Personalized Perfume Recommendations Based on User Descriptions Using Cosine and Jaccard Similarity

The research proposes a perfume recommendation system using Natural Language Processing (NLP) and content-based filtering to help consumers choose perfumes in a growing market. The system analyzes descriptive user input in Bahasa Indonesia to identify preferences through keyword matching and a question-answering approach. A dataset of 24,760 perfumes with 11 features, primarily sourced from Fragrantica, was used. The method includes exploratory data analysis, preprocessing, model design based on content-based filtering with Cosine and Jaccard similarity, and evaluation using nDCG@3, Precision@3, Recall@3, Diversity, and Coverage. The system achieved strong performance in Jaccard based recommendation with a Precision@3 and Recall@3 of 0.8889 and an nDCG@3 of 0.8978, while cosine based results are also performed well. A web- based interface using Streamlit was developed to demonstrate the system’s functionality. User feedback was gathered through a survey involving 31 participants, showing positive responses with average scores ranging from 4.2 to 4.6 out of 5 across usability criteria.