Causal Discovery and Reasoning for Continuous Variables with an Improved Bayesian Network Constructed by Locality...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Mining

Background:

• Structure learning in Bayesian networks (BNs) is vital for understanding variable dependencies in datasets.
• Existing methods often struggle with continuous data or computational efficiency.
• There is a need for robust, non-parametric BN learning techniques.

Purpose of the Study:

• To propose a novel, computationally efficient Bayesian network structure learning method for continuous data.
• To integrate kernel density estimation (KDE) with locality sensitive hashing (LSH) for improved performance.
• To develop a Bayesian network classifier (BNC) based on the enhanced learning method.

Main Methods:

• Employed a non-parametric, distribution-free conditional independence (CI) test using KDE.
• Constructed the BN skeleton using mutual information and conditional mutual information.
• Incorporated locality sensitive hashing (LSH) into KDE to accelerate computations.
• Utilized a conditional entropy scoring function and hill-climbing for causal relationship discovery.

Main Results:

• The proposed KDE with LSH significantly improved computational speed compared to traditional KDE.
• The method maintained high fitting accuracy, balancing speed and precision.
• Network structure learning accuracy surpassed traditional methods on benchmark networks.
• The resulting Bayesian network classifier (BNC) showed superior accuracy and interpretability.

Conclusions:

• The integration of LSH with KDE offers a substantial speed-up for BN structure learning with continuous data.
• The proposed method achieves a favorable balance between computational efficiency and learning accuracy.
• The developed BNC provides a competitive and interpretable alternative to conventional classifiers.