CausNet-partial : 'Partial Generational Orderings' based search for optimal sparse Bayesian networks via dynamic...

Area of Science:

• Computational Biology
• Machine Learning
• Statistical Modeling

Background:

• Developed CausNet, a dynamic programming algorithm for optimal Bayesian networks (BNs) with parent set constraints.
• CausNet efficiently searches BN space for continuous and categorical data with various outcomes.
• Introduced CausNet-partial, a variant optimizing search with partial generational orderings for sparse BNs.

Purpose of the Study:

• To develop and evaluate CausNet-partial for efficient discovery of smaller, sparse optimal Bayesian networks.
• To demonstrate the algorithm's scalability for datasets with thousands of variables.
• To compare CausNet-partial's performance against existing state-of-the-art algorithms.

Main Methods:

• Implemented a dynamic programming search algorithm based on partial generational orderings.
• Tested CausNet and CausNet-partial on synthetic continuous data and the ALARM benchmark discrete Bayesian network.
• Varied partial order parameters in CausNet-partial to assess impact on network discovery and runtime.

Main Results:

• CausNet-partial significantly reduces search space and runtime for finding small, sparse optimal BNs.
• The algorithm outperforms three extensively used state-of-the-art BN discovery algorithms.
• Demonstrated successful application to simulated data and the ALARM network, discovering smaller networks faster.

Conclusions:

• CausNet-partial offers an efficient and scalable approach for identifying optimal sparse Bayesian networks.
• The algorithm supports various data types (continuous, categorical, survival) and scoring options (BIC, Bge).
• Tunable parameters allow control over network size and density, making it suitable for high-dimensional data analysis.