Time Series Knowledge Mining

Time Series Knowledge Mining (TSKM) is a new method for the understandable description of local temporal relationships in multivariate data. The patterns are expressed with the Time Series Knowledge Representation (TSKR), a new language for expressing temporal knowledge. In comparison to patterns expressed with Allen's interval relations the TSKR has advantages in robustness, expressivity, and comprehensibility. The patterns can be discovered with efficient algorithms from itemset mining. Human interaction is used during the mining to analyze and validate partial results as early as possible and guide further processing steps.

The TSKR patterns have a hierarchical structure, each level corresponds to a single temporal concept. On the lowest level, intervals are used to represent duration (Tones). Overlapping parts of intervals represent coincidence on the next level (Chords). Several such blocks of intervals are connected with a partial order relation on the highest level (Phrases). The patterns are described in a language that is understandable for the expert analyzing the data and can be automatically processed by a computer at the same time. The patterns are very compact, but offer details for each element on demand. The hierarchical structure of the temporal concepts enables the mining with seperate processing steps. This leads to smaller search spaces for each algorithm.

The picture shows the processing steps of the TSKM for the preprocessing of numerical time series to obtain symbolic interval series and most importantly the mining of coincidence and partial order.

In an application to sports medicine the results were recognized as valid and useful by an expert of the field. The numerical time series describing muscle activation and limb movement during inline speed skating were transformed to a series of states represented by symbolic intervals. Chord patterns that describe coincidences of such intervals were searched and analyzed. Phrase patterns were searched from the most interesting Chords. The best Phrases explained the most dominant components of the skating movement cycle.

You can download the code for Matlab and try it on your own data.

This research was performed under the supervision of Prof. Dr. Alfred Ultsch in the Databionics Research Group, Philipps-University of Marburg, Germany. The skating data was provided and analyzed by Dr. Olaf Hoos from the Institute for Sports Medicine, Philipps-University of Marburg, Germany.

Mörchen, F., Fradkin, D.: Robust mining of time intervals with semi-interval partial order patterns, In Proceedings SIAM Conference on Data Mining, (2010)

Mörchen, F.: Unsupervised pattern mining from symbolic temporal data, SIGKDD Explorations 9(1)ACM, (2007), pp. 41-55 ACM

Mörchen, F., Ultsch, A.: Efficient Mining Of Understandable Patterns From Multivariate Interval Time Series, Data Mining and Knowledge Discovery 15(2)Springer, (2007), pp. 181-215 SpringerLink

Mörchen, F.: A better tool than Allen's relations for expressing temporal knowledge in interval data, In Proceedings Temporal Data Mining Workshop (TDM), The Twelveth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Philadelphia, PA, USA, (2006)

Mörchen, F.: Algorithms For Time Series Knowledge Mining, Tina Eliassi-Rad, Lyle H. Ungar, Mark Craven, Dimitrios Gunopulos (Eds), In Proceedings The Twelveth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Philadelphia, PA, USA, (2006), pp. 668-673 ACM Digital Library

Mörchen, F.: Time Series Knowledge Mining, Phd Thesis, Philipps-University Marburg, Germany, Görich & Weiershäuser, Marburg, Germany, (2006), pp. 180 ISBN 3-89703-670-3

Mörchen, F., Ultsch, A., Hoos, O.: Extracting interpretable muscle activation patterns with Time Series Knowledge Mining, International Journal of Knowledge-Based & Intelligent Engineering Systems 9(3)(2006), pp. 197-208

Mörchen, F., Ultsch, A.: Discovering Temporal Knowledge in Multivariate Time Series, Weihs, C., Gaul, W. (Eds), In Classification; The Ubiquitous Challenge, Proceedings 28th Annual Conference of the German Classification Society (GfKl 2004), Dortmund, Germany, Springer, Heidelberg, (2005), pp. 272-279

Mörchen, F., Ultsch, A., Hoos, O.: Discovering interpretable muscle activation patterns with the Temporal Data Mining Method, Jean-Francois Boulicaut, Floriana Esposito, Fosca Giannotti and Dino Pedreschi (Eds), In Knowledge Discovery in Databases: Proceedings 8th European Conference on Principles and Practice of Knowledge Discovery in Databases (PKDD 2004), Pisa, Italy, Springer, (2004), pp. 512-514

Mörchen, F., Ultsch, A.: Mining Hierarchical Temporal Patterns in Multivariate Time Series, Susanne Biundo, Thom W. Frühwirth, Günther Palm (Eds), In KI 2004: Advances in Artificial Intelligence, Proceedings 27th Annual German Conference in AI, Ulm, Germany, Springer, Heidelberg, (2004), pp. 127-140

A professional athlete performed a standardized indoor test on a large motor driven treadmill. Electromyography (EMG) and kinematic data were measured for 30 seconds, corresponding to 19 movement cycles.

The lattice structure of Chords describing the coincidence of states was presented to the analyst. A tradeoff between frequency and size of the patterns needed to be made. The most interesting Chords (shown bold) were kept for further processing.