Multi-Model Recommendation System for Amazon

---

---

Project Overview

This project involves the design, implementation, and deployment of an intelligent and modular recommendation system for Amazon products. It integrates four different recommendation approaches, along with a customer review sentiment analysis module, all accessible through an interactive web interface built with Streamlit.

The system is designed to handle various real-world scenarios like cold-start problems, sparse user data, and the need for diverse recommendations.

---

Key Features

• Four Recommendation Models:

  • Popularity-Based Filtering – For cold-start scenarios
  • Content-Based Filtering – Uses product descriptions and TF-IDF vectorization
  • Collaborative Filtering – Model-based using SVD matrix factorization
  • Hybrid Approach – Combines all three methods for robust recommendations

• Sentiment Analysis:

  • NLP-based classification of customer reviews
  • Bernoulli Naïve Bayes model achieving 70.43% accuracy
  • Converts ratings (1-5 stars) into sentiment labels (-1, 0, 1)

• Interactive Dashboard:

  • Real-time product recommendations
  • Visual exploration of customer reviews and ratings
  • Multiple filtering options and result visualization

• Scalable Architecture:

  • Big Data processing with PySpark
  • MongoDB integration for data storage
  • Modular Python pipeline for data processing

---

System Architecture

Data Pipeline:

1. Data Collection – Amazon Reviews 2023 dataset from McAuley Lab
2. Data Processing – Cleaning, feature engineering, and text preprocessing
3. Model Training – Multiple recommendation algorithms and sentiment analysis
4. Deployment – Streamlit web application with interactive interface

Technical Stack:

• Data Processing: PySpark, Pandas
• Database: MongoDB
• ML Libraries: Scikit-learn, Surprise, NLTK
• Web Framework: Streamlit
• Visualization: Matplotlib, Tableau (for analytics dashboard)

---

Model Performance

Recommendation Models Evaluation (on 1,000 products sample):

Method	Precision	Recall	Diversity
Content-Based	47.4%	47.4%	0.986
Collaborative	0.4%	0.4%	0.991
Popularity	0.04%	0.04%	0.984

Sentiment Analysis Models:

Model	Accuracy	Status
SVC (C=0.01)	66.08%	-
Multinomial Naïve Bayes	70.09%	-
Bernoulli Naïve Bayes	70.43%	Selected

Note: Collaborative and popularity-based models show better performance on full dataset (>30% precision) despite low scores on the 1,000-product sample due to data sparsity.

---

Installation & Usage

Prerequisites

• Python 3.8+
• MongoDB
• Java (for PySpark)

Installation Steps

1. Clone the repository:
2. Set up MongoDB:

Install and start MongoDB service Update connection URI in configuration files

3. Download dataset:

Obtain Amazon Reviews 2023 dataset from McAuley Lab McAuley Lab Datasets: https://cseweb.ucsd.edu/~jmcauley/datasets.html

Place in data/raw/ directory

3. Run data processing pipeline:

python data_processing/data_cleaning.py
python data_processing/feature_generation.py
python data_processing/data_merge.py

Train models:

python models/sentiment_analysis.py
python models/collaborative_model_based.py

Launch web application:

streamlit run appstreamlit.py

Methodology Details

Data Processing:

• Web Scraping Challenges: Addressed Amazon's anti-bot protections by using McAuley Lab's curated datasets
• Text Preprocessing: HTML cleaning, stopword removal, stemming with PorterStemmer
• Feature Engineering: TF-IDF vectorization, review counting, sentiment scoring

Recommendation Algorithms:

• Content-Based: Cosine similarity on product descriptions
• Collaborative: SVD matrix factorization with implicit feedback
• Popularity: Aggregate ratings and review counts
• Hybrid: Weighted combination of all methods

Evaluation Metrics:

• Precision@K – Relevance of top-K recommendations
• Recall@K – Coverage of relevant items
• Diversity – Variety in recommendations (1 - average similarity)

---

Results & Insights

• Content-based filtering performs best on small datasets with rich product descriptions
• Collaborative filtering requires substantial user interaction data to be effective
• Sentiment analysis adds valuable contextual understanding beyond star ratings
• Hybrid approach provides the most robust coverage across different scenarios

---

Usage Examples

Demo Video

Through Web Interface:

1. Select a recommendation model from the sidebar
2. Search and select a product from the dropdown
3. Click "Generate Recommendations" to see similar products
4. View detailed results with images, ratings, and prices

---

Feedback

Have fun exploring the system! If you have any questions, suggestions for improvement, or would like to contribute to enhancing the recommendation algorithms, feel free to reach out!