An advanced machine learning-powered fashion recommendation system that predicts customer purchases for H&M using hybrid collaborative filtering and customer segmentation techniques.
This project addresses the challenge of choice overload in fashion e-commerce by delivering highly personalized product recommendations. Unlike traditional recommendation systems that rely solely on purchase history, our solution combines multiple data-driven techniques to enhance accuracy and user experience.
- Hybrid Collaborative Filtering: Combines Singular Value Decomposition (SVD) and Alternating Least Squares (ALS) algorithms
- Customer Segmentation: K-Means clustering for targeted recommendations based on purchasing behavior
- RFM Analysis: Recency, Frequency, and Monetary value analysis for customer categorization
- Comprehensive EDA: Deep insights into customer behavior, seasonal trends, and product popularity
- Cold Start Problem Handling: Effective recommendations for new users and products
- Scalable Architecture: Designed to handle large datasets efficiently
- Programming Language: Python
- Machine Learning: scikit-learn, PyTorch
- Data Processing: pandas, numpy
- Visualization: matplotlib, seaborn
- Clustering: K-Means, PCA for dimensionality reduction
- Matrix Factorization: SVD, ALS with regularization
- GPU Acceleration: CUDA-enabled training
The project utilizes the H&M Personalized Fashion Recommendations dataset from Kaggle, containing:
- Customer transaction history
- Product metadata (categories, prices, descriptions)
- Customer demographics
- Article images and attributes
- Multi-faceted Approach: Integrates collaborative filtering with customer segmentation for enhanced personalization
- Feature Engineering: Time-based features, price sensitivity analysis, and repeat purchase behavior
- Advanced Clustering: Customer segmentation using RFM analysis and K-Means clustering
- Performance Optimization: GPU-accelerated training with regularization techniques
| Model | Precision@10 | Recall@10 | MSE | Speed |
|---|---|---|---|---|
| SVD | 0.72 | 0.68 | 0.015 | Fast |
| ALS | 0.78 | 0.74 | 0.012 | Moderate |
Best Performing Model: ALS achieved superior accuracy with better handling of sparse data and higher personalization capabilities.
- Top Product Categories: Trousers, T-shirts, and socks dominate purchases
- Customer Segments: Identified distinct groups including high-value shoppers, occasional buyers, and price-sensitive customers
- Seasonal Trends: Transaction patterns show periodic spikes during promotional periods
- Channel Distribution: Analysis of online vs. in-store purchasing behavior
- Enhanced User Experience: Personalized recommendations reduce choice overload
- Increased Conversion Rates: Targeted suggestions improve sales performance
- Inventory Optimization: Insights into product popularity aid inventory management
- Customer Retention: Segmentation enables targeted marketing strategies
- Deep Learning Integration: Implement Transformer-based models (BERT4Rec, SASRec)
- Real-time Recommendations: Session-based dynamic adaptation
- Hybrid Approach: Combine content-based and collaborative filtering
- Image Recognition: Visual similarity recommendations using CNNs
- A/B Testing Framework: Continuous model improvement and validation
- Precision@K: Relevance of top-K recommendations
- Recall@K: Coverage of relevant items
- Mean Squared Error (MSE): Prediction accuracy
- Silhouette Score: Clustering quality assessment
Supervisor: Dr. M. Krishna Siva Prasad, Assistant Professor, CSE Department, SRM University-AP
SRM University-AP
Department of Computer Science & Engineering
Mangalagiri, Guntur, Andhra Pradesh - 522502
This project is licensed under the MIT License - see the LICENSE file for details.
- H&M for providing the comprehensive dataset
- Kaggle community for dataset hosting and resources
- SRM University-AP for academic support and guidance
- PyTorch and scikit-learn communities for excellent documentation
⭐ Star this repository if you found it helpful! ⭐