This project follows this paper to implement a structure learning algorithm known as K2. Tests on it are done using Insurance network from the bnlearn package.

• src contains the Python code
  • assets folder used for input and output files of Python scripts
  • structure_learning main Python module for structure learning
    • networks sub-module for Bayesian networks
      • bayesian.py Bayesian network classes (Node and Network)
      • cpt_bayesian.py Same classes as in bayesian.py but with conditional probability tables for savings bnlearn's networks
      • net_reader.py Functions to read networks from .net files
    • heuristics.py Heuristics classes for K2 algorithm (Factorial Heuristic and Logarithmic Heuristic)
    • k2.py K2 algorithm implementation
    • samples.py Class for creating and managing samples from a Bayesian network with CPTs
  • main_comparison.py Script to compare K2 algorithm with different parameters and sample sizes
  • min_diff.py Script that tries to find the minimum colliders difference between a K2's generated network and the original network from bnlearn
• docs contains LaTeX source files for documentation

You might need to install Python 3.8, pip, graphviz and gzip for decompressing the bnlearn network. Then, run the following commands:

cd src
wget -O - https://www.bnlearn.com/bnrepository/insurance/insurance.net.gz  | gzip -d > assets/insurance.net
pip install -r requirements.txt
python main_comparison.py
python min_diff.py