M

About Moshkin

Moshkin is built by a group of graduates of Sharif University of Technology, with one goal: to combine the traditional, experiential art of choosing a fragrance with the language of data and mathematics. Instead of relying on brand names or advertising, we place perfumes next to each other by how they actually smell, so the choice is clearer and more accurate for everyone.

Our data is gathered from Fragrantica and several other reputable fragrance sites, forums and magazines, then cleaned, cross-checked and merged. This dataset is not static: it is continuously monitored and updated, so notes, ratings and rankings change and sharpen over time.

The science behind Moshkin

How does Moshkin find similar fragrances?
Each fragrance is turned into a 256-dimensional vector that encodes its accords, its notes (weighted differently for top, heart and base) and traits like rating, longevity and sillage. Rare notes count more (IDF), the vectors are normalized, and the closeness of two fragrances is measured with cosine similarity. It is a content-based recommender: 'similar' means genuinely close in scent-space, not just the same brand.
What is the Fragrance Map and how is it built?
The 256-dimensional scent-space is too high to see, so we project it to 2D with UMAP (and t-SNE as a fallback), methods that preserve both local and global structure. Fragrances that smell alike end up next to each other on the map, forming scent families and neighbourhoods you can explore visually.
What are the mathematics and data science behind it?
The pillars are a vector-space representation, feature hashing for thousands of notes, IDF weighting, L2 normalization, cosine similarity, and nonlinear dimensionality reduction (UMAP/t-SNE). Community ratings are used as a complementary signal. In short: linear algebra and machine learning, applied to scent.
Do you use neural networks?
Today Moshkin's core engine is vector-based with dimensionality reduction (classical machine learning), not a neural network. On our roadmap, we are adding neural networks and deep learning to learn scent representations directly; we are inspired by work such as the Principal Odor Map (Science, 2023), which uses graph neural networks to map molecular structure to odor perception.
Where does the data come from and how current is it?
It is aggregated from Fragrantica and several other reputable fragrance sources (sites, forums and magazines), then cleaned, cross-checked and merged. The data is continuously monitored and updated, and new records are placed in the same vector space as the old ones so comparisons stay consistent over time.

References and further reading

The methods and ideas our work builds on: