Semisupervised Learning Method to Adjust Biased Item Difficulty Estimates Caused by Nonignorable Missingness in a...

Area of Science:

• Educational Measurement
• Psychometrics
• Data Science

Background:

• Item response theory (IRT) models are crucial for measuring student ability in virtual learning environments (VLEs).
• Accurate IRT application depends on unbiased item parameter estimates, which are often compromised by significant missing data in VLEs.
• Existing methods for handling missing data are often inadequate for the large-scale missingness typical of VLE data.

Purpose of the Study:

• To introduce a novel semisupervised learning method for estimating unbiased item parameters in VLE data with substantial missingness.
• To address the negative impact of missing data on the accuracy of student ability estimates in VLEs.
• To provide a framework for reliable, ongoing student ability measurement in VLEs.

Main Methods:

• Exploration of factors contributing to missing data in VLE logs.
• Implementation of a semisupervised learning approach within the two-parameter logistic IRT model.
• Application of two bias adjustment techniques to refine item parameter estimates.

Main Results:

• The proposed semisupervised learning framework demonstrates potential for obtaining unbiased item parameter estimates.
• The method effectively handles large proportions of missing data common in VLEs.
• Improved accuracy in item parameter estimation leads to more reliable student ability estimates.

Conclusions:

• The developed semisupervised learning method offers a viable solution for the challenge of missing data in VLEs.
• This framework enables more accurate and dependable ongoing measurement of student abilities.
• The approach facilitates the operational use of IRT models in virtual learning environments.