VCG: A Multimodal Retrieval Framework for E-Commerce Video Feeds under Extreme Cold-Start Conditions

A new retrieval framework called Video Candidate Generation (VCG) aims to solve the extreme cold-start problem in e-commerce video feeds by using visual content instead of behavioral history to match users with short-form videos [1]. The digital commerce landscape is shifting from static, search-driven catalogs to dynamic, immersive video feeds, a transition that creates an extreme cold-start problem because new short-form videos lack the dense interaction history required for traditional collaborative filtering [1]. Immersive feeds also introduce strong position and duration biases that distort standard engagement signals [1]. The VCG system, detailed in a paper submitted on June 17, 2026, addresses these challenges by leveraging a domain-adapted vision-language model based on CLIP to map users and videos into a shared semantic space, enabling zero-shot retrieval [1]. The researchers conducted a rigorous evaluation comparing generative large language model embeddings against discriminative CLIP embeddings [1]. While generative models excelled at attribute prediction, they suffered from embedding space collapse in retrieval tasks [1]. Online A/B testing demonstrated that VCG effectively mitigates engagement biases, yielding a 50% uplift in deep video completion [1]. The system supports three bi-directional retrieval scenarios: Product-to-Video, Video-to-Product, and Zero-Shot Semantic Search [1]. The paper was authored by Katsiaryna Mirylenka and submitted to arXiv under the Information Retrieval category [1]. The manuscript totals 4,513 KB [1]. The work arrives as e-commerce platforms increasingly prioritize video-first experiences, where recommendation engines must surface relevant content without relying on prior user interactions [2]. The VCG architecture represents a scalable approach for large-scale environments where new inventory is continuously introduced and immediate relevance is critical [2].