What Is Decisioning Infrastructure for Consumer Platforms?
Consumer platforms make thousands or millions of decisions every second about what users see.
Spots 
A social platform decides which posts appear in
A social platform decides which posts appear in a feed. A marketplace decides which listings appear first. A creator platform decides which creators or content to recommend. A dating app decides which profiles to surface. A job board decides which jobs should appear at the top of a candidate's search.
Behind all of these experiences is a common
Behind all of these experiences is a common technical problem: given a set of possible items, which items should be shown, in what order, and under what business rules?
Traditionally, companies build this capability themselves using a
Traditionally, companies build this capability themselves using a combination of retrieval systems, ranking models, recommendation engines, business rules, and separate advertising infrastructure. As platforms grow, this layer becomes increasingly complex to operate. This is where decisioning infrastructure comes in.
Decisioning infrastructure is the layer that sits between
Decisioning infrastructure is the layer that sits between candidate generation and the user-facing product. It takes candidate items and contextual information, applies ranking and business logic, determines the final ordering and monetized placements, and records the decision. What Is Decisioning Infrastructure?
Decisioning infrastructure is a software layer responsible for
Decisioning infrastructure is a software layer responsible for making real-time decisions about what a consumer platform should display. A simplified architecture looks like this: User + Context → Candidate Generation → Decisioning → User Interface The candidate-generation layer answers: The decisioning layer answers: "What should we show, in what order, and where should monetized inventory appear?" "What should we show, in what order, and where should monetized inventory appear?" The distinction is important.
Modern recommendation systems commonly separate candidate generation, scoring
Modern recommendation systems commonly separate candidate generation, scoring, and re-ranking. Google describes candidate generation as narrowing a potentially huge corpus into a smaller set, followed by scoring and re-ranking to determine what ultimately appears to the user. Decisioning infrastructure extends this final part of the architecture into an explicit production layer. Decision logging and explanations
Instead of embedding all of this logic directly
Instead of embedding all of this logic directly into an application, a platform can expose it through a dedicated decisioning service. Why Consumer Platforms Need a Separate Decisioning Layer At first, ranking can look relatively simple.
A marketplace might sort products by relevance. A
A marketplace might sort products by relevance. A social network might sort posts by predicted engagement. A job board might sort jobs by relevance to a candidate. As the product grows, however, ranking becomes a multi-objective problem. The platform may simultaneously need to consider: What the user is likely to find relevant What is available right now What the user has already seen Whether an item is eligible for the surface Whether a business rule should override the model Whether sponsored inventory should be displayed How much sponsored inventory the experience can tolerate How monetization affects the final ordering This creates an infrastructure problem rather than simply a machine-learning problem.
LinkedIn, for example, describes a multi-stage ranking architecture
LinkedIn, for example, describes a multi-stage ranking architecture in which candidate generation first selects candidates from a very large inventory, followed by ranking stages that calibrate and reduce the candidate set against a common objective.
News
What Is Decisioning Infrastructure for Consumer Platforms?
Consumer platforms make thousands or millions of decisions every second about what users see.
@spots #dev
Source: Dev.to