Algorithms For Time Series Knowledge Mining Fabian Moerchen 沈奕聰

Outline Introduction Related work and motivation Knowledge representation Time series knowledge mining Mining coincidence Mining partial order Experiments Discussion

Introduction Backgroud Patterns mined from symbolic interval data can provide explanation for the underlying temporal processes or anomalous behavior Symbolic interval time series are an important data format for discovering temporal knowledge Numerical time series are often converted to symbolic interval time series

Introduction Problems Allen’s interval relations ‘s input usually consists of exact but incomplete data and temporal constraints Determining the consistency of the data Answering queries about scenarios satisfying all constraints Noisy and incorrect interval data

Introduction Propose Time Series Knowledge Representation(TSKR) Hierarchical language Based on interval time series Extends the Unification-based Temporal Grammar Using itemset techniques

Related work and motivation Allen’s relations have severe disadvantages Patterns from noisy interval data expressed with Allen’s interval relations are not robust

Related work and motivation Allen’s relations have severe disadvantages Patterns expressed with Allen’s interval relations are ambiguous

Related work and motivation Allen’s relations have severe disadvantages Patterns expressed with Allen’s interval relations are not easily comprehensible

Related work and motivation The TSKR extends these core ideas achieving higher robustness and expressivity The hierarchical structure of the UTG The separation of temporal concepts

Knowledge representation Tones : basic primitives of the TSKR representing duration Chord: a Chord pattern describes a time interval where k>0 Tones coincide Phrase: a paritial order of k>1 Chords

Time series knowledge mining ——Mining coincidence

Time series knowledge mining ——Mining partial order

Experiments

Discussion Advantages Hierarchical structure show the coinciding Tones and one Tone to show the original numerical time series with the thresholds for discretization The pruning by margin-closedness largely reduced the number of patterns Effects on search space Our novel data model conversion to itemset intervals greatly reduce the redundancy Search for phrases with our semantically motivated search space restrictions are much faster than sequential pattern