A better tool than allen's relations for expressing temporal knowledge in interval data

Abstract

Temporal patterns composed of symbolic intervals are commonly formulated with Allen's interval relations originating in temporal reasoning. We show that this representation has severe disadvantages for knowledge discovery. The patterns are not robust, in the sense that small disturbances of interval boundaries lead to different patterns for similar situations. The representation is ambiguous since the same pattern can have quantitatively widely varying appearances. For all but very simple cases the patterns are not understandable because the textual descriptions are lengthy and unstructured. We present the Time Series Knowledge Representation (TSKR), a new hierarchical language for interval patterns to express the temporal concepts of coincidence and partial order. We demonstrate the superiority of this novel form of representing temporal knowledge over Allen's relations for data mining. Results on a real data set support our claims and show a successful application.