TransE Implementation & Reproduction

Code that directly implements the algorithm from Translating Embeddings for Modeling Multi-relational Data (NeurIPS, 2013) in PyTorch and reproduces the performance of the Link prediction experiment

Implementation Details

• Initialization: Initialize entity/relation embeddings with $Unif(-6/\sqrt{k}, 6/\sqrt{k})$, normalize relations right after initialization
• Training loop: Every epoch, entity normalize → mini-batch sampling → negative sampling (corrupted triplet) → margin-based ranking loss
• Evaluation: Link prediction (raw / filtered settings), Mean Rank, Hits@10

Dataset

FB15k — a standard KGE benchmark based on Freebase

FB15k

A subset reduced to a scale suitable for experiments, out of the entire Freebase (approx. 1.2 billion triples, over 80 million entities)

• Only entities that also exist in the Wikilinks database are selected
• Only entities/relations that appear at least 100 times within Freebase are selected
• Duplicate relations that are just the reverse of another (head and tail swapped), like !/people/person/nationality being the reverse of /people/person/nationality, are removed

Statistics

Item	Count
Triplets	592,213
Entities	14,951
Relationships	1,345

How to Run

Run directly in Colab

Run locally

1. Git Clone

   git clone https://github.com/0hseunghwan/TransE.git
   cd TransE

2. Install Dependencies

   pip install -r requirements.txt

3. Run the Notebook

   jupyter notebook TransE.ipynb

※ The FB15k dataset is automatically downloaded/loaded when the notebook runs (no separate setup needed)

Key Hyperparameters

Parameter	Paper	This Implementation
Embedding dimension $k$	50	50
Margin $\gamma$	1.0	1.0
Distance function	L1	L1
Learning rate	0.01	0.001
Optimizer	SGD	Adam
Batch size	-	1,024
Epoch	1,000	200

Note on hyperparameter changes: When training with the paper's settings (SGD, 1000 epochs), the training error kept decreasing continuously even up to 1000 epochs, which was judged to be an underfitting state After changing the optimizer to Adam and reducing the number of epochs to 200 and retraining, convergence was reached much faster, and the performance evaluated at this state was similar to the performance reported in the paper

Reproduction Results

Setting	Metric	Paper result	Reproduction result
Raw	Mean Rank	243	227.8
Filtered	Mean Rank	125	110.6
Raw	Hits@10 (%)	34.9	37.1
Filtered	Hits@10 (%)	47.1	49.4

This is the result from Table 3 (Link prediction results) of the paper (Bordes et al., 2013), and this implementation shows similar performance

References

Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J., & Yakhnenko, O. (2013). Translating Embeddings for Modeling Multi-relational Data. NeurIPS.