Graph-Based I2I Recall for Short Video Recommendation at Kuaishou

Background and Challenges Short‑video recommendation serves as the first stage of the recommendation system, requiring efficient filtering of massive content pools while maintaining relevance. The domain faces three main pain points: high noise, multiple objectives, and cold‑start issues, compounded by the concentration and dynamic nature of user‑item interactions.

Why Use Graphs Traditional one‑hop (first‑order) relationships are sparse and lead to bias toward popular items. Modeling users and items as a graph captures higher‑order, non‑linear relationships, enriching positive samples, improving model generalization, and enhancing diversity and latent interest activation.

Graph‑Based I2I Recall Pipeline Kuaishou’s pipeline is built on internal platforms (IDP, KML) and open‑source tools (DGL, FAISS). It processes raw data, generates item‑pair samples, constructs the graph, performs neighbor and negative‑sample sampling, trains models to produce embeddings, and finally serves I2I recommendations in an online‑offline decoupled manner. The design balances flexibility, rapid experimentation, dynamic node updates, heterogeneous graph support, diverse sampling strategies, and optimized execution efficiency.

Practical Optimizations Three key optimizations are presented: 1. Similarity Measurement : Enhancing neighbour‑based similarity (e.g., Jaccard, Adamic‑Adar) with a trick that weights the numerator by the smaller of two edge capacities, yielding noticeable offline hit‑rate gains. 2. Graph Structure Learning : A collaborative effort with Tsinghua’s Cui‑Feng Lab introduces a confidence network to filter noisy user‑item interactions, followed by a VAE‑based encoder‑decoder that learns robust embeddings, demonstrating superior noise resistance (accepted at KDD). 3. Edge‑Weight Learning : An end‑to‑end approach learns edge confidence scores, enabling weighted neighbor sampling and reducing training iterations while improving robustness against noisy edges.

Online Characteristics In production, graph‑based recall shows slightly lower precision than pure target‑oriented recall but significantly improves diversity, delivering a broader set of videos and ads. Case studies illustrate how graph recall introduces varied product styles compared to homogeneous recommendations.

Future Outlook The authors discuss potential directions such as multi‑interest modeling via embedding distributions, constrained recall pools for commercial content, contrastive and unsupervised learning for noise robustness, interpretability, and dynamic graph optimization.

Overall, the presentation provides a comprehensive view of applying graph neural networks to large‑scale short‑video recommendation, covering system design, empirical improvements, and research opportunities.