Product Recommendation System : 🎮 Video Games

A comprehensive, end-to-end machine learning project that deploys a SVD-Powered User-to-Item Personalized Hybrid Recommender for video games. It predicts explicit user ratings using advanced matrix factorization, moving beyond simple item similarity to deliver unparalleled personalization. The entire pipeline, from MongoDB data ingestion to a production-ready Streamlit web app, is implemented for maximum efficiency and real-world impact.

---

📋 Table of Contents

• Overview
• Key Features
• Project Architecture
• Dataset
• Methodology
• Results & Insights
• Installation
• Usage
• Visualizations
• Technologies Used
• Business Recommendations
• Future Enhancements

---

🎯 Overview

In the highly competitive e-commerce landscape, personalized recommendations are the core engine for driving conversions. This project solves the personalization challenge by developing a sophisticated Dual-Hybrid Recommendation System for video games using 230,000+ Amazon reviews. The system's flagship feature is the User-to-Item Personalized Recommender, which uses the highly accurate SVD algorithm to predict exactly what a specific user will love.

Key Objectives

1. Build a Personalized User-to-Item Model using the optimal SVD Matrix Factorization algorithm.
2. Develop a Dual-Hybrid Model combining SVD prediction, content-based features, and popularity for robust suggestions.
3. Establish a fast data pipeline from MongoDB to a trained, serialized model (.joblib) for near-instantaneous inference.
4. Perform accurate sentiment analysis on review text using ML classifiers (LightGBM, XGBoost).
5. Deploy an interactive Streamlit application showcasing both Personalized and Item-to-Item results.

---

✨ Key Features

• SVD-Powered Personalized Recommender: Predicts explicit ratings for unrated products based on individual user latent factors.
• Dual-Hybrid Engine: Offers two modes: Personalized (User-to-Item) for engagement and Item-to-Item for product similarity.
• Matrix Factorization Optimality: SVD demonstrated superior accuracy (RMSE: 1.0823) with fast training time (Avg. 2.72s).
• Sentiment Analysis: Classifies review sentiment with high-accuracy models (LightGBM F1-Score $\approx$ 0.90).
• Fast Inference: All matrices and the SVD model are pre-computed and saved for near-real-time performance in the web application.
• Interactive Web App (3 Pages): Streamlit-based UI for real-time recommendation generation.
• Modular Architecture: Clean separation of concerns (mongo_connection, hybrid_personalized, etc.) for scalability.

---

🏗️ Project Architecture

Product_Recommendation/
│
├── data/
│   ├── video_games_reviews.csv          # Raw dataset
│   ├── cleaned_reviews.joblib           # Processed data
│   ├── svd_model.joblib                 # 🌟 TRAINED SVD USER-TO-ITEM MODEL
│   ├── all_products.joblib              # List of all ASINs (for SVD prediction)
│   ├── cf_sim_df.joblib                 # Item-to-Item CF similarity matrix
│   ├── tfidf_matrix.joblib              # Content-based TF-IDF matrix
│   └── ml_results.joblib                # ML model results & metrics
│
├── src/
│   ├── logger_config.py
│   ├── mongo_connection.py
│   ├── data_preprocessing.py
│   ├── baseline.py
│   ├── collaborative.py
│   ├── content_based.py
│   ├── hybrid.py
│   ├── hybrid_fast.py                   # Item-to-Item Hybrid Logic
│   └── hybrid_personalized.py           # 🌟 USER-TO-ITEM HYBRID LOGIC
│   └── ml_models.py
│
├── app/
│   └── streamlit_app.py                 # Web application (3 Pages)
│
├── notebooks/
│   └── Product_Recommendation_System.ipynb  # Main notebook
│
└── README.md

---

📊 Dataset

Source: Amazon Video Game Reviews
Size: 231,780 entries
Time Period: 2000-2014

Features

Column	Description
reviewerID	Unique identifier for the reviewer
asin	Unique product identifier
reviewerName	Display name of the reviewer
helpful	Helpfulness votes [helpful_votes, total_votes]
reviewText	Full review text
overall	Star rating (1-5)
summary	Review title/summary
unixReviewTime	Unix timestamp
reviewTime	Readable date format

Engineered Features

• helpful_ratio: Proportion of helpful votes
• helpful_votes: Total helpful votes received
• label: Binary sentiment (1=positive: $\ge 4$, 0=negative: $<4$)
• reviewTime: Standardized datetime format

---

🔬 Methodology

1. Data Pipeline

Raw Data → MongoDB → Preprocessing → Feature Engineering → Model Training → Serialization (.joblib) → Deployment

Key Steps: Data cleaning, parsing helpful votes, and creating a binary sentiment target label.

