The CAPIRE Curriculum Graph: Structural Feature Engineering for Curriculum-Constrained Student Modelling in Higher Education

TL;DR

This paper introduces the CAPIRE Curriculum Graph, a structural feature engineering approach that models curricula as directed graphs to improve student attrition prediction in higher education.

Contribution

It formalizes curricula as directed acyclic graphs and derives structural features that enhance predictive models beyond macro-context variables.

Findings

Structural curriculum features improve prediction accuracy.

Macro-context socioeconomic variables do not significantly enhance models.

All structural features contribute synergistically to performance.

Abstract

Curricula in long-cycle programmes are usually recorded in institutional databases as linear lists of courses, yet in practice they operate as directed graphs of prerequisite relationships that constrain student progression through complex dependencies. This paper introduces the CAPIRE Curriculum Graph, a structural feature engineering layer embedded within the CAPIRE framework for student attrition prediction in Civil Engineering at Universidad Nacional de Tucuman, Argentina. We formalise the curriculum as a directed acyclic graph, compute course-level centrality metrics to identify bottleneck and backbone courses, and derive nine structural features at the student-semester level that capture how students navigate the prerequisite network over time. These features include backbone completion rate, bottleneck approval ratio, blocked credits due to incomplete prerequisites, and graph distance to graduation. We compare three model configurations - baseline CAPIRE, CAPIRE plus macro-context variables, and CAPIRE plus macro plus structural features - using Random Forest classifiers on 1,343 students across seven cohorts (2015-2021). While macro-context socioeconomic indicators fail to improve upon the baseline, structural curriculum features yield consistent gains in performance, with the best configuration achieving overall Accuracy of 86.66% and F1-score of 88.08% and improving Balanced Accuracy by 0.87 percentage points over a strong baseline. Ablation analysis further shows that all structural features contribute in a synergistic fashion rather than through a single dominant metric. By making curriculum structure an explicit object in the feature layer, this work extends CAPIRE from a multilevel leakage-aware framework to a curriculum-constrained prediction system that bridges network science, educational data mining, and institutional research.