Amazon Product Recommendation System

This is a Flask-based web application that provides content-based product recommendations using Amazon product data.

Preview

Features

• Product browsing with dynamic recommendations
• Product detail pages with customer reviews
• Review pages showing all reviews for a product
• Advanced content-based recommendation system with MAP metrics
• Dynamic extraction of reviews from comma-separated data

Setup Instructions

1. Install Dependencies

   pip install flask pandas numpy nltk scikit-learn wordninja textblob
   

2. Download NLTK Components

   import nltk
   nltk.download('stopwords')
   nltk.download('wordnet')

3. Data Preparation

   • Place your amazon.csv file in the root directory
   • Generate the similarity matrix and ground truth by running:

     python export_similarity_matrix.py
     

     This will create two files:
     • cosine_similarity_matrix.pkl: Precomputed similarity matrix
     • ground_truth.pkl: Precomputed recommendation ground truth

4. Run the Application

   python app.py
   

   The application will be available at http://127.0.0.1:5000/

File Structure

• app.py: Main Flask application
• export_similarity_matrix.py: Script to generate the cosine similarity matrix and ground truth
• cosine_similarity_matrix.pkl: Precomputed similarity matrix (generated)
• ground_truth.pkl: Precomputed recommendation ground truth (generated)
• templates/: HTML templates for the web interface
  • index.html: Home page
  • product.html: Product detail page with recommendations and MAP metrics
  • reviews.html: Page showing all reviews for a product
• amazon.csv: Dataset with product information and reviews

How It Works

Recommendation System

The recommendation system uses the following approaches:

1. Content-Based Filtering:

   • Uses TF-IDF to convert product information into numerical vectors
   • Calculates cosine similarity between these vectors to find similar products
   • Products with higher similarity scores are recommended first

2. Ground Truth Generation:

   • For each product, we identify truly relevant recommendations
   • Uses content similarity with a threshold value
   • Falls back to category-based recommendations when needed

3. Evaluation Metrics:

   • Mean Average Precision (MAP) is calculated for each set of recommendations
   • This metric is displayed on the product page to indicate recommendation quality

Review Extraction

Reviews are extracted dynamically from comma-separated entries in the dataset, properly handling:

• Multiple users per product
• Multiple review titles
• Multiple review contents
• Commas within review content (using a smart parsing strategy)