Notification Reranking
“For each user, which notification type will drive the highest engagement in the next 7 days?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
For each user, which notification type will drive the highest engagement in the next 7 days?
Apps send the same notification mix to every user — promotional, social, transactional, content updates — at the platform's preferred frequency. Users who only engage with social notifications get bombarded with promotions. Notification opt-out rates climb to 40-60%, and each opt-out permanently kills a high-value engagement channel. For a consumer app with 20M users, reducing opt-outs by 10 points preserves $15-25M in lifetime engagement value.
Quick answer
Notification reranking predicts which notification type will drive the highest engagement for each user over the next 7 days. Graph-based models rank notification types by learning from tap rates, dismiss patterns, and cross-user engagement trends, reducing opt-outs by 30-50% and doubling tap rates compared to one-size-fits-all notification mixes.
Approaches compared
4 ways to solve this problem
1. Fixed Notification Mix
Send the same ratio of notification types to all users: 40% promotional, 30% social, 20% transactional, 10% content. The default for most consumer apps.
Best for
Early-stage apps where user behavior data is sparse and the notification channel is still being established.
Watch out for
Users who only engage with social notifications get bombarded with promotions. Opt-out rates climb to 40-60%, permanently killing a high-value engagement channel for those users.
2. Frequency Capping by Type
Limit each notification type to N per day per user. Prevents the worst spam behavior but does not optimize the mix. Implemented in most notification platforms.
Best for
Reducing notification fatigue without requiring ML infrastructure. A necessary hygiene measure regardless of the optimization approach.
Watch out for
Caps are uniform across users. A power user who taps on 5 social notifications per day gets the same cap as a casual user who only wants 1. The cap prevents harm but does not optimize engagement.
3. Bandit / Reinforcement Learning
Use multi-armed bandit algorithms to explore and exploit notification types per user. Gradually learn which types each user responds to. Popular in experimentation platforms.
Best for
Teams with experimentation infrastructure who want to continuously optimize notification types while measuring impact.
Watch out for
Slow exploration phase. Each user needs many notification interactions before the bandit converges on the optimal mix. For users who receive 1-2 notifications per day, convergence takes weeks. Also cannot anticipate emerging preferences since it only learns from explicit feedback.
4. KumoRFM (Graph Neural Networks on Relational Data)
Builds a graph connecting users, notifications, and interactions. Ranks notification types per user by learning from tap rates, dismiss patterns, time-to-tap, and cross-user engagement trends. Detects emerging preferences by propagating signals from graph neighbors.
Best for
Consumer apps with 10M+ users where notification opt-out has a high lifetime value cost and the notification mix significantly affects engagement.
Watch out for
Requires notification-level interaction data (tap, dismiss, ignore) with timestamps. If your platform only tracks whether a notification was sent without recording the user response, instrument the feedback loop first.
Key metric: Graph-based notification reranking reduces opt-outs by 30-50% and doubles tap rates, preserving $15-25M in lifetime engagement value for consumer apps with 20M+ users.
Why relational data changes the answer
User U001 (iOS, power user) taps on social notifications within 1.8 minutes but dismisses 78% of promotional notifications. A flat model sees these as two features and ranks social above promotional. But the relational graph adds a temporal dimension: U001's graph neighbors (users with similar behavior patterns) have recently started engaging with content_update notifications at a 40% higher rate than last month. This emerging preference has not yet shown up in U001's own data, but the graph propagates it as a weak but directional signal.
The graph also captures the notification budget constraint: U001 has 2 of 3 daily notification slots remaining. The ranking model must maximize engagement across these 2 remaining slots, not just rank types in isolation. This requires understanding the interaction between notification types: sending social and then content_update yields higher total engagement than sending two social notifications, because diminishing returns kick in within the same type. These cross-notification and cross-user dynamics are invisible to models that rank notification types independently. For consumer apps with 20M+ users, graph-based notification reranking reduces opt-outs by 30-50% and doubles tap rates, preserving $15-25M in lifetime engagement value.
