🚀 MixLinear

April 24, 2026 · View on GitHub

Welcome to the official repository of MixLinear, the ICLR 2026 paper:

MixLinear: Extreme Low-Resource Multivariate Time Series Forecasting with 0.1K Parameters
Aitian Ma, Dongsheng Luo, Mo Sha
¹Knight Foundation School of Computing and Information Sciences, Florida International University

📄 Paper • 🎥 Presentation & Slides • 👤 Profile • 🎓 Google Scholar

🌐 Additional Materials

🇨🇳 中文解读
📺 ICLR 预讲会（55:00）
https://www.bilibili.com/video/BV1XGoaBXEmr/
📝 CSDN 博客
https://blog.csdn.net/AITIME_HY/article/details/159808163

🎉 Published as a conference paper at ICLR 2026

🔥 TL;DR

MixLinear is a dual-domain forecasting model that achieves state-of-the-art long-term time series forecasting performance using just 0.1K parameters — up to 98% fewer parameters and 3.2× faster inference than existing efficient baselines.

By processing local trends in the time domain and global trends in the frequency domain, MixLinear reduces model complexity from O(n²) to O(n) while preserving accuracy, enabling deployment on edge devices, embedded systems, and low-resource environments.

✨ Why MixLinear?

Feature	MixLinear
Parameters	~0.1K
Time Complexity	O(n log n)
Space Complexity	O(n)
Architecture	Time + Frequency Dual Pathway
Target Scenarios	Edge / IoT / Small Data / Real-Time

🏆 Up to 16.2% MSE improvement over SparseTSF
⚡ Up to 3.2× faster inference
📦 81–98% parameter reduction vs. lightweight baselines
🌍 Strong cross-domain generalization

🧠 Core Idea

MixLinear is motivated by a key observation about time series structure:

Local trends are best modeled in the time domain, while global trends are sparse in the frequency domain.

Instead of forcing a single architecture to model both, MixLinear introduces a dual-domain framework:

🔹 Segment-Based Trend Extraction (Time Domain)

Local temporal patterns are captured using factorized linear transformations that disentangle:

Intra-segment correlations (local shape, slopes)
Inter-segment correlations (long-range drift)

This reduces dense forecasting layers from O(n²) parameters to O(n).

🔹 Adaptive Low-Rank Spectral Filtering (Frequency Domain)

Global trends are processed via learnable rank-constrained complex filters, compressing spectral representations into an ultra-low-dimensional latent space while preserving dominant frequency modes.

🔹 Unified Forecasting

Final prediction combines both pathways:

Y = F_segment(X) + F_frequency(X)

This additive fusion preserves domain-specific representations while enabling joint optimization, achieving an unprecedented efficiency–accuracy tradeoff.

Up to 3.2× speedup in low-dimensional datasets
Up to 2.6× speedup in high-dimensional datasets

⚡ Getting Started

Environment Setup

conda create -n MixLinear python=3.8
conda activate MixLinear
pip install -r requirements.txt

Data Preparation

Download datasets from Autoformer:

https://drive.google.com/drive/folders/1ZOYpTUa82_jCcxIdTmyr0LXQfvaM9vIy

Place CSV files in:

./dataset

Example:

./dataset/ETTh1.csv

Training

Run single dataset:

sh scripts/MixLinear/etth1.sh

🧪 Using MixLinear on Your Own Data

MixLinear does not require strong periodicity assumptions and generalizes well across domains.

Recommended hyperparameters:

segment_len — local granularity
rank — spectral compression level (default: 2)

The model is stable across wide ranges of both.

📚 Citation

@inproceedings{ma2026mixlinear,
  title={MixLinear: Extreme Low-Resource Multivariate Time Series Forecasting with 0.1K Parameters},
  author={Ma, Aitian and Luo, Dongsheng and Sha, Mo},
  booktitle={International Conference on Learning Representations (ICLR)},
  year={2026}
}

📬 Contact

Aitian Ma — ama003@fiu.edu
Dongsheng Luo — dluo@fiu.edu
Mo Sha — msha@fiu.edu

Knight Foundation School of Computing and Information Sciences
Florida International University

🙏 Acknowledgement

We thank the following repositories for datasets and baseline implementations: