Optimizing Coarse Ranking Models for Short Video Recommendation: From GBDT to Dual‑Tower DNN and Cascading

The industrial recommendation system typically consists of recall, coarse ranking, fine ranking, and re‑ranking stages, each acting as a funnel to filter massive item pools. The article introduces the iQIYI SuiKe basic recommendation team's practical improvements on the coarse‑ranking model for short‑video feeds.

Background : Traditional coarse‑ranking models prioritize performance and are categorized into simple score truncation, LR/decision‑tree based machine‑learning models, and the widely used dual‑tower DNN that computes user and item embeddings via deep networks. iQIYI originally used a GBDT model based on statistical features.

Dual‑Tower DNN Coarse‑Ranking Model : Chosen for its balance of computation and effectiveness, the model contains three fully‑connected layers on both user and item sides, outputting 512‑dimensional embeddings. Feature selection is heavily trimmed: user side uses a few basic profile, context, and historical behavior features; item side retains only video ID, uploader ID, and video tags.

Knowledge Distillation : To compensate for feature pruning, a teacher‑student framework is employed where the fine‑ranking wide&deep model serves as the teacher. The student loss, teacher loss, and distillation loss (MSE with a scaling hyper‑parameter λ that grows with training steps) are combined to train a lightweight yet expressive coarse‑ranking model.

Embedding Parameter Optimization : Embedding parameters are optimized with the sparse‑solver FTRL while other layers use AdaGrad. This reduces model size by 46.8%, cuts inference memory by 100%, and increases transmission speed nearly twofold.

Online Inference Optimizations : User‑side embedding computation is de‑duplicated so each user‑candidate pair passes the user network only once, reducing p99 latency by ~19 ms. Video‑side embeddings are cached for high‑frequency items, further accelerating scoring.

These optimizations allowed the dual‑tower DNN coarse‑ranking model to match the previous GBDT performance while delivering significant gains in user engagement metrics for both the iQIYI hotspot channel and the SuiKe homepage feed.

Cascade Model : To keep coarse‑ranking objectives aligned with evolving fine‑ranking goals, the team adopted a cascade architecture that directly learns the fine‑ranking model’s predictions as training targets, eliminating the distillation step and reducing training resources. This change yielded ~3 % improvements in exposure‑click rates and video watch metrics.

Future Plans : Continue exploring next‑generation coarse‑ranking systems (COLD), expand recall volume with more features, refine user/item embedding similarity calculations (potentially replacing cosine similarity with a shallow network), and further improve online performance.

References :

1. https://www.kdd.org/kdd2018/accepted-papers/view/modeling-task-relationships-in-multi-task-learning-with-multi-gate-mixture- 2. H.B. McMahan, Follow‑the‑regularized‑leader and mirror descent: Equivalence theorems and L1 regularization, AISTATS 2011 3. https://arxiv.org/abs/1503.02531 4. https://arxiv.org/abs/2007.16122