Comparison of TF-IDF and Sentence-Transformer NLP Methods for a Perfume Recommendation System Based on User Descriptions with Streamlit Visualization

Abstract

The global perfume industry has grown considerably in recent years, yet helping consumers identify products that genuinely align with their personal taste remains a non-trivial problem, especially when preferences are communicated through unstructured, free-form natural language rather than standardized scent terminology. This study addresses that gap by evaluating and comparing two Natural Language Processing (NLP) text-representation methods, Term Frequency-Inverse Document Frequency (TF-IDF) and Sentence-Transformer, within a cross-lingual content-based filtering perfume recommendation system. The novelty lies in the cross-lingual setup, where unstructured Indonesian-language queries are matched directly against an English fragrance dataset of 42,115 records without any external translation module. A Streamlit web interface was deployed to collect data from 30 respondents, who submitted 180 graded relevance judgments for Top-3 recommendations evaluated through Precision@3 and Normalized Discounted Cumulative Gain (nDCG@3). System stability was assessed by partitioning the corpus into a training set of 33,692 records and a test set of 8,423 records. Sentence-Transformer, using the pre-trained paraphrase-multilingual-MiniLM-L12-v2 model, outperformed TF-IDF on all metrics, achieving a mean Precision@3 of 0.867 against 0.833 and a mean nDCG@3 of 0.912 against 0.850, with substantially lower score degradation across data partitions. These findings confirm that semantic embedding methods provide superior ranking quality and cross-lingual robustness, offering a scalable and translation-free solution applicable to multilingual product recommendation systems in commercial settings.