ExpertBayes: Automatically refining manually built Bayesian networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Statistics

Background:

• Bayesian network structures are typically constructed from data alone, often starting from scratch or using a naive Bayes approach.
• Existing domain knowledge, particularly in fields like medicine, can provide a valuable prior structure for Bayesian networks.
• These prior structures can be refined manually or automatically to enhance model performance.

Purpose of the Study:

• To investigate the impact of minor structural perturbations on expert-defined Bayesian networks.
• To determine if small modifications to established Bayesian network structures can lead to improved classification accuracy.
• To assess the computational cost and preservation of meaning when refining expert-built Bayesian networks.

Main Methods:

• Utilized pre-existing Bayesian networks developed by domain specialists.
• Applied minor structural modifications to these expert-built networks.
• Evaluated the performance of the modified networks as classifiers.

Main Results:

• Minor perturbations to expert-built Bayesian networks resulted in enhanced classifier performance.
• The improvements in classification accuracy were achieved with a very small computational overhead.
• The refined networks largely retained the original intended meaning derived from specialist knowledge.

Conclusions:

• Slightly altering expert-built Bayesian networks is an effective strategy for developing better classifiers.
• This approach offers a computationally efficient method for improving model performance.
• The findings suggest a practical way to leverage and refine existing domain expertise within Bayesian network models.