A Multi-Source Unified Dataset for Student Retention Modeling (Education)
Description
The dataset serves as a high-fidelity, curated repository engineered for advanced predictive modeling and institutional research in higher education. Synthesized from three heterogeneous source systems risk_students, student_dropout_v3, and enrollment_outcomes the asset provides a longitudinal perspective on student persistence across 88,518 unique observations. The repository harmonizes demographic, academic, behavioral, and socio-economic variables into a unified flat-file architecture through rigorous ETL protocols. Technically, the dataset demonstrates 100% data density with zero null values across 22 engineered attributes, establishing a "model-ready" state. The cohort exhibits a mean age of 22.16 years and a standardized GPA mean of 2.21 (σ = 1.15). To optimize performance for gradient-based machine learning algorithms, continuous features such as Age, GPA, and Attendance Rate have been Min-Max normalized. The schema includes unique identifiers and lineage tracking (student_id, source), academic metrics (GPA, attendance_rate, attendance_band), and behavioral factors (study_hours_per_day, stress_index, engagement_score). It also incorporates environmental context through indicators for scholarships and part-time employment, culminating in derived predictive indicators: a composite_risk_score and a binary is_at_risk target label. The primary objective of this repository is to facilitate Evidence-Based Decision Making (EBDM) through Early Warning Systems (EWS) and retention analytics. All records have been strictly de-identified, with Personally Identifiable Information (PII) removed to comply with global academic data privacy standards. Delivered in UTF-8 encoded CSV format, the data is optimized for SQL, Python, R, and professional BI platforms. This asset integrates research from the UCI Machine Learning Repository, Kaggle, and institutional repositories to provide a robust foundation for identifying factors affecting student success.
Files
Steps to reproduce
The transition from raw ingestion to the Gold Layer involved a multi-stage engineering pipeline: Ingestion & Source Alignment: Data was extracted from three disparate origins: institutional risk reports (risk_students), historical dropout logs (student_dropout_v3), and administrative enrollment records (enrollment_outcomes). Schema Harmonization: A unified schema was defined to map inconsistent naming conventions and data types across sources into a single standard, using student_id as the primary key for record reconciliation. Data Scrubbing: Records were audited for integrity. Given the 100% completeness of the final output, missing values from upstream sources were either imputed using statistical means or filtered during the transformation phase to maintain a "model-ready" state. Feature Engineering & Scaling: Normalization: Continuous variables were subjected to Min-Max scaling to ensure that features like age and attendance_rate contribute equally to model convergence. Categorical Encoding: Behavioral attributes (e.g., attendance_band) were transformed into ordinal or categorical encodings for computational compatibility. Composite Risk Synthesis: A predictive risk engine was applied to generate the composite_risk_score. This continuous variable was subsequently discretized into the binary target is_at_risk based on validated institutional thresholds. Metadata Enrichment: Technical metadata, including source lineage (source) and processing timestamps (_gold_created_at), were appended to ensure full audibility and data lineage within the DWH.
Institutions
- New Mansoura UniversityDakahlia, Al Mansurah
- Mansoura UniversityDakahlia, Al Mansurah