TY - GEN
T1 - Developing Efficient Bayesian Estimation of IRT Models for Integrated STEM Education
AU - Sheng, Yanyan
AU - Welling, William S.
AU - Zhu, Michelle M.
N1 - Publisher Copyright: © 2023 IEEE.
PY - 2023
Y1 - 2023
N2 - Item response theory (IRT) is a popular approach used for addressing psychometric problems in educational and psychological measurement. Its use in large-scale assessments typically involves a calibration stage where a large and representative sample is needed to ensure the accuracy in estimating item parameters. This is, however, difficult to achieve in small-scale or classroom settings, especially when immediate feedback is desired. The problem can be resolved by combining existing and newly collected item response data to simultaneously estimate both item parameters and person abilities, which require a complex estimation procedure and an efficient algorithm. The complex estimation of IRT models via fully Bayesian approach is usually computationally expensive due to the large number of iterations, and a large amount of memory to store massive amount of data. This limits the use of the procedure in small-scale time sensitive or large-scale applications using traditional CPU architecture. In an effort to overcome such restrictions, previous studies focused on utilizing high performance computing using either distributed memory based message passing interface (MPI) or massive threads compute unified device architecture (CUDA) to achieve certain speedups with a simple IRT model where one latent trait is assumed. This study focuses on such models and aims at demonstrating the scalability of parallel algorithms integrating CUDA into MPI computing paradigm. Results of this study further sheds light on applications of IRT in integrated STEM education.
AB - Item response theory (IRT) is a popular approach used for addressing psychometric problems in educational and psychological measurement. Its use in large-scale assessments typically involves a calibration stage where a large and representative sample is needed to ensure the accuracy in estimating item parameters. This is, however, difficult to achieve in small-scale or classroom settings, especially when immediate feedback is desired. The problem can be resolved by combining existing and newly collected item response data to simultaneously estimate both item parameters and person abilities, which require a complex estimation procedure and an efficient algorithm. The complex estimation of IRT models via fully Bayesian approach is usually computationally expensive due to the large number of iterations, and a large amount of memory to store massive amount of data. This limits the use of the procedure in small-scale time sensitive or large-scale applications using traditional CPU architecture. In an effort to overcome such restrictions, previous studies focused on utilizing high performance computing using either distributed memory based message passing interface (MPI) or massive threads compute unified device architecture (CUDA) to achieve certain speedups with a simple IRT model where one latent trait is assumed. This study focuses on such models and aims at demonstrating the scalability of parallel algorithms integrating CUDA into MPI computing paradigm. Results of this study further sheds light on applications of IRT in integrated STEM education.
KW - CUDA-Aware MPI
KW - Gibbs sampling
KW - high performance computing
KW - item response theory
UR - http://www.scopus.com/inward/record.url?scp=85184851355&partnerID=8YFLogxK
U2 - 10.1109/ISEC57711.2023.10402348
DO - 10.1109/ISEC57711.2023.10402348
M3 - Conference contribution
T3 - 13th IEEE Integrated STEM Education Conference, ISEC 2023
SP - 267
EP - 270
BT - 13th IEEE Integrated STEM Education Conference, ISEC 2023
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 13th IEEE Integrated STEM Education Conference, ISEC 2023
Y2 - 11 March 2023
ER -