Skripsi
SISTEM REKOMENDASI BOARD GAME HYBRID MENGGUNAKAN JARINGAN SIAMESE DENGAN FUNGSI CONTRASTIVE DAN TRIPLET LOSS
This research addresses the challenge users face in discovering board games that align with their preferences, particularly when those preferences are expressed descriptively in text. Traditional recommendation systems often struggle to capture the semantic nuances within natural language queries and effectively balance these with specific attribute criteria. To overcome this limitation, this study develops and evaluates a hybrid recommendation system integrating deep semantic understanding with structured attribute filtering. The system employs a Siamese DistilBERT network architecture, fine-tuned to learn semantic representations from board game descriptions. Eight model configurations were systematically evaluated, comparing the effectiveness of Contrastive Loss and Triplet Loss functions using a holistic suite of custom evaluation metrics. Experimental results indicate that the model trained with Contrastive Loss using a margin of 0.5 (Configuration 2) significantly outperformed other variants, achieving the highest Hybrid Relevance Score of 0.9397. This configuration demonstrated robustness and stability across diverse query types, confirming its superior ability to interpret thematic user intent through a "semantic-first" design philosophy. This research validates an effective architecture for a text-based board game recommender system, yielding a contextually intelligent solution capable of delivering more satisfying recommendations.
No other version available