Predicting Significant Events via Sequence Learning

S. Das and H. Ruda

Proceedings of the Workshop on Knowledge Discovery from Temporal and Spatial Data, 15th European Conference on Artificial Intelligence, Lyon, France (July, 2002)

We present here an approach to sequence learning for predicting significant events at multiple levels of abstraction. The approach is capable of analyzing a collection of events over a period of time to identify trends in the form of temporal causal rules, thereby extending conventional static data mining techniques that discover only non-temporal rules. Taxonomical organization of contextual event types in hierarchies allows us to obtain higher-level abstractions of events corresponding to observed low-level events. Therefore, significant events can be generated at multiple levels of abstractions with associated confidences. The in-house implementation of the approach predicts significant events in widely different applications, including spacecraft anomalies using status and health telemetry data, and terrorist actions using member and organizational activities. We demonstrate the effectiveness of the approach by an evaluation that shows high hit rates and low false alarm rates on real events about activities of a terrorist group collected from open source literature.

Event sequence with prediction and evaluation markers; with partial event taxonomy

Request a Copy