Abstract
Observing a discrete-valued variable over a period of time yields a sequence of observation intervals: whenever the variable changes its value, the observation of its current value ends and a new interval starts. Measurements of real-valued variables over time (time series) are often partitioned in homogenous subsegments, such that every segment can be described by simple attributes, e.g. convexly increasing, constant, high-value, etc. Thus for both discrete and continuous data we obtain a sequence of labeled intervals. In this paper, a temporal pattern is defined as a set of labeled intervals together with their interval relationships in terms of Allen's interval logic. We consider the discovery and evaluation of informative rules in such sequences. In particular we focus on the problem of specializing temporal rules by imposing quantitative constraints on numerical attributes of the labeled intervals.