Douyin Upgrades Minor Mode Recommendation Engine with Multimodal AI
Douyin has upgraded its minor mode recommendation algorithm by integrating Multimodal Large Language Models (MLLMs) to evaluate and filter video content. The system analyzes text, audio, and video to determine age appropriateness and cognitive complexity for young users. This marks a significant real-world deployment of MLLMs for content moderation and personalized recommendations on a major social platform. It demonstrates how advanced AI can be used to address parental concerns and improve digital safety for minors. The workflow combines AI filtering with human expert review to build a dedicated minor content pool, which currently holds 7.7 million videos and adds 6,000 new ones daily. The algorithm also adapts to user feedback, allowing children to mark videos as "not interested" to refine recommendations.
## BACKGROUND
Multimodal Large Language Models (MLLMs) are deep learning algorithms capable of understanding and generating content across multiple formats, including text, images, audio, and video. Traditional recommendation systems often struggle to comprehend the deep semantic meaning of video content, making MLLMs a powerful tool for analyzing complex media.