Improved Scalability and Accuracy of Bayesian Network Structure Learning in the Score-and-Search Paradigm
MetadataShow full item record
A Bayesian network is a probabilistic graphical model that consists of a directed acyclic graph (DAG), where each node is a random variable and attached to each node is a conditional probability distribution (CPD). A Bayesian network (BN) can either be constructed by a domain expert or learned automatically from data using the well-known score-and-search approach, a form of unsupervised machine learning. Our interest here is in BNs as a knowledge discovery or data analysis tool, where the BN is learned automatically from data and the resulting BN is then studied for the insights that it provides on the domain such as possible cause-effect relationships, probabilistic dependencies, and conditional independence relationships. Previous work has shown that the accuracy of a data analysis can be improved by (i) incorporating structured representations of the CPDs into the score-and-search approach for learning the DAG and by (ii) learning a set of DAGs from a dataset, rather than a single DAG, and performing a technique called model averaging to obtain a representative DAG. This thesis focuses on improving the accuracy of the score-and-search approach for learning a BN and in scaling the approach to datasets with larger numbers of random variables. We introduce a novel model averaging approach to learning a BN motivated by performance guarantees in approximation algorithms. Our approach considers all optimal and all near-optimal networks for model averaging. We provide pruning rules that retain optimality while enabling our approach to scale to BNs significantly larger than the current state of the art. We extend our model averaging approach to simultaneously learn the DAG and the local structure of the CPDs in the form of a noisy-OR representation. We provide an effective gradient descent algorithm to score a candidate noisy-OR using the widely used BIC score and we provide pruning rules that allow the search to successfully scale to medium sized networks. Our empirical results provide evidence for the success of our approach to learning Bayesian networks that incorporate noisy-OR relations. We also extend our model averaging approach to simultaneously learn the DAG and the local structure of the CPD using neural networks representations. Our approach compares favourably with approaches like decision trees, and performs well in instances with low amounts of data. Finally, we introduce a score-and-search approach to simultaneously learn a DAG and model linear and non-linear local probabilistic relationships between variables using multivariate adaptive regression splines (MARS). MARS are polynomial regression models represented as piecewise spline functions. We show on a set of discrete and continuous benchmark instances that our proposed approach can improve the accuracy of the learned graph while scaling to instances with over 1,000 variables.
Cite this version of the work
Charupriya Sharma (2023). Improved Scalability and Accuracy of Bayesian Network Structure Learning in the Score-and-Search Paradigm. UWSpace. http://hdl.handle.net/10012/19450