Experimental code for QABBA

Setup

The software is configured in Linux. Before running the experimental code, please ensure Python3 and Cython are properly installed, and run the following commands in Bash:

>>> cd source
>>> python setup.py build_ext --inplace

>>> cd software
>>> python setup.py build_ext --inplace

We use some publicly open-source code to aid in our experiments. Some scripts in the folder source are downloaded from https://github.com/nla-group/fABBA and https://github.com/nla-group/ABBA. 

Experiments

Data

The Monash Regression Dataset downloads from Time Series ExtrinsicRegression (TSER) benchmark. The data from the UCR Archive and the UEA Archive can be downloaded from https://www.timeseriesclassification.com/index.php; other data are contained in this repo.

Implementations

1. Fine-Tune Mistral-7B on QABBA for Monash Regression Dataset

Using a single A100 40G GPU, we present the steps to fine-tune Mistral-7B (or Llama2-7B)on TSER using QLoRA.

2. Demos on Errors Analysis

The results of Table 2 and Figure 1 can be reproduced by running demo.ipynb

The results of Figure 3 can be reproduced by running quantize_err.ipynb

The results of Figures 4 and 5 can be reproduced by first running multithreading.ipynb, and then the figures are generated via running mthread_results.ipynb

3. Experiment on UCR datasets

The results of Figures 6 and 7 can be reproduced by UCRPP1.ipynb and UCRPP2.ipynb, and then the figures are generated via running run_variants_profiles.ipynb

The results of Figures 8, 9, and 10 simulated via qabba_uea_0.001.ipynb, qabba_uea_0.01.ipynb, corresponding to the figures generated by results1.ipynb and results1.ipynb, respectively.

Cite

@misc{carson2025quantizedsymbolictimeseries,
      title={Quantized symbolic time series approximation}, 
      author={Erin Carson and Xinye Chen and Cheng Kang},
      year={2025},
      eprint={2411.15209},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.15209}, 
}