How a New AI System Gives E-Commerce Video Feeds a 50% Boost — Without Any User History

E-commerce platforms are shifting from static search to dynamic video feeds, but new short-form videos face an 'extreme cold-start' problem — no interaction history for collaborative filtering, plus strong position and duration biases. Researchers present VCG (Video Candidate Generation), a scalable multimodal retrieval engine that uses a domain-adapted CLIP model to map users and videos into a shared semantic space, enabling zero-shot retrieval based on visual content alone.

In rigorous evaluation, VCG compared generative (LLM) vs. discriminative (CLIP) embeddings. Generative models excelled at attribute prediction but suffered from embedding space collapse in retrieval tasks. VCG's CLIP-based approach mitigated engagement biases, yielding a 50% uplift in deep video completion in online A/B tests. The system also features three interactive retrieval scenarios: Product-to-Video, Video-to-Product, and Zero-Shot Semantic Search, demonstrating practical deployment for live e-commerce feeds.