GTEE: A Global Timestamp Encoding Enhanced Method for Robust Time Series Imputation in Complex Missing Scenarios

This repository provides the official implementation of the paper:

GTEE: A Global Timestamp Encoding Enhanced Method for Robust Time Series Imputation in Complex Missing Scenarios

Requirements

We recommend creating the environment with Conda:

conda env create -f requirements.yaml
conda activate TSI

If you use NNI for automatic model hyperparameter tuning and encounter the error: ModuleNotFoundError: No module named 'pkg_resources'

Please downgrade the setuptools version with the following command:

pip install setuptools==68.2.2

Dataset Preparation

The missing-value datasets are constructed from raw time-series data through sample splitting, normalization, and artificial masking.

This project includes scripts for dataset generation with synchronized timestamp extraction and normalization.

For example, to generate the missing dataset for ETTh1, follow the steps below.

Generated dataset structure

The following directory structure covers three missingness patterns: Point 05, Block 05, and Subseq 05.

data/
    generated_datasets/
        ETTh1/
            ett_rate05_step48_point/
                train.h5
                test.h5
                val.h5
                TS_prior.h5
            ett_rate03_step48_block_blocklen6/
                train.h5
                test.h5
                val.h5
                TS_prior.h5
            ett_rate05_step48_subseq_seqlen36/
                train.h5
                test.h5
                val.h5
                TS_prior.h5

Generation Command

cd <your_project_path>
PYTHONPATH=. python -u data/data_factory/dataset_generating_ETTh1.py

Note: Replace <your_project_path> with the root directory of your local project.

Quick Test

To perform imputation on ETTh1 with SAITS_GTEE, first change to the project root directory, then run:

sh scripts/SAITS_GTEE/run_single.sh

The training process and evaluation results for the three missing patterns, Point 05, Block 05, and Subseq 05, can be found in the three log files under logs/imputation.

Hyper-parameter Optimization (Stage 1)

In Stage 1, we use the NNI hyper-parameter optimization framework to tune the Timestamp-Dominated Imputation Model (project alias: Pure_GTEE).

Using ETTh1 as an example, the procedure is as follows:

1. Open scripts/Pure_GTEE/run_nni.sh, update line 3 from export PYTHONPATH=/root/autodl-tmp/GTEE to your actual project path, and save the file.

2. Open auto_tuning/models_tuning_space_setting/Pure_GTEE/ETT_h1/Pure_GTEE_searching_config.yml, update line 8 from trialCodeDirectory: /root/autodl-tmp/GTEE to your actual project path, and save the file.

3. Run the following command in the terminal:

   sh scripts/Pure_GTEE/run_nni.sh

   This will launch NNI tuning according to the configuration file Pure_GTEE_searching_config.yml. For more advanced or customized tuning settings, please refer to the official NNI documentation: https://nni.readthedocs.io/en/stable/index.html.

4. Open the NNI web UI in your browser (default: http://127.0.0.1:8080). After tuning is complete, copy the best-performing hyper-parameter set, selected by the lowest validation loss, into configs/hpo_results/ett_h1.py under ETT_h1["Pure_GTEE"], for example:

   ETT_h1 = {
       "Pure_GTEE": {
           "ts_main_freq_num": 5,
           "ts_main_freq_solvers": ["ACF", "Burg_roots"],
           "ts_solvers_combine": False,
           "ts_sampling_num": 1,
           "ts_sampling_length": 256,
           "ts_scaling_factor": 0.6067428453718272,
           "ts_freqs_merge_value": 0.022596998462003477,
           "ts_refining_mode": "median",
           # ... other hyper-parameters
       },
       ...
   }

5. Stop the NNI process with:

   nnictl stop --all

6. Based on the Stage 1 hyper-parameters, compute the final main frequency mining result $\bm{F}_{main}$ using:

   python -u calculate_main_freqs.py --model Pure_GTEE --dataset ETT_h1 --dataset_fold_path data/generated_datasets/ETTh1/ett_rate05_step48_point

   The resulting $\bm{F}_{main}$ file will be saved to prior_knowledge/Pure_GTEE_ETTh1/main_freqs.h5.

Note: Under the experimental setting used in the paper, Stage 1 tuning is performed only on the dataset generated under the Point 05 missing pattern, and the resulting hyper-parameters are shared across all missing patterns.

