← AI HoodmanNo. 009 · Recommendations

Instagram Notifications

Which posts deserve a notification?

01of 07
Use the arrows, swipe or the arrow keys to turn the slides. Click a slide to read it page by page.
  1. Slide 1 of 7: Ali follows 900 accounts. Together they posted 2,000 times today. now @sara posted a photo 1m ago @nasa posted a photo 2m ago @chef_omar posted a reel +1,997 more Which ones should Ali be notified about?
  2. Slide 2 of 7: The obvious answer: The posts they would tap. CTR model ranking Chance they tap 1 @sara · photo 40% 2 @nasa · photo 22% 3 @chef_omar · reel 15% … 1,997 more But Sara's photo will be in their feed anyway. High CTR, no impact.
  3. Slide 3 of 7: The key idea: Notify what they would miss. Sara's photo Sara's story stays in their feed gone tonight If notified 40% If notified 25% If not 38% If not 5% Difference +2% Difference +20% That difference is uplift. Send the story.
  4. Slide 4 of 7: The trick: Skip some notifications on purpose. Training data Notified? at random Saw post? the label User Post Ali @sara's story yes 1 via notification Omar @nasa photo no 1 via feed Lina @chef_omar reel yes 0 … millions more Saw post: via notification or feed, within a day You can't both notify and skip the same person about the same post. Random skips give you both sides.
  5. Slide 5 of 7: The uplift model: One model, run twice. Training Inputs The user The post Model Output Chance they see the post Was notified? yes/no Label: saw the post (1) or not (0) Serving: Sara posts a story. Ask the model twice about Ali. Ali, Sara's story, was notified = yes 0.25 Ali, Sara's story, was notified = no 0.05 Uplift 0.25 − 0.05 = 0.20 The model predicts if they will see the post. Asking with and without a notification gives the uplift.
  6. Slide 6 of 7: The catch: Too many notifications, and users turn them off. Score = Gain − w × Risk Gain the uplift model's prediction Risk the chance the user turns notifications off w how much one opt-out hurts, raise it until opt-outs stay low. w = 30 Gain Risk Score Story 20% 0.1% +17% Send 6th today 5% 1% −25% Skip Bottom line Once a user turns notifications off, you can't reach them again.
  7. Slide 7 of 7: Takeaway: Don't notify what they would see anyway. Notify what they would miss.

Keep reading

Instagram Notifications · 1 / 7
  1. Slide 1 of 7: Ali follows 900 accounts. Together they posted 2,000 times today. now @sara posted a photo 1m ago @nasa posted a photo 2m ago @chef_omar posted a reel +1,997 more Which ones should Ali be notified about?
  2. Slide 2 of 7: The obvious answer: The posts they would tap. CTR model ranking Chance they tap 1 @sara · photo 40% 2 @nasa · photo 22% 3 @chef_omar · reel 15% … 1,997 more But Sara's photo will be in their feed anyway. High CTR, no impact.
  3. Slide 3 of 7: The key idea: Notify what they would miss. Sara's photo Sara's story stays in their feed gone tonight If notified 40% If notified 25% If not 38% If not 5% Difference +2% Difference +20% That difference is uplift. Send the story.
  4. Slide 4 of 7: The trick: Skip some notifications on purpose. Training data Notified? at random Saw post? the label User Post Ali @sara's story yes 1 via notification Omar @nasa photo no 1 via feed Lina @chef_omar reel yes 0 … millions more Saw post: via notification or feed, within a day You can't both notify and skip the same person about the same post. Random skips give you both sides.
  5. Slide 5 of 7: The uplift model: One model, run twice. Training Inputs The user The post Model Output Chance they see the post Was notified? yes/no Label: saw the post (1) or not (0) Serving: Sara posts a story. Ask the model twice about Ali. Ali, Sara's story, was notified = yes 0.25 Ali, Sara's story, was notified = no 0.05 Uplift 0.25 − 0.05 = 0.20 The model predicts if they will see the post. Asking with and without a notification gives the uplift.
  6. Slide 6 of 7: The catch: Too many notifications, and users turn them off. Score = Gain − w × Risk Gain the uplift model's prediction Risk the chance the user turns notifications off w how much one opt-out hurts, raise it until opt-outs stay low. w = 30 Gain Risk Score Story 20% 0.1% +17% Send 6th today 5% 1% −25% Skip Bottom line Once a user turns notifications off, you can't reach them again.
  7. Slide 7 of 7: Takeaway: Don't notify what they would see anyway. Notify what they would miss.