Summary: In knowledge discovery from time series abstractions (like piecewise linear representations) are often preferred over raw data. In most cases it is implicitly assumed that there is a single valid abstraction and that the abstraction method, which is often heuristic in nature, finds this abstraction. We argue that this assumption does not hold in general and that there is need for knowledge discovery methods that pay attention to the ambiguity of features: In a different context, an increasing segment may be considered as (being part of) a decreasing segment. It is not a priori clear which view is correct or meaningful. We show that the relevance of ambiguous features depends on the relevance of the knowledge that can be discovered by using the features. We combine techniques from multiscale signal analysis and interval sequence mining to discover rules about dependencies in multivariate time series.
For the entire collection see [Zbl 1011.00038].