![[2602.21043] T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.21043] T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation
Summary
The paper presents T1, a CNN-Transformer hybrid model for robust multivariate time-series imputation, achieving state-of-the-art performance by utilizing a novel Channel-Head Binding mechanism.
Why It Matters
Accurate imputation of missing values in multivariate time series is crucial for various applications, including finance and healthcare. The T1 model addresses common shortcomings in existing methods, offering significant improvements in performance, especially under extreme missing data conditions. This advancement could enhance decision-making processes reliant on time-series data.
Key Takeaways
- T1 model significantly reduces mean squared error (MSE) by 46% on average compared to existing methods.
- Introduces Channel-Head Binding for effective information transfer between CNN channels and attention heads.
- Demonstrates strong performance even with up to 70% missing data.
- Generalizes well to unseen missing patterns without requiring retraining.
- Utilizes a consistent hyperparameter configuration across multiple datasets.
Computer Science > Machine Learning arXiv:2602.21043 (cs) [Submitted on 24 Feb 2026] Title:T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation Authors:Dongik Park, Hyunwoo Ryu, Suahn Bae, Keondo Park, Hyung-Sin Kim View a PDF of the paper titled T1: One-to-One Channel-Head Binding for Multivariate Time-Series Imputation, by Dongik Park and 4 other authors View PDF HTML (experimental) Abstract:Imputing missing values in multivariate time series remains challenging, especially under diverse missing patterns and heavy missingness. Existing methods suffer from suboptimal performance as corrupted temporal features hinder effective cross-variable information transfer, amplifying reconstruction errors. Robust imputation requires both extracting temporal patterns from sparse observations within each variable and selectively transferring information across variables--yet current approaches excel at one while compromising the other. We introduce T1 (Time series imputation with 1-to-1 channel-head binding), a CNN-Transformer hybrid architecture that achieves robust imputation through Channel-Head Binding--a mechanism creating one-to-one correspondence between CNN channels and attention heads. This design enables selective information transfer: when missingness corrupts certain temporal patterns, their corresponding attention pathways adaptively down-weight based on remaining observable patterns while preserving reliable cross-variable connections through unaff...
