🛒 Amazon ML Challenge 2025 – Smart Product Pricing

📌 Overview

E-commerce pricing depends on a complex mix of textual attributes (brand, pack size, specifications) and visual cues (packaging, perceived quality). This project builds a multimodal machine learning pipeline to predict product prices using both text and image data.

Objective: Predict continuous product prices. Evaluation Metric: SMAPE (Symmetric Mean Absolute Percentage Error)

Dataset Link: https://www.kaggle.com/datasets/infiniper/amazon-ml-challenge-2025-dataset

Repository Link: https://github.com/Infiniper/Amazon-ML-Challenge-2025.git

---

📊 Dataset

Split	Size	Description
Train	75,000	Product details + images + price
Test	75,000	Product details (no price)

Data Fields

• sample_id – Unique product identifier
• catalog_content – Title + description + item pack quantity
• image_link – Public URL of product image
• price – Continuous float value (target variable, train only)

---

🧠 Multimodal Architecture

1️⃣ Raw Inputs

• Text: catalog_content
• Image: image_link
• Target: price

2️⃣ Feature Engineering

Unstructured data is converted into numerical embeddings:

• Regex Numeric Extraction → word count, char length, extracted numeric values & units
• S-BERT (Text Embeddings) → 384-dimensional semantic vector
• CLIP (Image Embeddings) → 512-dimensional visual feature vector

3️⃣ Feature Stacking

All features are concatenated into a unified matrix (~900 features per product).

4️⃣ Regression Model

A LightGBM Regressor learns non-linear relationships between stacked features and price.

---

⚙️ Training Strategy

🔹 Log Target Transformation

To handle skewed price distribution:

log_price = ln(1 + price)

Predictions are inverse-transformed after inference.

🔹 Stratified K-Fold (for Regression)

Log-prices are binned into quantiles to maintain balanced price distribution across folds.

🔹 Custom SMAPE Metric

SMAPE is implemented directly inside LightGBM for evaluation consistency.

🔹 Ensemble Bagging

Multiple models trained across different seeds and averaged to reduce variance.

---

💡 Design Decisions

• Why Multimodal? Text misses visual cues; images miss quantity/scale. Fusion removes blind spots.
• Why S-BERT & CLIP? Transfer learning provides rich features without heavy training cost.
• Why LightGBM? Tree-based models outperform deep networks on structured embedding features.

---

📈 Evaluation Metric

SMAPE = (1/n) * Σ |Actual - Predicted| / ((|Actual| + |Predicted|)/2) * 100%

• Range: 0% – 200%
• Lower is better

---

📂 Repository Structure

├── dataset/                     # Sample CSVs and dataset info
├── src/
│   ├── utils.py                 # Image downloading utilities
│   └── example.ipynb            # Starter notebook
├── Other_Notebooks/
│   ├── light-gbm/               # LightGBM pipeline & feature extraction
│   └── SVM/                     # Alternative approach
├── Results on test data/        # Prediction outputs
├── main_notebook.ipynb          # Final multimodal pipeline
├── download_dataset.py          # Dataset download script
└── calculate_smape.py           # Local SMAPE evaluation

---

🚀 Getting Started

1️⃣ Clone the Repository

git clone https://github.com/Infiniper/Amazon-ML-Challenge-2025.git
cd Amazon-ML-Challenge-2025

2️⃣ Download Dataset

Download from Kaggle and place:

dataset/train.csv
dataset/test.csv

3️⃣ Run Pipeline

Open:

main_notebook.ipynb

This notebook covers end-to-end:

• Data preprocessing
• Text & image embedding generation
• Feature stacking
• LightGBM training
• SMAPE evaluation

---

📌 Notes

• No external price lookup was used.
• Model size constraints follow competition guidelines.
• Designed for reproducibility and modular experimentation.

---