Documentation  |  Quickstart  |  Results  |  Roadmap

---

Mission

One place to find, run and compare entity-alignment models on DBP15K.

Entity Alignment (EA) research is scattered across dozens of repositories, each with its own data format, training loop and evaluation quirks. EntityAlignment-Nexus brings nine landmark methods together under a single, readable codebase with a shared data layer, shared metrics and a shared trainer pattern, so you can study a method, reproduce its numbers, and compare it to the others without re-learning a new project every time.

Every model ships twice:

• as an installable package (code/) driven by one YAML config per model, and
• as a self-contained notebook (Notebook/) that re-implements the whole engine inline, documented cell by cell, so you can read a method top to bottom in one file.

---

Highlights

Unified & faithful 9 models, 1 data loader, 1 metrics module, 1 trainer pattern DBP15K zh_en / ja_en / fr_en, the standard 30% seed split CSLS, MRR, Hit@1/5/10 reported the same way for everyone

Reproducible & transparent Timestamped run dirs: config snapshot, logs, CSV, checkpoints, curves Paper-level (or above) on several models, gaps documented honestly "Debugging lessons" notes explaining what actually made each model work

---

The model collection

flowchart LR
    subgraph S["Structural"]
        NAEA["NAEA<br/>IJCAI'19"]
        BootEA["BootEA<br/>IJCAI'18"]
        AliNet["AliNet<br/>AAAI'20"]
        GCN["GCN-Align<br/>EMNLP'18"]
    end
    subgraph R["Relation-aware GNN"]
        KECG["KECG<br/>EMNLP'19"]
        MRAEA["MRAEA<br/>WSDM'20"]
        RREA["RREA<br/>CIKM'20"]
    end
    subgraph A["Side information"]
        JAPE["JAPE<br/>attributes<br/>ISWC'17"]
        DGMC["DGMC<br/>entity names<br/>ICLR'20"]
    end
    DBP["DBP15K"] --> S & R & A
    classDef s fill:#1f6feb,stroke:#58a6ff,color:#fff;
    classDef r fill:#8957e5,stroke:#a371f7,color:#fff;
    classDef a fill:#bf4b8a,stroke:#f778ba,color:#fff;
    class NAEA,BootEA,AliNet,GCN s;
    class KECG,MRAEA,RREA r;
    class JAPE,DGMC a;

Loading

Model	Venue	Family	One-line idea	Page
NAEA	IJCAI 2019	structural + attention	Neighbourhood-aware GAT over translation-consistent messages	docs
BootEA	IJCAI 2018	structural + bootstrap	AlignE TransE + alignment-by-swapping + editable MWGM self-training	docs
AliNet	AAAI 2020	gated multi-hop GNN	1-hop GCN + attentional 2-hop, fused by a learned gate	docs
KECG	EMNLP 2019	GNN + TransE	Shared diagonal GAT cross-graph + knowledge-embedding loss	docs
GCN-Align	EMNLP 2018	GCN	Functionality-weighted adjacency + shared 2-layer GCN (SE)	docs
JAPE	ISWC 2017	TransE + attributes	Merged-seed TransE fused with a TF-IDF attribute channel	docs
DGMC	ICLR 2020	name features + GNN	GloVe name features + sparse top-k neighbourhood consensus	docs
MRAEA	WSDM 2020	meta-relation GNN	Relation + inverse aware GAT + iterative mutual-NN bootstrap	docs
RREA	CIKM 2020	relational reflection	Householder reflection aggregation + turn-based CSLS bootstrap	docs

---

Results

DBP15K zh_en, 30% seed. Bold = this repo matches or beats the paper. Every value is re-scored from the run's best checkpoint under one protocol (CSLS k=10, both directions averaged). Training curves live in the results page.

Model	Hit@1 (paper)	Hit@1 (here)	Hit@10 (paper)	Hit@10 (here)	MRR (paper)	MRR (here)
NAEA	0.650	0.628	0.867	0.860	0.720	0.711
BootEA	0.629	0.543	0.847	0.851	0.703	0.653
AliNet	0.539	0.513	0.826	0.812	0.628	0.621
KECG	0.477	0.497	0.835	0.855	0.598	0.619
GCN-Align (SE)	0.384	0.363	0.703	0.679	-	0.475
JAPE (SE+AE)	0.412	0.412	0.745	0.751	0.490	0.525
DGMC (names)	0.801	0.753	0.875	0.836	-	0.783
MRAEA (base)	0.638	0.698	0.882	0.923	0.729	0.780
MRAEA (+iter)	0.757	0.747	0.930	0.936	0.827	0.817
RREA (basic)	0.715	0.715	0.929	0.932	0.794	0.794
RREA (semi)	0.801	0.802	0.948	0.946	0.857	0.857

RREA (semi-supervised) is the current top performer in this repo. MRAEA (base) and KECG exceed their published Hit@1 under this protocol.

---

Quickstart

# 1. install the package (editable)
cd code
pip install -e .

# 2. train any model from its YAML config
python -m src.main --config ../configs/rrea_dbp15k.yaml          # RREA (semi-supervised)
python -m src.main --config ../configs/jape_dbp15k.yaml          # JAPE (SE + attributes)
python -m src.main --config ../configs/dgmc_dbp15k.yaml --lang zh_en   # DGMC (entity names)

# 3. override on the fly
python -m src.main --config ../configs/mraea_dbp15k.yaml --lang fr_en --epochs 2000

Each run writes a timestamped folder (config snapshot, training.txt, loss.csv, metrics.csv, dark-theme curves, checkpoints, embeddings).

Prefer reading over running? Open any notebook in Notebook/ - each one re-implements the full pipeline inline, documented cell by cell.

---

Repository layout

EntityAlignment-Nexus/
├── code/                  # installable package (pip install -e .)
│   └── src/
│       ├── data.py        # DBP15K loading + graph builders + samplers
│       ├── trainer.py     # all 9 trainers (shared pattern)
│       ├── main.py        # CLI entry point (YAML-driven)
│       ├── models/        # one file per model + its losses
│       └── utils/         # config, logging, metrics (MRR/Hit@k/CSLS), plotting
├── configs/               # one YAML per model
├── Notebook/              # 9 self-contained, documented notebooks
└── docs/                  # MkDocs Material documentation site
#  (training runs are written to a git-ignored experiments/ directory)

---

Documentation

A full MkDocs Material site (dark + light theme, Mermaid diagrams, math, training curves) covers every model in depth:

pip install -r requirements-docs.txt
mkdocs serve        # live preview at http://127.0.0.1:8000

It is deployed automatically to GitHub Pages on every push to main (see .github/workflows/docs.yml).

---

Roadmap

The repo is built to grow. Next up: transformer-based entity-alignment models (self-attention encoders, pre-trained language model initialisation, and large-model EA pipelines). See the roadmap for the shortlist and how to propose a model.

---

Contributing

Contributions are very welcome - a new model, a stronger config, a ja_en/fr_en run, or a doc fix. The shared data.py / trainer.py / metrics.py make adding a model mostly a matter of writing models/<your_model>.py, a trainer and a YAML. See the about page.

Citation

@software{ea_dbp15k,
  title  = {EntityAlignment-Nexus: A Unified Framework of Entity Alignment Models on DBP15K},
  author = {Nadjib ZAHAF},
  year   = {2026},
  url    = {https://github.com/Z-Nadjib/EntityAlignment-Nexus}
}

License

Released under the MIT License. The original papers and the DBP15K benchmark remain the property of their respective authors.