Deep Neural Networks for YouTube Recommendations
Two-stage retrieval + ranking architecture, dynamic user embeddings, watch-time optimization, and the elegance of softmax-to-ANN inference
Recommender Systems with Generative Retrieval (TIGER)
Semantic IDs via RQ-VAE, the straight-through estimator, and generative retrieval with an encoder-decoder Transformer
Multi-gate Mixture of Experts (MMoE): Multi-Task Learning
Multi-task learning through specialized expert networks and task-specific gating
Deep & Cross Network (DCN): Explicit Feature Crossing
Explicit feature crossing and deep learning for improved recommendations
Deep Interest Network (DIN): Attention for Recommendations
Attention-based recommendation models and dynamic user interest representation
OneSearch: From Search-Specific Semantic IDs to Generative Ranking and DPO
Search-aware Semantic IDs (RQ + OPQ), generative retrieval with beam search, and using DPO to distill a reward model’s ranking preferences into the generator