Concept‑aware Denoising Graph Neural Network (CONDE) for Short Video Recommendation

Background: short‑video platforms like Kuaishou face massive data streams, leading to challenges in recommendation due to noisy interactions, sparse feedback, and the Matthew effect.

To address these, the authors propose CONDE, a concept‑aware denoising graph neural network that builds a heterogeneous three‑part graph linking users, items (videos), and semantic concepts extracted from video titles and comments.

The model first propagates concept information to videos and users via attention‑based graph convolutions (warm‑up stage), then performs a personalized denoising process. Using a GRU to encode neighbor representations, edge sampling probabilities are obtained via Gumbel‑Softmax, retaining the most informative edges and discarding noisy ones, iterated over multiple sub‑graphs.

Finally, user and item embeddings from the pruned sub‑graphs are combined in a prediction layer that computes interaction probabilities.

Experiments on large Chinese and English datasets (including Kuaishou short‑video logs and Amazon e‑commerce data) show that CONDE improves AUC by nearly 6 % over state‑of‑the‑art baselines, with consistent gains across HR, MAP, and NDCG. Ablation studies confirm the importance of both denoising stages, especially concept‑level filtering.

Hyper‑parameter analysis demonstrates the impact of neighbor count, sub‑graph number, and initial temperature on performance. The model also exhibits robustness on both popular and long‑tail videos, and provides interpretable recommendations by revealing which concepts drive each user’s preferences.