DGMC¶

entity names + GNN

Deep Graph Matching Consensus Matthias Fey, Jan E. Lenssen, Christopher Morris, Jonathan Masci, Nils M. Kriege - ICLR 2020 Paper | models/dgmc.py | notebook

Idea in one sentence

Use entity-name word embeddings as node features, build a sparse top-k correspondence by local feature matching, then refine it so that matched neighbourhoods agree (consensus).

This is the strongest family

Unlike the purely structural models, DGMC leans on entity names (summed GloVe-300d vectors), which is why it reaches Hit@1 in the 0.77-0.94 range rather than 0.4-0.6.

Architecture¶

%%{init: {'theme':'base','themeVariables':{'fontSize':'14px','fontFamily':'Inter, sans-serif','lineColor':'#7d8590','primaryTextColor':'#e6edf3'}}}%%
flowchart TD
    X["node features<br/>sum of GloVe-300d name vectors"] --> P1["psi_1 (RelCNN GNN)"]
    P1 --> S0["sparse top-k correspondence<br/>S0 = softmax(top_k(h_s . h_t))"]
    S0 --> LOOP{"consensus loop<br/>L iterations"}
    LOOP --> RND["random colourings r_s -> r_t = S^T r_s"]
    RND --> P2["psi_2 (RelCNN) diffuse"]
    P2 --> D["disagreement D = psi_2(r_s) - psi_2(r_t)"]
    D --> UP["S_hat += MLP(D)"]
    UP --> LOOP
    LOOP --> SL["refined S_L"]
    classDef b fill:#7c2d54,stroke:#f778ba,stroke-width:2px,color:#ffe4f0;
    class X,P1,S0,LOOP,RND,P2,D,UP,SL b;

Components¶

Name features. Each node feature is the sum of the GloVe-300d embeddings of the words in its (translated) name.
Local matching (\(\psi_1\)). A 3-layer RelCNN produces L2-normalised embeddings; the initial correspondence keeps the top-k targets per source, scored by cosine with a temperature.
Neighbourhood consensus (\(\psi_2\) + MLP). Random node colourings are pushed through the current correspondence; their disagreement after diffusion drives an additive re-ranking so that matched neighbourhoods become consistent.
Two training phases. Phase 1 trains only the local matching; phase 2 enables the consensus.

Results¶

DBP15K, 30% split. This repo beats the paper on fr_en.

Hits@1	zh_en	ja_en	fr_en
DGMC (paper)	0.801	0.848	0.933
This repo	0.767	0.814	0.939

Hits@10	zh_en	ja_en	fr_en
DGMC (paper)	0.875	0.897	0.960
This repo	0.840	0.874	0.965

Debugging lessons

Cosine + temperature is critical: the summed-name features have wildly varying norms, so a raw inner product gives recall@10 ~0.13 vs ~0.69 with cosine. Apply the temperature only inside the softmax, or it saturates and kills the refinement gain.
dropout 0.2 (not the paper's 0.5): only ~100 full-batch steps, 0.5 under-fits.
k=25 (not 10) gives more contrastive negatives -> better embeddings; k=50 diverges.
detach_refine=false: keep training \(\psi_1\) during refinement; its recall keeps rising.
The Hits@1 ceiling in the sparse model is the top-k recall of the initial matching (~0.84 zh vs ~0.87 paper) - that is the ~3 points left on zh/ja.

References¶

Fey et al., Deep Graph Matching Consensus, ICLR 2020.
Xu et al., Cross-lingual KG Alignment via Graph Matching NN (GMNN features), ACL 2019.
Lample et al., CSLS, ICLR 2018.