Self-improvement for Computerized Adaptive Testing
Research output: Contributions to collected editions/works › Article in conference proceedings › Research › peer-review
Standard
Machine Learning and Knowledge Discovery in Databases - Research Track: European Conference, ECML PKDD 2025, Porto, Portugal, September 15–19, 2025, Proceedings. ed. / Rita P. Ribeiro; Alípio M. Jorge; Carlos Soares; João Gama; Bernhard Pfahringer; Nathalie Japkowicz; Pedro Larrañaga; Pedro H. Abreu. Vol. 2 Cham: Springer International Publishing, 2026. p. 70-86 (Lecture Notes in Computer Science; Vol. 16014 LNCS).
Research output: Contributions to collected editions/works › Article in conference proceedings › Research › peer-review
Harvard
APA
Vancouver
Bibtex
}
RIS
TY - CHAP
T1 - Self-improvement for Computerized Adaptive Testing
AU - Rudolph, Yannick
AU - Neubauer, Kai
AU - Brefeld, Ulf
N1 - Publisher Copyright: © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.
PY - 2025/9/23
Y1 - 2025/9/23
N2 - Computerized adaptive testing (CAT) allows for assessing latent traits and abilities of students with fewer items and in less time due to an individualized item selection algorithm based on previous responses. Following recent machine learning solutions to CAT, we study learning both the underlying response model for cognitive diagnosis and a policy for the item selection algorithm jointly from offline training data. While the task of the response model is to predict performances on all unseen items for a user, the goal of the policy is to select the subset of items which maximizes information for the response model. Since subset selection is a combinatorial problem, we propose to leverage an iterative self-improvement approach to policy learning from the field of neural combinatorial optimization while accounting for interdependencies between response model and policy. We specifically focus on the generalization capabilities of transformer-based models and, in contrast to related work, do not rely on optimization of local variables during inference. We report on empirical results.
AB - Computerized adaptive testing (CAT) allows for assessing latent traits and abilities of students with fewer items and in less time due to an individualized item selection algorithm based on previous responses. Following recent machine learning solutions to CAT, we study learning both the underlying response model for cognitive diagnosis and a policy for the item selection algorithm jointly from offline training data. While the task of the response model is to predict performances on all unseen items for a user, the goal of the policy is to select the subset of items which maximizes information for the response model. Since subset selection is a combinatorial problem, we propose to leverage an iterative self-improvement approach to policy learning from the field of neural combinatorial optimization while accounting for interdependencies between response model and policy. We specifically focus on the generalization capabilities of transformer-based models and, in contrast to related work, do not rely on optimization of local variables during inference. We report on empirical results.
KW - Computerized adaptive testing
KW - Educational data mining
KW - Neural combinatorial optimization
KW - Self-improvement
UR - http://www.scopus.com/inward/record.url?scp=105019304024&partnerID=8YFLogxK
U2 - 10.1007/978-3-032-05981-9_5
DO - 10.1007/978-3-032-05981-9_5
M3 - Article in conference proceedings
SN - 978-3-032-05980-2
VL - 2
T3 - Lecture Notes in Computer Science
SP - 70
EP - 86
BT - Machine Learning and Knowledge Discovery in Databases - Research Track
A2 - Ribeiro, Rita P.
A2 - Jorge, Alípio M.
A2 - Soares, Carlos
A2 - Gama, João
A2 - Pfahringer, Bernhard
A2 - Japkowicz, Nathalie
A2 - Larrañaga, Pedro
A2 - Abreu, Pedro H.
PB - Springer International Publishing
CY - Cham
ER -