Thinking Fast & Slow for Personalized Notification Systems

Netflix sends hundreds of millions of personalized notifications (Push, email, in-app notifications), and has struggled to balance short-term click-through with long-term user fatigue.

The existing system, based on a single Causal Model, was effective at optimizing short-term performance metrics, but had the following limitations.

• Focus on short-term rewards: By concentrating on reactions immediately after a message is sent, it failed to predict user fatigue from accumulated messages or the long-term risk of subscription cancellation.
• Coupled decisions: Because the decision of whether to send a message and the decision of content selection (Ranking) were made simultaneously within a single system, it was difficult to finely control per-user send frequency (Pacing).

To address this, Netflix introduced a hierarchical Slow-Fast architecture. This system mimics human cognitive processes and is divided into two layers.

The Slow Policy handles strategic decision-making. It establishes a weekly messaging plan for each member (send frequency and pacing per channel) to manage the long-term user experience.

The Fast Policy handles tactical decision-making. When a send opportunity arises, it selects the most relevant specific content in real time, within the bounds set by the plan, to drive an immediate response.