A Semi-supervised Learning Method for Q-Matrix Specification Under the DINA and DINO Model With Independent Structure

Area of Science:

• Educational Measurement
• Psychometrics
• Artificial Intelligence

Background:

• Cognitive diagnosis assessment (CDA) provides formative feedback on student cognitive strengths and weaknesses.
• Accurate Q-matrix specification is crucial for effective CDA, but often challenging in practice.
• Misspecified Q-matrices can significantly impact the accuracy of classifying examinees.

Purpose of the Study:

• To propose a novel semi-supervised learning approach for Q-matrix specification in cognitive diagnosis.
• To introduce an optimal examinee sampling design to improve Q-matrix accuracy.
• To address challenges in specifying the Q-matrix under conjunctive and disjunctive models with independent structures.

Main Methods:

• A semi-supervised learning approach based on the relationship between Q-matrix and R-matrix.
• Utilizing the logical OR operation to express Q-matrix columns via R-matrix columns.
• Implementing an optimal examinee sampling design for efficient Q-matrix specification.

Main Results:

• The proposed method simplifies Q-matrix specification by requiring expert input only for a subset of items.
• Simulation and real data analysis demonstrated high correct recovery rates for Q-matrix entries.
• The optimal examinee sampling design proved effective in conjunction with the new specification method.

Conclusions:

• The novel semi-supervised learning approach offers a promising solution for accurate Q-matrix specification in CDA.
• This method reduces the burden on subject matter experts while improving diagnostic accuracy.
• The findings support the use of this method for enhancing formative assessment and instructional modification.