Explainable causal inference results
summary
Claims a graph-only algorithm — the extended Markov Blanket — finds which observable variables actually matter for a node’s posterior, implemented using a C++ BN engine with Python bindings and compared to PyMC/pgmpy. On an industrial model of ~6000 nodes with 250 sensors it shrinks the relevant sensor set to ~25 on average (≈90% reduction). The method is cheap (graph-walk with caching), fails when observability is sparse or cross‑linked, and needs dynamic BNs to capture time effects.
AI-written summary · about
chapters (13)
- 0:00Explainable causal inference
- 1:13AZ high-tech systems mission
- 4:43Causal inference with Bayesian nets
- 6:30Bayes nets intro
- 10:06Reasoning with partial networks
- 11:32Markov blankets explained
- 13:22Demo on small Bayesian network
- 15:28Scaling to larger networks
- 17:16Reduced observability impact
- 18:43High-tech diagnostics method
- 20:51Implementation tools
- 22:14Industrial use case results
- 23:47Q&A
related talks
32mNew Approaches To Learning Nonparametric (Latent) Causal Graphical Models
33mSimple Yet Efficient Estimators For Network Causal Inference...
31mIntroduction to Causal Inference using pgmpy
29m🧠🧹 Causality - Mental Hygiene for Data Science
41mBayesian Network Modeling using R and Python
27mUsing Causal thinking to make Media Mix Modeling
01h 27mFlexible Statistical Modeling
01h 06mIntroduction to Causal Graphical Models: Graphs, d-separation, do-calculus