Another technique for extracting information from the training data is to evaluate some aggregate values and use them as propositional attributes. This would cover features of a stream that are not localised. For example, for continuous channels, the global maximum and minimum value of a channel, or the mean of each channel may be useful. Such attributes are examples of aggregate global attributes - they measure some property of each training instance as a whole, rather than looking at the temporal structure. Such features can be surprisingly useful for classification. For example, [Kad95] shows that in the sign language domain task discussed in Section 6.3.2 almost 30 per cent accuracy can be obtained in classifying 95 signs based solely on the maxima and minima of they x, y, and z positions of the right hand.
Global features may be more complicated than simple maxima, minima and averages. For example, we could count the number of local maxima and minima, we could measure the total ``distance'' covered by each channel, or use some measure of energy. For discrete binary channels, it might be something like how many changes there are, the percentage of time for which the value is true and so on. Duration is another example of an aggregate temporal feature, which may be useful for some classification purposes.
Note that global features may be associated with either single or multiple channels. For example, in the Auslan domain, we could measure the distance using a Euclidean metric on the x, y and z channels, or on each channel separately. For simplicity below, we consider global feature calculation for a single channel.