate time series, based on metafeatures, is proposed.

• The performance of the new system, TClass, is evaluated on two artificial and two real-world datasets: sign recognition and electrocardiograph (ECG) diagnosis.

• Extensive avenues for future research are discussed. This includes both direct extensions to the current work and also some very different approaches to doing temporal classification.

• Conclusions on the work are presented.