ETD

• explainability
• fairness
• learning to rank
• machine learning

The thesis proposes an interpretable-by-design Learning-to-Rank framework tailored for sensitive decision-making scenarios, where fairness and auditability are essential. The proposed architecture combines a Shallow Regression Tree, which provides an initial estimation of the relevance, with a Pairwise Distance Tree that refines the ordering through interpretable pairwise comparisons. This design ensures that every ranking decision can be explained through explicit decision rules.
The model is evaluated on popular benchmark datasets and on synthetic HR datasets that simulate job-candidate ranking. A comprehensive analysis of hyperparameter interactions is included, followed by comparisons with state-of-the-art methods such as LambdaMART and LambdaRank. Results show that the proposed model achieves competitive ranking performance while maintaining a high degree of interpretability.