DiffRuleEA: Diffusion-Guided Rule Induction for Entity Alignment

AAAI 2027

A rule-centered LLM-based Entity Alignment framework that converts alignment evidence and verified decisions into explicit, executable, and reusable rules.

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    DiffRuleEA Pipeline                       │
├─────────────────────────────────────────────────────────────┤
│ Module 1: Multi-View Encoder (Simple-HHEA)                  │
│   ├── BERT + whitening → semantic embeddings (300d)         │
│   ├── Relation-aware Random Walk + Skip-gram (300d)         │
│   ├── Time2Vec temporal encoding (64d)                      │
│   └── CSLS candidate retrieval (K_cand=50)                  │
├─────────────────────────────────────────────────────────────┤
│ Module 2: Mixed-State Diffusion Rule Induction              │
│   ├── D3PM discrete diffusion for predicate masks           │
│   ├── DDPM continuous diffusion for decision thresholds     │
│   ├── Per-predicate token + FiLM conditioning               │
│   ├── CLS-token quality prediction                          │
│   └── Two-phase: SL pre-training + RL fine-tuning           │
├─────────────────────────────────────────────────────────────┤
│ Module 3: Self-Evolving Rule-Based Alignment                │
│   ├── Rule Store (Jaccard > 0.8 merge)                      │
│   ├── Rule Retrieve (domain filter + ρ·log(1+use) + MMR)    │
│   ├── QNorm score normalization by mask length              │
│   ├── Confidence-adaptive routing (τ_high=0.85, τ_low=0.5) │
│   ├── Rule Evolve (decay, distill, rerun Phase B)           │
│   └── Rule Transfer (cross-KG reuse)                        │
└─────────────────────────────────────────────────────────────┘

Key Hyperparameters

Parameter	Value	Description
K	20	Number of predicates (8 core + 12 domain)
T	100	Diffusion steps
D	256	Hidden dimension
L_d	4	Transformer layers
heads	8	Attention heads
λ_θ	0.5	Threshold loss weight
λ_q	0.1	Quality loss weight
κ	10	Threshold sharpness
B	200	RL batch size
λ_H	0.01	Entropy coefficient
τ_high	0.85	High-confidence threshold
τ_low	0.5	Low-confidence threshold
E_SL	50	Supervised epochs
E_RL	100	RL epochs

Installation

pip install torch numpy openai httpx

Usage

Full Pipeline Training

python -m diffrule_ea.train_evolve \
    --data_dir data/dbp15k_en_fr \
    --output_dir outputs \
    --num_predicates 20 \
    --diffusion_steps 100 \
    --epochs_sl 50 \
    --epochs_rl 100 \
    --rl_batch_size 200 \
    --evolution_iterations 3 \
    --outer_iterations 3

Evaluation

python -m diffrule_ea.evaluate \
    --data_dir data/dbp15k_en_fr \
    --model_path outputs/diffrule_ea_final.pt \
    --run_ablation \
    --run_efficiency

Quick Test

python -m diffrule_ea.train_evolve \
    --data_dir data/dbp15k_en_fr \
    --output_dir outputs_test \
    --num_predicates 20 \
    --diffusion_steps 10 \
    --epochs_sl 5 \
    --epochs_rl 5 \
    --skip_llm \
    --outer_iterations 1

Module Details

Module 2: Diffusion Rule Induction

• diffusion_rule_induction.py: D3PM + DDPM denoising with per-predicate tokens, FiLM, CLS head
• Paper reference: Appendix F

Module 3: Rule-Based Alignment

• rule_based_alignment.py: Rule base with Store/Retrieve/Evolve/Transfer, QNorm, confidence-adaptive routing
• llm_verifier.py: LLM verification with rule traces, token budgets (~200 VERIFY, ~800 REASON)
• Paper reference: Appendix G

Citation

@inproceedings{diffruleea2027,
  title={DiffRuleEA: Diffusion-Guided Rule Induction for Entity Alignment},
  author={Anonymous},
  booktitle={AAAI},
  year={2027}
}