causal bayesian networks

14
Causal Bayesian Networks Fl´ avio Code¸ co Coelho Basic graph theory Bayesian Networks Inference Causal Bayesian Networks Fl´ avio Code¸ co Coelho Oswaldo Cruz Foundation November 1, 2006

Upload: flavio-codeco-coelho

Post on 13-May-2015

3.464 views

Category:

Education


2 download

DESCRIPTION

Introduction to Causal Bayesian Networks, based on Judea Pearl's book.

TRANSCRIPT

Page 1: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Causal Bayesian Networks

Flavio Codeco Coelho

Oswaldo Cruz Foundation

November 1, 2006

Page 2: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Graphs

Sets of elements called vertices, V , that may or may not beconnected to other vertices in the same set by a set of edges,EA graph may be defined uniquely by its set of edges, wich implythe set of vertices, e.g.E = {W ,X ,Y ,Z}:

G : E = {(W ,Z ), (Z ,Y ), (Y ,X ), (X ,Z )}

1

by running the above code, you’ll get the following output:

[′Y ′,′ X ′,′ Z ′,′ W ′][(′Y ′,′ X ′), (′Y ′,′ Z ′), (′X ′,′ Z ′), (′Z ′,′ W ′)]

1https://networkx.lanl.gov/

Page 3: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Some properties of graphs

Graphs can be directed or undirected;

The order of a graph corresponds to its number of vertices;

The size of a graph corresponds to its number of edges;

Vertices connected by an edge are neighbors or adjacent;

The order of a vertex corresponds to its number ofneighbors;

A path is a list of edges connecting two vertices;

A cycle is a path starting and ending in the same vertex;

A graph with no cycles is termed acyclic.

Page 4: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Visualizing the graph

From the code above we get the following picture:

Page 5: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Directed Acyclic Graph (DAG)

In directed acyclic graphs we use arrows to represent edges.

The output:

Page 6: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

DAG properties

Parents,children,descendants,ancestors, etc.

Root node

sink node

Every DAG has at least one root and one sink

Tree graph: every node has at most one parent

Chain graph: every node has at most on child

Complete graph: All possible edges exist.

Page 7: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Bayesian Networks

Advantages:

1 Convenient means of expressing assumptions

2 economical representation of Joint probabilit functions

3 Facilitate efficient inferences from observations

Why Bayesian?

1 Subjective nature of input information

2 Reliance on Bayes conditioning for updating information

3 The distinction between causal and evidential reasoning

Page 8: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Definitions

Markovian parents (PAj) P(xj | paj) = P(xj | x1, . . . , xj−1)such that no subset of PAj satisfies the aboveequation.

Markov compatibility If a probability function admits thefactorization P(xi | x1, . . . , xn) = P(xi | pai )relative to a DAG G we say that G and P arecompatible or that P is Markov relative to G .

d-separation Z d-separates X and Y iff Z blocks every pathfrom a node in X to a node in Y .

Page 9: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Theorems

Probabilistic Implications of d-separationIf sets X and Y are d-separated by Z in a DAG G , then X isindependent of Y conditional on Z every distributioncompatible with G . Conversely, if X and Y are not d-separatedby Z in a DAG G , then X and Y are dependent conditional onZ in at least one dist. compatible with G .

Ordered Markov ConditionA necessary and sufficient condition for a probabilitydistribution P to be markovian relative a DAG G is that everyvariable be independent of all its predecessors in someoredering of the variables that agrees with the arrows of G .

Page 10: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Theorems, cont.

Parental Markov ConditionA necessary and sufficient condition for a probabilitydistribution P to be markovian relative a DAG G is that everyvariable be independent of all its nondescendants (in G ),conditional on its parents.

Observational EquivalenceTwo DAGs are observationally equivalent if and only if theyhave the same skeletons and the same sets of v-structures, thatis, two converging arrows whose tails are not connected by anarrow.

Page 11: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Inference with Bayesian Networks

In the presence of a set of observations X the posteriorprobability:

P(y | x) =

∑s P(y , x , s)∑

y ,s P(y , x , s)

can be calculated from a DAG G and the conditionalprobabilities P(xi | pai ) defined on the families of G

Page 12: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Causal Bayesian Networks

DAGs constructed around Causal, instead of associationalinformation is mor intuitive and more reliable.

Causal relationships are a direct representations of ourbeliefs

Direct representation of mechanisms

Simple to represent interventions thanks to modularity inthe network

Definition: Causal bayesian networkLet P(v) be a probability distribution on a set of V variables,and let Px(v) denote the distributionresulting from theintervention do(X = x) that sets a subset X of variables toconstants x.

Page 13: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

Inference

Causal Bayesian Networks

Definition (cont.): Causal bayesian networkDenote by P∗ the set of all interventional distributionsPx(v), X ⊆ V , including P(v), which represents nointervention (i.e., X = ∅). A DAG G is said to be a causalbayesian network compatible with P∗ if and only if thefollowing three conditions hold for every Px ∈ P∗:

1 Px(v) is Markov relative to G;

2 Px(vi ) = 1 for all Vi ∈ X whenever vi is consistent withX = x ;

3 Px(vi | pai ) = P(vi | pai ) for all Vi ∈ X whenever pai isconsistent with X = x .

Page 14: Causal Bayesian Networks

CausalBayesianNetworks

Flavio CodecoCoelho

Basic graphtheory

BayesianNetworks

InferenceThank you!