This project explores the application of Bayesian Networks (BNs) for diabetes risk assessment. Various Bayesian structures, including Naïve Bayes, Hill Climbing (BIC, K2, BDeu), Simulated Annealing, and a Domain Knowledge-Based model, are developed and evaluated to determine the most effective framework for prediction.
The dataset used in this study can be accessed here: Diabetes Health Indicators Dataset.
- Feature Selection: SHAP values from a trained XGBoost classifier are used to identify important features. Additionally, features are ranked based on Mutual Information (MI) to gain deeper insights into their importance.
- Structure Learning: Bayesian Networks are constructed using both data-driven and domain knowledge-based approaches.
- Parameter Estimation: Maximum Likelihood Estimation (MLE) is applied to estimate Conditional Probability Distributions (CPDs).
- Evaluation: Models are assessed using the AUC-ROC evaluation metric to identify top-performing models.
- Inference: The top-performing Bayesian Networks are employed for probabilistic reasoning, enabling predictions and risk assessments based on given evidence, with Variable Elimination used as the inference method.
-
feature selection
shap_and_mi.py: Trains an XGBoost classifier on the dataset and analyzes feature importance using SHAP values with TreeExplainer. Then, applies Scikit-learn's mutual information to rank features and finally examines the correlation between the top 10 features identified by each method.
-
bayesian network modeling
ai_based_structure_learning.py: Data-driven Bayesian Network construction using various AI techniques.domain_knowledge_driven.py: Manually designs the Bayesian Network structure using insights from the research literature.- For both Bayesian Modeling approaches,
pgmpylibrary is used.
-
evaluation & inference
evaluation_and_inference.py: Performs the parameter estimation using MLE, evaluates models using AUC-ROC metric, and then performs probabilistic inference using Variable Elimination.
-
models
- Contains saved checkpoints (.pkl) of the learned models required for the
diabetes_risk_analysis_(lightweight).ipynbnotebook.
- Contains saved checkpoints (.pkl) of the learned models required for the
-
dataset
-
diabetes_indicators.csv: Includes the cleaned dataset before feature selection. -
df_selected.csv: Provides the dataset containing only the selected features.
Clone the repository and install the required packages:
git clone https://github.com/faezesarlakifar/Unibo-FAIKR-M3-project
cd "Unibo-FAIKR-M3-project"
pip install -r requirements.txtRun Inference
cd "evaluation & inference"
python inference.py- You can directly access the main notebook with all the experiment results here: diabetes_risk_analysis_(lightweight).ipynb
- A notebook containing all existing evaluation steps (each variable elimination) is
Diabetes_Risk_Prediction.ipynb. Since it has a large file size, a lightweight version is provided for easier exploration."
Top features are selected based on this result.
This plot suggests that individuals with higher education levels have better general health. (Education levels range from 1 lowest to 6 highest, while GenHealth values are in reverse order, with 5 indicating the poorest health and 1 the best). More insightful plots are available in the feature engineering section of the main notebook. 🙂