Institut Supérieur de Gestion de Tunis

Integrating Ontological Knowledge for Iterative Causal Discovery & Visualization Presented by: Montassar Ben Messaoud

Pr. Philippe Leray (Ecole Polytechnique de l’Université - Nantes) Dr. Nahla Ben Amor (Institut Supérieur de Gestion -Tunis)

2008-2009

Outline ▪ Introduction ▪ Causal Theory ▪ Causal Bayesian Networks ▪ Ontology ▪ SEMantical CAusal DiscOvery (SEMCADO) Approach ▪ Experimental Study ▪ Conclusion & Perspectives

Introduction

Bayesian Networks (Pearl, 1988) P(C=F) P(C=T) 0.5 0.5

Causality ? C F T

P(S=F) P(S=T) 0.5 0.9

Cloudy

C F T

P(R=F) P(R=T) 0.8 0.2 0.2 0.8

Rain

Sprinkler

0.5 0.1

WetGrass

Numerical component

+

S F T T T

R F F

P(W=F) P(W=T) 1.0 0.0 0.1 0.9 0.1 0.9 T 0.01 0.99 Probabilistic

Inference Graphical component

F

Causal Theory (1) ▪ David Hume ▪ 1711-1776 ▪ British empricist ▪ Modern study of causation ▪ Establishment of an empirical experimental science

?

Natural right Limited ressources

Private property

« We may define a cause to be an object followed by another…, where, if the first object had not been, the second never had existed. » Hume, 1748

Causal Theory (2)

Causal discovery, what affects ? …Biology

…Health

…Climate changes …Chemistry …Physics

… Economy

Causal Bayesian Networks

+

BNs

Causality

Causal Bayesian Networks X1

Probabilistic Inference X2

X3

+ Causal Inference X4

Problematic

How can we learn such causal networks ?

Causal Bayesian Network (1) What is the Cause ? Markov Equivalence External intervention Swine Flu (H1N1)

Influenza Pondemic Alert Level

Learning Causal networks Structure Learning Algorithms: IC, IC*, SGS, PC, BN-PC K2, MWST, GES, GS

Observational data

Interventional data

Structure Learning Approaches: Learning from mixture of obs. & exp. Data (Cooper & Yoo 1999) Active Learning (Tong & Koller 2001) Theoretical Study (Eberhardt 2005) Learning CBNs from observations an experiments (Leray & al. 2006)

Objective

Causal graphs

+

Ontologies ?

Integrating Ontological Knowledge for Iterative Causal Discovery & Visualization SEMCADO (SEMantical CAusal DiscOvery) algorithm

Ontology

Generalization

Transport is-a

Vehicle is-a

Bus

Car is-a

Hummer

Plane is-a

▪ edges  is-a relations.

Airbus is-a

A320

▪ nodes  concepts.

is-a

is-a

is-a

Megane

□ An ontology is a DAG where:

is-a

is-a

□Dist Hierarchical categorization = 6=of2terms Megane) rmbb (Hummer, A380)

A380

□ Semantic measures are used to evaluate the strength of the semantic link between 2 concepts. □ Semantic distance evaluates the disaffection between two concepts. □ Rada & al. distance (Rada, 1989): is based on the shortest path between 2 concepts.

SEMCADO Approach

Learning phase

System

Observational data

S-rule CPDAG S-CPDAG

Ontology CPDAG

Structure Learning

SEMCADO Approach Performing experiment System

Interventional data

PC rules

Analysing results

CPDAG S-CPDAG

Selecting best experiment

Ontology

Enriched Visualization

Causal Discovery

CBN

SEMCADO Approach

How to select appropriate interventions ? Cost = Cost Intervention + Cost Observation

We use commensurable scale: X1

X3

Cost Table: X2

X4

X5

X

X1

X2

X3

X4

X5

Cost (perf. X) $

100

300

1200

∞

500

Cost (meas. X) $

10

25

15

20

5

Experimental Study (1) Random CBN Generator

Ontology

CBN

Generate Rada & al. Distance Matrix

DAG to CPDAG algorithm

SEMCADO algorithm CPDAG

Experimental Study (2) Original structure

Mutilated structure

Sampling algorithm

Perform experiment

SEMCADO algorithm

Experimental data

Experimental Study (3) SemCaDo (3 ont. Sets) MyCaDo

Average of directed edges

100 80 60 40 20 0

1

2

3

4

5

Experiment number

Conclusion(1)

■ Novel approach for integrating ontological knowledge for causal discovery & visualization. ■ Experimental study shows that we maximize the total number of inferred edges after each experiment. ■ We guide the manipulation process to the graph variables with high semantical inertia. ■ Our approach leads indirectly to a decrease in the number of interventions.

Conclusion(2)

The major results of this work are developed in: M. Ben Messaoud, P. Leray and N. Ben Amor Integrating Ontological Knowledge for Iterative Causal Discovery & Visualization Are accepted for: Publication in ECSQARU’2009 (Verona-Italy) & Oral presentation in the Machine Learning and Visualization Workshop (Hammamet-Tunisia)

Perspectives

□ Implement the visualization tools using the Tulip Software. □ Propose an experimentation strategy for umperfect observational data. □ Extend the SemCaDo algorithm to deal with latent variables. □ Consider the case of incomplete interventional data.

Intervention Selection Available ontology

CPDAG

Root X2

X3

SF1

X1 X4

X5

X6

Intervention selection criterion: Semantical #NeU(Xinertia i)

F1

X1

SF2 F2

X2

X3

X4

F3

X5

X6

mscs: the most specific common subsumer

Visualization Force Directed Placement

▪ Edges  Springs ▪ Nodes  Electrically charged particles ▪ Force application  Equilibrium State

Visualization ZOOM IN / ZOOM OUT

0 1 2 3

SEMCADO Approach PC Rules (SGS 1993) R1: Directing without creating new v-structure R2: Directing without creating cycles

X1

X1

X2 X3

X2 X4

X3

X5 X4

SEMCADO Approach Semantical-Rule S-CPDAG CPDAG X1

Available ontology X2

X3

X4

Root X5

F1 X6

F3

X7 X1

X8

F2 X2

X3

X4

X5

X6

X7

X8

Mutilation Causal inference