project 04
Core Shopping Recommendation System
Industrial-grade recommendation system framework with both retrieval and ranking models. Includes state-of-the-art deep learning architectures and reusable neural network layers for building production-ready e-commerce recommendations.
Python · PyTorch · TensorFlow · XGBoost · LightGBM · FAISS
The problem
Building scalable recommendation systems requires implementing multiple complex models (retrieval + ranking) with proper feature engineering, evaluation metrics, and production considerations.
The approach
Created a modular framework with reusable layers (attention, cross-net, embeddings) and implementations of state-of-the-art ranking models (DIN, DIEN, DCNv2, BST) plus retrieval models (Two-Tower, Collaborative Filtering).
The outcome
Accelerated recommendation system development with battle-tested implementations based on industrial-grade patterns.
Key features
- Ranking Models: DIN, DIEN (AUGRU), DCNv2, BST, XGBoost/LightGBM Rankers
- Retrieval Models: Two-Tower (Dual Encoder), Matrix Factorization + BPR
- Reusable Layers: DIN attention, Multi-head attention, CrossNet, FM, SENet
- Custom Activations: Dice, GELU for CTR prediction
- Comprehensive Metrics: AUC, NDCG, MRR, gAUC, MAP
- Feature Processor for encoding and normalization