guyshilo12/synthetic-ecommerce-price-estimation

Synthetic E-Commerce Product Price Dataset

Overview

This dataset contains a fully synthetic collection of e-commerce product listings. Each product is represented by an AI-generated textual description and a deterministically computed fair price.

The dataset was created to support tasks such as:

• Price estimation and regression
• Embedding-based product similarity
• Feature importance analysis
• Recommendation systems

All text data was generated using a pre-trained Hugging Face language model, while prices were calculated using predefined pricing rules to ensure internal consistency and interpretability.

---

Dataset Structure

Each row represents a single product listing.

Columns

• category Product category (e.g., Gaming Laptop, Smartphone, Smartwatch, Headphones)

• features Comma-separated list of product features (e.g., OLED Display, Pro Camera System)

• generated_text AI-generated product description created by a Hugging Face LLM

• fair_price Deterministically computed fair price (USD), based on the selected product features

---

Synthetic Data Generation

The dataset was generated using the following process:

1. A product category is sampled.

2. A base price is assigned per category.

3. A random subset of features is selected.

4. Each feature contributes a fixed additive price.

5. The final fair price is computed as:

   Base Price + Sum of Feature Prices

6. A Hugging Face pre-trained language model generates a short marketing-style product description based on the selected features.

This approach guarantees:

• Full control over price logic
• Realistic price distributions
• Ground-truth labels suitable for supervised learning

---

Exploratory Data Analysis (EDA)

1. Price Distribution by Category

The following figure shows the distribution of prices across product categories.

Each category exhibits a distinct and realistic price range. The clear separation between categories reflects typical market segmentation and validates the pricing logic used during synthetic data generation.

---

2. Reverse Engineering Feature Price Impact

To validate dataset consistency, we performed a reverse-engineering experiment.

For each product category:

• Product features were converted into binary indicators
• A linear regression model was trained to predict the final price
• Learned coefficients represent the estimated price contribution of each feature

Example: Feature price impact for Gaming Laptops

The estimated feature prices closely match the predefined pricing rules, demonstrating that the dataset follows a clear and learnable structure.

---

3. Text Length Analysis

The following figure shows the distribution of generated text lengths.

The generated product descriptions show a consistent and well-controlled length distribution. Most descriptions fall between 60 and 120 characters, with a small right tail extending up to approximately 250 characters.

This behavior indicates that the language model produced concise, marketing-style descriptions rather than long-form text. Such consistency is well-suited for embedding-based similarity models, as it reduces noise and avoids extreme outliers.

---

Files in This Repository

• price_estimator_10k.parquet Full synthetic dataset (10,000 samples).

• synthetic_data_generation.ipynb Notebook used to generate the dataset.

• eda.ipynb Exploratory Data Analysis notebook, including all figures shown above.

• figures/ Directory containing exported EDA visualizations.

---

Intended Use

This dataset is designed for educational and research purposes, including:

• Machine learning coursework
• Embedding and similarity pipelines
• Price prediction models
• Demonstrations of synthetic data generation

---

License

This dataset is released under the MIT License.