+ All Categories
Home > Documents > Using Markov Blankets for Causal Structure Learning

Using Markov Blankets for Causal Structure Learning

Date post: 22-Feb-2016
Category:
Upload: kura
View: 75 times
Download: 0 times
Share this document with a friend
Description:
Using Markov Blankets for Causal Structure Learning. Jean-Philippe Pellet Andre Ellisseeff Presented by Na Dai. Motivation. Why structure l earning? What are Markov blankets? Relationship between feature selection and Markov blankets?. Previous work. Score-based approaches - PowerPoint PPT Presentation
Popular Tags:
19
Using Markov Blankets for Causal Structure Learning Jean-Philippe Pellet Andre Ellisseeff Presented by Na Dai
Transcript
Page 1: Using Markov Blankets for Causal Structure Learning

Using Markov Blankets for Causal Structure Learning

Jean-Philippe PelletAndre Ellisseeff

Presented by Na Dai

Page 2: Using Markov Blankets for Causal Structure Learning

Motivation

• Why structure learning?• What are Markov blankets?• Relationship between feature selection and

Markov blankets?

Page 3: Using Markov Blankets for Causal Structure Learning

Previous work

• Score-based approaches• Constraint-based approaches• Hybrid approaches

Page 4: Using Markov Blankets for Causal Structure Learning

Central Ideas

• Building up local structures from Markov blankets.

• Generating global graph structure from local structure.

• How to generate Markov blankets?

Page 5: Using Markov Blankets for Causal Structure Learning

Background

• Feature selection– Conditional independence

– Strong relevance

– Weak relevance

– Irrelevance

– Feature selection task

Page 6: Using Markov Blankets for Causal Structure Learning

Background

• Causal structure learning– Goal: learn the full structure of the network– D-separation:

1) A --> C --> B 2) A <-- C <-- B 3) A <-- C --> B 4) A --> C <-- B

Page 7: Using Markov Blankets for Causal Structure Learning

Background

• Perfect map

• Causal Markov condition

• Faithfulness condition

Page 8: Using Markov Blankets for Causal Structure Learning

Background

• Causal sufficiency assumption

• V-structure

Page 9: Using Markov Blankets for Causal Structure Learning

Causal Network Construction

• Properties of Markov blankets

Page 10: Using Markov Blankets for Causal Structure Learning

Recovering Local Structure

• Remove possible spouse links– Find d-separation set

• Orient the arcs

Page 11: Using Markov Blankets for Causal Structure Learning

Algorithm 1

Page 12: Using Markov Blankets for Causal Structure Learning

Example of Local Causal Structure

Page 13: Using Markov Blankets for Causal Structure Learning

Potential Improvements

• Two passes becomes one pass– Combine spouse link detection and edge orientation.

• If can find S to make X and Y conditionally independent, then X and Y are spouse.

• If Z \in Mb(X) and Mb(Y) is not in S is a mutual child, the direction between X, Y, Z is determined.

• Transform the problem to identify d-separation set.

Page 14: Using Markov Blankets for Causal Structure Learning

Algorithm 2

Page 15: Using Markov Blankets for Causal Structure Learning

Generic Algorithm based on Feature Selection

• Find the conjectured Markov blanket of each variable with feature selection.

• Build the moral graph.• Remove spouse links and orient V-structure.• Propagate orientation constraints.

Page 16: Using Markov Blankets for Causal Structure Learning

Algorithm 3

Page 17: Using Markov Blankets for Causal Structure Learning

Algorithm 4

Page 18: Using Markov Blankets for Causal Structure Learning

Algorithms for Causal Feature Selection

• RFE based approach• TC and TCbw algorithm

Page 19: Using Markov Blankets for Causal Structure Learning

Conclusion

• Causal discovery is close to feature selection• Three steps to build up the causal structure

from Markov blankets. More efficient, and even better than previous methods.


Recommended