Abstract:This work introduces the LLM Online Spatial-temporal Reconstruction (LLM-OSR) framework, which integrates Graph Signal Processing (GSP) and Large Language Models (LLMs) for online spatial-temporal signal reconstruction. The LLM-OSR utilizes a GSP-based spatial-temporal signal handler to enhance graph signals and employs LLMs to predict missing values based on spatiotemporal patterns. The performance of LLM-OSR is evaluated on traffic and meteorological datasets under varying Gaussian noise levels. Experimental results demonstrate that utilizing GPT-4-o mini within the LLM-OSR is accurate and robust under Gaussian noise conditions. The limitations are discussed along with future research insights, emphasizing the potential of combining GSP techniques with LLMs for solving spatiotemporal prediction tasks.

What problem does this paper attempt to address?

### Problems the paper attempts to solve The paper aims to solve the problem of signal reconstruction in dynamic spatio - temporal networks, especially in the presence of noise and missing values. Specifically, the paper proposes a framework named LLM Online Spatial - temporal Reconstruction (LLM - OSR), which combines Graph Signal Processing (GSP) and large - scale language model (LLMs) techniques for online spatio - temporal signal reconstruction. ### Main contributions 1. **Propose the LLM - OSR algorithm**: This is a new method that realizes the effective reconstruction of dynamic spatio - temporal signals in the presence of noise and missing values by seamlessly combining GSP techniques and LLM - driven predictors. 2. **Reverse embedding method**: LLM - OSR adopts a sophisticated reverse embedding method to convert spatio - temporal signals into coherent and context - meaningful natural - language expressions, enabling LLMs to process these signals more effectively. ### Method overview 1. **GSP - based Spatial - temporal Signal Handler**: - **Denoising and feature enhancement**: Use GSP techniques to denoise and enhance the features of spatio - temporal signals to prepare the data for LLM prediction. - **Training process**: In the training phase, optimize the denoising effect by iteratively applying graph convolution operations, calculating the estimation error, and updating the filter parameters. 2. **LLM - based Spatial - temporal Signal Predictor**: - **Reverse embedding**: Convert the denoised signal into a natural - language expression so that the LLM can understand and process it. - **Task generation**: Generate tasks for each node, and the task content includes the neighbor information of the node and its past signal values. - **Prediction**: Use the LLM to predict missing values based on spatial and temporal information. ### Experimental setup - **Datasets**: - **Traffic data**: Use the Seattle Loop Detector Dataset, which contains traffic flow data on highways in the Seattle area. Different levels of Gaussian noise are added in the experiment to evaluate the robustness of the model. - **Meteorological data**: Obtain hourly wind speed and temperature data from NOAA. Due to the different behaviors of wind speed and temperature, they are analyzed as two independent datasets respectively. - **Experimental results**: The experimental results show that LLM - OSR using GPT - 4 - o mini has high accuracy and robustness under Gaussian noise conditions. ### Future research directions - **Expand the application range**: Explore the possibility of applying LLM - OSR to other fields, such as medical diagnosis, financial prediction, etc. - **Improve the algorithm**: Further optimize the combination of GSP and LLMs to improve the performance and robustness of the model. ### Summary The paper proposes a new online spatio - temporal signal reconstruction method, LLM - OSR, by combining GSP and LLMs techniques, which effectively solves the problem of signal reconstruction in dynamic complex networks, especially in the presence of noise and missing values.

LLM Online Spatial-temporal Signal Reconstruction Under Noise

LLM-based Online Prediction of Time-varying Graph Signals

STONet: A Neural-Operator-Driven Spatio-temporal Network

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

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

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

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

Online Signal Estimation on the Graph Edges via Line Graph Transformation

Unraveling Complexity: an Exploration into the Large-Scale Multi-Modal Signal Processing

Strada-LLM: Graph LLM for traffic prediction

Spatial-Temporal Large Language Model for Traffic Prediction

Graph-guided Bayesian matrix completion for ocean sound speed field reconstruction.

GATGPT: A Pre-trained Large Language Model with Graph Attention Network for Spatiotemporal Imputation

How Can Large Language Models Understand Spatial-Temporal Data?

Spatio-Temporal Signal Recovery Based on Low Rank and Differential Smoothness

LLMSense: Harnessing LLMs for High-level Reasoning Over Spatiotemporal Sensor Traces

A spatio-temporal LSTM model to forecast across multiple temporal and spatial scales

Reconstruction of Time-varying Graph Signals via Sobolev Smoothness

Advancing Large Language Models for Spatiotemporal and Semantic Association Mining of Similar Environmental Events

UrbanGPT: Spatio-Temporal Large Language Models

E-LMC: Extended Linear Model of Coregionalization for Spatial Field Prediction