Alraj/ScentByte_Perfume_SVC_Dataset

Dior Perfume SVC Dataset: Full Methodology and Reference

The Dior Perfume SVC Dataset provides structured scent vector codes (SVCs) for Dior perfumes. Each perfume is represented numerically, enabling quantitative fragrance analysis, AI applications in perfumery, and digital scent comparison.

This document describes not only the dataset but also the full methodology, including mathematical formulation, feature selection, and encoding procedures.

1. Introduction

Perfumes are traditionally analyzed through subjective human perception, which varies across individuals and environments. Unlike colors, which can be represented in RGB or HEX codes, fragrances lack a standardized numeric representation.

The Scent Vector Code (SVC) is designed to quantify perfume scent profiles into reproducible vectors. This allows:

Objective fragrance comparison

Clustering and classification of perfumes

Integration into AI and machine learning pipelines

The Dior dataset contains 335 perfumes, each encoded with 43 fragrance attributes and their corresponding SVCs.

2. Components and Necessary Inputs

To create an SVC, the following inputs and components are necessary:

Perfume Composition Data

List of ingredients and notes for each perfume (top, middle, base notes)

Concentration or relative intensity of each note (if available)

Feature Space Definition

A standardized list of 43 primary fragrance attributes

Categories include floral, woody, citrus, spicy, animalic, balsamic, etc.

Attribute Weighting Rules

Top notes: weight = 0.4

Middle/heart notes: weight = 0.35

Base notes: weight = 0.25

Normalization Constraints

Each attribute is scaled to a 0–100 range

Ensures consistency across different perfumes

3. Mathematical Formulation of SVC

The SVC of a perfume can be represented as a vector:

SVC = [a1, a2, ..., a43]

Where:

n = 43 (number of fragrance attributes)

ai is the normalized weighted intensity of the i-th attribute

Step 1: Weighted Intensity Calculation

For each attribute i, the intensity is calculated as:

Ii = sum over j of (wj * cij) for j = 1 to m

Where:

m = number of notes contributing to attribute i

cij = raw concentration/intensity of note j in attribute i

wj = weight of note j (top, middle, base)

Step 2: Normalization

The raw intensity vector is normalized to a 0–100 scale:

ai = ((Ii - Imin) / (Imax - Imin)) * 100

Where Imin and Imax are the minimum and maximum intensities observed for that attribute across all perfumes.

Step 3: Vector Assembly

After normalization, the SVC is assembled as:

SVC = [a1, a2, ..., a43]

Each SVC is a 43-dimensional vector

Can be stored as a string, JSON array, or serialized object

Facilitates vector operations such as cosine similarity or Euclidean distance

4. Data Extraction and Preprocessing

Source Extraction

Perfume notes collected from Fragrantica

Extracted using automated scraping pipelines or manual curation

Categorization into Attributes

Each note is mapped to one or more fragrance attributes

Example: “Bergamot” → Citrus, Fresh

Weight Assignment

Top notes: w = 0.4

Middle notes: w = 0.35

Base notes: w = 0.25

Handling Missing Data

Notes missing intensity values are imputed using the average intensity of that attribute across the dataset

SVC Generation

Weighted, normalized intensities are combined into a final 43-dimensional vector

5. Computational Methodology

Vector Operations

SVCs can be compared using cosine similarity:

Similarity(SVC1, SVC2) = (SVC1 · SVC2) / (||SVC1|| * ||SVC2||)

Clustering and Analysis

Perfumes can be clustered using k-means, hierarchical clustering, or PCA

Enables visualization of fragrance families

AI Applications

Train ML models for perfume recommendation

Generate synthetic SVCs to predict new fragrance combinations

6. Dataset Structure Column Description name Perfume name url Fragrantica source URL svc_code 43-dimensional SVC vector (JSON or array) fragrance_attributes Numerical intensities (0–100) of 43 attributes
7. Example Usage from datasets import load_dataset from scipy.spatial.distance import cosine

Load dataset

dataset = load_dataset("Alraj/ScentByte_Perfume_SVC_Dataset") entry = dataset["train"][0]

svc_vector = entry["svc_code"] # 43-dimensional SVC attributes = entry["fragrance_attributes"]

Cosine similarity between two perfumes

similarity = 1 - cosine(svc_vector, dataset["train"][1]["svc_code"]) print("Similarity:", similarity)

8. Applications of SVC

Objective Perfume Comparison – measure similarity numerically

AI Recommendations – suggest perfumes with similar SVCs

Digital Perfume Libraries – catalog scents digitally

Fragrance Research – analyze patterns and trends

9. License

CC BY 4.0: Free for research, educational purposes, and derivative works with proper credit

10. Conclusion

The Scent Vector Code (SVC) methodology provides a reproducible, quantitative, and computationally accessible representation of perfumes. By combining data extraction, weighted feature encoding, normalization, and vector operations, SVCs create a scalable digital language for fragrances, bridging the gap between human olfaction and computational analysis.

For more detailed analysis, methodology diagrams, and examples, please refer to scentbyte.pdf in the main directory.