Implementation of a Hybrid Two-Tower Architecture for Music Recommendation. Solves cold-start problems by integrating content-based multimodal encoders with sequential user modeling.
Production-grade ML recommendation engine: Kafka → Flink → Two-Tower PyTorch + FAISS ANN + LinUCB Bandit → sub-100ms FastAPI serving with live React dashboard
Two-stage music recommender: PyTorch two-tower retrieval and an XGBoost LambdaRank ranker on Music4All-Onion, scored on a split by time against ALS, with a popularity-bias and genre-fairness analysis.
Multi-stage restaurant recommender system on 1.3M Google Maps ratings: two-tower retrieval with logQ correction, Semantic IDs, SASRec Transformer, LightGBM learning-to-rank, leakage-safe evaluation and a serving API.
Recipe recommender using Two-Tower retrieval, LightGBM ranking, and MMR diversification
This repo is made for Football analytics and is based on FIFA 24 data
Production-ready personalized recommender: two-tower retrieval, CatBoost ranking, optional LLM rerank. FastAPI + Qdrant + MLflow. Workshop by learnwithparam.com
Two-tower recommender (users - crypto tokens) in PyTorch, with a feature-based user tower and cold-start.
Production-grade MovieLens-25M recommender: two-stage retrieval + re-ranking with a full MLOps lifecycle (Spark features, MLflow, FAISS, FastAPI on Kubernetes, A/B testing, drift detection, Prometheus/Grafana).
🤖 Explore and optimize rewards with Bandexa, a PyTorch-native library for Neural-Linear Thompson Sampling in contextual bandits.
Two-tower retrieval with a from-scratch C++20 HNSW index and filter-aware vector search. Filters are applied during graph traversal and the strategy adapts to filter selectivity. Python bindings, benchmarked against FAISS and hnswlib.
Two-tower dual-encoder recommendation/retrieval with in-batch negatives and FAISS ANN (Recall@10/50, MRR)
A distributed ML recommendation system — real-time streaming, multi-node distributed training, and fault-tolerant, scalable serving.
The ‘recommended for you’ problem, step one: a two-tower model and FAISS narrow thousands of movies to the best 500 in under a millisecond, beating popularity on MovieLens 1M (recall@10 0.083 vs 0.048).
Short-video recommendation reproduction: Two-Tower recall, Faiss TopK, time split, negative sampling, Recall@50/NDCG@50, ablation and badcases.
2-Stage Recommendation on MovieLens-1M.
Two-stage recommender on MovieLens 100K: two-tower retrieval + FAISS + wide-and-deep ranker behind FastAPI, measured Recall/NDCG.
Two-tower neural recommender on MovieLens 100K (PyTorch, FAISS, FastAPI): Hit@10 of 56.6% vs 27.7% for matrix factorization under the same temporal-split protocol.
A recommender on 25M real MovieLens ratings: temporal split, two-tower retrieval, and the logQ correction that took it from losing to a popularity counter (0.018) to beating every baseline (0.085), served through this series' own HNSW
Two-tower reading recommender plus an off-policy 'opportunity lab' — retrieval, LambdaMART reranking and causal evaluation.
To associate your repository with the two-tower topic, visit your repo's landing page and select "manage topics."