1140731 meeting
目錄
前言
此處修改前次實驗 2 與實驗 3 ，使實驗結果的 autoFRK 也預測已知地點，而非僅預測缺失地點。結果如下：
實驗 2
gallery_made_with_nanogallery_exp2
| 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 | 122.088806 | 123.208755 | 117.288971 | 122.455757 | 121.838974 | 125.099060 | 122.074120 | 123.263573 | 116.976578 |
| RMSPE | 11.049380 | 11.099944 | 10.830004 | 11.065973 | 11.038070 | 11.184769 | 11.048716 | 11.102413 | 10.815571 |
| MAPE | 8.474641 | 8.495713 | 8.384321 | 8.760569 | 8.725639 | 8.910264 | 8.463192 | 8.486518 | 8.363283 |
| MSPE% | 0.434309 | 0.438404 | 0.416762 | 0.429368 | 0.427416 | 0.437735 | 0.434507 | 0.438843 | 0.415924 |
| RMSPE% | 0.659021 | 0.662121 | 0.645571 | 0.655262 | 0.653770 | 0.661615 | 0.659172 | 0.662452 | 0.644921 |
| MAPE% | 0.030226 | 0.030313 | 0.029852 | 0.030770 | 0.030664 | 0.031227 | 0.030204 | 0.030299 | 0.029797 |
實驗 3
gallery_made_with_nanogallery_exp3
| 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 | 17.35709501 | 17.18139008 | 18.11011615 | 24.97269111 | 24.97548082 | 24.96073520 | 17.05247116 | 16.86962645 | 17.83609138 |
| RMSPE | 4.16618471 | 4.14504404 | 4.25559821 | 4.99726836 | 4.99754748 | 4.99607198 | 4.12946379 | 4.10726508 | 4.22327970 |
| MSPE% | 0.06271946 | 0.06212193 | 0.06528030 | 0.08875877 | 0.08882456 | 0.08847681 | 0.06167789 | 0.06105383 | 0.06435244 |
| RMSPE% | 0.25043854 | 0.24924272 | 0.25550009 | 0.29792410 | 0.29803450 | 0.29745052 | 0.24835033 | 0.24709072 | 0.25367782 |
| MAPE | 3.23188170 | 3.21607107 | 3.29964151 | 4.06021896 | 4.05973676 | 4.06228551 | 3.19874821 | 3.18232445 | 3.26913575 |
| MAPE% | 0.01160292 | 0.01155222 | 0.01182024 | 0.01437071 | 0.01437764 | 0.01434101 | 0.01149221 | 0.01143920 | 0.01171941 |
結論
相較於實驗 2 ，實驗 3 先使用 SSSDS4 再使用 autoFRK ，使填補結果較為保守。
結語
運行環境
- 本機作業系統：Windows 11 24H2
- 程式語言：Python 3.12.9
- 計算平臺：財團法人國家實驗研究院國家高速網路與計算中心臺灣 AI 雲
- 作業系統：Ubuntu
- Miniconda
- GPU：NVIDIA Tesla V100 32GB GPU
- CUDA 12.8 driver
- 程式語言：Python 3.10.16 for Linux
延伸學習
- 我測試此項目的 Github 儲存庫 。
參考資料
Juan Lopez Alcaraz 、 Nils Strodthoff（2022）。Diffusion-based time series imputation and forecasting with structured state space models。Transactions on Machine Learning Research。參考自 https://openreview.net/forum?id=hHiIbk7ApW
SSSD（2022）。GitHub。參考自 https://github.com/AI4HealthUOL/SSSD
SSSD_CP（2024）。GitHub。參考自 https://github.com/egpivo/SSSD_CP
近兩年小時值查詢（無日期）。環境部 - 空氣品質監測網。參考自 https://airtw.moenv.gov.tw/CHT/Query/InsValue.aspx
