Graph-Based I2I Recall for Short Video Recommendation at Kuaishou

The presentation introduces the short‑video recommendation background at Kuaishou, highlighting business characteristics such as high noise, multiple objectives, and cold‑start issues, and explains why recall is a critical first stage in the recommendation pipeline.

It argues that graph neural networks can capture high‑order relationships among users, items, and behaviors, overcoming the sparsity of one‑hop interactions and improving diversity and robustness of recommendations.

The core of the solution is a graph‑based I2I recall pipeline built on internal platforms (IDP, KML, DGL, FAISS). The pipeline processes raw data, generates item pairs, constructs graphs, samples neighbors and negative edges, trains models, produces embeddings, and iterates with online feedback.

Three main optimization methods are described:

Similarity metric refinement using a weighted Adar/Jaccard approach to improve offline hit‑rate.

Graph structure learning in collaboration with Tsinghua, employing confidence networks, set‑level resampling, and a VAE to denoise data and produce robust embeddings.

Edge‑weight learning that iteratively refines edge confidence, reducing noise and accelerating convergence.

Online observations show that graph recall, while slightly less precise than target‑driven recall, significantly enhances diversity, especially in commercial advertising scenarios.

The future outlook includes multi‑interest modeling via embedding distributions, constrained recall pools for ads, contrastive and unsupervised learning for denoising, interpretability, and dynamic graph optimization.