Research map / Vision-language and multimodal / Cross-modal learning
MM-DyGraph: A Dataset, Benchmark, and Model for Multimodal Dynamic Graphs
Zongyuan Wu, Pengjie Wang, Lingshan Chen, Tianhang Wan, Yuan Meng, Xin Wang, Ling Feng, Wenwu Zhu
Read the paperCheck your idea against 137,217 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.
Introduces MM-DyGraph, a dataset, benchmark, and model for multimodal dynamic graphs where nodes and edges carry text and image signals.
Problem
Benchmarks and methods explicitly studying multimodal learning on dynamic graphs (nodes and edges with text and images) are scarce, limiting use of multimodal information for downstream tasks.
Approach
Provides a multimodal dynamic graph dataset and benchmark together with a proposed model for learning over text- and image-attributed temporal graphs; specific method details are not stated in the available abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add text embeddings instead of random features | TST-MF · GDELT LP transductive | AP (%): 92.83 -> 94.03 | Improved |
| add image embeddings on top of text (naive concatenation) | DyGFormer · Amazon-Office LP inductive | AP (%): 73.50 -> 68.64 | Hurt |
| add image embeddings on top of text | TST-MF · Amazon-Office LP inductive | AP (%): 77.47 -> 78.16 | Improved |
| add image embeddings on top of text | TST-MF · Yelp node retrieval transductive | Hits@10 (%): 75.19 -> 79.11 | Improved |
| add image embeddings on top of text | CAWN · Amazon-Industrial LP transductive | AP (%): 91.94 -> 92.21 | Improved |
| replace naive concatenation with target-aware fusion | DyGFormer · Amazon-Industrial LP inductive (Text+Image) | AP (%): 76.33 -> 81.95 | Improved |
| replace naive concatenation with target-aware fusion | DyGFormer · GMCT LP transductive (Text+Image) | AP (%): 95.54 -> 89.08 | Hurt |
| use target-aware fusion instead of gated fusion | DyGFormer backbone · GDELT node retrieval transductive | Hits@10 (%): 62.18 -> 81.62 | Improved |
| use target-aware fusion instead of modality attention | DyGFormer backbone · Yelp node retrieval transductive | Hits@10 (%): 57.09 -> 79.11 | Improved |
| use target-aware fusion instead of naive concatenation | DyGFormer · Amazon-Industrial edge classification | Precision (%): 59.38 -> 50.32 | Hurt |
| add visual modality with misaligned source (edge/node mismatch) | dynamic graph methods · MM-DyGraph | various | Mixed |
| add multimodal features on saturated structural datasets | strong backbones · GDELT / Amazon-Office link prediction | AP/AUC | No clear effect |
Closest papers by meaning
- Mosaic of Modalities: A Comprehensive Benchmark for Multimodal Graph Learning (CVPR 2025 · similarity 0.93)
- Graph4MM: Weaving Multimodal Learning with Structural Information (ICML 2025 · similarity 0.90)
- Mario: Multimodal Graph Reasoning with Large Language Models (CVPR 2026 · similarity 0.89)
- DTGB: A Comprehensive Benchmark for Dynamic Text-Attributed Graphs (NeurIPS 2024 · similarity 0.88)
- GraphVLM: Benchmarking Vision Language Models for Multimodal Graph Learning (CVPR 2026 · similarity 0.88)
- Quality Measures for Dynamic Graph Generative Models (ICLR 2025 · similarity 0.87)
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.
