Problem
A product with many notification sources (direct messages, mentions, followed content, platform recommendations) and several channels (email, push, in-app, inbox) needs to decide what reaches each user, where and when. A notification helps only when it arrives at the right time, in the right place, about the right thing, and user reactions come back late. The article gives a reference architecture for this in a social product.
Example from the article: a direct message may be urgent, important but not urgent enough to interrupt, or promotional spam. For a million mentions of one person, the channel depends on how much each notification interrupts.
Solution in brief
Put one ML decision service, the Smart Notification Service (SNS), between all notification sources and all channels. It drops or forwards each notification and picks the channel, and it learns from logged user reactions. Move re-consideration of old items into "smart inputs" and channel-specific optimization into "smart channels", so separate teams can build each part.
Questions this case answers
- How to architect a decision layer that chooses, per event and per recipient, whether to deliver a message and on which surface
- How to keep a delayed-feedback learning loop when the deliver-or-drop decision also determines which outcomes are ever observed
- How to decide for each notification whether to send it and through which channel?
- How to design a notification system so that ML models can learn from delayed user reactions?
Tools and methods
- Smart Notification Service (SNS) — ML decision point: drop a notification or send it, and to which channel.
- Interaction logs — delayed user reactions, training data for SNS.
- Smart inputs — periodic generators (email digest, notable-post push) that choose what, whom and when to send.
- Smart channels — channels such as the notification inbox that receive data and optimize their own experience.
- Sent-already log — past sends that smart inputs read.
Google or a one-time email code. No password.