AI BriefingLG AI Research: Adaptive Information Routing Technology for Multimodal Time Series Forecasting
Key point
LG AI Research has proposed 'AIR,' a multimodal time series forecasting technique that uses text information to control the behavior of time series models.
Details
Recently at NeurIPS 2024, as the use of LLM and large models in the time series forecasting field has increased, the TSALM (Time Series in the Age of Large Models) workshop was held. LG AI Research's Data Intelligence (DI) Lab presented research on multimodal time series forecasting at this workshop.
Existing multimodal time series forecasting has been limited to either directly utilizing language models or separately using text embeddings. However, while time series data is structural and specific, text has global and ambiguous characteristics, which limits simply combining the two types of data.
To address this, LG AI Research proposed Adaptive Information Routing (AIR) technology. This technology uses text data as a controller for the time series model, regulating the Information Pathway through which information flows and mixes within the time series model.
The specific mechanism works as follows:
- Information Routing module: Integrates text embeddings to generate Latent weights between FC Layers.
- Behavior control: Adjusts the strength of Temporal Mixing and Featural Mixing in the time series model according to text information.
- Base model application: In this research, performance was validated based on TSMixer, a time series forecasting model composed of FC Layers.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.