Projects
Research directions and representative systems at the intersection of HCI, AI, and visual analytics.
Our research develops human-centered AI methods and interactive systems that help people understand intelligent technologies, make better decisions, and improve human performance. The projects below are organized into three connected directions.
AI for Mental Health and Human Understanding
We combine large language models, behavioral science, and visual analytics to study human behavior and build AI systems that support responsible interpretation, reflection, and professional training.
MIAdvisor
Counselor trainees need to understand not only which motivational interviewing techniques were used, but also how those techniques affected a client's behavior, cognition, and emotions. MIAdvisor uses LLMs to annotate therapy conversations with the Motivational Interviewing Skill Code and turns the annotations into linked visual views. Trainees can inspect interaction dynamics, recurring skill-response patterns, and differences across sessions to support reflective and vicarious learning.
RealCBT and emotional fidelity
Synthetic therapy dialogues are widely used because authentic counseling data are difficult to obtain, yet fluent language does not guarantee realistic therapeutic dynamics. This project introduces RealCBT, a dataset of authentic cognitive behavioral therapy sessions, and adapts an emotion-dynamics framework to compare valence, arousal, and dominance trajectories in real and LLM-generated dialogues. The results reveal important gaps in emotional variability, reactivity, regulation, and counselor-client alignment.
PersonalityLens
LLMs can infer personality traits from language, but their outputs are difficult to interpret without the underlying dialogue context and psychological mechanisms. PersonalityLens connects traits and facets to supporting utterances, shows how they evolve over time, and aligns them with established personality theory. Case studies in fictional and therapy dialogue demonstrate how the system supports multi-level, context-aware interpretation rather than treating personality assessment as a single prediction.
Visual Analytics for Model Reasoning
We design interactive visualizations that make statistical models, assumptions, uncertainty, and performance easier to inspect - particularly for people without advanced statistical expertise.
Argus
Conventional power-analysis tools return a sample-size calculation after researchers specify fixed assumptions, offering little support for reasoning about uncertain design choices. Argus instead simulates experimental outcomes and visualizes how power changes with sample size, effect size, replications, order effects, and other confounds. Its coordinated views and exploration history help researchers compare design scenarios, understand trade-offs, and make more defensible decisions before collecting data.
RegLine
Refining even a simple linear regression model requires repeated transformations, residual checks, assumption verification, and comparisons with earlier models - a process that is especially difficult for novices. RegLine integrates these activities into an iterative visual workspace. It exposes the effects of transformations and influential observations, links residual and assumption diagnostics, and preserves alternative models so users can compare how each refinement changes validity and accuracy.
Structured time-series exploration
Evaluating time-series forecasts requires moving among charts of aggregate accuracy, cross-validation folds, and predicted versus observed values, but conventional tools leave this sequence unstructured. This project constructs a navigation graph from the syntax and semantics of the charts and exposes relevant paths through a contextual menu. The approach helps analysts systematically uncover relationships among metrics, folds, and forecasts instead of examining disconnected plots in an arbitrary order.
Human Performance, Motivation, and Decision-Making
We investigate how interactive systems and representations influence motivation, confidence, physical activity, and judgment.
Exercise-linked game rewards
Traditional exergames require people to exercise while playing, which can limit accessibility and long-term use. This project studies pervasive accumulated context exergames, where physical activity completed earlier is translated into rewards in a later sedentary game. A controlled study shows that explicitly linking exercise performance to game rewards can improve motivation and increase exercise intensity, suggesting a flexible way to connect everyday activity tracking with engaging play.
Multi-state uncertainty visualization
A single statistical chart exposes only part of an uncertain distribution and may support one reasoning task while making another difficult. Multi-state visualizations let people switch between complementary representations, pairing error bars with violin plots, quantile dot plots, or hypothetical outcome plots. A crowdsourced experiment shows that access to multiple coordinated states improves both the accuracy and confidence of probability estimates, informing the design of uncertainty displays for decision-making.
