LLM-TS Integrator: Integrating LLM for Enhanced Time Series Modeling
Previous Works
LLM for TS modeling
Previous works: Primiarility position LLM as the predictive backbone
\(\rightarrow\) Problematic!! omit the modeling within traditional TS models (e.g., periodicity)
LLM-TS Integrator
Effectively integrates the capabilities of (1) LLMs into (2) traditional TS modeling
Two modules
- (1) Mutual information (MI) module
- Maximizing MI btw ..
- (1) traditional model’s representation
- (2) LLM’s textual representation
- Maximizing MI btw ..
- (2) Sample reweighting module
- Recognize that samples vary in importance for two losses
- a) prediction loss
- b) MI loss
- Dynamically optimize these weight via bi-level optimization
- Recognize that samples vary in importance for two losses