Uncovering Latent Patterns in Program-Associated Library Attendance through Background Adjustment
Abstract
Routinely collected library entrance records offer an opportunity to understand attendance patterns associated with library programs, but ordinary daily variation and detailed administrative user status categories complicate their interpretation. Building on a project-based learning collaboration with Kyushu University Central Library, we analyzed entrance records from October 2018 to July 2026. Descriptive analysis revealed relatively modest differences in attendance across the main weather categories, higher attendance on days with mean temperatures of 30°C or above, and pronounced differences between weekdays and weekends. To account for such background variation, we estimated expected attendance for each user status category and applied nonnegative matrix factorization (NMF) to identify common patterns in positive background-adjusted responses from 92 events across 39 user status categories. Three interpretable response modes were identified: two concentrated on temporary external visitors and undergraduates, respectively, and one representing a broader combination of graduate student and research-related categories. The three-mode solution was selected by jointly considering held-out reconstruction error and basis stability across year-excluded fits. The identified modes also remained stable after excluding events from the selected COVID-19 period. The modes appeared in different combinations across individual events, with their dominance varying by program category. These findings illustrate how existing library records can reveal interpretable patterns of program-associated attendance without predefined grouping of user status categories.
Related
- Attendance Information System Success in Mandatory Use Contexts: An Integrated TAM–DeLone and McLean–TPB Framework in Higher Education
- Evaluating a candidate predictor beyond statistical significance: Atmospheric pressure and psychiatric day-care attendance as a worked example
- Regression adjustment in completely randomized experiments with many covariates
- Automated, Efficient, and Model-Free Inference for Randomized Clinical Trials via Data-Driven Covariate Adjustment
- Consumption adjustment with mean-reverting risk premia: A singular stochastic control approach
BibTeX
@misc{kim2026uncovering,
title = {{Uncovering Latent Patterns in Program-Associated Library Attendance through Background Adjustment}},
author = {Keunsu Kim and Hisatoshi Kodani and Motoki Masada and Nam Tuan Nguyen},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.34150},
eprint = {2610.34150},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.34150},
}
RIS
TY - UNPB
TI - Uncovering Latent Patterns in Program-Associated Library Attendance through Background Adjustment
AU - Keunsu Kim
AU - Hisatoshi Kodani
AU - Motoki Masada
AU - Nam Tuan Nguyen
PY - 2026
DA - 2026/10/25
PB - arXiv
DO - 10.48550/arxiv.2610.34150
UR - https://arxiv.org/abs/2610.34150
AB - Routinely collected library entrance records offer an opportunity to understand attendance patterns associated with library programs, but ordinary daily variation and detailed administrative user status categories complicate their interpretation. Building on a project-based learning collaboration with Kyushu University Central Library, we analyzed entrance records from October 2018 to July 2026. Descriptive analysis revealed relatively modest differences in attendance across the main weather categories, higher attendance on days with mean temperatures of 30°C or above, and pronounced differences between weekdays and weekends. To account for such background variation, we estimated expected attendance for each user status category and applied nonnegative matrix factorization (NMF) to identify common patterns in positive background-adjusted responses from 92 events across 39 user status categories. Three interpretable response modes were identified: two concentrated on temporary external visitors and undergraduates, respectively, and one representing a broader combination of graduate student and research-related categories. The three-mode solution was selected by jointly considering held-out reconstruction error and basis stability across year-excluded fits. The identified modes also remained stable after excluding events from the selected COVID-19 period. The modes appeared in different combinations across individual events, with their dominance varying by program category. These findings illustrate how existing library records can reveal interpretable patterns of program-associated attendance without predefined grouping of user status categories.
ER -