As digital commerce evolves from static, search-driven catalogs to dynamic, immersive video feeds, enterprises encounter a critical challenge: the "extreme cold-start" problem. New short-form videos, unlike traditional product items, lack the dense interaction history required for collaborative filtering. Additionally, immersive video feeds introduce strong position and duration biases that distort standard engagement signals, making it difficult for recommendation systems to surface relevant content.
A team of researchers from an undisclosed large-scale e-commerce environment (the paper does not name the company) has proposed the Video Candidate Generation (VCG) system, a scalable multimodal retrieval engine designed specifically for these conditions. The framework is detailed in a preprint on arXiv (arXiv:2606.19627).
The Extreme Cold-Start Problem
Traditional recommendation systems rely on user interaction data — clicks, watches, purchases — to train collaborative filtering models. In a video feed context, new videos have zero history, leading to the extreme cold-start scenario. Moreover, the layout of feeds creates position bias: videos at the top of a scrollable list are far more likely to be watched than those further down, regardless of relevance. Duration bias also skews signals because longer videos may accumulate more partial views. According to the paper, these biases distort the engagement metrics typically used to train recommenders.
VCG: Architecture and Approach
VCG addresses these issues by leveraging a domain-adapted vision-language model based on CLIP. It maps users and videos into a shared semantic space, enabling zero-shot retrieval based on visual content rather than behavioral history. This means a new video can be recommended immediately using only its visual features, without any past interaction data.
The system architecture is built around a multimodal encoder that processes both video frames and associated metadata. The embedding space is trained on e-commerce-specific data, adapting the general CLIP model to product categories, styles, and visual patterns relevant to shopping.
Evaluation: Generative vs. Discriminative Embeddings
The paper presents a rigorous comparison between two embedding approaches: generative models (LLMs) and discriminative models (CLIP) . The researchers found that while generative models excel at attribute prediction — for example, correctly identifying product features — they suffer from embedding space collapse in retrieval tasks. This means that semantically different items end up with nearly identical embeddings, degrading recall. In contrast, discriminative CLIP embeddings maintained a well-structured space, allowing precise matching between users and videos.
Online A/B Testing Results
To validate the system in a production environment, the team conducted online A/B testing. The results showed that VCG effectively mitigates engagement biases — specifically, position and duration biases — leading to a 50% uplift in deep video completion, a key metric measuring how much of a video a user watches. This indicates that users are finding more relevant content, even for brand-new videos.
Bi-Directional Retrieval Scenarios
VCG supports three bi-directional retrieval scenarios, showcased in an interactive demonstration:
- Product-to-Video: Given a product image, retrieve relevant videos.
- Video-to-Product: Given a video, retrieve relevant product listings.
- Zero-Shot Semantic Search: Search for videos using natural language queries without any prior training on those terms.
These capabilities enable enterprises to build more dynamic and responsive video feeds that can adapt to user intent in real time.
Implications for Enterprise Technology Leaders
For CTOs and digital commerce directors, the VCG approach offers a concrete path to deploying video-based shopping experiences without sacrificing recommendation quality. By replacing behavioral history with visual understanding, companies can reduce the onboarding lag for new content and lower the barrier for user-generated video contributions. The 50% improvement in deep completion suggests significant gains in user engagement and, potentially, conversion rates. While the paper does not disclose the exact infrastructure stack, the reliance on CLIP — an open-source model — implies that the solution can be built with existing cloud AI services and custom fine-tuning pipelines.