ACD: A Lossless Compression Based on Dimensionality Reduction for Multi-dimensional Time Series Data
Haolong Chen, Liang Liu · 2024
Multi-dimensional time series data are generated daily in lots of domains. Compressing these data to reduce storage overhead is an important topic. Existing compression algorithms resort to converting multi-dimensional time series data into multiple separate one-dimensional data to compress. However, these algorithms fail to exploit both intra- and inter-correlation among dimensions and the compression is not efficient. To solve the problem, this paper introduces a novel multi-dimensional time series compression framework called ACD, a lossless compression based on dimensionality reduction. Initially, the n-dimensional time series data undergo dimensionality reduction via space-filling curves(SFCs) encoding. Subsequently, we partition the encoded data heuristically into several segments based on distinctive patterns in different time windows. Finally, suitable compression is conducted according to segment patterns. We implement our algorithm in an out-of-box commercial time series database engine VictoriaMetrics. Our experiments demonstrate that ACD achieves at least 40% improvement over existing methods such as Chimp, Gorilla, and DeltaZstd as evaluated through the Time Series Benchmark Suite(TSBS).