Sending the same notification mix to everyone is like a radio station playing the same playlist for all listeners. Graph-based reranking is like a streaming service that learns you skip ads but tap on new episode alerts, notices that listeners with your taste profile are discovering a new podcast genre, and adjusts your notification mix to include one genre suggestion per day alongside your core preferences.
How KumoRFM solves this
Relational intelligence for true personalization
Kumo ranks notification types for each user by learning from the full interaction-notification-user graph. It discovers that User U001 taps on social notifications within 2 minutes but ignores promotional ones, while users in U001's graph neighborhood have started engaging with content update notifications. The model balances engagement probability, fatigue signals, and cross-user patterns to produce a ranked notification queue per user.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
USERS
| user_id | platform | segment |
|---|---|---|
| U001 | iOS | power_user |
| U002 | Android | casual |
| U003 | iOS | new_user |
NOTIFICATIONS
| notif_id | user_id | notification_type | sent_at |
|---|---|---|---|
| N001 | U001 | social | 2025-02-18 09:00 |
| N002 | U001 | promotional | 2025-02-18 14:00 |
| N003 | U002 | content_update | 2025-02-19 10:00 |
INTERACTIONS
| interaction_id | user_id | notification_type | action | timestamp |
|---|---|---|---|---|
| INT001 | U001 | social | tap | 2025-02-18 09:02 |
| INT002 | U001 | promotional | dismiss | 2025-02-18 14:15 |
| INT003 | U002 | content_update | tap | 2025-02-19 10:08 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT LIST_DISTINCT(INTERACTIONS.NOTIFICATION_TYPE, 0, 7, days) RANK TOP 3 FOR EACH USERS.USER_ID
Prediction output
Every entity gets a score, updated continuously
| USER_ID | CLASS | SCORE | TIMESTAMP |
|---|---|---|---|
| U001 | social | 0.94 | 2025-03-12 |
| U001 | content_update | 0.71 | 2025-03-12 |
| U002 | content_update | 0.83 | 2025-03-12 |
Understand why
Every prediction includes feature attributions — no black boxes
User U001 (iOS, power_user segment)
Predicted: Social notification ranked #1 — score 0.94
Top contributing features
Social notification tap rate (7 days)
0.91 (taps 91% of social notifs)
36% attribution
Time-to-tap (social)
Avg 1.8 minutes (immediate)
24% attribution
Promotional dismiss rate
0.78 (high fatigue signal)
18% attribution
Graph neighbors' emerging preference
Content_update engagement rising 40%
14% attribution
Daily notification budget remaining
2 of 3 slots available
8% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about notification reranking
How much do notification opt-outs cost?
Each notification opt-out permanently kills a high-value engagement channel. For consumer apps, the average user who opts out of notifications has 40-60% lower retention over 12 months. For an app with 20M users, each 1% reduction in opt-outs preserves $1.5-2.5M in lifetime engagement value.
How do you detect notification fatigue before opt-out?
Graph models track dismiss rates, time-to-dismiss (rapid dismissal indicates annoyance), and the trend in tap rates per notification type. A user whose promotional tap rate dropped from 30% to 5% over 2 weeks is showing fatigue even though they have not opted out yet. The model reduces promotional frequency before the opt-out happens.
Should you send fewer notifications or better notifications?
Better notifications, up to a point. Reducing total notification volume by 30% while increasing relevance per notification typically produces higher total engagement than the original high-volume approach. The optimal frequency varies by user: power users tolerate 5-8 per day, casual users prefer 1-2.
Can notification reranking work across channels (push, email, in-app)?
Yes. The graph model can rank notification types across all channels simultaneously, learning that a specific user prefers push for social alerts but email for promotional content. Cross-channel reranking prevents the same message from hitting multiple channels and optimizes the channel-type combination per user.
Bottom line: 30-50% reduction in notification opt-outs and 2x improvement in notification tap rates. For consumer apps with 20M+ users, this preserves $15-25M in lifetime engagement value.
Related use cases
Explore more personalization use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
