Skip to content

Latest commit

 

History

History
186 lines (146 loc) · 8.38 KB

README.md

File metadata and controls

186 lines (146 loc) · 8.38 KB
Raw

Reproducibility Repository for ORBITS

  • Paper: Mourad Khayati, Ines Arous, Zakhar Tymchenko and Philippe Cudré-Mauroux: ORBITS: Online Recovery of Missing Values in Multiple Time Series Streams. PVLDB 2021.
  • Algorithms: The benchmark evaluates all the algorithms mentioned in the paper: ORBITS, SPIRIT, SAGE, OGDImpute, pcaMME, TKCM and M-RNN.
  • Datasets: The benchmark evaluates all the datasets used in the paper: gas (drfit10), motion, bafu, and soccer.
  • Scenarios: The benchmark will execute the full set of 11 recovery scenarios and report the recovery error (RMSE, MSE and MAE) and correlation (Pearson and Spearman). A detailed description of the recovery scenarios can be found here.
  • News: This code is mentioned in the VLDB reproducibility highlights.

Prerequisites | Build | Execution | Citation

Prerequisites

Build

  • Build the Testing Framework using the installation script located in the root folder (takes a few minutes):
    $ sh install_linux.sh

Execution

  • We provide three different scripts to reproduce the experiments. In case you are using the same machine to reproduce the results, the three scripts should be executed sequentially.
  1. The results and plots of all techniques, except MRNN, can be obtained by running the following script (takes ~3 days):
    $ cd TestingFramework/bin/Debug/
    $ sh repro.sh
  1. The results of MRNN can be obtained by running the following script (takes 2-7 days, depending on the machine):
     $ cd TestingFramework/bin/Debug/
     $ sh repro_extra_mrnn.sh
  1. The runtime results of ssv (Fig 13a and 13b: Soccer dataset) take a considerable amount of time and are, thus, not reported. These results can be obtained by running the following script (takes ~3 days):
     $ cd TestingFramework/bin/Debug/
     $ sh repro_extra_ssv.sh

Results

Figure 5

  • Fig. 5a:
    • Plot: Results/streaming_end/missingpercentage/drift10/error/plots/drift10_rmse.eps
    • RMSE: Results/streaming_end/missingpercentage/drift10/error/rmse
  • Fig. 5b:
    • Plot: Results/streaming_end/missingpercentage/motion/error/plots/motion_rmse.eps
    • RMSE: Results/streaming_end/missingpercentage/motion/error/rmse
  • Fig 5c:
    • Plot: Results/streaming_end/missingpercentage/bafu/error/plots/bafu_rmse.eps
    • RMSE: Results/streaming_end/missingpercentage/bafu/error/rmse
  • Fig 5c:
    • Plot: Results/streaming_end/missingpercentage/soccer/error/plots/soccer_rmse.eps
    • RMSE: Results/streaming_end/missingpercentage/soccer/error/rmse

Figure 6

Manual aggregation over mutiple results. The reproducibility of this figure is very hard to automate.

Figure 7

  • Fig 7a:
    • Plot: Results/streaming_end/ts-block-fix/drift10/error/plots/drift10_rmse.eps
    • RMSE: Results/streaming_end/ts-block-fix/drift10/error/rmse
  • Fig 7b:
    • Plot: Results/streaming_end/ts-block-fix/motion/error/plots/motion_rmse.eps
    • RMSE: Results/streaming_end/ts-block-fix/motion/error/rmse
  • Fig 7c:
    • Plot: Results/streaming_end/ts-block-fix/bafu/error/plots/bafu_rmse.eps
    • RMSE: Results/streaming_end/ts-block-fix/bafu/error/rmse
  • Fig 7d:
    • Plot: Results/streaming_end/ts-block-fix/soccer/error/plots/soccer_rmse.eps
    • RMSE: Results/streaming_end/ts-block-fix/soccer/error/rmse

Figure 8

  • Fig 8a:
    • Plot: Results/streaming_end/mcar-ts-multiblock/drift10/error/plots/drift10_rmse.eps
    • RMSE: Results/streaming_end/mcar-ts-multiblock/drift10/error/rmse
  • Fig 8b:
    • Plot: Results/streaming_end/mcar-ts-multiblock/motion/error/plots/motion_rmse.eps
    • RMSE: Results/streaming_end/mcar-ts-multiblock/motion/error/rmse
  • Fig 8c:
    • Plot: Results/streaming_end/mcar-ts-multiblock/bafu/error/plots/bafu_rmse.eps
    • RMSE: Results/streaming_end/mcar-ts-multiblock/bafu/error/rmse
  • Fig 8d:
    • Plot: Results/streaming_end/mcar-ts-multiblock/soccer/error/plots/soccer_rmse.eps
    • RMSE: Results/streaming_end/mcar-ts-multiblock/soccer/error/rmse

Figure 9

We compute the AVG for each technique over three runs. The absolute correlation values might be slightly different as, in each run, we randomly pick a different time series to recover.

  • Fig 9a:
    • Correlation: Results/streaming_end/mcar-ts-multiblock/drift10/error/correlation.dat
  • Fig 9b:
    • Correlation: Results/streaming_end/mcar-ts-multiblock/motion/error/correlation.dat
  • Fig 9c:
    • Correlation: Results/streaming_end/mcar-ts-multiblock/bafu/error/correlation.dat
  • Fig 9d:
    • Correlation: Results/streaming_end/mcar-ts-multiblock/soccer/error/correlation.dat

Figure 10

  • Fig 10a:
    • Plot: Results/streaming_end/mcar-length/bafu/runtime/plots/bafu_rt_log.eps
    • Runtime: Results/streaming_end/mcar-length/bafu/runtime/values
  • Fig 10b:
    • Plot: Results/streaming_end/mcar-length/socccer/runtime/plots/socccer_rt_log.eps
    • Runtime:Results/streaming_end/mcar-length/socccer/runtime/values

Figure 11

  • Fig 11a:
    • Plot: Results/streaming_end/mcar-columns/motion/runtime/plots/drift10_rt_log.eps
    • Runtime: Results/streaming_end/mcar-columns/motion/runtime/values
  • Fig 11b:
    • Plot: Results/streaming_end/mcar-columns/motion/runtime/plots/drift10_rt_log.eps
    • Runtime: Results/streaming_end/mcar-columns/motion/runtime/values

Figure 12

  • Fig 12a:
    • Plot: ResultsSynth/streaming_end/mcar-columns/bafu_full/runtime/plots/bafu_full_rt_log.eps
    • Runtime: ResultsSynth/streaming_end/mcar-columns/bafu_full/runtime/values
  • Fig 12b:
    • Plot: ResultsSynth/streaming_end/mcar-columns/soccer_full/runtime/plots/soccer_full_rt_log.eps
    • Runtime: ResultsSynth/streaming_end/mcar-columns/bafu_full/runtime/values

Figure 13

  • Fig 13a:

    • Runtime: ResultsSVector/f13_len.txt

  • Figure 13b:

    • Runtime: ResultsSVector/f13_col.txt

Note: In case you run the MRNN script, the results will be added to the folder ResultsMRNN/ under the same subfolders as decribed above.

Citation

@inproceedings{orbits2021vldb,
 author    = {Mourad Khayati and Ines Arous and Zakhar Tymchenko and Philippe Cudr{\'{e}}{-}Mauroux},
 title     = {ORBITS: Online Recovery of Missing Values in Multiple Time Series Streams},
 booktitle = {Proceedings of the VLDB Endowment},
 volume    = {14},
 number    = {3},
 year      = {2021}
}