Abstract:In this paper, we propose a novel framework that leverages large language models (LLMs) for predicting missing values in time-varying graph signals by exploiting spatial and temporal smoothness. We leverage the power of LLM to achieve a message-passing scheme. For each missing node, its neighbors and previous estimates are fed into and processed by LLM to infer the missing observations. Tested on the task of the online prediction of wind-speed graph signals, our model outperforms online graph filtering algorithms in terms of accuracy, demonstrating the potential of LLMs in effectively addressing partially observed signals in graphs.

What problem does this paper attempt to address?

This paper attempts to address the problem of predicting missing values in time-varying graph signals. Specifically, the authors propose a new framework based on large-scale language models (LLM) that leverages spatial and temporal smoothness to predict missing values in graph nodes. By inputting the neighboring nodes and previous estimates of each missing node into the LLM for processing, the missing observations are inferred. Experimental results show that this method achieves higher accuracy than traditional online graph filtering algorithms in the online prediction task of wind speed graph signals. ### Main Contributions: 1. **Proposed New Framework**: Designed a new framework utilizing the powerful learning capabilities of LLM to predict missing values in graph signals through a message-passing mechanism. 2. **Spatial and Temporal Smoothness**: Combined spatial and temporal smoothness assumptions to improve prediction accuracy. 3. **Online Prediction**: Achieved online prediction, enabling real-time updates of prediction results at each time point. 4. **Experimental Validation**: Conducted experiments on a wind speed graph signal dataset to validate the effectiveness of the method, and compared it with traditional methods, demonstrating better performance. ### Problems Addressed: - **Partial Observational Data**: How to accurately predict missing values in graph signals when some observational data is missing. - **Online Prediction**: How to achieve real-time, dynamic prediction to adapt to continuously changing data environments. - **Improving Accuracy**: How to leverage the powerful learning capabilities of LLM to improve prediction accuracy, especially in complex time-varying graph signals. ### Experimental Results: - **MSE Comparison**: Compared to traditional GLMS and G-Sign algorithms, the GPT-3.5-turbo model shows significant advantages in predicting missing values, with its MSE value being much lower than other methods. ### Limitations: - **Occasional NaN Values**: In rare cases, the model may output NaN values, the cause of which is not yet clear. - **Batch Input Issues**: When inputting multiple node tasks in batches, the number of returned results may be inconsistent. Overall, the method proposed in this paper performs excellently in handling partially observed time-varying graph signals, providing new ideas and tools for research in related fields.

LLM-based Online Prediction of Time-varying Graph Signals

LLM Online Spatial-temporal Signal Reconstruction Under Noise

Adaptive Least Mean Squares Graph Neural Networks and Online Graph Signal Estimation

Online Signal Estimation on the Graph Edges via Line Graph Transformation

Adaptive Spatio-temporal Estimation on the Graph Edges via Line Graph Transformation

LLM4DyG: Can Large Language Models Solve Spatial-Temporal Problems on Dynamic Graphs?

LLM4DyG: Can Large Language Models Solve Problems on Dynamic Graphs?

Strada-LLM: Graph LLM for traffic prediction

Time Series Forecasting with LLMs: Understanding and Enhancing Model Capabilities

Time-varying graph learning from smooth and stationary graph signals with hidden nodes

An Evaluation of Standard Statistical Models and LLMs on Time Series Forecasting

Time-LLM: Time Series Forecasting by Reprogramming Large Language Models

A Comprehensive Evaluation of Large Language Models on Temporal Event Forecasting

Online Graph Learning for Time‐varying Graphs

Adaptive Least Mean pth Power Graph Neural Networks

Multi-Patch Prediction: Adapting LLMs for Time Series Representation Learning

Large Language Model with Graph Convolution for Recommendation

Online Learning for Time Series Prediction of AR Model with Missing Data

LPNL: Scalable Link Prediction with Large Language Models

Chain of History: Learning and Forecasting with LLMs for Temporal Knowledge Graph Completion

Revisited Large Language Model for Time Series Analysis through Modality Alignment