Research map / Graph learning / Knowledge and temporal graphs
LLM-based Online Prediction of Time-varying Graph Signals (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Uses GPT-3.5-turbo in a message-passing-style online loop to predict missing node values in time-varying graph signals (wind speed), beating GLMS and G-Sign baselines in MSE.
Problem
Predicting missing observations in time-varying graph signals online, exploiting spatial and temporal smoothness when many nodes are unobserved.
Approach
Treats each missing node as an LLM task: its previous estimate and currently observed neighbor values are serialized into a prompt, and the LLM infers the current value, mimicking message passing without any weight updates.
Key results
- GPT-3.5-turbo achieves MSE 1.560 vs 3.396 for GLMS and 3.718 for G-Sign on wind-speed graph signals with 30% missing nodes.
- Each method repeated 5 runs; MSE averaged over nodes and time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GPT-3.5-turbo for online missing-value prediction instead of GLMS | Wind-speed graph signals · Wind-speed graph signals | MSE: 3.396 -> 1.560 | Helped |
| Use GPT-3.5-turbo for online missing-value prediction instead of G-Sign | Wind-speed graph signals · Wind-speed graph signals | MSE: 3.718 -> 1.560 | Helped |
Surprising findings
- Occasional NaN outputs from the LLM on a single node, with no clear cause; replaced by average of previous time points.
- Batch-feeding node tasks to the LLM often returns a number of outputs different from N.
Limitations
- LLM occasionally outputs NaN with unclear cause, requiring a fallback heuristic.
- Batched prompts return inconsistent output counts.
- Evaluated only on a single wind-speed dataset with one LLM and one-step prediction.
Experimental setup
Datasets: hourly wind speed graph signals (N=197 nodes, T=95 time points)
Benchmarks: hourly wind speed graph signals
Metrics: MSE
Baselines: GLMS G-Sign
Builds on: graph signal processing smoothness message passing
Closest papers by meaning
- How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension (ICLR 2025 · similarity 0.86)
- LLaGA: Large Language and Graph Assistant (ICML 2024 · similarity 0.85)
- LoReC: Rethinking Large Language Models for Graph Data Analysis (NeurIPS 2026 · similarity 0.85)
- Designing Specialized Two-Dimensional Graph Spectral Filters for Spatial-Temporal Graph Modeling (AAAI 2025 · similarity 0.85)
- State Space Models on Temporal Graphs: A First-Principles Study (NeurIPS 2024 · similarity 0.84)
- Can Large Language Models Analyze Graphs like Professionals? A Benchmark, Datasets and Models (NeurIPS 2024 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.