The widespread adoption of Competency-Based Education (CBE) in higher education faces a fundamental scalability paradox: while CBE promises unprecedented personalization and mastery-driven learning, its implementation is severely bottlenecked by the enormous manual effort required to map individual exercises to competency networks. Current automated approaches fail because they rely on untrustworthy direct similarity matching between exercises and often inconsistent competency descriptions, creating systems that break down across different instructors and institutions. This thesis challenges current practice by asking: can unsupervised machine learning extract the real pedagogical structure directly from the content instructors naturally create?
This work introduces AtlasML, an automated microservice that turns competency mapping into a discovery problem. By clustering exercises based on their underlying semantics and aligning these emergent themes with course competencies, AtlasML builds a robust, objective mapping completely independent of the limitations of manual definitions. The system goes further, continually improving through instructor feedback, evolving as networks change, and revealing previously invisible relationships between competencies.
Methodologically, AtlasML encodes course content into high dimensional vectors, clusters exercises via k-means to figure out latent themes, and aligns these with competencies using optimal matching algorithms based on semantic similarity. Real time adaptation comes via incremental centroid updates from instructor actions, full reclustering if the networks evolves, and on-demand inter competency pattern discovery, all seamlessly integrated into the Artemis learning platform.