Hyper-parameter Optimization (Stage 2)

Stage 2 performs joint hyper-parameter tuning for the timestamp-local joint imputation framework.

Using SAITS_GTEE on ETTh1 as an example, the procedure is as follows:

1. Since Stage 2 tuning does not modify the Stage 1 hyper-parameters, you can directly reuse the Stage 1 $\bm{F}_{main}$ to accelerate optimization. For example, on Linux:

   mkdir -p ./prior_knowledge/SAITS_GTEE_ETTh1 && cp ./prior_knowledge/Pure_GTEE_ETTh1/main_freqs.h5 ./prior_knowledge/SAITS_GTEE_ETTh1/

2. Open scripts/SAITS_GTEE/run_nni.sh, update line 3 from export PYTHONPATH=/root/autodl-tmp/GTEE to your actual project path, and save the file.

3. Open auto_tuning/models_tuning_space_setting/SAITS_GTEE/ETT_h1/SAITS_GTEE_searching_config.yml, update line 8 from trialCodeDirectory: /root/autodl-tmp/GTEE to your actual project path, and save the file.

4. Run the following command in the terminal:

   sh scripts/SAITS_GTEE/run_nni.sh

   This will launch NNI tuning according to the configuration file SAITS_GTEE_searching_config.yml.

5. After tuning is complete, copy the best-performing hyper-parameter set, selected by the lowest validation loss, into configs/hpo_results/ett_h1.py under ETT_h1["SAITS_GTEE"]. Note that all hyper-parameters prefixed with ts_ must remain identical to the Stage 1 settings rather than using the Stage 2 tuning results. For example:

   ETT_h1 = {
       "Pure_GTEE": {
           "ts_main_freq_num": 5,
           "ts_main_freq_solvers": ["ACF", "Burg_roots"],
           "ts_solvers_combine": False,
           "ts_sampling_num": 1,
           "ts_sampling_length": 256,
           "ts_scaling_factor": 0.6067428453718272,
           "ts_freqs_merge_value": 0.022596998462003477,
           "ts_refining_mode": "median",
           # ... other hyper-parameters
       },
       "SAITS_GTEE": {
           # Stage 1 settings
           "ts_main_freq_num": 5,
           "ts_main_freq_solvers": ["ACF", "Burg_roots"],
           "ts_solvers_combine": False,
           "ts_sampling_num": 1,
           "ts_sampling_length": 256,
           "ts_scaling_factor": 0.6067428453718272,
           "ts_freqs_merge_value": 0.022596998462003477,
           "ts_refining_mode": "median",
   
           # Stage 2 tuned hyper-parameters
           "TS_ch_embedding_dim": 32,
           "TS_hidden_dim": 8,
           "TS_reg_weight": 0.014487810264286003,
           "TS_fusion_conv_kernels": [5, 7, 9],
           "TS_fusion_hidden_dim": 8,
           "TS_loss_weight": 0.004579005749274037,
           "TS_dropout": 0.2,
           "n_steps": 48,
           "n_features": 7,
           "epochs": 100,
           "patience": 10,
           "n_layers": 2,
           "d_model": 64,
           "d_ffn": 256,
           "n_heads": 1,
           "d_k": 128,
           "d_v": 256,
           "dropout": 0,
           "attn_dropout": 0.2,
           "lr": 0.0011752356815298876
       },
       ...
   }

6. Stop the NNI process with:

   nnictl stop --all

7. Refer to the Quick Test section for subsequent analysis and evaluation.

Acknowledgements

The overall project framework is built upon PyPOTS: https://github.com/WenjieDu/PyPOTS

The hyper-parameter search spaces for the baseline models are primarily based on Awesome_Imputation: https://github.com/WenjieDu/Awesome_Imputation

We sincerely thank the authors, especially Wenjie Du and collaborators, for their valuable contributions.

Citation

If this work is helpful for your research, please cite our paper:

@article{feng2026gtee,
  title={GTEE: A global timestamp encoding enhanced method for robust time series imputation in complex missing scenarios},
  author={Feng, Yuan and Liu, Zhenyu and Liu, Hui and Jiang, Xishan and Liu, Xin and Sun, Jiacheng and Tan, Jianrong},
  journal={Information Fusion},
  pages={104220},
  year={2026},
  publisher={Elsevier}
}