Advances in Short Video Recommendation: Multi‑Objective Optimization and Model Enhancements

Short video has become a major growth driver for information‑flow platforms, offering high user engagement but posing challenges for frame‑level analysis and structured extraction. Improving distribution efficiency and personalized recommendation is a core capability of modern recommender systems.

Business and System Overview – UC embeds video recommendation into its browser, turning the product from a tool‑oriented app into a content distribution platform. The recommendation pipeline consists of three stages: recall, coarse ranking, and fine ranking, with the candidate set shrinking and accuracy increasing at each step.

Technical Evolution – Early solutions used Logistic Regression (LR), which required extensive feature engineering. Subsequent attempts with GBDT and FM showed weaker generalization. The current production model is a Wide & Deep (WnD) architecture, combining a fast LR side (optimized with FTRL) and a deep neural network side.

Multi‑Objective Optimization – Traditional multi‑objective models share a hidden layer and treat positive and negative samples equally, which conflicts with ranking‑oriented goals. The presented approach introduces a positive‑sample‑weighted loss that emphasizes ranking, handles sample imbalance, and incorporates additional signals such as watch time and continuity. This design yields noticeable online gains.

Graph Embedding Fusion – User click sequences are transformed into an undirected graph where items are vertices and adjacent clicks form edges. Random walks generate “sentences” that are fed to a CBOW Word2Vec model (DeepWalk) to obtain low‑dimensional item embeddings. These embeddings are then used to re‑weight the WnD loss with factors derived from position, watch length, and continuity, improving recommendation precision.

Logit Optimization – Because the LR side dominates training in high‑traffic scenarios, the DNN logit is scaled by a coefficient and added to the final prediction, increasing loss for poorly recognized samples while leaving well‑identified items unchanged.

Time‑Weighted Loss – A weight \(\text{weigh}=\frac{u_{play\_len}}{\text{avg\_item\_history\_view\_len}}\) (capped at 1) is applied to the loss, giving higher importance to items watched longer than average, thereby boosting both click‑through rate and consumption time.

Continuity Modeling – Sequences with high continuity scores (computed via the same graph‑embedding pipeline) are up‑weighted, encouraging users to continue watching related videos.

Boosting‑Based WnD – To overcome the limitation of a single decision hyperplane in binary classification, five parallel WnD models share an embedding layer and are combined using an AdaBoost‑style algorithm. The error term is weighted by AUC rather than the standard indicator, and only positive samples receive weight updates, resulting in a strong classifier with only a 6% increase in model size.

Future Plans – Move from point‑wise scoring to list‑wise recommendation, develop a retention model for new users, and improve gender balance in recommendations.