20250729 meeting

Introduction

This week’s experiment focuses on integrating the autoFRK and SSSDS4 models. All experiments use the full version of autoFRK for prediction, rather than the MRTS variant.

The experiments this week are designed around three scenarios:

1. All known locations are re-predicted using autoFRK, utilizing the current locations to predict current values. These results are then directly passed to the SSSDS4 model for future forecasting.

   This experiment is designed to verify whether, assuming autoFRK predictions are accurate, the subsequent future predictions using SSSDS4 can maintain high accuracy.

   

2. Unknown locations are first imputed each time period using autoFRK, and then passed to SSSDS4 for future forecasting.

   This experiment is used to compare with the previous one and verify whether imputing unknown locations still results in high accuracy.

   

3. Known locations are first forecasted into the future using SSSDS4, and then unknown locations are imputed for each time period using autoFRK.

   This experiment is used to compare with the previous one, in order to examine whether the order of using autoFRK and SSSDS4 affects the prediction results. Since this experiment does not impute past missing values, it is expected to be faster than the previous one.

   

Experiment 1

• autoFRK predicted values (250 days)

  Method	Value
  MSE	2.208425763
  RMSE	2.208425763
  MSE%	0.007850246
  RMSE%	0.007850246
  MAE	1.058653390
  MAE%	0.003772132

• SSSDS4 predicted values (250 + 10 days)

  Method	Value
  MSPE (All)	25.643074
  MSPE (Future)	22.201488
  MAPE (All)	3.847528
  MAPE (Future)	3.777283
  MSPE% (All)	0.092841
  MSPE% (Future)	0.078886
  MAPE% (All)	0.013823
  MAPE% (Future)	0.013367

Experiment 2

• autoFRK predicted values (250 days, predicting only missing locations)

  Method	Value
  MSPE	3.876077622
  RMSPE	1.968775666
  MSPE%	0.013798632
  RMSPE%	0.117467580
  MAPE	1.406543020
  MAPE%	0.005010491

• SSSDS4 predicted values (250 + 10 days)

  Method	ALL Locs & All Time	Known Locs & All Time	Unknown Locs & All Time	ALL Locs & Future	Known Locs & Future	Unknown Locs & Future	ALL Locs & Past	Known Locs & Past	Unknown Locs & Past
  MSPE	125.489044	124.578705	129.390610	124.660667	122.997314	131.789291	125.522163	124.641937	129.294662
  RMSPE	11.202189	11.161483	11.374999	11.165154	11.090415	11.479952	11.203668	11.164315	11.370781
  MAPE	8.593054	8.562331	8.724644	8.801373	8.733807	9.090941	8.584720	8.555482	8.709991
  MSPE%	0.446431	0.443675	0.458241	0.437042	0.431520	0.460707	0.446807	0.444162	0.458142
  RMSPE%	0.668155	0.666090	0.676935	0.661092	0.656902	0.678754	0.668436	0.666455	0.676862
  MAPE%	0.030651	0.030567	0.031012	0.030911	0.030694	0.031840	0.030640	0.030561	0.030979

Experiment 3

(Not yet started)

Conclusion

After examining autoFRK, its imputation method first selects the 3 nearest neighboring coordinates via KNN and takes the average as the imputed value. To improve SSSDS4 by incorporating spatial characteristics, one possible approach is to borrow the KNN method: form a new data frame from neighboring points for training, then fill in the missing values. Alternatively, after forming the new data frame from neighbors, train using a method similar to random forests—train multiple times and select the best-fitting model. However, note that training a single SSSDS4 model currently takes too long; if this approach is adopted, the existing model must be improved to reduce its training time.

Epilogue

Environment

• Local Operating System: Windows 11 24H2
  • Programming Language: Python 3.12.9
• Computing Platform: National Center for High-Performance Computing (NCHC) – Taiwan AI Cloud
  • Operating System: Ubuntu
  • Miniconda
  • GPU: NVIDIA Tesla V100 32GB GPU
  • CUDA 12.8 driver
  • Programming Language: Python 3.10.16 for Linux

Further Learning

• I tested this project on the GitHub repository.

References

• Juan Lopez Alcaraz, Nils Strodthoff. (2022). Diffusion-based time series imputation and forecasting with structured state space models. Transactions on Machine Learning Research. Retrieved from https://openreview.net/forum?id=hHiIbk7ApW

• SSSD. (2022). GitHub. Retrieved from https://github.com/AI4HealthUOL/SSSD

• SSSD_CP. (2024). GitHub. Retrieved from https://github.com/egpivo/SSSD_CP

• 近兩年小時值查詢. (n.d.). 環境部 - 空氣品質監測網. Retrieved from https://airtw.moenv.gov.tw/CHT/Query/InsValue.aspx