A newly launched product catalog with thousands of SKUs faced a classic cold-start problem with AI-driven recommendation: sparse interactions per item meant recommendations were weak. The cold start problem occurs when a recommender lacks sufficient data to make accurate suggestions, especially when a system is launched.
Cold starts reduce recommendation relevance, lower engagement, and slow personalization until enough signals accumulate. Companies try to minimize cold start period by content-based features and popularity and trends. So does Quarticon. But to accelerate learning and improve early personalization, Quarticon introduced precomputed candidates into the pipeline, what helped to quickly increse conversions.
What does “precomputed candidate” mean?
“Precomputed candidate” means artificially generated user-item interaction records created to resemble real behavioral data but not copied from any actual user. We train on a small seed of real interactions and metadata, then sample new interaction sequences conditioned on item features, user cohorts and session context.
Generated synthetic pool filled missing item-user vectors but the model was trained only on real behavioral data. As real interactions arrived they replaced corresponding synthetic records and the synthetic pool shrinked.
How does it work?
We started from a small but representative seed of real interactions – sessions, clicks, add-to-cart events and purchases paired with rich item metadata (category, price, brand, attributes) and coarse user cohorts. Generation combined a conditional VAE with rule-based session augmentations so the synthetic records respected key dependencies on item features, user cohorts and session context while adding controlled variability to session length, click patterns and price sensitivity.
Each synthetic record was tagged and designed to be replaced as soon as corresponding real interactions arrived. The operational flow kept synthetic data as a temporary coverage layer rather than a training signal. Downstream consumers read from a unified dataset. When a real interaction appeared for an item or user, it replaced the synthetic records for that cell.
Replacement rules were simple and conservative: precomputed records for an item were removed once that item reached a threshold of real sessions and purchases. This approach preserved training purity, avoided generator artifacts influencing models, and still supplied full-coverage inputs for systems that require dense data.
The precomputed coverage layer was rolled out in stages. We generated the pool, validated synthetic feature distributions and sequence statistics against the seed. The replacement mechanism and tagging were tested. In production, precomputed coverage was enabled for new and long-tail items while models continued training only on real interactions. Real data gradually decreased the precomputed pool as items accumulated genuine signals.
Results: a threefold increase
Within few days of staged rollout the targeted cohort’s conversion rate rose from 0.78% (popularity based recommendations) to 2.4% (precomputed candidates based recommendations) – a threefold increase in a very short period of time. Engagement metrics also improved. The gains came from presenting more relevant items early (thanks to plausible precomputed coverage for cold items) while ensuring recommender models learned only from authentic behavior.
It is extremely important for smaller sites (less interactions, many products), where cold start period could be counted not in weeks, but in months. Few interactions per user/item make real-time collaborative signals weak – precomputed candidates models in Quarticon provide immediate, relevant recommendations.
Using precomputed candidates models as a temporary, replaceable coverage layer enabled rapid improvement in cold-start performance – delivering a clear, short-term conversion uplift while preserving model reliability and auditability.
Interested in AI Product Recommendations from Quarticon? Request more info.
