Empowered User Control: Learnings from Building "Event Level Deletion"
summary
Meta implements event-level deletion for ad interactions by storing each click as a single heterogeneous blob keyed with a unique ID and a lightweight deletion-status map injected via API, then parsing data at consume time so deleted attributes never reach ranking storage. Feature aggregation moves into a sequence model (Transformer-like) that builds attribute embeddings (grid search over K attributes and n lookback) and drops an event embedding when its attributes are deleted. This reduces duplicated copies and infra cost but increases model size and can cause QPS regressions.
AI-written summary · about
chapters (6)
- 0:00Introduction
- 0:33Problem Space
- 7:08Current Challenges
- 9:02Key Terms
- 11:33Transformer Model
- 12:38Pros Cons
related talks
14mReal-Time Data Processing for ML Feature Engineering
13mAlgorithmic Transparency in Personalized Advertising: Learnings from Building "Why Am I
17mCreating & operating ML models from event-based data using feature stor...
31mGetting AI to Forget About You: Data Deletion From Machine Learning Models
32mOptimizing Machine Learning Training Infrastructure: A Governance Approach
46mIncreasing Participant Agency by Understanding the Impact of Training Data
40mOvercoming data infrastructure limitations with a new paradigm for ML
18mBehavioral Data Creation for AI - Delivering Value and Insights