Graph-Based I2I Recall for Short Video Recommendation at Kuaishou

Short‑video recommendation faces challenges such as high noise, diverse targets, and cold‑start problems; Kuaishou's platform must efficiently retrieve relevant items from massive pools while ensuring precision and diversity.

The recall stage is the first step of the recommendation system, requiring fast, accurate, and diverse candidate selection. Traditional one‑hop methods suffer from data sparsity and bias, prompting the use of high‑order relationships captured by graph structures.

Kuaishou implements a graph‑based I2I recall pipeline built on internal platforms (IDP, KML) and open‑source tools (DGL, FAISS). The pipeline processes raw data, generates item pairs, constructs a heterogeneous graph, samples neighbors and negative examples, trains embeddings, and iterates with user feedback.

Optimization techniques include:

Similarity‑measure improvement by weighting the numerator of the Adamic‑Adar score with the smaller of two edge capacities, yielding higher offline hit‑rate.

Graph‑structure learning in collaboration with Tsinghua's Cui‑Feng lab, using confidence networks and set‑level resampling to denoise data, followed by a VAE‑based encoder‑decoder that reconstructs purified user‑item sets, improving robustness to injected noise.

Edge‑weight learning that predicts edge confidence, enabling weighted sampling during embedding training, which accelerates convergence and enhances model robustness.

Online observations show that graph recall slightly reduces precision compared with pure target‑oriented I2I but significantly improves diversity, especially in commercial ad scenarios where it surfaces varied product styles.

Future work aims to model multi‑interest users via embedding distributions, handle constrained recall pools for ads, incorporate contrastive and unsupervised learning for denoising, and improve interpretability and dynamic graph updates.