Feizhu Information Flow Content Recommendation: Architecture, Cold-Start Strategies, Multi-Modal Understanding, and Ranking Mechanisms

Feizhu information flow is a mixed‑material recommendation scenario embedded in the homepage’s “Guess You Like” module, combining videos, images, text, products, ads, and POI themes to present travel‑related content.

The technical framework consists of selection, recall (item‑based, user‑based, and cold‑start), coarse ranking, fine ranking, and mixing, with multi‑objective models tailored to different material types such as CTR, CVR, dwell time, likes, and completion rate.

The content delivery chain includes a creation platform, a content middle‑platform for acquisition, moderation, and tagging, content algorithms for classification and vector representation, a first‑guess distribution pool for pre‑selection and recall, and a feedback loop with human review for post‑delivery quality control.

Cold‑start recall addresses the scarcity of interaction data for new items by constructing graph‑based similarity transfers from old to new content, using both attribute‑based and multimodal similarity graphs, while coarse ranking employs a dual‑tower MLP model that emphasizes side‑information rather than historical exposure statistics.

Multi‑modal understanding leverages CLIP pre‑training on public image‑text pairs followed by UNITER‑style fine‑tuning, incorporating masked language modeling, masked image modeling, and multi‑label classification to generate robust joint embeddings for travel content.

Fine ranking incorporates rich side information, a user‑state gate that adjusts weights for different user groups and material categories, and a bias‑correction network to mitigate artificial click inflation from human‑portrait cover images.

Mixing combines heterogeneous materials by first aligning their feature spaces and then applying a lightweight expert‑gating network at the top level, using per‑material CTR estimates to preserve order while keeping latency low.

The traffic‑mechanism design introduces tiered pools with quality‑assessment classifiers, bandit‑based trial allocation, and incentive‑driven promotion, enabling transparent flow control and efficient content scaling.

The presentation concludes with a summary of the year‑long explorations and practical implementations in Feizhu’s content recommendation system.