Causal Data Integration
summary
Proposes Causal Data Integration: an automated pipeline that mines unobserved attributes from knowledge graphs, joins hundreds of candidates to an input table, and then selects a small subset by minimizing conditional mutual information to explain surprising aggregates. Demonstrates on flight data: a three‑variable explanation (population size, low temperature, carrier) explains most city‑delay differences. Uses Pearl identification to estimate average treatment effects, builds a summarized causal DAG (k‑node summary preserving conditional independences), and finds LLMs help label nodes but fail to reliably infer causal edges; robustness to missingness and selection bias remains unresolved.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:00Correlation vs. Causality
- 4:12Explaining Query Results
- 7:58Causal Explanation Goals
- 11:58Causal Data Integration
- 15:15Core Challenges
- 20:44Conciseness and Summary
- 23:04Conclusion and Future Work
- 25:05Q&A Session
related talks
47mApproximate Lifted Inference with Probabilistic Databases
36mVibing with Data -- Multi-Agent Data Modeling and Constructi...
23mKnowledge Graphs and LLMs in Action
38mKnowledge Graphs : Adding a New Dimension to Vector Search and Generative AI
37mEngineering data quality
22mThe Nexus of Innovation: How Knowledge Graphs Fuel AI Scientific Breakthroughs
21mOvercoming Data Barriers for Successful AI Projects
46mBuilding Knowledge Graphs