Promotion and Offer Recommenders

Why Promotions Need a Different Set of Signals

Promotion catalogs behave differently from product catalogs in several ways.

• Offers have lifecycles. Offers launch, run for a specific window of time, and then retire. Engagement history is sparse by nature.
• User engagement is limited. Most users engage with far fewer offers than products.
• Relevance depends on eligibility. An offer that fits one user doesn't always apply to another.

Because of these differences, you map offer-specific engagement signals and the promotion catalog to the recommender rather than the product catalog.

How a Promotion Recommender Works

DLPR follows the same recommend-observe-reward-learn feedback loop as item recommenders, but learns from how users interact with promotions and the items those promotions target. The recommender draws on the user's recent activity, the promotion catalog, and the relationships between offers and the items they apply to. The recommender then ranks offers in a way that reflects both relevance to the user and alignment with your business goal.

Engagement Signals

To get the most out of a promotion recommender, provide signals across two related areas:

• Offer engagement: How users interact with promotions themselves, such as views, clicks, redemptions, and custom signals tied to your offer lifecycle.
• Targeted item engagement: How users interact with the items each offer applies to, such as product views, add-to-cart actions, and purchases.

Mapping signals for both areas gives the recommender a clearer picture of which offers actually drive the behavior tied to your goal.

You can also provide additional signals that help shape relevance, such as engagement with related content (articles, campaigns, landing pages) and profile metadata (preferences, segments, loyalty status). The richer the signal mix, the better the recommender can tailor offers to each customer.

Cold Start for Promotions

Promotion engagement is often sparse. For this reason, cold start is often the norm rather than the exception. The recommender uses one or more content-based strategies to surface relevant offers even when engagement data is limited.

Promotion-to-Promotion Similarity

DLPR learns from promotion metadata, including offer name, description, type, category, and terms, to generate semantic embeddings for every promotion in your catalog. At inference, the recommender identifies the user's most recently engaged promotion and returns similar ones.

You can launch offer recommendations with minimal engagement history, provided your promotion catalog has descriptive metadata.

Promotion-to-Product Similarity

The recommender can also recommend promotions based on the products a user has engaged with, bridging product and promotion metadata to surface offers that relate to items the user already cares about.

Model Inputs

A promotion recommender uses the following inputs to rank offers:

• Promotion catalog metadata. Offer names, descriptions, offer types, categories, eligibility, and any attributes you provide.
• Promotion engagement signals. Views, clicks, redemptions, and other interactions with offers.
• Objective configuration. The business goal that recommendations optimize for.

Guidance for Your Promotion Catalog

A well-structured promotion catalog is the foundation of cold start quality. Follow these best practices to maximize recommendation relevance from day one.

• Provide complete, descriptive names and descriptions for every offer.
• Provide consistent offer types and categories.
• Provide eligibility attributes where relevant.