You are here:
How Recommendation Models Work
Objective-Based Recommenders use the Deep Learning for Personalized Recommendations (DLPR) model to deliver personalized recommendations that align with your business goals. DLPR powers every recommender.
How Recommenders Use DLPR
A recommender is what you configure to serve a specific personalization need, such as recommending products on a web page, presenting offers in an email, or returning items in an Agentforce conversation. Each recommender defines:
- Where it gets its item or catalog data from
- Where it gets its customer profile data from
- What customer engagements it uses to gather data
- What filters it uses and how it applies them to results
- Whether to use a rules-based strategy or an objective-based approach
When using an objective-based recommender, an incoming request triggers the use of a machine learning model to score and rank personalized items to return as recommendations.
The only automatic switch the recommender makes is between cold start and full personalization. When the recommender doesn't have enough engagement data, it uses cold start strategies based on catalog content. After data accumulates, it transitions to full personalization.
You can configure recommenders for different personalization needs. Examples include:
- Item recommenders that use product engagement signals and your product catalog
- Promotion recommenders that use offer-related engagement signals and your promotion catalog
- Intent-based recommenders that use conversational intent as a signal alongside engagement history
The signals available for training depend on what's flowing into Data 360 and what you map to the recommender.
The Recommender Model Foundation
To ensure consistency, every recommender shares a common foundation model.
- Trained only on your data: Each recommender gets a customized version of DLPR that's trained only on your data. The data of one Salesforce customer doesn't affect recommender performance or personalized output for another Salesforce customer.
- Driven by your business objective: By defining the goal, such as maximize revenue, increase add-to-cart actions, or lift offer redemption, the DLPR aligns its output to that outcome.
- Continuously improving: DLPR retrains regularly. As engagement data grows, recommendations improve.
- Privacy by design: DLPR doesn't use demographic data, personally identifiable information beyond engagement interactions, or data purchased from third parties.
- Deep Learning for Personalized Recommendations
Deep Learning for Personalized Recommendations (DLPR) is the core model that powers every objective-based recommender you configure. It analyzes the sequence of engagement actions a user takes, and predicts the item that the user is most likely to engage with next, weighted by the business goal that you've defined. - Item Recommenders
An item recommender presents items from your catalog based on each user's engagement history. An item can be a product, an article, a property listing, a course, a service, or any other entity you serve to users. The item recommender is the default recommender configuration for personalized recommendations across web, email, Agentforce, and other channels. - Promotion and Offer Recommenders
A promotion recommender recommends the most relevant promotions, offers, or discounts to each user. The underlying model is the same DLPR, but mapping different engagement signals to the recommender, and ranking it against a different catalog. - Conversational Intent-Based Recommenders
A conversational recommender brings personalization into Agentforce by interpreting a user's intent in real time and feeding it into DLPR alongside engagement history. Live conversation becomes a primary engagement signal. - 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.
