Penerapan Explainable Boosting Machine Pada Hasil Pencarian Best Matching 25 Sebagai Rekomendasi Kualitas Buku Dengan Penjelasan Fitur

Abstract

The rapid growth of digital book collections creates information overload, complicating book discovery for readers. Traditional BM25 search systems retrieve keyword-relevant documents but completely ignore book quality signals. Furthermore, conventional machine learning re-ranking models act as black-boxes, obscuring their decision-making processes, while post-hoc explanation methods only provide approximations that do not represent the model's exact logic. This study proposes a two-stage ranking architecture integrating an Explainable Boosting Machine (EBM) as a glass-box re-ranker over BM25 initial retrievals. Trained on 13.4 million Goodreads interaction records, the EBM predicts general book quality based on metadata. Evaluation using 50 queries demonstrates that EBM with feature interactions achieves an NDCG@10 of 0.9373, outperforming the BM25 baseline (0.9284) specifically on exploratory genre queries by prioritizing popular, high-quality books. Global interpretability analysis reveals that popularity contributes most significantly to prediction scores, whereas local explanations precisely map individual feature attributions. Ablation testing confirms the faithfulness of these explanations; removing dominant features consistently reduces prediction scores matching their exact shape function values. In conclusion, integrating EBM successfully enhances BM25 ranking quality through community-based signals while providing fully transparent and auditable feature explanations.