Designing Trustworthy Recommender Systems: A Glass-Box, Interpretable, and Auditable Approach

Recommender systems are widely deployed across digital platforms, yet their opacity raises concerns about auditability, fairness, and user trust. To address the gap between predictive accuracy and model interpretability, this study proposes a glass-box architecture for trustworthy recommendation, designed to reconcile predictive performance with interpretability. The framework integrates interpretable tree ensemble model (Random Forest, XGBoost), an NLP sub-model for tag sentiment, prioritising transparency from feature engineering through to explanation. Additionally, a Reality Check mechanism enforces strict temporal separation and removes already-popular items, compelling the model to forecast latent growth signals rather than mimic popularity thresholds. Evaluated on the MovieLens dataset, the glass-box architectures demonstrated superior discrimination capabilities, with the Random Forest and XGBoost models achieving ROC-AUC scores of 0.92 and 0.91, respectively. These tree ensembles notably outperformed the standard Logistic Regression (0.89) and the neural baseline (MLP model with 0.86). Beyond accuracy, the design implements governance through a multi-layered Governance Stack: (i) attribution and traceability via exact TreeSHAP values, (ii) stability verification using ICE plots and sensitivity analysis across policy configurations, and (iii) fairness audits detecting genre and temporal bias. Dynamic threshold optimisation further improves recall for emerging items under severe class imbalance. Cross-domain validation on Amazon Electronics test dataset confirmed architectural generalisability (AUC = 0.89), demonstrating robustness in sparse, high-friction environments. These findings challenge the perceived trade-off between accuracy and interpretability, offering a practical blueprint for Safe-by-Design recommender systems that embed fairness, accountability, and auditability as intrinsic properties rather than post hoc add-ons.