Results

All numbers are on DBP15K with the 30% seed split, the standard protocol. "This repo" values come from the training runs that produced the curves shown below. Evaluation uses CSLS unless noted (L1 for GCN-Align, top-k for DGMC).

Headline: zh_en

Model	Family	Hit@1 (paper)	Hit@1 (here)	Hit@10 (paper)	Hit@10 (here)	MRR (paper)	MRR (here)
GCN-Align (SE)	structural	0.384	~0.38	0.703	~0.68	-	~0.49
JAPE (SE+AE)	attributes	0.412	0.425	0.745	0.761	0.490	0.537
KECG	rel. GNN	0.477	~0.42	0.835	~0.73	0.598	~0.52
AliNet	structural	0.539	~0.53	0.826	~0.81	0.628	~0.63
BootEA	structural	0.629	~0.56	0.847	~0.85	0.703	~0.66
MRAEA (base)	rel. GNN	0.638	0.659	0.882	0.898	0.729	0.746
NAEA	structural	0.650	~0.62	0.867	~0.86	0.720	~0.70
RREA (basic)	rel. GNN	0.715	0.712	0.929	0.934	0.794	0.793
MRAEA (+iter)	rel. GNN	0.757	0.746	0.930	0.930	0.827	0.814
DGMC	names	0.801	0.767	0.875	0.840	-	-
RREA (semi)	rel. GNN	0.801	0.805	0.948	0.950	0.857	0.859

Sorted by paper Hit@1. Bold = this repo matches or beats the paper.

Reading the landscape

%%{init: {'theme':'base','themeVariables':{'fontSize':'15px','fontFamily':'Inter, sans-serif','lineColor':'#8b949e'}}}%%
flowchart LR
    A["<b>structure only</b><br/>0.38 - 0.62 Hit@1"]:::a --> B["<b>+ relations (GNN)</b><br/>0.66 - 0.80 Hit@1"]:::b2
    B --> C["<b>+ side info</b><br/>attributes / names<br/>0.43 - 0.94 Hit@1"]:::c
    classDef a fill:#0c2d6b,stroke:#58a6ff,stroke-width:2px,color:#dbeafe;
    classDef b2 fill:#3b0764,stroke:#a371f7,stroke-width:2px,color:#f3e8ff;
    classDef c fill:#7c2d54,stroke:#f778ba,stroke-width:2px,color:#ffe4f0;

• Purely structural models (GCN-Align, AliNet, NAEA, BootEA) top out around 0.4-0.65 Hit@1 and are the hardest to reproduce exactly - independent benchmarks land well below the papers too.
• Relation-aware GNNs (MRAEA, RREA) are the strongest structural family; RREA semi is the best model in this repo and matches/beats its paper.
• Side information changes the game: JAPE's attributes push it above the structure-only models, and DGMC's entity-name embeddings reach 0.77-0.94 Hit@1.

Training curves

The figures below are the curves produced by the runs in logs/.

RREAMRAEA / JAPENAEA / BootEAAliNet / KECG / GCN-Align

Reproduce a number

Every run directory keeps config_used.yaml, so any number above can be reproduced with python -m src.main --config <that config>. See Getting started.