Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Optimal Transport for Time Series Imputation
Authors: Hao Wang, zhengnan li, Haoxuan Li, Xu Chen, Mingming Gong, BinChen, Zhichao Chen
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that PSW-I effectively accommodates temporal patterns and non-stationarity, outperforming prevailing time-series imputation methods. Code is available at https://github.com/FMLYD/PSW-I. ... 4 EMPIRICAL INVESTIGATION ... Table 1 compares the performance of PSW-I with baseline methods, averaged over four missing ratios: 0.1, 0.3, 0.5, and 0.7. ... Table 2 presents an ablation study dissecting the contributions of the Pairwise Spectrum Distance (PSD) and Selective Matching Regularization (SMR)... |
| Researcher Affiliation | Academia | Hao Wang1 Zhengnan Li2 Haoxuan Li3 Xu Chen4 Mingming Gong5,6 Bin Chen7 Zhichao Chen1, 1Zhejiang University 2Communication University of China 3Peking University 4Renmin University of China 5The University of Melbourne 6Mohamed bin Zayed University of Artificial Intelligence 7University of Chinese Academy of Sciences EMAIL EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 The imputation workflow of PSW-I. Input: X(obs): the incomplete data; M: the mask matrix. Parameter: κ: the matching strength; ε the entropic regularization strength; ℓmax: the max number of iterations to solve PSW; Tmax: the max number of epochs; B: the batch size; η: the update rate. Output: X(imp): the imputed dataset. 1: t 0. 2: Xt Pre-Impute(X(obs)) 3: while t < Tmax do 4: α, β, Mα, Mβ Sample(Xt, M; B). 5: α(F) F(α), β(F) F(β) 6: D(F) ij = α(F) i β(F) j 1, 1 i, j B 7: P Algorithm1(D(F); ε, κ, ℓmax). 8: α α η αP M. 9: β β η βP M. 10: Xt+1 Update(α , β , Xt). 11: t t + 1. 12: if Xt Xt 1 F < 1e 4 then 13: break. 14: X(imp) Xt. |
| Open Source Code | Yes | Code is available at https://github.com/FMLYD/PSW-I. |
| Open Datasets | Yes | Datasets: Experiments are performed on public time-series datasets (Wu et al., 2021; Liu et al., 2024), including ETT, Electricity, Traffic, Weather, PEMS03, Illness, and Exchange. |
| Dataset Splits | Yes | We exclude 5% of the indices from the training data to form the validation set. |
| Hardware Specification | Yes | The experiments are conducted on a platform with two Intel(R) Xeon(R) Platinum 8383C CPUs @ 2.70GHz and a NVIDIA Ge Force RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions "The Adam optimizer, known for its adaptive update rate and effective convergence, is employed for training, with an update rate η = 0.01." but does not provide specific version numbers for Adam, any programming languages, or other libraries used for the implementation. |
| Experiment Setup | Yes | To ensure consistency in experimental conditions, the batch size B is fixed at 256. The Adam optimizer... is employed for training, with an update rate η = 0.01. We exclude 5% of the indices from the training data to form the validation set. The key hyperparameters involved in PSW-I are tuned to minimize the MSE on the validation set. The patch size is tuned within {24, 36, 48}; the matching strength is tuned within {1, 10, 100, 1000}. The experiments are conducted on a platform with two Intel(R) Xeon(R) Platinum 8383C CPUs @ 2.70GHz and a NVIDIA Ge Force RTX 4090 GPU. Performance is evaluated using modified mean absolute error (MAE) and mean squared error (MSE)... We set Tmax = 200 and ℓmax = 1, 000 to ensure convergence, applying early stopping on the validation dataset with a patience of 10. |