2. Recommendation Models (Core Logic)

Personalized User-to-Item Hybrid

• Model: $\text{SVD Prediction} \times \mathbf{\alpha} + \text{Popularity Score} \times \mathbf{\beta} + \text{Content Score} \times \mathbf{\gamma}$
• CF Core: SVD Matrix Factorization predicts the user's rating for unrated items.
• Content-Based: Item similarity calculated based on the user's highest-rated game.
• Weights: Optimized as $\mathbf{\alpha=0.5}$ (SVD Prediction), $\mathbf{\beta=0.3}$ (Popularity), $\mathbf{\gamma=0.2}$ (Content).

Benchmarking & Item-to-Item Models

• SVD, BaselineOnly, NMF: Evaluated using RMSE and MAE to select the optimal algorithm for the personalized model.
• Item-to-Item Hybrid: A faster fallback model using item similarity on pre-computed matrices.

3. Sentiment Analysis

Models Trained: RandomForest, XGBoost, LightGBM (using TF-IDF on reviewText) Evaluation Metrics: Accuracy, Precision, Recall, F1-Score, ROC AUC Goal: Validate review sentiment to provide granular market intelligence alongside recommendations.

---

📈 Results & Insights

Collaborative Filtering Performance (Matrix Factorization)

Algorithm	Mean RMSE (Error)	Mean MAE (Error)	Mean Fit Time
SVD	1.0823	0.8344	$\approx$ 2.72s
BaselineOnly	1.0875	0.8473	$\approx$ 1.10s
NMF	1.2749	0.9809	$\approx$ 6.64s

Conclusion: SVD is the optimal model, providing the lowest predictive error with superior efficiency compared to NMF.

Sentiment Classification Performance

Model	Accuracy	Precision	Recall	F1-Score	ROC AUC
LightGBM	0.840	0.851	0.955	0.900	0.874
XGBoost	0.838	0.847	0.959	0.899	0.872
RandomForest	0.822	0.822	0.977	0.892	0.860

Conclusion: LightGBM is the top-performing classifier. The high Recall for the positive class ($\approx$ 95.5%) suggests the models are excellent at identifying positive reviews but struggle slightly more with the less frequent negative class (due to data imbalance).

---

🚀 Installation

Prerequisites

Python 3.8+
MongoDB (local or cloud instance)
Google Colab (recommended) or Jupyter Notebook

Setup

1. Clone the repository

git clone https://github.com/KBhardwaj-007/Product-Recommendation-System.git
cd Product-Recommendation-System

2. Install dependencies

pip install -r requirements.txt

Required packages:

pandas
numpy
scikit-learn
matplotlib
seaborn
nltk
pymongo
wordcloud
surprise
xgboost
lightgbm
streamlit
joblib
pyngrok

3. Configure secrets (in Google Colab)

Add to Colab Secrets:

• MONGO_URI: Your MongoDB connection string
• NGROK_TOKEN: Your ngrok authentication token

4. Download dataset

Place video_games_reviews.csv in the data/ directory.

---

💻 Usage

Running the Notebook

Open Product_Recommendation_System.ipynb in Google Colab and run all cells. The pipeline automatically:

1. Loads data into MongoDB.
2. Preprocesses data.
3. Trains and serializes the final SVD model and all necessary matrices.
4. Generates and displays comparison results for all models.

Launching the Web App

The final notebook cells deploy the application via Streamlit and expose it via a public ngrok URL.

# In the notebook, execute:
!streamlit run app/streamlit_app.py &>/dev/null &

# Create public tunnel
from pyngrok import ngrok
public_url = ngrok.connect(addr="8501")
print(f"🎉 App live at: {public_url}")

Using the Streamlit App (3 Pages)

1. 👤 Personalized Recommender: Select a User ID to receive a ranked list of games predicted to be rated $\mathbf{\ge 4.5}$ stars by that specific user.
2. 🔗 Item-to-Item Recommender: Select a Product ID to find similar games based on combined user behavior and review content.
3. 📊 Model Performance: View and analyze all classification and collaborative filtering results, heatmaps, and the Confusion Matrix.

