Amazon ML Hackathon 2025 | Team AVIS | Product Price Prediction | Rank 1693 |Total Registered:82,790 | SMAPE 51.4

Amazon Machine Learning Hackathon 2025 — Product Price Prediction

This repository contains Team AVIS’s solution for the Unstop ML Hackathon 2025, where the task was to predict product prices from structured and unstructured catalog data.

🏆 Team Name: AVIS 📈 Final SMAPE: 51.4 🥇 Leaderboard Rank: #1693 ⚙️ Frameworks: LightGBM + Sentence Transformers (MiniLM-L6-v2) 🧠 Hardware: GPU-accelerated training on Google Colab

Project Overview

The goal was to build a regression model that accurately predicts product prices using both structured features (brand, quantity, unit) and unstructured text (titles, bullet points, and product descriptions).

Component Description Text Encoder SentenceTransformer – all-MiniLM-L6-v2 Model LightGBM (GPU, regression_l1 objective) Metric SMAPE (Symmetric Mean Absolute Percentage Error) Optimization Early stopping, feature scaling, lemmatization, unit normalization Model Architecture

Text Cleaning — remove emojis, punctuation, and stopwords

Embedding Generation — use SentenceTransformer (MiniLM-L6-v2)

Feature Fusion — combine embeddings + categorical + numeric features

Training — GPU-based LightGBM regressor

Evaluation — SMAPE metric

📊 Results Metric Score Validation SMAPE 47.43 Public Leaderboard SMAPE 51.4 Final Rank #1693 / 82,790 🧩 Tech Stack

🐍 Python 3.12

💡 LightGBM (GPU)

🤖 SentenceTransformers (MiniLM-L6-v2)

🧹 NLTK for text preprocessing

📦 scikit-learn, pandas, numpy, joblib

🧪 How to Run

1️⃣ Install dependencies

!pip install -q lightgbm sentence-transformers emoji nltk

2️⃣ Run the training script

python amazon_price_prediction.ipynb

3️⃣ Predict on new data

python inference_script.py --input test.csv --output predicted_prices.csv

Highlights

✅ Preprocessed 95K+ records combining structured and unstructured data ✅ Generated 384-dimensional text embeddings using MiniLM ✅ Optimized LightGBM with GPU acceleration ✅ Achieved SMAPE 51.4 — Top 2% out of 82,790 global participants