Research

Main Research Area

Current Ongoing Research Projects

Statistical Network Analysis

We examine how individual traits interact with surrounding social networks using a mediation framework, aiming to disentangle the simultaneous mechanisms of social influence and social selection.
We develop a mediation analysis framework for high-dimensional spatial transcriptomics data to characterize how spatial exposures are associated with biological outcomes through coordinated transcriptional variation.
We extend network mediation analysis by embedding the latent network structure directly into the causal pathway. By incorporating it within a hierarchical data structure, this framework enables the separate identification of mediation effects at both the group level and the overall population level.

We develop a hyperbolic latent space item response model that embeds both respondents and items in hyperbolic space, for bipartite binary data in which items are organized in a hierarchical, tree-like structure. An exact Gromov-product decomposition separates each item's depth in the hierarchy from a residual pairwise term that captures the item's branching relations to other items, and simulations show that the model recovers this hierarchical structure more accurately than a Euclidean model.
We develop an asymmetric dynamic latent space model for directed weighted networks that separates changes in relational geometry from changes in each actor’s inward and outward reach. Applied to international trade and global value chain flow networks, the framework characterizes how economies’ relational positions and directional roles evolve across layers and over time.
We combine the Weisfeiler–Lehman kernel values from statistically projected item graphs with latent distances that capture similarity between respondent network positions, providing interpretable insights into homophily and changes in edge states.

- We propose a novel phylogenetic inference framework that compares model-based Bayesian phylogenetics with the geometric perspective of a Hyperbolic Latent Space Joint Model (HLSJM), reconstructing evolutionary trees from inter-language distance structures derived from hyperbolic embeddings.

We develop scalable and fast graph neural network methods for link prediction and node classification, with particular emphasis on solving statistical problems on current evaluation methods.

We focus on an enhanced item response model that incorporates log data from computer-based testing. The current research involves embedding methods, dimensionality reduction, and statistical analysis of the resulting embedding vectors.
We investigate cross-national differences in behavioral transition structures using problem-solving log data. Multi-state survival models are applied to estimate transition probabilities, which are then embedded into network representations to analyze behavioral similarities and differences across countries.
We develop nonparametric two-sample hypothesis tests for sequential data by embedding sequences into a signature-kernel-based RKHS, enabling rigorous comparison of complex temporal patterns such as clickstreams.
We develop an objective-free Bayesian nonparametric clustering framework for process log data by combining a Mixture of Finite Mixtures (MFM) prior with a Beta stochastic block model. The method identifies latent behavioral groups from similarities in problem-solving action sequences while allowing the number of clusters to be inferred from the data. We apply the framework to PIAAC data to identify distinct latent behavioral groups and characterize their problem-solving strategies.

We develop a novel Phase I clinical trial methodology by accounting for both dose-limiting and low-grade toxicities, refining dose escalation strategies and enhancing patient safety.

We develop a manifold-constrained Bayesian singular value decomposition (MC-BSVD) framework for analyzing single-cell RNA sequencing data. The model combines a zero-inflated negative binomial likelihood with low-dimensional gene and cell embeddings, accounting for overdispersion and excess zeros. Through a sampling algorithm that preserves orthogonality and centering constraints, the framework jointly estimates the embeddings and their latent dimension while quantifying uncertainty.
We develop a Bayesian Gram model for continuous bipartite data that separates covariate-explained interactions from orthogonal residual structure while accounting for additive node effects and incomplete observations. The framework incorporates shrinkage and data-driven rank selection to recover latent interaction structures and provide robust prediction under complex dependency and missingness patterns.

Collaborative Projects

We investigate how prefrontal functional connectivity relates to loneliness, perceived stress, and social support across population-scale HCP resting-state fMRI and independent fNIRS data. Using Bayesian sparse reduced-rank regression, we identify a reproducible low-dimensional axis linking greater social distress with lower perceived support and assess its robustness across modalities.
We develop a statistical framework for modeling temporal heterogeneity in multivariate neuroimaging data by combining hierarchical latent-state modeling with individual-level sequence analysis. This framework enables simultaneous characterization of population-level dynamic structure and subject-specific variation in latent state trajectories.
We use fNIRS data collected during the Iowa Gambling Task (IGT) and apply a Hidden Markov Model (HMM) to identify latent brain states and examine how their temporal dynamics are associated with decision-making behavior and individual psychological characteristics.

We investigate how large vision-language models (LVLMs) understand social interactions depicted in images and how their judgments compare with those of humans. Beyond assessing whether their judgments agree, we seek to identify the visual and relational factors that lead humans and models to interpret the same scene differently. To this end, we use network-based and attention-based approaches to analyze the relational structure of each scene and the elements that models attend to. Ultimately, this research seeks to provide a more nuanced account of where human and LVLM perceptions of social interaction converge and diverge.

Page updated

Google Sites

Report abuse