Supervised classification has been one of the most active areas of machine learning research. However, in most of the algorithms and techniques presented, the model of the domains has been static; that is, the training and test instances have been described in terms that don't allow for changes over time.
This is surprising, because some of the most useful and interesting domains are dynamic in nature; the description of the instances inherently changes over time. Furthermore, it is only the nature of these changes that makes classification possible. Examples of such domains include recognition of gestures, speech recognition, medical signal analysis, robots detecting temporal events from sensor readings [RC98], data mining in temporal databases and many more.
Consider the classification of spoken words. We can measure what are called the Cepstral coefficients (which try to model the human aural system) of the incoming speech signal. There might be 22 such coefficients, with the values of these coefficients updated every 20 milliseconds. Looking at the value of these coefficients at a single point of time tells us very little about the word that it comes from. Only by looking at how the Cepstral coefficients have changed can the utterance be classified as a particular word. Consider a robot that is watching a door. While the robot might be able, using existing learning techniques, to learn from sensors when the door is open or closed; it would be a great deal harder for the robot to learn from sensors when the door was opening or closing.
No doubt, these problems can be converted to conventional classification tasks. However, techniques for doing so have generally been ad-hoc, labour-intensive and domain-specific. For example, Kadous [Kad95] does so for the sign language domain. In addition, such techniques do not make good use of the special temporal properties and heuristics that are likely to apply in such domains. Nor do they take into account the special ways in which temporal data can vary.
This work presents a general architecture that can be used for supervised classification in these domains, which can be described as multivariate (in the sense that there is more than one attribute being analysed and learnt) time series.