---

📊 Visualizations

Distribution of Ratings

Analysis: Strong positive skew with 58% 5-star and 27% 4-star reviews, indicating high customer satisfaction.

---

Helpfulness Ratio Distribution

Analysis: Bimodal distribution with peaks at 0.0 (unvoted) and 1.0 (unanimously helpful), suggesting polarized community engagement.

---

Reviews Over Time

Analysis: Exponential growth from 2012-2014, peaking at 6,000+ monthly reviews, with strong seasonal patterns.

---

Top Products & Reviewers

Analysis: High concentration with top product receiving 800 reviews and most active reviewer contributing 780 reviews.

---

Review Summary Word Cloud

Analysis: Dominant positive terms ("Great", "Good", "Best", "Awesome") with gaming-specific vocabulary ("Game", "Play", "PS3").

---

Sentiment Model Comparison

Analysis: LightGBM leads with 84.3% accuracy; all models show high recall (>95%) but lower precision due to class imbalance.

---

Collaborative Filtering Comparison

Analysis: SVD achieves lowest error rates (RMSE: 1.09); BaselineOnly offers best speed-accuracy tradeoff.

---

Confusion Matrix

Analysis: LightGBM correctly classifies 33,422 positive reviews but generates 5,832 false positives due to 3:1 class imbalance.

---

🛠️ Technologies Used

Category	Technologies
Languages	Python 3.8+
Data Processing	Pandas, NumPy
Machine Learning	Scikit-learn, XGBoost, LightGBM, Surprise
NLP	NLTK, TF-IDF Vectorizer
Database	MongoDB, PyMongo
Visualization	Matplotlib, Seaborn, WordCloud
Web Framework	Streamlit
Deployment	ngrok, Google Colab
Utilities	Joblib, tqdm

---

💼 Business Recommendations

1. 🥇 Maximize Sales with SVD Personalization

Action: Immediately deploy the SVD-Powered Personalized Hybrid Model (User-to-Item) for real-time inference on the homepage, checkout, and email campaigns.

Impact: Maximize revenue by showing each user the few items they are most likely to purchase (based on predicted high rating), leading to conversion rates significantly higher than generic top-seller lists.

2. 🛡️ Implement a Dynamic Cold-Start Strategy

Action: Use a conditional system:

• New Users ($\le 1$ review): Default to the weighted_popularity_based model.
• New Items (No reviews): Use the Content-Based module based on product description/metadata.
• Active Users: Use the SVD-powered Personalized Hybrid.

Impact: Guarantees a relevant recommendation experience from the first interaction, retaining new users who lack history.

3. 💸 Trigger High-Confidence Bundling

Action: Use the SVD prediction score as a campaign trigger. If a user's predicted rating for a new or high-margin game is $\mathbf{\ge 4.5}$, automatically create and send a targeted bundle offer.

Impact: Converts high-confidence intent into higher-value sales, improving Average Order Value (AOV).

4. 📉 Real-Time Sentiment & Inventory

Action: Apply the trained LightGBM model to incoming reviews in real-time. Create an alert system for any product whose Negative sentiment exceeds a 20% threshold for quick review and potential inventory adjustment.

Impact: Provides early warning for product issues, mitigating financial risk and protecting brand reputation.

---

🔮 Future Enhancements

• Real-Time Retraining Pipeline: Automate the SVD model re-training nightly using new data on a scalable cloud resource (e.g., AWS Lambda/GCP Cloud Functions).
• A/B Test Integration: Build a logging framework to compare conversion rates between the old Item-to-Item and the new Personalized Hybrid model in a live environment.
• Multi-Modal Features: Integrate game metadata (e.g., Genre, Developer, Release Year) into the SVD feature matrix for deeper latent factor modeling.
• Mobile Optimization: Deploy a lighter-weight, mobile-friendly Streamlit interface.

⭐ If you found this project useful, please consider giving it a star! ⭐

© 2025 Product Recommendation System | Powered by Python 🐍