Identification of Conditional Causal Effects under Markov - - PowerPoint PPT Presentation
Identification of Conditional Causal Effects under Markov Equivalence Amin Jaber, Jiji Zhang, Elias Bareinboim jaber0@purdue.edu , jijizhang@ln.edu.hk , eb@cs.columbia.edu 33rd Conference on Neural Information Processing Systems (NeurIPS) 1
1
jaber0@purdue.edu , jijizhang@ln.edu.hk , eb@cs.columbia.edu
33rd Conference on Neural Information Processing Systems (NeurIPS) Vancouver, Canada — December, 2019
2
1 2
2
1 2
Solution: Causal Diagram
Age Exercise Cholesterol
2
1 2
More exercise → Higher cholesterol
Correlation ✕ Solution: Causal Diagram
Age Exercise Cholesterol
2
More exercise → Lower cholesterol (for each age group)
Causation
1 2
More exercise → Higher cholesterol
Correlation ✕ Solution: Causal Diagram
Age Exercise Cholesterol
3
1
Query
Px(y|z)
Q=
3
1
Query
2
Px(y|z)
Q=
Diagram Z Y X
3
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q=
Diagram Z Y X
3
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q= Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable?
Diagram Z Y X
3
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q= Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable? computable
Diagram Z Y X
3
Causal Inference Engine
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no? Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable?
Diagram Z Y X
3
Causal Inference Engine
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Px(y|z) = P(y|z, x)
Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable?
Diagram Z Y X
3
Causal Inference Engine
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Px(y|z) = P(y|z, x) Association (Data) Causation (Query)
Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable?
Diagram Z Y X
3
Causal Inference Engine
1
Query
2 3
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Px(y|z) = P(y|z, x) Association (Data) Causation (Query)
Based on the current knowledge about the problem (2) and the available data (3), is the research question (1) identifiable?
Diagram Z Y X
3
4
4
1 2 3
Z Y X Z Y X Z Y X
4
1 2 3
Non-causal Probability Distribution
Z Y X Z Y X Z Y X
4
1 2 3
Non-causal Probability Distribution Diagrams impose the same constraints over the observational distribution.
Z Y X Z Y X Z Y X
4
1 2 3
Non-causal Probability Distribution Diagrams impose the same constraints over the observational distribution. Markov Equivalence Class
Z Y X Z Y X Z Y X
4
1 2 3
Non-causal Probability Distribution Diagrams impose the same constraints over the observational distribution. Markov Equivalence Class
Z Y X Z Y X Z Y X
Partial Ancestral Graph (PAG)
Summary Graph
4
Relaxed CI Engine
1
Query
3 Diagram 2
Data
P(x, z, y) Px(y|z)
Q=
Causal Learning
5
?
Relaxed CI Engine
1
Query
3 Diagram 2
Data
P(x, z, y) Px(y|z)
Q=
Causal Learning
5
Relaxed CI Engine
1
Query
3 Diagram 2
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Based on the qualitative causal diagram (3) learned from data (2), is the causal effect (1) computable?
Causal Learning
5
Relaxed CI Engine
1
Query
3 Diagram 2
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Based on the qualitative causal diagram (3) learned from data (2), is the causal effect (1) computable?
Causal Learning
5
Relaxed CI Engine
1
Query
3 Diagram 2
Data
P(x, z, y) Px(y|z)
Q=
Solution
yes / no?
Based on the qualitative causal diagram (3) learned from data (2), is the causal effect (1) computable?
Causal Learning
Effect is identifiable in every diagram in the equivalence class, and with same expression!
Px(y|z) = P(y|z, x) 👎
5
6
6
6
6
6