Hier finden Sie wissenschaftliche Publikationen aus den Fraunhofer-Instituten.

A distributed online learning approach for pattern prediction over movement event streams with apache flink

: Qadah, Ehab; Mock, Michael; Alevizos, Elias; Fuchs, Georg

Volltext urn:nbn:de:0011-n-4842507 (922 KByte PDF)
MD5 Fingerprint: 8ee8c85f74888b3077128e3c921ba5fb
(CC) by-nc-nd
Erstellt am: 13.2.2018

Workshops of the EDBT/ICDT 2018 Joint Conference. Proceedings. Online resource : Vienna, Austria, March 26, 2018
Vienna, 2018 (CEUR Workshop Proceedings 2083)
International Conference on Extending Database Technology (EDBT) <21, 2018, Vienna>
International Conference on Database Theory (ICDT) <21, 2018, Vienna>
European Commission EC
H2020; 687591; datAcron
Konferenzbeitrag, Elektronische Publikation
Fraunhofer IAIS ()

In this paper, we present a distributed online prediction system for user-defined patterns over multiple massive streams of movement events, built using the general purpose stream processing framework Apache Flink. The proposed approach is based on combining probabilistic event pattern prediction models on multiple predictor nodes with a distributed online learning protocol in order to continuously learn the parameters of a global prediction model and share them among the predictors in a communication-efficient way. Our approach enables the collaborative learning between the predictors (i.e., "learn from each other"), thus the learning rate is accelerated with less data for each predictor. The underlying model provides online predictions about when a pattern (i.e., a regular expression ove r the event types) is expected to be completed within each event stream. We describe the distributed architecture of the proposed system, its implementation in Flink, and present experimental results over real-world event streams related to trajectories of moving vessels.