TF-LLM: Enhanced time series analysis with time-frequency large language models.

Area of Science:

• Artificial Intelligence
• Data Science
• Signal Processing

Background:

• Large language models (LLMs) show potential in time series analysis, particularly for complex symbolic sequences.
• Effectively leveraging LLMs' contextual reasoning for time series data presents a significant challenge.
• Existing methods struggle to fully utilize LLMs for diverse time series tasks.

Purpose of the Study:

• To propose the TF-LLM framework for advanced time series analysis tasks.
• To enhance LLM capabilities in time series forecasting, classification, imputation, and anomaly detection.
• To improve the understanding and handling of complex time series data using LLMs.

Main Methods:

• The TF-LLM framework integrates time and frequency domain representations.
• Frequency representations simplify data complexity and capture periodic patterns.
• Time modeling addresses fine-grained dependencies and non-stationarity.
• Prompt learning is employed to enrich input context and improve LLM understanding.

Main Results:

• Extensive experiments were conducted on seven benchmark datasets.
• TF-LLM demonstrated superior performance across multiple time series tasks.
• The proposed framework outperformed several existing state-of-the-art methods.

Conclusions:

• The TF-LLM framework effectively leverages LLMs for complex time series analysis.
• Integrating time and frequency domains enhances performance in forecasting, classification, imputation, and anomaly detection.
• Prompt learning further boosts the reasoning capabilities of LLMs for time series data.