Building Machine Learning Models with Strict Privacy Boundaries
summary
Slack builds a single global search ranking model that preserves team boundaries by replacing high‑dimensional sensitive attributes with aggregated functions, avoiding per‑team models for over half a million teams. The pipeline uses attribute parameterization to turn user ids, channel ids and terms into non‑sensitive features like term counts and CTR features (clicks over impressions), precomputes behavioral aggregates in a data warehouse and serves them from a feature store so the model learns dense, non‑leaky signals at inference.
AI-written summary · about
chapters (20)
- 0:00Intro
- 0:15Agenda
- 0:44Data Slack
- 1:08Public vs Private Data
- 1:32Slack Data
- 1:50New Users
- 2:00Machine Learning
- 2:26Data Privacy
- 2:52Data Privacy Example
- 3:17Machine Learning Privacy
- 3:40Search
- 4:07Learn to Rank
- 5:31Privacy Preserving
- 5:42Learning Individual Models
- 6:59Extracting Attributes
- 8:35Traditional Approach Example
- 9:35CTR Function
- 11:33Terms
- 12:10Inference Time
- 13:28Outro
related talks
24mJosh Wills, Head of Data Engineering, Slack
29mSeveral People Are Tuning: Data and Machine Learning at
44mPrivacy by Design Machine Learning
12mStriking the Balance: Safeguarding Customer Privacy While Empowering Employees
28mPrivacy preserving machine learning
31mScaling privacy embedded distributed ML systems
56mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
29mMachine Learning in Production: What to do after you’ve trained the models?