Unicorn framework cracks high-dimensional time series forecasting with universal correlation modeling
New model scales to multivariate data by learning reusable correlation patterns across domains.
Modern time series forecasting faces a fundamental trade-off: channel-independent models scale well with data volume but ignore inter-channel dependencies, while channel-dependent models are expressive but struggle to generalize across heterogeneous datasets. To bridge this gap, researchers from Shanghai Jiao Tong University propose Unicorn (Universal Correlation Network), a framework designed for scalable, multi-dataset pretraining on high-dimensional time series.
Unicorn's key innovation is a latent prototype codebook that projects heterogeneous channels into a shared latent space, learning identity-agnostic, reusable interaction patterns that transfer across domains with diverse dimensionalities and semantics. This allows the model to capture complex correlations without being tied to specific channel labels. Extensive experiments show Unicorn significantly outperforms state-of-the-art forecasting architectures, particularly in few-shot transfer scenarios, offering a scalable path toward multivariate time series foundation models.
- Unicorn uses a latent prototype codebook to decouple correlation modeling from specific channel identities.
- The framework enables scalable multi-dataset pretraining on high-dimensional time series with diverse dimensionalities and semantics.
- Outperforms state-of-the-art forecasters, especially in few-shot transfer, moving toward foundation models for time series.
Why It Matters
Unicorn unlocks scalable multivariate time series forecasting, crucial for finance, weather, and IoT analytics.