A1945
Title: Model-based clustering with item response and textual data
Authors: Thomas Chan - Hong Kong University of Science and Technology (Hong Kong) [presenting]
Mike So - The Hong Kong University of Science and Technology (Hong Kong)
Amanda Chu - The Education University of Hong Kong (China)
Chin Sum Shui - National Yang Ming Chiao Tung University (Taiwan)
Abstract: A flexible model-based clustering framework is developed that integrates mixed-type data, including ordinal, nominal, and textual items, into a unified Item Response Theory (IRT) model. By assuming that latent traits follow a mixture of elliptical distributions, the framework provides a robust alternative to standard Gaussian mixtures, effectively accommodating heavy-tailed data and outliers. The model captures complex Response patterns through Item group factors and spline smoothing, while qualitative textual data are incorporated by embedding topic models directly within the latent trait space. To ensure an interpretable and parsimonious cluster structure, a repulsive normal prior is employed to mitigate the risk of redundant or overlapping clusters. Model estimation is conducted via a variational Bayes approach optimized with the Adam algorithm, facilitating efficient computation for high-dimensional data. Simulation results indicate that the proposed model outperforms classical multidimensional IRT and latent Dirichlet allocation (LDA) in both clustering accuracy and adjusted mutual information. The methodology is applied to a real-world dataset of shuttle bus service evaluation, demonstrating its ability to identify distinct respondent profiles and provide data-driven recommendations for service optimization.
