Supervised classification has been one of the most active areas of machine learning research. However, the domains where it has been applied are relatively limited. In particular, much of the work has focused on classification in static domains, where the attributes of the training examples are assumed not to change over time. In many domains, attributes are not static; in fact, it is the way they vary temporally that can make classification possible. Examples of such domains include speech recognition, event recognition from sensors in robotics and analysis of electrocardiographs.
So far, researchers tackling these domains have used ad-hoc techniques for converting the problem to a standard classification task. This fails to take into account both the special problems and special heuristics applicable to temporal data.
This paper proposes a general architecture for classification of multivariate time series. Training proceeds in five steps: extraction of events from the data training based on parametrised event primitives; clustering of the events in their parameter space to create synthetic events; event attribution of the training data and finally building a classifier with a conventional learner. Recognition takes two steps: selective event searching for synthetic events within the test instance (usually only a small subset of the synthetic events generated in training need to be searched for), then feeding through the classifier created in the training stage.
An example implementation of this general architecture is presented. Some preliminary results of its application to recognition of signs from Australian Sign Language are also discussed.
Keywords: machine learning, classification, temporal classification, gesture recognition, time series.