Key research themes
1. How can new algorithmic frameworks improve exact and approximate inference efficiency in discrete probabilistic graphical models?
This research area focuses on developing novel algorithms that enable efficient exact and approximate inference in probabilistic graphical models—especially discrete Bayesian networks and Markov chains—by exploiting structural properties and computational mechanisms such as generators and join-tree propagation. The motivation is to overcome computational bottlenecks in conventional inference methods to better handle complex, large-scale discrete domains.
2. How do independence criteria and conditional independence tests affect the accuracy of network structure inference from observational data?
This theme examines the role of independence measures—ranging from classic linear-Gaussian tests to more generalized statistics like Hilbert-Schmidt Independence Criterion (HSIC) and Distance Covariance Criterion (DCC)—in inferring network structures, notably directed acyclic graphs, from data. Central to this theme is improving causal or functional dependency identification by exploiting nonlinear dependencies and non-Gaussian noise, addressing limitations in traditional methods that assume linearity and Gaussianity.
3. What are the challenges and approaches in learning and validating Bayesian network structures from domain-specific expert knowledge, arguments, or passively collected data?
This research area investigates methods to derive Bayesian network structures either from structured arguments or passive data collection, including adversary perspective modeling and network reconstruction with feasibility guarantees. Challenges include bridging expert knowledge represented as arguments to formal probabilistic models, ensuring inferred networks accurately explain input data or traces, and constructing interpretable decision support systems that model belief formation.