Yoo Video Bottom‑Page Recommendation System: From Zero to One Practice

The talk presented by Tencent senior researcher Qian Dingding introduces the architecture and practice of the Yoo video bottom‑page recommendation system, covering business background, recall and ranking pipelines, and engineering implementation.

Business background : describes the bottom‑page scenario where videos following a clicked video must be recommended based on relevance to the main video and user context.

Recall techniques : outlines four recall directions—popularity, attribute‑based, content‑based, and behavior‑based—detailing collaborative‑filtering optimizations, image embedding with Inception+FAISS, title/tag embedding, and network/graph embedding methods.

Ranking techniques : explains the evolution from LR to Deep and Wide&Deep models, the use of sparse feature IDs, embedding layers, DNN, Wide&Deep, and Wide&DCN, and the tree‑based deep match approach for full‑library retrieval.

Engineering practice : discusses the service flow (profile service, trigger, merge, ranking, exposure), the underlying infrastructure (Faiss, Redis, Hippo, Hadoop, Spark, Angel, TensorFlow), feature handling (plain vs TFRecord, ID‑based features), offline training, online inference with TFServing vs TensorFlow API, and performance‑tuning techniques.

Future directions : mentions plans to fuse behavior and content embeddings, apply graph neural networks, explore sequence models (LSTM/GRU), bandit and reinforcement learning for exploration, multi‑task learning, and multimodal integration.