Using machine learning algorithms to personalize curricula in technical universities with a focus on engineering specialties

Abstract

The research is aimed at evaluating the effectiveness of machine learning algorithms for personalizing curricula in technical universities and justifying the transition from a unified model to adaptive trajectories that take into account the heterogeneity of students' training and learning styles. The empirical base included 874 students of 2-3 courses in the fields of Software Engineering, Mechanical Engineering and Radio Engineering from three leading universities; Digital traces were collected in LMS Moodle/Canvas (timestamps, navigation, interactions with content, task results), entrance test data and academic indicators for two semesters. Methodologically, clustering (K Means, DBSCAN) has been applied to typologize learning profiles, a hybrid recommendation system (collaborative SVD filtering and a content-oriented approach) for selecting materials and assignments, as well as an LSTM model for predicting final grades and dropout risk based on time series of activity; data purification and normalization (Min Max), TF IDF for texts and statistical checks (t criterion, correlations, ANOVA, PCA) were performed. In the experimental groups, a statistically significant increase in academic performance was recorded: in software engineering +15,14%, in mechanical engineering +12,36%, in radio engineering +12,40% at comparable starting levels; the accuracy of the LSTM forecast reached ~85% 4-5 weeks before the end of the semester. Behavioral metrics indicate a qualitative restructuring of learning activity: time in the system +18,06%, interactions with content +42,09%, the share of optional tasks +153,29%, forum activity +118,18%, with strong links optional tasks ↔ final score (r≈0.74) and interactions ↔ final score (r≈0.69). The survey showed high satisfaction with personalization (overall 4.62/5; rate 4.71; relevance 4.58; interface 4.23), and PCA identified an exploratory pattern of behavior as a marker of success. The discussion highlights that personalized presentation reduces cognitive overload, enhances intrinsic motivation, and develops self-regulation; scaling, integration of NLP and generative models for fine diagnosis and task construction, as well as the development of predictive analytics and UX improvements while respecting ethics and data privacy are recommended.