EvoRM: Evolutionary Rule Mining for Knowledge Graph Entity Alignment

This repository contains the implementation and experimental results of EvoRM, a neuro-symbolic reasoning framework for knowledge graph entity alignment, integrated into the AdaCoAgentEA (ICDE 2025) framework.

Paper

EvoRM: Neuro-Symbolic Evolutionary Rule Mining for Knowledge Graph Reasoning
Submitted to TKDE

Overview

EvoRM enhances LLM-based entity alignment through 4 key components:

RuleEncoding - Symbolic rule extraction from LLM reasoning traces
HypergraphStorage - Weighted hypergraph for rule organization and retrieval
TwoStageInferenceController - Two-stage inference with Mlight/Mheavy separation
RuleMaintenance - Dynamic rule pruning and evolution

Key Features

MLP Gate (gϕ): Lightweight MLP for survival pair gating (Section III-E)
Mlight/Mheavy Separation: Rationale elicitation separated from decision-making (Section III-C)
Entity Embeddings: 1024-dim feature hashing for cosine similarity (Section III-D)
SemanticEquiv/SemanticConflict: LLM identifies semantically equivalent/conflicting attribute values
5 Ablation Modes: w/o Stage-1, w/o Maintenance, w/o MLP Gate, w/o Hypergraph, w/o Mlight

Results

RQ1: Main Experiment (ICEWS-WIKI)

Method	Hits@1	Precision	Recall	F1
AdaCoAgentEA (Baseline)	0.9003	0.9235	0.9003	0.9117
AdaCoAgentEA + EvoRM	0.9508	0.99	0.9508	0.97

RQ3: Ablation Study

Configuration	Hits@1	Precision	Recall
Full System (EvoRM)	0.9267	0.9929	0.9267
w/o Stage-1 (Symbolic)	0.95	0.9965	0.95
w/o Maintenance	0.9433	0.993	0.9433
w/o MLP Gate	TBD	TBD	TBD
w/o Hypergraph	TBD	TBD	TBD
w/o Mlight	TBD	TBD	TBD

RQ5: Cold-Start Scaling

N_warmup	Hits@1	Tokens/Pair	Rules
100	0.97	858.1	-
200	0.935	881.1	-
500	0.936	884.9	-
1000	0.938	877.9	-

Code Structure

AdaCoAgentEA/
├── evorm_plugin.py          # Core EvoRM framework (1599 lines)
├── evorm_mlight.py          # Mlight/Mheavy two-step pipeline (608 lines)
├── evorm_entity_embedding.py # Entity embeddings + feature hashing (285 lines)
├── evorm_config.py          # Centralized hyperparameter config (195 lines)
├── evorm_mlp_gate.py        # MLP Gate for survival pair gating (351 lines)
├── baselines/               # All baseline implementations
│   ├── evorm_wrappers/      # EvoRM wrappers for baselines
│   │   ├── chat_ea.py       # ChatEA + EvoRM
│   │   ├── zero_cot.py      # ZeroCoT + EvoRM
│   │   ├── self_consistency.py # Self-Consistency + EvoRM
│   │   ├── cohard.py        # Collaboration-Hard + EvoRM
│   │   ├── matchgpt.py      # MatchGPT/Anymatch + EvoRM
│   │   └── lela_el.py       # LELA/Schema Matching + EvoRM
│   └── original/            # Original baseline source code
├── Area2/                   # AdaCoAgentEA core
│   ├── LLM1_label_selector.py       # AdaCoAgentEA + EvoRM
│   └── LLM1_label_selector_baseline.py # AdaCoAgentEA baseline
├── experiments/             # Experiment scripts
│   ├── run_comprehensive.py # RQ1-RQ5 comprehensive experiments
│   ├── run_ablation_v2.py   # 6-mode ablation study
│   └── run_baselines_v2.py  # Baseline comparison
└── results/                 # All experiment results (JSON)

Baselines

Task	Baselines	EvoRM-Enhanced
Entity Alignment (EA)	ChatEA, ZeroCoT, Self-Consistency, CoHard, AdaCoAgentEA	✅ All
Entity Resolution (ER)	MatchGPT (4 backends), Anymatch	✅ All
Entity Linking (EL)	LELA (4 datasets)	✅ All
Schema Matching (SM)	LLM-DP, ReMatch, Matchmaker	✅ All

Usage

# Run ablation study
python experiments/run_ablation_v2.py

# Run baseline comparison
python experiments/run_baselines_v2.py

# Run comprehensive experiments
python experiments/run_comprehensive.py --dataset icews_wiki --all

Citation

@article{evorm2025,
  title={EvoRM: Neuro-Symbolic Evolutionary Rule Mining for Knowledge Graph Reasoning},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2025}
}

License

This project is for research purposes only.