Unicorn: Universal Correlation Network for High-Dimensional Time Series Forecasting
A new framework called Unicorn (Universal Correlation Network) has been developed by researchers for efficient pretraining across multiple datasets involving high-dimensional time series. This model tackles the challenge posed by channel-independent models, which scale effectively but overlook inter-channel relationships, and channel-dependent models, which are more expressive yet have difficulty generalizing across diverse datasets. By employing a latent prototype codebook, Unicorn separates correlation modeling from specific channel identities, allowing it to map various channels into a unified latent space and learn reusable, identity-agnostic interaction patterns. Experimental results indicate that Unicorn significantly surpasses leading forecasting architectures, especially in few-shot transfer situations.
Key facts
- Unicorn is a framework for scalable, multi-dataset pretraining on high-dimensional time series.
- It uses a latent prototype codebook to decouple correlation modeling from specific channel identities.
- The model projects heterogeneous channels into a shared latent space.
- It learns identity-agnostic, reusable interaction patterns that transfer across domains.
- Unicorn outperforms state-of-the-art forecasting architectures.
- It is particularly effective in few-shot transfer scenarios.
- The approach bridges the gap between channel-independent and channel-dependent models.
- The paper is available on arXiv under ID 2605.30376.
Entities
Institutions
- arXiv