probabilistic relational models for link prediction problem

24
Probabilistic Relational Models for Link Prediction Problem By: Sina Sajadmanesh Advisor: Dr. Hamid Reza Rabiee

Upload: sina-sajadmanesh

Post on 06-Apr-2017

67 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Probabilistic Relational Models for Link Prediction Problem

Probabilistic Relational Models for Link Prediction Problem

By: Sina SajadmaneshAdvisor: Dr. Hamid Reza Rabiee

Page 2: Probabilistic Relational Models for Link Prediction Problem

2 DMLLink Prediction DMLDMLLink Prediction2

Outline

Introduction Probabilistic Relational Models Directed vs. Undirected Networks

Relational Bayesian NetworksRelational Markov Networks

Page 3: Probabilistic Relational Models for Link Prediction Problem

3 DMLLink Prediction DMLDMLLink Prediction3

Introduction

Probabilistic Relational Models Combines probabilistic graphical models

with entity-relationship models Defines a joint probability distribution over

attributes of entities and relations A method for describing probabilistic

relationships among attributes of entities and related entities

Page 4: Probabilistic Relational Models for Link Prediction Problem

4 DMLLink Prediction DML4

Introduction

Probabilistic Relational Network Directed Graphical Model

• Relational Bayesian Network (RBN)• Focus on causal interactions

Undirected Graphical Model• Relational Markov Network (RMN)• Focus on symmetric interactions

Page 5: Probabilistic Relational Models for Link Prediction Problem

5 DMLLink Prediction DMLDMLLink Prediction5

Outline

IntroductionRelational Bayesian Networks

Relational Schema Probabilistic Model Relational Skeleton Attribute Uncertainty Semantics Link Prediction

Relational Markov Networks

Page 6: Probabilistic Relational Models for Link Prediction Problem

6 DMLLink Prediction DMLDMLLink Prediction6

Relational Bayesian Networks

Page 7: Probabilistic Relational Models for Link Prediction Problem

7 DMLLink Prediction DMLDMLLink Prediction7

Relational Bayesian Networks

Relational Schema

AuthorGood Writer

Author ofHas Review

Review

PaperQualityAccepted

Mood

LengthSmart

Page 8: Probabilistic Relational Models for Link Prediction Problem

8 DMLLink Prediction DMLDMLLink Prediction8

Relational Bayesian Networks

Probabilistic Model

Length

MoodAuthor

Good Writer

Paper

Quality

Accepted

ReviewSmart

Page 9: Probabilistic Relational Models for Link Prediction Problem

9 DMLLink Prediction DMLDMLLink Prediction9

Relational Bayesian Networks

Relational Skeleton

Fixed relational skeleton σ: set of objects in each class relations between them

Author A1

Paper P1 Author: A1 Review: R1

Review R2

Review R1

Author A2

Paper P2 Author: A1 Review: R2

Paper P3 Author: A2 Review: R2

Review R2

Page 10: Probabilistic Relational Models for Link Prediction Problem

10

DMLLink Prediction DMLDMLLink Prediction10

Relational Bayesian Networks

Attribute Uncertainty

RBN defines distribution over instantiations of attributes

Page 11: Probabilistic Relational Models for Link Prediction Problem

11

DMLLink Prediction DMLDMLLink Prediction11

Relational Bayesian Networks

Aggregate DependenciesReview R1

Length

Mood

Review R2

Length

Mood

Review R3

Length

Mood

Paper P1

Accepted

Quality

Page 12: Probabilistic Relational Models for Link Prediction Problem

12

DMLLink Prediction DMLDMLLink Prediction12

Relational Bayesian Networks

Aggregate Dependencies

sum, min, max, avg, mode, count

Review R1

Length

Mood

Review R2

Length

Mood

Review R3

Length

Mood

Paper P1

Accepted

Quality

mode

3.07.04.06.08.02.09.01.0

,,,,

,

ttfttfff

P(A | Q, M) MQ

Page 13: Probabilistic Relational Models for Link Prediction Problem

13

DMLLink Prediction DMLDMLLink Prediction13

Relational Bayesian Networks

Semantics

Probability distribution over instantiation I

Author

Paper

Review

Author A1

Paper P2

Paper P1

ReviewR3

ReviewR2

ReviewR1

Author A2

Paper P3

𝑃 ( 𝐼|𝜎 ,𝑆 ,𝜃 )=∏𝑥∈𝜎

∏𝑥 . 𝐴

𝑃 (𝑥 .𝐴|𝑝𝑎𝑟𝑒𝑛𝑡𝑠(𝑥 .𝐴)¿ ¿

Page 14: Probabilistic Relational Models for Link Prediction Problem

14

DMLLink Prediction DMLDMLLink Prediction14

Relational Bayesian Networks

Link Prediction

? ??

