Models overview¶

Nine entity-alignment methods, grouped by the kind of signal they exploit. Click any card to open its dedicated page (idea, architecture diagram, losses, training recipe, results and the debugging lessons that actually made it reach paper level).

structural NAEA - IJCAI 2019

Neighbourhood-aware GAT over translation-consistent neighbour messages, with hard negatives and recomputed bootstrapping.
structural BootEA - IJCAI 2018

AlignE (limit-based TransE) + alignment-by-swapping + editable MWGM self-training.
structural AliNet - AAAI 2020

Gated multi-hop GNN: 1-hop GCN fused with attentional 2-hop, anchored by a relation-aware loss.
structural GCN-Align - EMNLP 2018

Functionality-weighted adjacency + a shared 2-layer GCN (structure channel).
relation-aware KECG - EMNLP 2019

Shared diagonal multi-head GAT cross-graph + a knowledge-embedding (TransE) loss, alternated.
relation-aware MRAEA - WSDM 2020

Meta-relation-aware GAT (a relation and its inverse differ) + iterative mutual-NN bootstrap.
relation-aware RREA - CIKM 2020

Relational reflection (Householder) aggregation + turn-based CSLS bootstrap. Top performer.
attributes JAPE - ISWC 2017

Merged-seed TransE fused with a TF-IDF attribute channel. The paper that introduced DBP15K.
entity names DGMC - ICLR 2020

GloVe entity-name features + sparse top-k neighbourhood consensus. Beats the paper on fr_en.

At a glance¶

Model	Encoder	Alignment signal	Loss	Self-training	Eval
NAEA	GAT (neighbour messages)	structure	limit-based margin	recomputed	CSLS
BootEA	embedding (AlignE)	structure	limit-based TransE + pull	MWGM	CSLS
AliNet	gated multi-hop GNN	structure + relations	margin + TransE anchor	optional	CSLS
KECG	diagonal multi-head GAT	structure + relations	triplet + TransE	-	CSLS
GCN-Align	shared 2-layer GCN	structure	L1 margin	-	L1 / CSLS
JAPE	TransE (merged seeds)	structure + attributes	margin + fused AE	-	CSLS
DGMC	RelCNN + consensus	entity names	sparse NLL	consensus	top-k
MRAEA	meta-relation GAT	structure + relations	L1 margin	mutual-NN	cosine/CSLS
RREA	relational-reflection GAT	structure + relations	L1 margin	CSLS mutual-NN	CSLS

Shared building blocks¶

All models reuse the same engine, so reading one makes the rest easy:

%%{init: {'theme':'base','themeVariables':{'fontSize':'14px','fontFamily':'Inter, sans-serif','lineColor':'#7d8590','primaryTextColor':'#e6edf3'}}}%%
flowchart LR
    C["<b>configs/your_model.yaml</b><br/><i>hyper-parameters</i>"]:::cfg
    M["<b>models/your_model.py</b><br/><i>encoder + loss</i>"]:::model
    subgraph ENGINE["shared engine - reused by all nine models"]
        direction LR
        D["<b>data.py</b><br/><i>DBP15K · graphs · negatives</i>"]:::data
        T["<b>trainer.py</b><br/><i>train · eval · bootstrap · log</i>"]:::train
        E["<b>utils/metrics.py</b><br/><i>MRR · Hit@k · CSLS</i>"]:::metric
        D --> T --> E
    end
    D --> M --> T
    C -.->|drives| T
    style ENGINE fill:#0d1117,stroke:#30363d,color:#e6edf3
    classDef cfg    fill:#3a2a05,stroke:#d29922,stroke-width:2px,color:#fde68a;
    classDef data   fill:#0c2d6b,stroke:#58a6ff,stroke-width:2px,color:#dbeafe;
    classDef model  fill:#3b0764,stroke:#a371f7,stroke-width:2px,color:#f3e8ff;
    classDef train  fill:#7c2d54,stroke:#f778ba,stroke-width:2px,color:#ffe4f0;
    classDef metric fill:#14532d,stroke:#3fb950,stroke-width:2px,color:#dcfce7;

See the results page for the full benchmark tables and training curves.