Sarcouncil Journal of Multidisciplinary

Abstract: Machine learning interpretability has emerged as a critical requirement in high-stakes applications where transparent decision-making processes are essential for regulatory compliance and expert validation. This article presents a novel framework for implementing monotonicity constraints in multiclass XGBoost systems through the one-vs-rest decomposition strategy, addressing the fundamental challenge of maintaining interpretable feature-prediction relationships across multiple class outputs simultaneously. The proposed framework introduces a constraint matrix representation that enables class-specific monotonicity specifications while ensuring global consistency through sophisticated coordination algorithms. The implementation modifies the traditional XGBoost tree-building process to incorporate per-class constraint-aware splitting criteria that balance predictive accuracy with interpretability requirements. Advanced loss function modifications augment standard multiclass cross-entropy with adaptive penalty terms that quantify constraint violations across all classes. This is a modification of the current XGBoost multi-class implementation, where only one constraint relationship can be defined for all the classes, limiting accurate business-aligned interpretability. Benchmarking results demonstrate that the framework successfully maintains interpretability without significant accuracy degradation, particularly excelling in domains with inherent monotonic relationships such as credit scoring and medical diagnosis. The constraint enforcement during training provides stronger guarantees about model behavior compared to post-hoc interpretability methods, ensuring that monotonic relationships are preserved throughout the model lifecycle rather than approximated after training completion.