Bayesian Networks: Directed Markov Properties (Cont d) and Markov Equivalent DAGs

Similar documents
10.2 Graph Terminology and Special Types of Graphs

Chapter 4 Fuzzy Graph and Relation

Section 2.3 Functions. Definition: Let A and B be sets. A function (mapping, map) f from A to B, denoted f :A B, is a subset of A B such that

V = set of vertices (vertex / node) E = set of edges (v, w) (v, w in V)

Lecture 8: Graph-theoretic problems (again)

COMMON FRACTIONS. or a / b = a b. , a is called the numerator, and b is called the denominator.

Chapter 9. Greedy Technique. Copyright 2007 Pearson Addison-Wesley. All rights reserved.

Outline. CS38 Introduction to Algorithms. Graphs. Graphs. Graphs. Graph traversals

MITSUBISHI ELECTRIC RESEARCH LABORATORIES Cambridge, Massachusetts. Introduction to Matroids and Applications. Srikumar Ramalingam

If you are at the university, either physically or via the VPN, you can download the chapters of this book as PDFs.

Lecture 13: Graphs I: Breadth First Search

Lecture 12 : Topological Spaces

Greedy Algorithm. Algorithm Fall Semester

CS 241 Week 4 Tutorial Solutions

FEEDBACK: The standard error of a regression is not an unbiased estimator for the standard deviation of the error in a multiple regression model.

Distance vector protocol

Adjacency. Adjacency Two vertices u and v are adjacent if there is an edge connecting them. This is sometimes written as u v.

12/9/14. CS151 Fall 20124Lecture (almost there) 12/6. Graphs. Seven Bridges of Königsberg. Leonard Euler

Introduction to Algebra

Graph theory Route problems

Tight triangulations: a link between combinatorics and topology

Lesson 4.4. Euler Circuits and Paths. Explore This

XML and Databases. Outline. XPath. Outline - Lectures. XPath Data Model. Outline - Assignments. XPath. Sebastian Maneth NICTA and UNSW

3D convex hulls. Convex Hull in 3D. convex polyhedron. convex polyhedron. The problem: Given a set P of points in 3D, compute their convex hull

Ma/CS 6b Class 1: Graph Recap

A Tautology Checker loosely related to Stålmarck s Algorithm by Martin Richards

Duality in linear interval equations

Ma/CS 6b Class 1: Graph Recap

Class Overview. Database Design. Database Design Process. Database Design. Introduction to Data Management CSE 414

Containers: Queue and List

1.5 Extrema and the Mean Value Theorem

Lecture 7: Integration Techniques

Graph Contraction and Connectivity

MTH 146 Conics Supplement

Premaster Course Algorithms 1 Chapter 6: Shortest Paths. Christian Scheideler SS 2018

COMP108 Algorithmic Foundations

MATH 25 CLASS 5 NOTES, SEP

Final Exam Review F 06 M 236 Be sure to look over all of your tests, as well as over the activities you did in the activity book

2 Computing all Intersections of a Set of Segments Line Segment Intersection

Cooperative Routing in Multi-Source Multi-Destination Multi-hop Wireless Networks

CS311H: Discrete Mathematics. Graph Theory IV. A Non-planar Graph. Regions of a Planar Graph. Euler s Formula. Instructor: Işıl Dillig

The Fundamental Theorem of Calculus

COMP 423 lecture 11 Jan. 28, 2008

1.1. Interval Notation and Set Notation Essential Question When is it convenient to use set-builder notation to represent a set of numbers?

CSc 453 Compilers and Systems Software. 6 : Top-Down Parsing I

10.5 Graphing Quadratic Functions

Can Pythagoras Swim?

CS553 Lecture Introduction to Data-flow Analysis 1

Graphing Conic Sections

Introduction. Example

TOPIC 10 THREE DIMENSIONAL GEOMETRY

CS 340, Fall 2016 Sep 29th Exam 1 Note: in all questions, the special symbol ɛ (epsilon) is used to indicate the empty string.

6.045J/18.400J: Automata, Computability and Complexity. Quiz 2: Solutions. Please write your name in the upper corner of each page.

Calculus Differentiation

A dual of the rectangle-segmentation problem for binary matrices

Partitioning for Parallelization Using Graph Parsing

Distributed Systems Principles and Paradigms. Chapter 11: Distributed File Systems

WORKSHOP 19 GLOBAL/LOCAL MODELING USING FEM FIELDS

PROBLEM OF APOLLONIUS

Tries. Yufei Tao KAIST. April 9, Y. Tao, April 9, 2013 Tries

Journal of Combinatorial Theory, Series A

Table-driven look-ahead lexical analysis

Paradigm 5. Data Structure. Suffix trees. What is a suffix tree? Suffix tree. Simple applications. Simple applications. Algorithms

[SYLWAN., 158(6)]. ISI

CMPUT101 Introduction to Computing - Summer 2002

Generating Editors for Direct Manipulation of Diagrams

5 ANGLES AND POLYGONS

UTMC APPLICATION NOTE UT1553B BCRT TO INTERFACE PSEUDO-DUAL-PORT RAM ARCHITECTURE INTRODUCTION ARBITRATION DETAILS DESIGN SELECTIONS

A matching algorithm for measuring the structural similarity between an XML document and a DTD and its applications $

Graph Searching & Perfect Graphs

WORKSHOP 9 HEX MESH USING SWEEP VECTOR

ASTs, Regex, Parsing, and Pretty Printing

Introduction to Integration

Section 10.4 Hyperbolas

Solutions to Tutorial 2 (Week 9)

