This project focuses on building a machine learning model to predict the sale price of online product listings using structured and unstructured data provided by sellers. Accurately estimating product prices is a challenging task influenced by subtle differences in product descriptions, conditions, and categories—particularly in competitive online marketplaces.
The goal is to predict the final sale price of items based on features such as item condition, category, brand name, shipping information, and text descriptions. This task simulates a real-world application faced by e-commerce platforms when setting or suggesting optimal prices.
The dataset consists of two tab-delimited files:
train.tsv: Contains ~700,000 labeled listings with sale prices.test.tsv: Contains ~3.5 million listings without prices (to be predicted).
Each listing includes:
name: Product title (with price mentions removed and replaced by[rm]).item_condition_id: Condition of the item.category_name: Hierarchical category of the product.brand_name: Brand of the item.price: Target variable (only intrain.tsv).shipping: Indicates who pays for shipping (0 = buyer, 1 = seller).item_description: Detailed product description (also cleaned of prices).
The predictive model is trained on the train.tsv data and evaluated on test.tsv. To account for computational complexity, model development begins with a representative sample of the training data.
Model performance is evaluated using Root Mean Squared Logarithmic Error (RMSLE) to penalize underestimation more than overestimation and handle large-scale price ranges effectively.
RMSLE Formula:
ε = sqrt( (1/N) * Σ[ log(p̂ᵢ + 1) - log(pᵢ + 1) ]² )
Where:
- p̂ᵢ is the predicted price
- pᵢ is the actual price
- N is the number of predictions