One Batch Is Enough: A Unified Dataset Condensation Framework for General Time Series Analysis

May 17, 2026 · View on GitHub

In this paper, we propose UniTSC, a unified dataset condensation framework for general time series analysis.

The framework of UniTSC consists of three stages: (i) Task-Adaptive Synthetic Initialization: We initialize latent parameters $\mathcal{X}_{syn}$ by utilizing task-adaptive priors. (ii) Multi-View Hybrid Encoder: A shared backbone extracts holistic representations by integrating temporal, spectral, and topological views. (iii) Tri-Space Alignment: We optimize $\mathcal{X}_{syn}$ by minimizing the divergence between student and expert models across parameter, frequency, and topology spaces, yielding the condensed dataset $\mathcal{S}$ .

Extensive experiments show that UniTSC retains up to 97.9% of downstream performance using as little as 0.01% of the original training data, revealing that a single batch-equivalent budget ( $\textless$ 128 samples) is sufficient to capture the essential dynamics of complex time-series data.

Overview of UniTSC

Environment

Python.version = 3.9.21
PyTorch.version = 2.5.1
Other dependencies are listed in requirements.txt.

Datasets

We evaluate UniTSC on across four mainstream time-series tasks on 22 datasets, with detailed statistics summarized in the table below.

Tasks	Datasets	Dim	Length	Train	Test	Information
Forecasting	ETTm1, ETTm2	7	96	34,369	11,425	Electricity
	ETTh1, ETTh2	7	96	8,449	2,785	Electricity
	Weather	21	96	36,696	10,444	Weather
	Electricity	321	96	18,221	5,165	Electricity
	Traffic	862	96	12,089	3,413	Transportation
Imputation	ETTm1, ETTm2	7	96	34,369	11,425	Electricity
	ETTh1, ETTh2	7	96	8,449	2,785	Electricity
	Weather	21	96	36,696	10,444	Weather
	Electricity	321	96	18,221	5,165	Electricity
Classification (UEA)	EthanolConcentration	3	1751	261	263	Alcohol
	FaceDetection	144	62	5890	3524	Face
	Handwriting	3	152	150	850	Handwriting
	Heartbeat	61	405	204	205	Heart Beat
	JapaneseVowels	12	29	270	370	Voice
	PEMS-SF	963	144	267	173	Transport
	SelfRegulationSCP1	6	896	268	293	Health
	SelfRegulationSCP2	7	1152	200	180	Health
	SpokenArabicDigits	13	93	6599	2199	Voice
	UWaveGestureLibrary	3	315	120	320	Gesture
Anomaly Detection	SMAP	25	100	135,183	427,617	Spacecraft
	PSM	25	100	132,481	87,841	Server
	SMD	38	100	708,405	708,420	Server
	SWaT	51	100	495,000	449,919	Infrastructure
	MSL	55	100	73,729	58,317	Spacecraft

Please refer to TSLib for the download links of each dataset.

Code Structure

UniTSC/
├── README.md                                   # Official README
├── requirements.txt                            # pip dependency list for quick environment setup
├── run.py                                      # Unified entry that parses args and dispatches tasks
├── framework.jpg                               # README figure
├── buffer_forecast.py                          # Forecasting experts training
├── buffer_imputation.py                        # Imputation experts training
├── buffer_cls.py                               # Classification experts training
├── buffer_ad.py                                # Anomaly detection experts training
├── distill_forecast.py                         # Forecasting dataset condensation
├── distill_imputation.py                       # Imputation dataset condensation
├── distill_cls.py                              # Classification dataset condensation
├── distill_ad.py                               # Anomaly detection dataset condensation
├── data_provider/                              # Dataset loaders and splits
│   ├── data_factory.py                         # Chooses the proper DataLoader per task
│   ├── data_loader.py                          # Generic TS reader with sliding-window logic
│   └── uea.py                                  # Parsers for UEA formats
├── models/                                     # All model implementations
│   ├── backbone.py                             # Main forecasting models
│   ├── layers.py                               # Reusable attention / conv / embedding blocks
│   └── embed.py                                # Shared primitives
├── tools/                                      # Utility toolbox
│   ├── augmentation.py                         # Augmentations for classification / detection
│   ├── data_generator.py                       # Task-adaptive synthetic initialization
│   ├── dtw.py                                  # DTW calculator
│   ├── exp.py                                  # Training utilities for four tasks
│   ├── masking.py                              # Imputation masking
│   ├── metrics.py                              # MSE / MAE / DTW and other metrics
│   ├── reparam_module.py                       # Reparameterizes modules
│   ├── timefeatures.py                         # Time feature generators
│   └── utils.py                                # General helpers such as EarlyStopping
├── scripts_buffer/                             # Bash scripts for generate expert buffer
│   ├── forecast/                               # Forecasting scripts
│   ├── imputation/                             # Imputation scripts
│   ├── classification/                         # Classification scripts
│   └── anomaly_detection/                      # Anomaly detection scripts
└── scripts_distill/                            # Bash scripts for dataset condensation
    ├── forecast/                               # Forecasting scripts
    ├── imputation/                             # Imputation scripts
    ├── classification/                         # Classification scripts
    └── anomaly_detection/                      # Anomaly detection scripts

Quick Start

bash scripts_buffer/forecast/etth1.sh
bash scripts_distill/forecast/etth1.sh