anonymous-ai-researcher/kdd2026: v1.0.1: Official Implementation of DF-EL++ (KDD 2026)

DF-EL++ 🚀 Fast and Faithful: Scalable Neuro-Symbolic Learning and Reasoning with Differentiable Fuzzy EL++ KDD 2026 This repository provides a from-scratch PyTorch implementation of the DF-EL++ framework, designed to be a faithful reproduction of the paper's methodology. 📋 Overview Figure 1: The DF-EL++ framework operates as a principled refinement engine, using logical constraints to improve the output of neural perception models. DF-EL++ is the first end-to-end differentiable framework that unifies PTIME-complete reasoning with neural learning, resolving the persistent trade-off between logical rigor and computational scale in neuro-symbolic AI. 🎯 Key Features Pure Python & PyTorch: A clean, modern implementation with no external language dependencies Principled Refinement Engine: Neural networks produce initial fuzzy knowledge bases, refined through gradient-based optimization Massive Scale: Validated on ontologies like SNOMED CT with 377K concepts State-of-the-art Performance: Achieves up to 42% relative improvement in Hits@1 on KBC tasks 🛠️ Installation & Setup Prerequisites The framework is built entirely in Python. All experiments were designed to be run on servers equipped with NVIDIA GPUs. | Component | Version | Purpose | |:----------|:--------|:--------| | Hardware | NVIDIA RTX 3090+ | Recommended for GPU acceleration | | Python | 3.7.0+ | Core framework language | | PyTorch | 1.12+ | Deep learning backend | | CUDA | 11.6+ | For GPU support | Dependencies # 1. Clone the repository git clone https://github.com/your-username/DF-EL-plus-plus.git cd DF-EL-plus-plus # 2. Install Python dependencies pip install -r requirements.txt 📦 Full Dependency List torch>=1.12.0 numpy>=1.21.4 pandas>=1.1.3 matplotlib>=3.3.2 scikit-learn>=0.24.0 tqdm>=4.62.0 loguru>=0.6.0 pyparsing>=3.0.6 🔧 Data Preprocessing Normalization of EL++ Ontologies The input EL++ ontology must first be normalized. Our Python script systematically decomposes complex axioms into elementary "normal forms" (NF1-NF6), preserving the original semantics while enabling stable gradient-based learning. # Run the Python-based normalization script python preprocess.py \ --input_ontology ontologies/snomed.owl \ --output_dir data/snomed_normalized 📁 Output Structure The script will create a directory containing the processed data, split into training, validation, and test sets. data/snomed_normalized/ ├── concepts.txt # Concept names ├── roles.txt # Role names ├── individuals.txt # Individual names ├── train.txt # 80% of normalized axioms ├── valid.txt # 10% of normalized axioms └── test.txt # 10% of normalized axioms 🚀 Training & Evaluation 📊 Knowledge Base Completion (KBC) This task evaluates the model's ability to predict missing axioms in an ontology. 1. Training Use train_kbc.py to train a model on a preprocessed dataset. The best model checkpoint will be saved based on validation MRR. # Example: Training DF-EL++ on SNOMED CT python train_kbc.py \ --data_path data/snomed_normalized \ --embedding_dim 200 \ --batch_size 512 \ --save_dir checkpoints/snomed 🔧 Advanced Training Options python train_kbc.py \ --data_path data/snomed_normalized \ --embedding_dim 200 \ --batch_size 512 \ --learning_rate 2e-4 \ --l2_lambda 1e-5 \ --negative_samples 50 \ --max_epochs 100 \ --patience 10 \ --save_dir checkpoints/snomed 2. Evaluation Use evaluate_kbc.py to evaluate a trained checkpoint on the test set. python evaluate_kbc.py \ --model_path checkpoints/snomed/best_model.pt \ --data_path data/snomed_normalized 🖼️ Semantic Image Interpretation (SII) This task refines noisy outputs from perception models to ensure logical consistency. 1. Training (Refinement) The train_sii.py script simulates the output of a perception model and uses the DF-EL++ framework to refine the noisy beliefs according to an ontology. # Example: Refining beliefs for the PASCAL_PART ontology python train_sii.py \ --data_path data/pascal_part_normalized 2. Evaluation The evaluate_sii.py script assesses the quality of the refined beliefs using classification and logical consistency metrics. python evaluate_sii.py \ --data_path data/pascal_part_normalized ⚙️ Hyperparameter Configuration The following tables summarize the optimal hyperparameters found in the paper. 📈 KBC Task Settings | Parameter | Tuning Range | Selected Value (DF-EL++) | Description | |:----------|:-------------|:------------------------------|:------------| | Optimizer | - | Adam | Adaptive learning rate optimization | | Learning Rate | {1e-4, 2e-4, 5e-4} | 2e-4 | Step size for gradient descent | | Embedding Dim | {100, 200, 400} | 200 | Dimensionality of embeddings | | Batch Size | {256, 512, 1024} | 512 (SNOMED) / 1024 (others) | Samples per iteration | | L2 Regularization | - | 1e-5 | Weight decay coefficient | | Negative Samples | - | 50 | Number of negative samples | | Early Stopping | - | 10 epochs (patience on val MRR) | Prevents overfitting | 🎨 SII Task Settings | Parameter | Configuration | Description | |:----------|:-----------------|:------------| | Optimizer | Adam | For refinement optimization | | Learning Rate | 1e-4 | Refinement step size | | Max Epochs | 50 | Per image refinement iterations | | Perception Model | Fast R-CNN (ResNet-50) | Pre-trained backbone | 📄 License This project is licensed under the MIT License - see the LICENSE file for details. ❓ FAQ How to handle OOM errors? Reduce the batch size or embedding dimension: python train_kbc.py --batch_size 256 --embedding_dim 100 Can I use my own ontology? Yes! Just ensure it's in OWL format and run: python preprocess.py --input_ontology your_file.owl Built with ❤️ for the Neuro-Symbolic AI Community