Research map / Vision-language and multimodal / Cross-modal learning
Mixed-Curvature Multi-Modal Knowledge Graph Completion
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 MCKGC, a mixed-curvature multi-modal knowledge graph completion model that embeds structural, image and text modalities into hyperbolic, hyperspherical and Euclidean spaces and fuses them with modality-level and space-level gates, using ComplEx as decoder.
Problem
Existing multi-modal knowledge graph completion methods fuse modalities in Euclidean space and struggle to model hierarchical and circular graph structures.
Approach
Embed each modality into three single-curvature spaces (hyperbolic, hyperspherical, Euclidean) with learnable curvature, then progressively fuse first across modalities within each space and then across spaces using gates.
Key results
- On DB15K, MCKGC reaches MRR 39.79 and Hits@10 54.66, a 5.4% MRR gain over MyGO (37.72).
- Ablations on DB15K show removing any of the three curvature spaces, the gates, or MIMCM lowers MRR (worst: w/o MIMCM at 38.48 and w/o S at 38.45 vs 39.79 full).
- On MKG-W and MKG-Y, MCKGC reports MRR 38.92 and 36.88 respectively, competitive with or above prior multi-modal baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use mixed-curvature multimodal fusion instead of MyGO | MCKGC · DB15K | MRR: +5.4% | Helped |
| use mixed-curvature multimodal fusion instead of MyGO | MCKGC · MKG-W | MRR: +2.2% | Helped |
| use mixed-curvature multimodal fusion instead of AdaMF | MCKGC · MKG-Y | MRR: +2.3% | Helped |
| add image information (ablation) | MCKGC · DB15K | MRR: 39.44 -> 39.79 | Helped |
| add text information (ablation) | MCKGC · DB15K | MRR: 39.00 -> 39.79 | Helped |
| add hyperbolic space (ablation) | MCKGC · DB15K | MRR: 38.72 -> 39.79 | Helped |
| add hyperspherical space (ablation) | MCKGC · DB15K | MRR: 38.45 -> 39.79 | Helped |
| add Euclidean space (ablation) | MCKGC · DB15K | MRR: 38.76 -> 39.79 | Helped |
| use modality-level and space-level gates instead of mean fusion (ablation) | MCKGC · DB15K | MRR: 38.60 -> 39.79 | Helped |
| add MIMCM before progressive fusion (ablation) | MCKGC · DB15K | MRR: 38.48 -> 39.79 | Helped |
Surprising findings
- Modality weights consistently follow structure > image > text across spaces, and relation types map to spaces (ActedIn to hyperbolic/hyperspherical, IsAffiliatedTo to Euclidean).
Limitations
- Results on MKG-W and MKG-Y are not uniformly best; MyGO is not reported on MKG-W and MKG-Y.
- Training time limited the study to averages of five random repetitions, with no variance reported.
Experimental setup
Datasets: DB15K MKG-W MKG-Y
Benchmarks: DB15K MKG-W MKG-Y
Metrics: MRR Hits@1 Hits@3 Hits@10
Baselines: TransE DistMult ComplEx RotatE GC-OTE IKRL TBKGC TransAE MMKRL RSME VBKGC OTKGE
Builds on: ComplEx MyGO ATTH MURP
Closest papers by meaning
- LAFA: Multimodal Knowledge Graph Completion with Link Aware Fusion and Aggregation (AAAI 2024 · similarity 0.86)
- APKGC: Noise-enhanced Multi-Modal Knowledge Graph Completion with Attention Penalty (AAAI 2025 · similarity 0.86)
- Mixed Geometry Message and Trainable Convolutional Attention Network for Knowledge Graph Completion (AAAI 2024 · similarity 0.86)
- Tokenization, Fusion, and Augmentation: Towards Fine-grained Multi-modal Entity Representation (AAAI 2025 · similarity 0.86)
- Multiple Heads are Better than One: Mixture of Modality Knowledge Experts for Entity Representation Learning (ICLR 2025 · similarity 0.85)
- SPAC: Sparse Partitioning and Adaptive Core Tensor Pruning Model for Knowledge Graph Completion (AAAI 2025 · similarity 0.85)
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.