One prominent area of machine learning research is supervised classification and in particular, attribute-value learning. Attribute-value learners typically assume that each training instance is defined by a set of attributes, that for the purposes of the learner, do not change. In many domains however, it is this change in attributes over time that makes classification possible. Examples of such domains include recognition of gestures, medical signal analysis, robot sensor analysis and mining in temporal databases.
To make the distinction clear, consider the speech recognition task, where an audio signal representing a word must be classified. The audio signal consists of 22 spectral frequency coefficients, updated 50 times a second. Looking at the values of the coefficients at a single instant of time is not very useful for classification; rather, classification can only be performed by looking at the changes of these coefficients over time.
It is possible to extract features from time series, then apply a conventional learner. There are several drawbacks to such an approach. Firstly, the techniques tend to be ad hoc, domain-specific and labour-intensive; the current work arises from the author's frustrating experiences trying to do this for the sign recognition task [Kadous, 1995]. Secondly, there are special heuristics applicable to temporal domains that are difficult to capture by such a conversion process (for example, different parts of the instance being slightly stretched or delayed in time). Thirdly, descriptions built using these extracted features can be hard to understand.
A description of the problem, followed by two examples of temporal classification tasks are given. A general architecture is proposed for the classification and description of multivariate time series. An implementation of the system, called TClass, is presented; as is its application to the two temporal classification tasks.