How iOS 15 Ranks Notification Summaries With ML

Apple introduced the Notification Summary in iOS 15 to help users minimize distractions by grouping non-urgent alerts into scheduled daily digests. Rather than displaying these notifications chronologically, the system uses on-device machine learning to rank and format them based on perceived relevance to the user. This article explains the underlying mechanisms iOS 15 uses to prioritize these non-priority notifications, including on-device behavioral analysis, developer relevance signals, and rich media weighting.

On-Device Machine Learning and Privacy

All ranking within the Notification Summary occurs locally on the device using Apple’s Core ML framework and the Apple Neural Engine. Apple does not send notification contents or user behavior metrics to cloud servers to determine rank. This on-device model continuously learns from individual device usage while preserving user privacy.

User Engagement Signals

The system evaluates how frequently and how quickly a user interacts with specific applications. Key behavioral factors include:

• App Usage Frequency: Notifications from apps that the user opens frequently are pushed toward the top of the summary.
• Direct Interaction History: The algorithm tracks whether a user typically taps, clears, or ignores notifications from a specific app. Alerts from apps that are regularly tapped receive higher priority.
• Routine and Context: The model considers temporal patterns, recognizing which apps a user typically interacts with at different times of the day (e.g., news apps in the morning versus entertainment apps in the evening).

Developer-Defined Relevance Scores

With iOS 15, Apple introduced the relevanceScore property within the UserNotifications framework (UNNotificationContent). App developers can assign a floating-point score between 0.0 and 1.0 to individual notifications.

• A score of 1.0 signifies the most critical non-urgent information within that app (such as an order status update).
• A score of 0.0 represents standard low-priority updates (such as promotional marketing).

The machine learning model combines this developer score with its own behavioral heuristics to determine the final order, ensuring that high-value updates within a single app stand out over generic promotional alerts.

Visual Formatting and Media Prioritization

The machine learning model also dictates the visual presentation within the summary card:

• Top Position (Hero Thumbnail): Notifications deemed most relevant, particularly those featuring rich media such as large thumbnails, images, or detailed previews, are given the prominent "hero" slot at the top of the summary.
• Stacked Alerts: Notifications with lower engagement likelihood and low relevance scores are condensed into smaller, collapsed rows toward the bottom of the digest.