Chrishtina Tharushi Fernandopulle, a Master of Science in Statistics candidate, will defend the thesis titled Structure Learning in Bayesian Networks: From Personality Modeling to Scalable Algorithms on Monday, September 21, 2026 from 1:00 PM to 3:30 PM in Plaza 601C.
The examination committee includes Melanie Pilkington, Chair; Jingjing Wu (University of Calgary), External Examiner; Xiaojian Xu, Supervisor; Pouria Ramazi, Supervisory Committee Member.
Abstract: Bayesian network structure learning provides a powerful framework for uncovering conditional dependency relationships in complex multivariate systems. However, its application presents two major challenges: interpreting dependency structures in real-world domains and achieving computational scalability for high-dimensional networks. This thesis addresses both challenges through an application study in personality psychology and the development of a scalable structure-learning framework. First, Bayesian network structure learning is applied to investigate age-related differences in the organization of HEXACO personality traits. Using a dataset of 98,917 respondents aged 3 to 100 years, separate network structures are learned for five developmental stages: childhood, adolescence, early adulthood, middle adulthood, and older adulthood. The resulting networks reveal substantial age-dependent variation in the conditional dependencies among personality traits. In particular, trait interactions become increasingly complex from childhood to early adulthood before simplifying in later life stages, while the relationship between Honesty-Humility and Agreeableness remains consistently prominent across the lifespan. These findings suggest that personality development involves not only changes in individual traits but also changes in how personality domains cohere as an integrated profile. Second, this thesis introduces Shield-Induced Partitioning (SIP), a recursive framework for scalable Bayesian network structure learning. SIP decomposes the variable set into two main blocks and a shield, enabling independent learning on smaller local domains while preserving the conditional dependency structure of the original network. Theoretical results establish conditions under which the proposed recursive learning procedure recovers the target P-map equivalence class under the standard Markov and faithfulness assumptions. Experimental evaluation on benchmark Bayesian networks demonstrates that SIP achieves runtime improvements relative to baseline structure-learning approaches, while maintaining structural accuracy comparable to PC. Together, these contributions demonstrate the utility of Bayesian network structure learning for uncovering complex dependency patterns and the potential of node partitioning for improving its scalability.
Keywords: Bayesian Networks; Structure Learning; HEXACO Personality Traits; Personality Development; Conditional Independence; Node Partitioning; Scalable Learning Algorithms
