SimplE implementation for PyG

[AMMAS1]

Add SimplE Knowledge Graph Embedding Model

Summary

Implements the SimplE model for knowledge graph embedding, addressing the independence issue in CP decomposition by leveraging inverse relations.

Changes

• New model: torch_geometric.contrib.nn.SimplE - A bilinear KGE model that uses two embeddings per entity (head/tail) and two per relation (forward/inverse)
• Scoring function: 0.5 * (⟨h_e, v_r, t_e⟩ + ⟨h_t, v_r⁻¹, t_h⟩) - Average of CP scores for forward and inverse relations
• Example script: examples/contrib/simple_fb15k_237.py - Training example on FB15k-237 dataset

Files Changed

• torch_geometric/contrib/nn/kge/simplE.py - SimplE model implementation
• torch_geometric/contrib/nn/kge/__init__.py - Export SimplE
• examples/contrib/simple_fb15k_237.py - Usage example
• test/contrib/nn/kge/test_simple.py - Unit tests
• CHANGELOG.md - Added entry under [Unreleased]

Implementation Details

• Extends KGEModel base class following the same pattern as TransE and RotatE
• Uses binary cross-entropy loss with logits (same as ComplEx)
• Xavier uniform initialization for all embeddings

Testing

• Ran the example on FB15k-237 dataset with successful training and evaluation
• Results: MRR 0.1969, Hits@10 0.3627 after 5 epochs (improves with more training)

References

• Paper: SimplE Embedding for Link Prediction in Knowledge Graphs