Optimasi model Hierarchical Transformer for Recommendation untuk meningkatkan metric precision

Marsaputra, Rifqi Syekhi (2025) Optimasi model Hierarchical Transformer for Recommendation untuk meningkatkan metric precision. Sarjana thesis, UIN Sunan Gunung Djati Bandung.

Preview	Text 1_cover.pdf Download (62kB) | Preview
Preview	Text 2_abstrak.pdf Download (155kB) | Preview
Preview	Text 3_skbebasplagiarism.pdf Download (67kB) | Preview
Preview	Text 4_daftarisi.pdf Download (76kB) | Preview
Preview	Text 5_bab1.pdf Download (251kB) | Preview
	Text 6_bab2.pdf Restricted to Registered users only Download (268kB) | Request a copy
	Text 7_bab3.pdf Restricted to Registered users only Download (144kB) | Request a copy
	Text 8_bab4.pdf Restricted to Registered users only Download (368kB) | Request a copy
	Text 9_bab5.pdf Restricted to Registered users only Download (25kB) | Request a copy
	Text 10_daftarpustaka.pdf Restricted to Registered users only Download (167kB) | Request a copy
	Text 11_lampiran.pdf Restricted to Repository staff only Download (342kB) | Request a copy

Abstract

Penelitian ini bertujuan untuk meningkatkan akurasi sistem rekomendasi item dalam game Multiplayer Online Battle Arena (MOBA) Dota 2 dengan mengoptimasi model Hierarchical Transformer for Recommendation (HT4Rec). Meskipun model HT4Rec telah menunjukkan kinerja unggul dalam metrik recall dan Mean Reciprocal Rank (MRR), precision yang rendah tetap menjadi tantangan signifikan. Permasalahan ini penting karena dalam sistem rekomendasi top-N, precision merepresentasikan sejauh mana item relevan muncul di urutan teratas dan berdampak langsung terhadap efektivitas sistem. Penelitian ini dilakukan dengan pendekatan CRISP-DM, yang meliputi tahapan Business Understanding, Data Understanding, Data Preparation, Modeling, dan Evaluation. Dataset yang digunakan adalah “Dota 2 Matches” dari Kaggle, yang berisi data urutan pembelian item dan komposisi hero dalam pertandingan. Proses optimasi dilakukan melalui tuning hyperparameter seperti dropout dan weight decay, tanpa menambahkan kompleksitas fitur input. Evaluasi dilakukan menggunakan metrik precision@K, recall@K, dan MRR@K. Hasil eksperimen menunjukkan bahwa konfigurasi hyperparameter tidak serta-merta meningkatkan performa model secara keseluruhan. Rata-rata nilai precision, recall, dan MRR justru mengalami penurunan pada sebagian besar skenario. Skenario 1 menunjukkan penurunan nilai metrik, namun tetap stabil tanpa perubahan signifikan, sehingga dinilai cocok untuk mempertahankan performa yang seimbang. Sementara itu, Skenario 2 menghasilkan precision yang tinggi pada Top-3, namun disertai penurunan pada nilai recall dan MRR. Hal ini menunjukkan bahwa tuning hyperparameter tidak sepenuhnya efektif untuk meningkatkan relevansi rekomendasi tanpa mengorbankan cakupan dan generalisasi. Penelitian ini memberikan kontribusi empiris dalam memahami trade-off antara peningkatan presisi pada posisi teratas dengan stabilitas performa model secara keseluruhan, sekaligus membuka arah baru untuk studi terkait optimasi arsitektur model rekomendasi. This study aims to improve the accuracy of item recommendation systems in the Multiplayer Online Battle Arena (MOBA) game Dota 2 by optimizing the Hierarchical Transformer for Recommendation (HT4Rec) model. Although HT4Rec has demonstrated superior performance in recall and Mean Reciprocal Rank (MRR) metrics, low precision remains a significant challenge. This issue is crucial because, in top-N recommendation systems, precision reflects the extent to which relevant items appear at the top of the list and directly affects the system’s effectiveness. The study was conducted using the CRISP-DM approach, which includes the stages of Business Understanding, Data Understanding, Data Preparation, Modeling, and Evaluation. The dataset used is “Dota 2 Matches” from Kaggle, which contains data on item purchase sequences and hero compositions in matches. Optimization was performed through hyperparameter tuning, specifically dropout and weight decay, without adding complexity to the input features. Evaluation was carried out using precision@K, recall@K, and MRR@K metrics. The experimental results show that hyperparameter configurations do not necessarily improve the overall model performance. The average values of precision, recall, and MRR decreased in most scenarios. Scenario 1 showed a decline in metric values but remained stable without significant changes, making it suitable for maintaining balanced performance. Meanwhile, Scenario 2 achieved high precision in Top-3 predictions but with decreases in recall and MRR. These findings indicate that hyperparameter tuning is not entirely effective in improving recommendation relevance without sacrificing coverage and generalization. This study provides empirical contributions to understanding the trade-off between increasing top-ranked precision and maintaining overall model stability, while also opening new directions for studies related to optimizing recommendation model architectures.

Actions (login required)

View Item