Page 15: Probabilistic Relational Models for Link Prediction Problem

15

DMLLink Prediction DMLDMLLink Prediction15

Relational Bayesian Networks

Link Prediction

Cites

PaperTopicWords

PaperTopicWords

Exists

Citer.Topic Cited.Topic

0.995 0.005 Theory Theory

False True

AI Theory 0.999 0.001

AI AI 0.993 0.008 AI Theory 0.997 0.003

Page 16: Probabilistic Relational Models for Link Prediction Problem

16

DMLLink Prediction DMLDMLLink Prediction16

Relational Bayesian Networks

Link Prediction

Paper#2 Topic Paper#3Topic

WordN

Paper#1

Word1

Topic

... ... ...

Author #1Area Inst

#1-#2

Author #2Area Inst

Exists

#2-#3Exists

#2-#1Exists

#3-#1Exists

#1-#3Exists

WordN

Word1WordN

Word1

Exists

#3-#2

Page 17: Probabilistic Relational Models for Link Prediction Problem

17

DMLLink Prediction DMLDMLLink Prediction17

Outline

IntroductionRelational Bayesian NetworksRelational Markov Networks

Advantages of Undirected Models Markov instead of Bayesian network Relational Clique Templates Formal Definition Link Prediction

Page 18: Probabilistic Relational Models for Link Prediction Problem

18

DMLLink Prediction DMLDMLLink Prediction18

Relational Markov Networks

Advantages of Undirected Models Cycles are not a problem Easy to learn discriminatively Symmetric, non-causal interactions Handles patterns involving multiple entities

• Triangle patterns• Transitive patterns

Devised for collective classification

Page 19: Probabilistic Relational Models for Link Prediction Problem

19

DMLLink Prediction DMLDMLLink Prediction19

Relational Markov Networks

Markov instead of Bayesian network

Author2 Paper2TopicArea

Venue

Paper1Topic

Author1

SubArea

Area

1.8AI

THTH

0.3

1.50.2

AIAITHAI

TH

T2T1 (T1,T2)

Template potential

Page 20: Probabilistic Relational Models for Link Prediction Problem

20

DMLLink Prediction DMLDMLLink Prediction20

Relational Markov Networks

Relational Clique Templates Specify tuples of variables in the network

instantiation using relational query language Components:

• F a set of entity variables (FROM)• W condition about the attributes (WHERE)• S subset of attributes (SELECT)

Query results to cliques

SELECT p1.topic, p2.topicFROM Paper p1, Paper p2, Cite cWHERE c.citer=p1.key AND c.cited=p2.key

Page 21: Probabilistic Relational Models for Link Prediction Problem

21

DMLLink Prediction DMLDMLLink Prediction21

Relational Markov Networks

Formal Definition Set of clique templates C Set of potential functions Ф

Defines a conditional distribution over labels of an instantiation

𝑃 ( 𝐼|Ф ,𝐶 )= 1𝑍 ∏

𝐶∏

𝑐∈𝐶(𝐼 )Φ𝑐 ( 𝐼 .𝑣𝑐 )

Page 22: Probabilistic Relational Models for Link Prediction Problem

22

DMLLink Prediction DMLDMLLink Prediction22

Relational Markov Networks

Link Prediction Factors affecting the relations of different

entities

Entity’s attributes:• Properties, Labels

Entity’s structural properties:• Similarity, Transitivity

More complex patterns can be captured using cliques that represents dependencies and correlations

Page 23: Probabilistic Relational Models for Link Prediction Problem

23

DMLLink Prediction DMLDMLLink Prediction23

References[1] Getoor, Lise. Introduction to statistical relational learning. MIT press, 2007.

[2] Pfeffer, Avrom J., and Daphne Koller. Probabilistic reasoning for complex systems. Stanford: Stanford University, 2000.

[3] Koller, Daphne, and Avi Pfeffer. "Probabilistic frame-based systems."AAAI/IAAI. 1998.

[4] Taskar, Ben, Pieter Abbeel, and Daphne Koller. "Discriminative probabilistic models for relational data." Proceedings of the Eighteenth conference on Uncertainty in artificial intelligence. Morgan Kaufmann Publishers Inc., 2002.

[5] Taskar, Ben, et al. "Link prediction in relational data." Advances in neural information processing systems. 2003.

Page 24: Probabilistic Relational Models for Link Prediction Problem

Q&A