HoloClean and Kamino - Structured Learning for Private Data Generation

Explore the intersection of data privacy and synthetic data generation in this 36-minute conference talk from MLOps World: Machine Learning in Production. Dive into the challenges of publishing synthetic data that preserves individual privacy while maintaining the utility of the original sensitive data. Learn about the limitations of existing differentially private data synthesis methods in preserving crucial data properties, such as correlations and dependencies among tuples and attributes. Discover how probabilistic database models can be leveraged to privately learn and sample new synthetic private instances. Gain insights into the HoloClean framework for structured data prediction and its application in learning underlying data distributions. Examine the technical challenges of learning these models privately and understand how Kamino, a system built on HoloClean, addresses these issues to synthesize useful private data instances. Presented by Ihab Ilyas, Professor at the University of Waterloo, this talk offers valuable knowledge for data scientists, privacy experts, and machine learning practitioners working with sensitive data.