5.2. Limitations of Current Predictive Architectures and Research Gaps
The synthesis of contemporary scholarship underscores fundamental tensions between data-driven retention modeling and learner governance. While educational data mining tools offer customized insights into student behaviors (2023), their deployment within distance learning frequently suffers from fragmented, non-repeatable architectures that fail to generalize across diverse academic environments (2026). The implementation of automated data and machine learning pipelines provides a structured mechanism to capture multi-layered learner attributes (2026); nonetheless, existing predictive frameworks exhibit critical methodological vulnerabilities when learner autonomy alters the underlying data corpus. A primary limitation of prevailing predictive systems is the unaddressed sampling distortion induced by voluntary participation. Most algorithmic models assume comprehensive baseline records, yet institutional pipelines struggle to sustain predictive validity when cohorts exercise opt-out rights. Furthermore, the persistent reliance on black-box machine learning algorithms impedes actionable institutional interventions, creating an acute research gap at the intersection of algorithmic transparency and predictive performance (2025). Although combining process mining with explainable artificial intelligence has been proposed to enhance model interpretability (2025), current research rarely evaluates how explainability frameworks behave when data pipelines are truncated by selective user consent. Consequently, higher education institutions face an unresolved trade-off: machine learning pipelines can automate early attrition identification (2026), but their analytical utility remains constrained without models specifically calibrated for missing-not-at-random consent patterns. Future research must bridge this empirical gap by developing pipeline architectures that integrate explainable artificial intelligence directly into robust, consent-aware predictive frameworks.