Research map / Graph learning / Knowledge and temporal graphs
ID-GMLM: Intelligent Decision-Making with Integrated Graph Models and Large Language Models
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.
Proposes ID-GMLM, combining GNN-based multi-task learning with LLM-derived criterion weights (via gpt-3.5-turbo and text-embedding-ada-002) to rank alternatives for multi-criteria decision making and preference learning.
Problem
Combining multi-criteria decision making (MCDM) and preference learning (PL) struggles with large data volumes, complex relationships among alternatives and criteria, and dynamic decision-maker preferences.
Approach
Maps decision problems onto graphs (alternative preference graph and criterion relationship graph), uses residual GNNs in a multi-task setup with an auxiliary criterion-relationship loss, and converts natural-language criterion descriptions into criterion weights via an LLM and text embeddings, which feed a parameter tuning network and attention network.
Key results
- ID-GMLM achieves average rank 1.17 across six datasets, best on most, with NDCG 0.9979 and C-index 0.9975 on BCC versus baselines around 0.83 C-index.
- On BCC, ID-GMLM improves NDCG, AP, and C-index by up to 20.3%, 25.2%, and 24.2% relative to the best-performing baselines' worst cases.
- Replacing LLM-generated weights with equal or random weights reduces performance only marginally (e.g., NDCG change of 0.27% to 0.34%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ID-GMLM instead of the strongest baseline | ID-GMLM · BCC | NDCG | Helped |
| Use ID-GMLM instead of the strongest baseline | ID-GMLM · BCC | AP | Helped |
| Use ID-GMLM instead of the strongest baseline | ID-GMLM · BCC | C-index | Helped |
| Add the criterion relationship network (CRN) (ablation) | ID-GMLM · six benchmarks | C-index | Helped |
| Add the parameter tuning network (PTN) (ablation) | ID-GMLM · six benchmarks | C-index | Helped |
| Add the attention network (AN) (ablation) | ID-GMLM · six benchmarks | C-index | Helped |
| Use original LLM-generated criterion weights instead of equal weights | ID-GMLM · six benchmarks | NDCG: +0.27% | Helped |
| Use original LLM-generated criterion weights instead of random weights | ID-GMLM · six benchmarks | NDCG: +0.34% | Helped |
| Use original LLM-generated criterion weights instead of equal weights | ID-GMLM · six benchmarks | AP: +0.55% | Helped |
| Use original LLM-generated criterion weights instead of equal weights | ID-GMLM · six benchmarks | C-index: +0.56% | Helped |
| Use original LLM-generated criterion weights instead of random weights | ID-GMLM · six benchmarks | AP: +0.81% | Helped |
| Use original LLM-generated criterion weights instead of random weights | ID-GMLM · six benchmarks | C-index: +0.69% | Helped |
Surprising findings
- Using equal or random LLM-generated criterion weights causes only minimal performance decline, suggesting robustness and limited contribution of the LLM weights.
Limitations
- LLM contribution appears minimal under equal or random weights, so the LLM component's value is not strongly demonstrated.
- Preferences are static and based on dataset descriptions; alignment of DM preferences with LLM outputs is left to future work.
- Evaluated only on six public UCI/WEKA benchmarks with monotonic criteria.
Experimental setup
Datasets: DBS BCC MPG ESL LEV CEV
Benchmarks: DBS BCC MPG ESL LEV CEV
Metrics: NDCG AP C-index
Baselines: NEUR-HCI ELECTRE TRI-rC PLMCS ANN-UTADIS
Builds on: graph neural networks graph convolutional networks residual GNN
Closest papers by meaning
- DeLLMa: Decision Making Under Uncertainty with Large Language Models (ICLR 2025 · similarity 0.83)
- Enhancing Decision Making Through the Integration of Large Language Models and Operations Research Optimization (AAAI 2025 · similarity 0.83)
- Logic-Logit: A Logic-Based Approach to Choice Modeling (ICLR 2025 · similarity 0.83)
- Advancing Tool-Augmented Large Language Models: Integrating Insights from Errors in Inference Trees (NeurIPS 2024 · similarity 0.82)
- TPO: Aligning Large Language Models with Multi-branch & Multi-step Preference Trees (ICLR 2025 · similarity 0.82)
- Group Robust Preference Optimization in Reward-free RLHF (NeurIPS 2024 · similarity 0.82)
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.