------ **Self-Supervised Learning of Time Series Representation via Diffusion Process and Imputation-Interpolation-Forecasting Mask**

Usage • Examples • Checkpoints • Processed Datasets • Citation

:triangularflagon_post:News (2024.06) We have added scripts to run forecasting experiments on the Electricity dataset provided by TimesNet.

:triangularflagon_post:News (2024.05) This work is accepted at KDD 2024, to main research track.

Time Series Diffusion Embedding (TSDE), bridges the gap of leveraging diffusion models for Time Series Representation Learning (TSRL) as the first diffusion-based SSL TSRL approach. TSDE segments time series data into observed and masked parts using an Imputation-Interpolation-Forecasting (IIF) mask. It applies a trainable embedding function, featuring dual-orthogonal Transformer encoders with a crossover mechanism, to the observed part. We train a reverse diffusion process conditioned on the embeddings, designed to predict noise added to the masked part. Extensive experiments demonstrate TSDE’s superiority in imputation, interpolation, forecasting, anomaly detection, classification, and clustering.

Usage

We recommend to start with installing dependencies in an virtual environment. conda create --name tsde python=3.11 -y conda activate tsde pip install -r requirements.txt

Datasets

Download public datasets used in our experiments:

python src/utils/download_data.py [dataset-name] Options of [dataset-name]: physio, pm25, electricity, solar, traffic, taxi, wiki, msl, smd, smap, swat and psm

Imputation

To run the imputation experiments on PhysioNet dataset: python src/experiments/train_test_imputation.py --device [device] --dataset PhysioNet --physionet_classification True --testmissingratio [test_missing_ratio] In our experiments, we set [testmissingratio] to: 0.1, 0.5, and 0.9

To run the imputation experiments on PM2.5 dataset, first set train-epochs to 1500, and finetuning-epochs to 100 in src/config/base.yaml, and run the following command: python src/experiments/train_test_imputation.py --device [device] --dataset Pm25

Interpolation

To run the imputation experiments on PhysioNet dataset: python src/experiments/train_test_interpolation.py --device [device] --dataset PhysioNet --physionet_classification True --testmissingratio [test_missing_ratio] In our experiments, we set [testmissingratio] to: 0.1, 0.5, and 0.9

Forecasting

Please first set the number of pretraining and finetuning epochs for each dataset in src/config/baseforecasting.yaml, and set the number of features for subsampling training in the TSDEforecasting model in src/model/mainmodel.py. Run the following command: ``` python src/experiments/traintestforecasting.py --dataset [dataset-name] --device [device] --samplefeat ``` Please remove the flag --sample_feat to disable the sub-sampling of features during training.

Anomaly Detection

Please first set the number of features, the number of pretraining and finetuning epochs for each dataset in src/config/basead.yaml. Run the following command: ``` python src/experiments/traintestanomalydetection.py --dataset [dataset-name] --device [device] --seed [seed] --anomalyratio [anomalyratio] ```

The values of [dataset-name], [seed] and [anomaly_ratio] used in our experiments are available in our paper.

Classification on PhysioNet

Run the following command: python src/experiments/train_test_classification.py --seed [seed] --device [device] --testmissingratio [test_missing_ratio]

Benchmarking TSDE against Time Series Library models

1. Upload the Electricity dataset following their guidelines available here. The dataset folder should be in the root directory.
2. Run the following command: python src/experiments/train_test_tslib_forecasting.py --device [device] --pred_length [pred_length] --hist_length [hist_length] The values of [predlength], and [histlength] used in our experiments are available in our paper.

Examples

Examples of imputation, interpolation and forecasting

Example of clustering

Example of embedding visualization

Checkpoints

To run the evaluation using a specific checkpoint, follow the instructions below. Ensure your environment is set up correctly for running and the datasets are downloaded first.

Imputation

1. Download the checkpoints: Access and download the required checkpoints from here. Place the content of this folder under [root_dir]/save.
2. Run the evaluation command by setting [path_to_checkpoint_folder] accordingly. The path should excludes '[rootdir]/save' and 'model.pth'. ``` python src/experiments/traintestimputation.py --device [device] --dataset PhysioNet --physionetclassification True --testmissingratio 0.1 --modelfolder [pathtocheckpointfolder] --run [runnumber] python src/experiments/traintestimputation.py --device [device] --dataset Pm25 --modelfolder [pathtocheckpointfolder] --run [runnumber] ```

Interpolation

1. Download the checkpoints: Access and download the required checkpoints from here. Place the content of this folder under [root_dir]/save.
2. Run the evaluation command by setting [path_to_checkpoint_folder] accordingly. The path should excludes '[rootdir]/save' and 'model.pth'. ``` python src/experiments/traintestinterpolation.py --device [device] --dataset PhysioNet --physionetclassification True --testmissingratio 0.1 --modelfolder [pathtocheckpointfolder] --run [runnumber] ``` ### Forecasting
3. Download the checkpoints: Access and download the required checkpoints from here. Place the content of this folder under [root_dir]/save.
4. Run the evaluation command by setting [path_to_checkpoint_folder] accordingly. The path should excludes '[rootdir]/save' and 'model.pth'. ``` python src/experiments/traintestforecasting.py --device [device] --dataset [dataset-name] --modelfolder [pathtocheckpointfolder] --run [run_number] ```

Anomaly Detection

1. Download the checkpoints: Access and download the required checkpoints from here. Place the content of this folder under [root_dir]/save.
2. Run the evaluation command by setting [path_to_checkpoint_folder] accordingly. The path should excludes '[rootdir]/save' and 'model.pth'. ``` python src/experiments/traintestanomalydetection.py --device [device] --dataset [dataset-name] --modelfolder [pathtocheckpointfolder] --run [runnumber] --disable_finetune ```

Classification

1. Download the checkpoints: Access and download the required checkpoints from here. Place the content of this folder under [root_dir]/save.
2. Run the evaluation command by setting [path_to_checkpoint_folder] accordingly. The path should includes '[rootdir]/save' and excludes 'model.pth'. ``` python src/experiments/traintestclassification.py --device [device] --modelfolder [pathtocheckpointfolder] --run [runnumber] --disablefinetune ## Citation @article{senane2024tsde, title={{Self-Supervised Learning of Time Series Representation via Diffusion Process and Imputation-Interpolation-Forecasting Mask}}, author={Senane, Zineb and Cao, Lele and Buchner, Valentin Leonhard and Tashiro, Yusuke and You, Lei and Herman, Pawel and Nordahl, Mats and Tu, Ruibo and von Ehrenheim, Vilhelm}, year={2024}, eprint={2405.05959}, archivePrefix={arXiv}, primaryClass={cs.LG} } ```