Achieving Understandability and Privacy With Provenance-Based Explanations
summary
Describes two concrete systems: a pipeline that turns value-level provenance into concise natural-language justifications by using NaLIR plus the provenance engine SelP2, producing factorizations guided by the dependency tree to get provenance factorization that is T-compatible and then summarized (counts like “63 authors, 170 papers” are computed); and a privacy module that abstracts tuple IDs using an NX semiring model and abstraction trees to achieve a k-anonymity–style privacy K, proves the optimal abstraction problem NP-hard, and implements an optimized exhaustive search whose runtime stays practical up to K≈20.
AI-written summary · about
chapters (6)
- 0:00Provenance explanations intro
- 1:48Natural language provenance
- 10:37Factorizing provenance
- 14:25Privacy in provenance
- 26:05Balancing clarity and privacy
- 27:29Future directions & Q&A
related talks
26mDebugging Large-Scale Datalog: A Scalable Provenance Evaluation Strategy
20mProvenance Without Surveillance: Privacy Engineering for AI Content Transparency
01h 06mAlgebraic Structures in Logic and Query Evaluation 1
01h 02mFormal Methods and Proofs of Privacy Properties, Part III
01h 05mAlgebraic Structures in Logic and Query Evaluation 2
14mToward Provably Private Insights into AI Use
38mTowards Privacy-Respectful Cyber-Physical Systems: Two Perspectives
56mFormal Methods and Proofs of Privacy Properties, Part II