🧠 Smart Product Pricing Challenge – ML Challenge 2025

---

👥 Team Name: TomanAI

Team Members: Harikarthik A, K V Raghul, K V Magesh
Submission Date: October 2025

---

🚀 Executive Summary

Our project tackles the Smart Product Pricing Challenge 2025, predicting product prices using textual catalog descriptions.
We built a TF-IDF + Ridge Regression model that extracts semantic patterns from product text to estimate pricing efficiently — without depending on image data or external sources.
This solution provides a robust, interpretable, and scalable baseline for large e-commerce datasets.

---

🧩 Methodology Overview

🔍 Problem Analysis

The problem is formulated as a text regression task, where a product’s title, description, and quantity information determine its optimal price.
We performed exploratory data analysis to understand text-price relationships.

Key Observations:

• Descriptive terms such as “premium”, “pack”, “organic”, and “bundle” indicate higher pricing.
• Sparse or generic product texts result in lower predictive accuracy.
• Outliers correspond to large pack sizes or combo products.

🧠 Solution Strategy

Approach Type: Single Model (Text-only Regression)
Core Innovation: Leveraging TF-IDF vectorization (unigrams + bigrams, 30K features) to capture contextual and frequency-based text signals, paired with Ridge Regression for robust numeric prediction.
We used 5-Fold Cross-Validation to minimize overfitting and ensure generalization across unseen data.

---

⚙️ Model Architecture

catalog_content ↓ Text Cleaning (Lowercase, Stopword Removal) ↓ TF-IDF Vectorization (1–2 grams, 30K features) ↓ Ridge Regression Model (α = 1.0) ↓ Predicted Product Price

🧰 Model Components

Text Processing Pipeline:

• Preprocessing: Lowercasing, removing nulls, and filtering English stopwords
• Vectorizer: TF-IDF (max_features=30,000, ngram_range=(1,2))
• Model: Ridge Regression (alpha=1.0)
• Validation: 5-Fold K-Fold (shuffle=True, random_state=42)

Image Processing Pipeline:

• Not included in this baseline version
• Future plan: Use CNN or CLIP embeddings for multimodal fusion

---

📊 Model Performance

Metric	Description	Value
RMSE	Root Mean Squared Error (5-Fold CV)	~12.45
SMAPE	Symmetric Mean Absolute Percentage Error	Evaluated during leaderboard testing

✅ The model achieved consistent fold-level performance and stable generalization, proving that textual patterns alone can predict price trends effectively.

---

🏁 Conclusion

This project demonstrates a text-only baseline for product price prediction using a simple yet effective TF-IDF + Ridge Regression model.
It provides strong initial performance and serves as a foundation for multimodal extensions that include image embeddings and transformer-based text encoders (e.g., BERT).

---

📁 Repository & Code Artefacts

GitHub Repository:
🔗 TomanAI – Amazon ML Challenge 2025

Main Files:

• M_L_PROJECT.ipynb → Main training and prediction notebook
• train.csv / test.csv → Dataset files
• submission.csv → Final model output (predicted prices)
• README.md → Documentation (this file)

---

📈 Additional Results

• Cross-fold RMSE consistency across 5 splits
• Top TF-IDF keywords most correlated with high product prices
• Planned visualization: price distribution vs. textual richness

---

💡 Future Enhancements

• Integrate product image embeddings using CLIP or ResNet
• Fine-tune BERT/RoBERTa for semantic understanding of descriptions
• Combine text + image predictions via ensemble learning
• Apply price normalization and outlier-aware regression

---

🧾 License

This project is licensed under the MIT License.

---

Built with ❤️ by Team TomanAI for Amazon ML Challenge 2025
“Learning the price of intelligence.”