Many behavioral processes are observed as categorical trajectories rather than ordinary numeric time series. This work develops methods for representing and clustering densely observed categorical longitudinal data and applies them to social-media behavior.
A social media user’s day-to-day behavior — posting, replying, lurking, resharing — is naturally a sequence of categorical states rather than a continuous measurement. Standard time-series and functional-data tools are built for numeric-valued curves, which makes them a poor fit for this kind of densely observed categorical data.
Treating each category as an arbitrary numeric code, or reducing a user’s trajectory to a handful of summary counts, throws away the sequential and functional structure of the data — the order and duration of behavioral states, not just their frequency.
This project develops a functional-data framework specifically for categorical longitudinal data, using GAM-based modeling together with functional principal component analysis to represent each trajectory as an underlying latent smooth process rather than a raw category sequence.
Once trajectories are represented in this latent functional space, standard clustering methods can be applied to group users (or other units of observation) by the shape of their behavioral trajectories rather than by simple aggregate counts.
The methodology was applied to cluster social media users by their categorical behavioral trajectories, and to compare behavioral patterns across different social media communities.
The methods are implemented in catfda, an open-source R package available on CRAN, co-developed by Chathura.
Chathura co-developed the catfda package and is a co-author on the resulting Journal of the American Statistical Association publication and the related HICSS proceedings paper.
This project’s treatment of behavioral trajectories as functional data connects to the broader statistical-methods theme, and its social-media application draws on the same datasets and diffusion questions as the information-diffusion and influence work.
My contribution
Chathura co-developed the catfda R package, implementing latent-process estimation and clustering methods for categorical longitudinal data, and is a co-author on the resulting JASA publication and HICSS proceedings paper.
Collaborators
Publications
2024
Champon, X., Jayalath, C., Rand, W., Jasser, J., Garibay, I., Garibay, O.
57th Annual Hawaii International Conference on System Science (HICSS)
Applies functional data analysis methods to compare behavioral patterns across different social media communities.
2026
Champon, X., Staicu, A., Weishampel, A., Jayalath, C., Rand, W.
Journal of the American Statistical Association
Develops categorical-valued functional data analysis methods for latent-process estimation and clustering of social media users' behavioral trajectories.
Verification note: The CV's author list spells the fourth author "Jayalah, C." — this appears to be a typo for "Jayalath, C." and has been corrected here; confirm against the official publication. Page range is listed as "1–11" in the CV, which may be an online-first placeholder rather than final pagination — confirm before publishing citation details externally.
Software
catfda
R package
Co-developed
An R package for categorical-valued functional data analysis, implementing methods for latent-process estimation and clustering of densely observed categorical longitudinal data using GAM-based modeling and functional principal component analysis.
View ↗Related