Driving a Higher Level of Customer Support with Machine Learning
summary
Airbnb replaces brittle rules with an ML pipeline that classifies “trip vs non‑trip” tickets using escalations as gold labels, tackles 10:1 class imbalance, missing online features and noisy logs, and moves from binary labels to a classification score plus a real‑time balancing tuner that throttles volume to meet SLA targets. Offline gains cut escalations and promise millions in savings, but full rollout required months of logging fixes, agent‑side A/B testing and capacity work before the model actually improves wait times.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:05What we do at Airbnb
- 2:05What are the applications we use
- 4:25What is customer support
- 5:30What is segmentation
- 7:05National language processing
- 7:32Selfhelp
- 8:23Challenges
- 8:57Segmentation Framework
- 12:03Rulebased classification
- 13:26Binary classification
- 14:02No trip
- 16:48Collecting features
- 18:41Offline results
- 19:58Experimentation
- 27:37Lessons Learned
related talks
22mBefore the Model: How Machine Learning Products Start...
40mML in Production: Why Analysts and Backend Struggle to Agree
32mHarnessing Large Language Models for Enhanced Customer Experience
40mMachine Learning
39mOptimizing User Support Services by Introducing Machine Learning Technologies
28mHow Machine Learning Can Achieve Data Quality Monitoring At Scale
24mChallenges in building Machine Learning models in production
19mUnveiling the Potential: Exploring Upselling, pLTV, and ML for Operations