This work targets persistent hurdles in practice-oriented instruction at the university level by combining a Naive Bayes student classifier with an upgraded k-Nearest Neighbors (kNN) recommender. It stratifies learners via Naive Bayes and forms cohorts from the resulting classes. Using these cohorts, the paper matches topic vectors to student interest vectors through a refined kNN procedure. Topics are encoded using an optimal binary tree scheme that steers recommendations toward each cohort’s collective preferences. In controlled experiments, the framework consistently selected discussion topics that fit heterogeneous groups and measurably increased participation in practical classes. Compared with user- and item-based collaborative filtering baselines, it achieved higher accuracy, precision, recall, and F1 across all test settings.