Reproduce “Towards Cross-Modality Modeling for Time Series Analytics: A Survey in the LLM Era” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chenxi Liu, Shaowen Zhou, Qianxiong Xu, Hao Miao, Cheng Long, Ziyue Li, Rui Zhao
The authors survey methods for connecting text with time series and test selected combinations in forecasting. They classify text as numerical, statistical, contextual, or word-token embeddings, and distinguish alignment from fusion. In climate forecasting, numerical-prompt retrieval reports MSE 0.287 versus 0.371 for a time-series-only model, a 22.6% reduction; results vary by text type and domain.
Key insights
- Contextual prompts are the most frequently used textual data type (56%) in cross-modality time series modeling, followed by word token embeddings (19%), statistical prompts (18%), and numerical prompts (7%).
- Fusion strategies (54%) are slightly more common than alignment strategies (46%) for integrating textual and time series data.
- A comprehensive framework for cross-modality time series modeling integrates pre-trained LLMs, projection layers, time series processing, cross-modality integration, and task-specific adapters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.