Multi‑Interest Recall Techniques in iQIYI Short‑Video Recommendation

The recommendation system fundamentally acts as an information filter, progressively presenting the most relevant content to users through multiple funnels; the recall stage is the first funnel that extracts potentially interesting videos from massive inventories for downstream ranking.

Technical Background : Effective recall is likened to early-stage athlete selection, requiring diverse sources to avoid limiting the quality of items fed to ranking models. Traditional recall strategies include item‑based collaborative filtering, item2vec/node2vec embeddings, and graph‑based methods.

01. Clustering Multi‑Interest Recall : Utilizes existing video embeddings (e.g., node2vec, item2vec) and hierarchical clustering (PinnerSage) to form multiple interest vectors without training complex neural networks, reducing time and space costs while mitigating information loss from simple pooling.

02. MOE Multi‑Interest Recall : Extends the dual‑tower model by replacing the user tower with a Mixture‑of‑Experts (MOE) architecture that outputs several user‑interest vectors; negative sampling is performed within each batch and focal loss improves hard‑sample discrimination.

03. Single‑Activation Multi‑Interest Recall : Inspired by MIND, replaces capsule networks with a transformer‑based interest extractor; during training only the most activated interest vector participates in sampled softmax loss, while inference retrieves candidates from all interest vectors.

04. Disagreement‑Regularization : Adds regularization terms on video embeddings, attention weights, and interest vectors to prevent learned interests from becoming overly similar, improving diversity.

05. Dynamic Capacity Multi‑Interest Recall : Introduces an activation record table that tracks which interest vectors are used during training; at inference time, rarely activated vectors are pruned, allowing the number of interests to adapt to each user.

06. Multimodal Feature Multi‑Interest Recall : Incorporates uploader and tag embeddings alongside video IDs, using pooling for multiple tags and retaining video‑ID embeddings for negative sampling to enhance precision while enriching feature representation.

Conclusion & Outlook : Multi‑interest recall transforms the “one‑size‑fits‑all” paradigm into a “thousands‑of‑faces” approach, improving both relevance and diversity, while future work may explore richer behavior signals, negative feedback integration, and static user profile fusion.