APPLICATION OF EXPLAINABLE BOOSTING MACHINE TO BEST MATCHING 25 SEARCH RESULTS FOR BOOK QUALITY RECOMMENDATION WITH FEATURE-LEVEL EXPLANATIONS

Wicaksono, Bagus Satrio (2026) APPLICATION OF EXPLAINABLE BOOSTING MACHINE TO BEST MATCHING 25 SEARCH RESULTS FOR BOOK QUALITY RECOMMENDATION WITH FEATURE-LEVEL EXPLANATIONS. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 22081010252.-cover.pdf Download (3MB)
	Text (BAB 1) 22081010252.-bab1.pdf Download (348kB)
	Text (BAB 2) 22081010252.-bab2.pdf Restricted to Repository staff only until 17 July 2029. Download (605kB)
	Text (BAB 3) 22081010252.-bab3.pdf Restricted to Repository staff only until 17 July 2029. Download (1MB)
	Text (BAB 4) 22081010252,-bab4.pdf Restricted to Repository staff only until 17 July 2029. Download (2MB)
	Text (BAB 5) 22081010252.-bab5.pdf Download (335kB)
	Text (Daftar Pustaka) 22081010252.-daftarpustaka.pdf Download (347kB)

Abstract

The rapid growth of digital book collections creates significant challenges for readers identifying high-quality books during exploratory search. Although Best Matching 25 (BM25) effectively retrieves textually relevant books, it ignores quality indicators like popularity and ratings, while existing re-ranking models remain uninterpretable black boxes. This study proposes a two-stage ranking architecture deploying an Explainable Boosting Machine (EBM) as a glass-box reranker to optimize BM25 search results while providing intrinsic explanations. The EBM was trained on 13.4 million user-book interactions from the UCSD Goodreads dataset. To evaluate ranking quality and feature impacts, EBM variants—evaluating main effects (int0) and five feature interactions (int5)—were tested. Prediction evaluations showed the int0 model achieved an RMSE of 0.9782 and an MAE of 0.7632. Ranking evaluations across 50 test queries demonstrated that the int5 model achieved an NDCG@10 of 0.9373, outperforming the BM25 baseline (0.9284) on exploratory queries. Specifically, the framework successfully enhanced ranking quality by elevating candidate books that exhibit strong popularity signals. Human judgment evaluation revealed a marginal 0.05 decrease in Precision@1 (0.80 for EBM vs. 0.85 for BM25), presenting a valid trade-off where EBM filters out textheavy box sets in favor of community-validated literature. Global explanations identified popularity as the most influential individual feature, while local explanations directly corresponded to global shape functions. An ablation study further confirmed explanation faithfulness. This framework enhances BM25 retrieval quality through verified quality re-ranking while maintaining transparent, interconnected explanations to support exploratory search. Keywords: book search, Goodreads, Best Matching 25, Explainable Boosting Machine, interpretable machine learning.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Purbasari, Intan Yuniar	19800602 202521 2 029	intanyuniar.if@upnjatim.ac.id
Thesis advisor	Wahanani, Henni Endah	NIP. 19780922 202121 2 005	henniendah@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA76.6 Computer Programming

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Bagus Bagus Satrio Wicaksono Wicaksono

Date Deposited:

22 Jul 2026 07:03

Last Modified:

22 Jul 2026 08:04

URI:

https://repository.upnjatim.ac.id/id/eprint/55990