GMT: Gzip-based Memory-efficient Time-series classification

Sungmin Lee, Kichang Lee, Jaeyeon Park, JeongGil Ko

Abstract: The deployment of embedded time-series sensing devices enabled better understanding of user environments and contexts. However, classifying them solely on extremely limited devices under data-scarce conditions is still a remaining challenge. We introduce GMT, a memory-efficient parameter-free classifier that uses gzip compressor and k-nearest neighbors (kNN) for classifying multi-channel time-series data. GMT tackles issues due to high data fidelity, multi-channel characteristics, and numerical properties of sensor data using techniques such as floating point quantization, channel-wise compression, and hybrid distance. Experiments show that GMT provides superior accuracy and memory efficiency compared to other classifiers across various tasks and applications.

External IDs:dblp:journals/ict-express/LeeLPK25