The Future of Education

Poor housing conditions can seriously affect students’ wellbeing and academic engagement [1], but universities rarely have early warnings about who’s struggling badly with accommodation and might need targeted help. We developed a practical early-warning system to predict serious housing dissatisfaction from survey responses — deliberately excluding any questions too close to the outcome to keep results realistic and useful. Using data from a Koblenz university housing survey (N=158), we defined “at-risk” as an overall experience score ≤3 on a 1–5 scale (32.3% positive rate; 51/158). We tested four models with 5-fold stratified cross-validation: Logistic Regression, Random Forest, histogram-based gradient boosting, and MLP. Random Forest performed best (ROC-AUC = 0.803), followed closely by Logistic Regression (0.792). Tuning the decision threshold for real-world use boosted performance: F1 improved from 0.635 to 0.733 for Random Forest and 0.607 to 0.695 for Logistic Regression, with recall reaching 0.802 for both. We also provide deployment-style confusion matrices at fixed thresholds (0.40 for Random Forest; 0.27 for Logistic Regression) to show outreach workload and missed cases. Permutation importance and sparse Logistic Regression highlight the main drivers: mismatched expectations, affordability concerns, housing instability, privacy/space shortages, and poor service responsiveness.