LTNcoder
LTNcoder explores neuro-symbolic process anomaly detection by combining classical process mining signals with neural anomaly detection models. The project centers around a notebook-driven workflow for data preparation, Declare rule mining, model training, and evaluation.
What it does
- Generates and prepares process event logs (including BPIC and synthetic variants).
- Mines Declare constraints and transforms them into features usable by downstream models.
- Trains anomaly detectors across multiple datasets and compares strategies/heuristics.
- Evaluates models with reproducible notebook pipelines and database-backed experiment tracking.
How it is built
- Uses a python package structure for datasets, filesystem conventions, evaluators, and anomaly detection modules.
- Integrates TensorFlow/Keras-based models with Logic Tensor Network concepts in custom encoder modules.
- Stores and tracks event logs/models/results with structured naming and SQLite-backed metadata.
- Provides dataset-specific training/evaluation notebooks (e.g., paper, small/large, BPIC variants) for controlled experiments.
Tech stack
Python, TensorFlow/Keras, Logic Tensor Network tooling, SQLAlchemy, scikit-learn, Jupyter notebooks, process-mining libraries.