You are here:
Cold Start Strategies for Recommendations
Cold start describes any scenario where a recommender doesn't have enough engagement data to personalize confidently. Cold start is the only automatic switch the recommender makes. Every recommender starts in cold start when engagement is sparse and transitions to full personalization as engagement data accumulates.
Cold Start Scenarios
- New Site
- A brand-new customer onboards or introduces a new catalog, and no engagement history exists.
To handle this scenario, the recommender builds an understanding of your catalog from its content, including titles, descriptions, categories, and other attributes. These attributes are converted into semantic embeddings that capture the meaning of each item. When a user interacts with any item, the recommender returns semantically similar items.
- New Item
- An established site adds items that no user has interacted with yet. Traditional recommenders can't surface items without engagement history.
DLPR is trained to use both engagement history and catalog content. For items with engagement, both signals contribute to the item's representation. For new items, catalog content acts as a stand-in until engagement data arrives. New items are eligible for recommendation immediately.
- Promotion and Offer Without Engagement History
- When you recommend promotions or offers but have limited engagement data on user responses, the recommender uses promotion metadata to recommend semantically similar promotions. For details, see Promotion and Offer Recommenders.
How a Recommender Exits Cold Start
As engagement data accumulates, the recommender continuously evaluates whether it has enough signal to switch from cold start to full personalization. When the threshold is met, future training runs use engagement data, and cold start logic steps back.
The transition is automatic. You don't configure a switch. As your site matures, users see a smooth progression from content-based recommendations to fully personalized recommendations.
Guidance for Your Catalog
A well-structured catalog makes early recommendations effective.
- Provide complete titles, descriptions, categories, and attributes for every item.
- Keep catalog metadata current as you add new products and offers.
- Don't wait for traction on new items - items with good metadata are eligible to recommend immediately.
