Bridging Global Pretraining and Similarity-Based Local Fine-Tuning in GAIN for Imputing Sparse Learner Performance Data

Learner performance data (e.g., correct or incorrect responses) from the interaction logs of Intelligent Tutoring Systems (ITSs) are often sparse, hindering accurate predictions of learner performance and the delivery of effective, adaptive feedback. To address the issue of data sparsity in learner performance logs, we introduce a two-stage imputation framework that integrates Bayesian Knowledge Tracing (BKT) with a Generative Adversarial Imputation Network (GAIN). First, we use BKT to estimate four learner parameters including initial mastery, learning rate, guess, and slip, and cluster learners with similar parameter vectors. Subsequently, we train a single GAIN model on all records, then fine-tune it within each cluster to capture local patterns. We evaluated our approach using three lesson datasets from MATHia, a representative ITS online math learning software that personalizes instruction for middle and high school learners, across varying levels of data sparsity ranging from 10% to 90% missing data. Across all sparsity levels, it achieved the lowest Root Mean Square Error (RMSE) compared with classical statistical and recent deep-learning baselines. Cluster-level finetuning provided an additional accuracy gain, underscoring the critical importance of leveraging cognitively similar learner characteristics for robust generative imputation. This proposed GAIN-based global-to-local data imputation approach effectively addresses the challenge of sparse learner performance data, offering substantial potential improvements in learner modeling within ITSs and enabling more precise and contextually informed educational interventions.