Orthogonality

The parameters in a given metafeature should be orthogonal to one another. For instance, if we were to represent an increasing event by start time, end time and duration, then this would be redundant. Any two of these would be sufficient to uniquely represent such an event. Fewer parameters reduce processing time of the instantiated features; furthermore, some learning and clustering algorithms applied assume that the parameters of the metafeature are independent. This may be an appropriate approximation at times, but if there is a known and clearly defined dependence (eg, in this case $\mathit{duration} = \mathit{endtime} - \mathit{starttime}$) it might affect the performance of the algorithm.