Deep Learning for Personalized Recommendations

How DLPR Learns

DLPR learns from the full sequence of user actions, not isolated clicks, using collective behavior patterns to predict engagement.

From User Journeys, Not Single Clicks

User behavior is sequential. Viewing a category leads to clicking an item, which leads to an action such as adding to cart or redeeming an offer. DLPR treats the full journey as a single unit of learning.

This approach is similar to how large language models (LLMs) predict the next word in a sentence. Given a user's engagement history, DLPR calculates the probability that the user will engage with each candidate item. It learns these probabilities from the browsing and purchasing patterns of all users on a site, not just an individual user.

From Collective Behavior

DLPR doesn't learn from only a single user's path. It analyzes the journeys of all users on a site and identifies which sequences of actions tend to lead to which outcomes. Every new user benefits from patterns learned across your entire user base.

Aligned to Your Business Goal

DLPR factors in both how likely a user is to engage with an item, and how much that item contributes to your business objective. For a maximize-revenue goal, items that drive higher order value carry more weight. The result is recommendations that are both personally relevant and aligned with your goal.

Model Inputs

DLPR learns from several sources.

Engagement signals

Interactions tied to your objective, such as product views and purchases for an item recommender, or offer clicks and redemptions for a promotion recommender. Engagement signals also include user intent indicators such as category browsing, search queries, and what was shown versus what was clicked. Each recommender feeds DLPR a different set of engagement signals.

Catalog metadata

Item attributes such as title, description, category, and brand. Catalog metadata helps DLPR understand items that don't yet have a rich engagement history.

Model Output

For every user, DLPR computes a representation that summarizes the user's journey. It scores every item in the relevant catalog against that representation and returns the top candidates, ordered by a combined likelihood and reward calculation.

DLPR generates recommendations in real time during inference, and retrains regularly to remain up to date with shifts in user behavior and catalog changes.

Considerations for DLPR Recommendations

Recommendation quality from DLPR depends on the engagement data it has learned from and how often the model retrains. Understand these dynamics so you can set expectations and plan around retraining cycles.

• DLPR learns from historical engagement data to identify what drives the behavior tied to your goal, such as purchases triggered by promotions. When the underlying data changes, such as a catalog overhaul, a major campaign shift, or new buying patterns, the model takes time to realign to what now drives the desired behavior. Recommendation quality improves with each retraining cycle as the model adjusts to the new patterns.
• DLPR retrains on a scheduled cadence. The specific frequency varies based on data volume and business unit configuration. To find your refresh schedule, contact your Salesforce representative.