Skip to main content
TR EN

MSc. Thesis Defense: Oğulcan Günaydın

SCENTSENSE: MULTI-MODAL PERFUME RECOMMENDATION VIA

ASPECT-MINED REVIEWS

 

 

Oğulcan Günaydın
MSc. Thesis, 2025

 

Thesis Jury

Prof. Dr. Enes Eryarsoy(Thesis Advisor)

Prof. Dr. Can Akkan

Assoc. Prof. Ayla Gülcü.

 

 

Date & Time: 18th July, 2025 – 11:30AM

Place: YBF 1127

Keywords: perfume recommendation, recommender systems, multi-modal

learning, review aspect mining, scent ontology, matrix completion, cold-start

problem

 

Abstract

 

ScentSense is a multi-modal perfume recommender system that integrates structured metadata with semantic insights from user reviews. In this thesis, we present its design to build a recommender system tailored for the unique and personal world of perfumes. Unlike traditional products, fragrances lack a standardized numeric representation of their properties rather they are often subjectively described in language rather than data. This makes the task of modeling preferences particularly challenging. To bridge this gap, we combine large scale user reviews with structured metadata, using a large language model to extract key facets such as mood, seasonality, projection, and longevity from free form text. These extracted dimensions are then combined with user profiles and processed through a lightweight neural network to generate recommendations. Our model is designed to perform well even under extreme sparsity, which is the nature of fragrances, delivering predictions for both cold-start and low-data users. To the best of our knowledge, this is the first large-scale attempt to apply deep learning to aspect-enriched perfume recom- mendation. While our approach is simple by design, the results are promising an demonstrating that meaningful personalization is possible even in such a subjective domain. We believe this work offers both a foundation and an invitation for future exploration at the intersection of language, emotion, and smell based experience.