This repository contains the implementation of a hybrid recommender system that combines Alternating Least Squares (ALS) collaborative filtering with Two-Tower content-based filtering for Amazon e-commerce recommendations.
Title: "A Hybrid Recommender System for Amazon E-commerce Combining ALS Collaborative Filtering and Two-Tower Content-Based Filtering"
Abstract: Recommender Systems are vital for enhancing user experience on e-commerce platforms like Amazon, but traditional approaches face challenges such as data sparsity. This paper proposes a hybrid recommender system combining Alternating Least Squares (ALS) collaborative filtering with a two-tower content-based filtering model for Amazon e-commerce. The ALS model utilizes user-item interactions, while the two-tower model incorporates product description features. The system adaptively combines predictions from both models using a weighted approach. Experiments on a large-scale Amazon dataset demonstrate that the proposed method outperforms state-of-the-art baselines, offering an improved solution for recommendation quality and diversity on the Amazon e-commerce platform.
- Input: Amazon e-commerce dataset (10,000 products, 17 attributes)
- Process:
- Probability-based imputation for missing values
- Label encoding for categorical features (manufacturer, price, categories, ratings)
- Feature engineering (userId from uniq_id, itemId from product_name)
- Output: Clean, processed dataset with 0 missing values
- Algorithm: Alternating Least Squares (PySpark implementation)
- Parameters: rank=10, maxIter=10, regParam=0.1
- Purpose: Captures user-item interaction patterns through matrix factorization
- Output: User and item latent factor matrices
- Architecture: Neural embedding-based two-tower design
- Implementation: TensorFlow/Keras with 50-dimensional embeddings
- Features: Product metadata (manufacturer, categories, descriptions)
- Output: Content-based similarity scores
- Approach: Adaptive weighted combination using F1-score based weighting
- Logic: 80-20 weighting favoring the better-performing model
- Decision Rule: IF ALS_F1 > TwoTower_F1: 80% ALS + 20% Two-Tower, ELSE: 20% ALS + 80% Two-Tower
- Final Output: Top-5 personalized recommendations
- User 1 (ID: 462): Hybrid F1=0.6396, NDCG=0.9775, MAE=1.0372
- User 2 (ID: 9435): Hybrid F1=0.6177, NDCG=0.9826, MAE=1.7688
- Performance: Consistent outperformance over individual models across multiple evaluation metrics
├── src/ # Core implementation modules
│ ├── init.py
│ ├── data_preprocessing.py # Data cleaning and feature engineering
│ ├── als_model.py # ALS collaborative filtering implementation
│ ├── two_tower_model.py # Two-Tower neural content-based filtering
│ ├── hybrid_system.py # Hybrid model combining both approaches
│ ├── evaluation.py # Comprehensive evaluation metrics
│ └── utils.py # Utility functions and helpers
├── data/ # Dataset documentation and access instructions
│ └── README.md # Detailed dataset information
├── docs/ # Additional documentation
│ └── methodology.md # Complete methodology documentation
├── requirements.txt # Python dependencies
├── LICENSE # MIT license
└── README.md # This file
git clone https://github.com/yourusername/hybrid-als-twotower-recommender.git
cd hybrid-als-twotower-recommender
pip install -r requirements.txt
Follow instructions in data/README.md to obtain the Amazon dataset Place the CSV file in the data/ directory
Data preprocessing
python src/data_preprocessing.py
Train individual models (optional - for experimentation)
python -c "from src.als_model import ALSModel; print('ALS module ready')"
python -c "from src.two_tower_model import TwoTowerModel; print('Two-Tower module ready')"
Run hybrid system
python -c "from src.hybrid_system import HybridRecommendationSystem; print('Hybrid system ready')"
Comprehensive evaluation:
python -c "from src.evaluation import RecommenderEvaluator; print('Evaluation module ready')"
from src.data_preprocessing import main
Run complete preprocessing pipeline:
processed_data = main()
if processed_data:
print("Preprocessing completed successfully!")
from src.hybrid_system import HybridRecommendationSystem
Initialize and train hybrid system:
hybrid_system = HybridRecommendationSystem()
hybrid_system.train_models(training_data)
Get recommendations for a user:
recommendations = hybrid_system.get_hybrid_recommendations(
user_id=462,
all_items=item_list,
top_k=5
)
from src.evaluation import RecommenderEvaluator
Comprehensive evaluation:
evaluator = RecommenderEvaluator()
results = evaluator.comprehensive_evaluation(actual_ratings, predicted_scores)
Generate evaluation plots:
evaluator.plot_precision_recall_at_k(results, model_name="Hybrid System")
- PySpark 3.5.1: For ALS collaborative filtering (Note: PySpark requires Java 8/11 and proper Spark configuration)
- TensorFlow 2.8.0: For Two-Tower neural network
- scikit-learn 1.0.2: For evaluation metrics and preprocessing
- pandas 1.4.2: For data manipulation
- numpy 1.21.5: For numerical computations
- matplotlib 3.5.1: For visualization
- Methodology: Complete technical methodology
- Data Documentation: Dataset information and access instructions
- Source Code: Well-documented Python modules with examples
This implementation provides:
- Complete source code with comprehensive documentation
- Modular design allowing individual component testing
- Reproducible preprocessing pipeline with probability-based imputation
- Consistent evaluation framework with multiple metrics< 6B0C /li>
- Clear methodology documentation for academic replication
Source Code: Openly available in this GitHub repository under MIT License
Dataset: Amazon e-commerce data available from https://github.com/aksharpandia/miniamazon_data. Due to licensing restrictions, the dataset cannot be redistributed. Complete access instructions are provided in data/README.md.
@article{hazra2025hybrid,
title={A Hybrid Recommender System for Amazon E-commerce Combining ALS Collaborative Filtering and Two-Tower Content-Based Filtering},
author={Hazra, Soumi},
journal={Submitted to PeerJ Computer Science},
year={2025}
}
This repository is part of academic research. For questions or collaboration inquiries, please open an issue or contact the author.
MIT License - see LICENSE file for details.
Note: This implementation is designed for academic research and educational purposes. The hybrid approach demonstrates novel techniques for combining collaborative and content-based filtering in e-commerce recommendation systems.