Sentiment-driven community detection in a network of perfume preferences

The rapid growth of online perfume retail platforms has generated vast amounts of user-generated content, including reviews, ratings, and feedback, which contain valuable insights into consumer preferences. However, effectively analyzing this information to uncover hidden structures and provide accurate recommendations remains a major challenge. The primary objective of this study is to apply community detection techniques to group similar perfumes liked by users into meaningful clusters, thereby supporting more accurate perfume recommendations. This study proposes a novel methodology that integrates sentiment analysis and community detection within a Perfume Co-Preference Network constructed from Persian perfume reviews. First, a transformer-based Persian language model was employed to classify user comments into positive and negative sentiments with high accuracy. To enhance classification performance, emojis were mapped to their Persian semantic equivalents, capturing subtle emotional cues often overlooked in text-based analysis. In addition, user ratings for scent, longevity, and sillage were incorporated to refine sentiment classification results and improve network edge weighting. Based on the positive reviews, a bipartite user–perfume graph was constructed and projected into the Perfume Co-Preference Network, where perfumes are connected if they receive positive feedback from the same users. Community detection algorithms, including Louvain, Fastgreedy, Walktrap, and Spin-Glass were then applied to identify clusters of similar perfumes. The results show that the Spin-Glass algorithm achieved the highest modularity score, producing semantically meaningful clusters that aligned with established fragrance families. Furthermore, the inclusion of sentiment, emoji-based refinement, and rating integration significantly improved network modularity compared to baseline models. As an illustrative example to demonstrate interpretability beyond modularity, we included a section analyzing the 21 communities detected by Louvain, showing that they also correspond to attributes such as perfume gender, fragrance groups, and olfactory nature (e.g., warm, fresh). This research represents the first application of community detection techniques in the analysis of perfume networks and contributes a unique, sentiment-annotated dataset of Persian perfume reviews for future network science research. The proposed approach offers practical implications for personalized perfume recommendation, targeted marketing, and consumer behavior analysis in the fragrance industry.