Welcome to the Fragrance Recommender System! This project is designed to help users discover new perfumes based on their preferred scent profiles. By leveraging a recommendation algorithm based on euclidian distance, the engine suggests fragrances that match the user's specified perfume accords, making it easier for users to find scents they will love.
| Feature | Type | Description |
|---|---|---|
| brand | string | The brand or maker of a perfume |
| perfume | string | The name of a particular perfume |
| launch_year | integer | The year a perfume was released |
| main_accords | string | The fragrance family(s) that a perfume belongs to; Example: Floral |
| notes | string | The specific notes in a fragrance; Example: Rose |
The objectives of this project are as follows:
- Input Collection: Users enter their preferred perfume accords they want in a fragrance.
- Feature Extraction: The system processes the input and extracts perfumes corresponding to the specified accords.
- Similarity Calculation: Using Euclidian Distance, the system calculates the similarity between the user's input and each perfume in the database.
- Recommendation Generation: The top 5 most similar perfumes are selected and presented as recommendations to the user.
Data Source:
After cleaning, the dataset contained:
- 144 fragrance families
- 1,302 notes
- 2,536 unique brands
- 33,146 unique perfumes
I created 3 recommendation models in total:
- A baseline model that recommends perfumes randomly
- A model that recommends based on Cosine Similarty
- A model that recommends based on Euclidian Distance
I compared these three models against eachother by choosing the accord "wine" and comparing the perfumes that each recommender ouput.
- The Euclidian Distance model and the Cosine Similrity model recommended the exact same perfumes, albeit in a slightly different order.
- The Random Recommender model did not recommend any relevant perfumes containing the specified "wine" accord.
- In the end, I chose to use the Euclidian Distance model because I found it to be less computationally intensive and had a faster run time.
- I am using streamlit to deploy my model.
- Here, the user can choose up to three fragrance accords that they wish to find in a perfume.
- Then, they will be shown 5 fragrances that are similar, including the perfume name, brand, and the notes within each fragrance.
- For a faster runtime, I am only using a small subset of 5,000 perfumes to deploy on streamlit.
- Therefore, the recommendations given are somewhat limited, when compared to utilizing the full dataset of over thirty thousand perfumes.