ΕΠΛ323 - Θεωρία και Πρακτική Μεταγλωττιστών

1 Disjoint-set data structure.

Problem Final Exam Set 2 Solutions

Realising Dead Path Elimination in BPMN

From Indexing Data Structures to de Bruijn Graphs

GENG2140 Modelling and Computer Analysis for Engineers

such that the S i cover S, or equivalently S

Internet Routing. Reminder: Routing. CPSC Network Programming

Some necessary and sufficient conditions for two variable orthogonal designs in order 44

COMPUTER EDUCATION TECHNIQUES, INC. (WEBLOGIC_SVR_ADM ) SA:

Computational geometry

Comparing Hierarchical Data in External Memory

EXPONENTIAL & POWER GRAPHS

box Boxes and Arrows 3 true 7.59 'X' An object is drawn as a box that contains its data members, for example:

Advanced Programming Handout 5. Enter Okasaki. Persistent vs. Ephemeral. Functional Queues. Simple Example. Persistent vs.

s 1 t 4 s 2 4 t 2 a b r 2 r 8 r10 g 4

Single-Layer Trunk Routing Using 45-Degree Lines within Critical Areas for PCB Routing

COMPUTER EDUCATION TECHNIQUES, INC. (XML ) SA:

Midterm Exam CSC October 2001

Fundamentals of Engineering Analysis ENGR Matrix Multiplication, Types

called the vertex. The line through the focus perpendicular to the directrix is called the axis of the parabola.

Enumerating pseudo-triangulations in the plane

Distributed Systems Principles and Paradigms

Chapter 16. Shortest Paths Shortest Weighted Paths

Hash-based Subgraph Query Processing Method for Graph-structured XML Documents

Transcription:

Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGs Huizhen Yu jney.yu@s.helsinki.fi Dept. Computer Siene, Univ. of Helsinki Proilisti Moels, Spring, 2010 Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 1 / 22

Outline Direte Mrkov Properties n Mrkov Equivlent DAGs Other Direte Mrkov Properties n Equivlene DAG, Unirete Grphs, n Deomposle Grphs Mrkov Equivlene of DAGs Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 2 / 22

Other Direte Mrkov Properties n Equivlene Outline Direte Mrkov Properties n Mrkov Equivlent DAGs Other Direte Mrkov Properties n Equivlene DAG, Unirete Grphs, n Deomposle Grphs Mrkov Equivlene of DAGs Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 3 / 22

Other Direte Mrkov Properties n Equivlene For DAG G = (V, E), G m : the morl grph. Nottion G A : the sugrph inue y suset A of verties. p(α): the set of prents of α. h(α): the set of hilren of α. n(α): the set of nestors of α. e(α): the set of esennts of α. An nestrl set: set A suh tht p(v) A, v A. An(A): the miniml nestrl set ontining A; An(A) = A α A n(α). A well-orering: (α, β) E numer(α) < numer(β). pr(β): the set of preeessors of β, i.e., pr(β) = α numer(α) < numer(β). l(v): the Mrkov lnket of v, equl to ne(v), the set of neighors of v in G m ; i.e., l(v) = ne(v) = p(v) h(v) {w h(w) h(v) }. Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 4 / 22

Other Direte Mrkov Properties n Equivlene Min Results from Lst Leture For istriution P tht ftorizes reursively oring to DAG G, P ftorizes oring to the morl grph G m ; the mrginl istriution P A (X A ) ftorizes reursively oring to G An(A), from whih we otin the glol irete Mrkov property (DG): for ny isjoint susets A, B, S, A B S in `G An(A B S) m X A X B X S. We hve lso shown The -Seprtion onept; (DG) is equivlent to the -seprtion se Mrkov property. Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 5 / 22

Other Direte Mrkov Properties n Equivlene Lol Direte Mrkov Property Denote n(v) the set of non-esennts of v in G, i.e., n(v) = V \ (e(v) {v}). Consier v, its prents p(v), n its non-esennts n(v). Beuse An`{v} p(v) n(v) is the miniml nestrl set ontining these verties, h(v) An`{v} p(v) n(v) =. Eges in the morl grph `G An({v} p(v) n(v)) m re either from G An({v} p(v) n(v)) or from mrriges. The two fts ove imply tht v is not onnete iretly to n(v) \ p(v) in `G An({v} p(v) n(v)) m. So y (DG), X v X n(v) X p(v). (1) The property in (1) is lle the lol irete Mrkov property (DL). The isussion ove shows lso (DG) (DL). Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 6 / 22

Other Direte Mrkov Properties n Equivlene Illustrtion of (DL) Exmple Stu-Frm: L Ann Brin Ceily K p`dorothy = {Ann, Brin} e`dorothy = {Henry, John} Fre Dorothy Eri Gwenn Given the genotypes of Ann n Brin, Dorothy s genotype is inepenent of the genotype of ll the other noes in the grph elow. Henry John Irene L Ann Brin Ceily K Fre Dorothy Eri Gwenn Irene Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 7 / 22

Other Direte Mrkov Properties n Equivlene Equivlene etween Direte Mrkov Properties Rell tht for n unirete grph, (F) (G) (L) (P). By ontrst, for DAG, the Mrkov properties re ll equivlent: (DF) (DG) (DL) (DO), where these properties re s efine erlier: (DF) P ftorizes reursively oring to G, i.e., p(x) = Y p`x v x p(v). (DG) P oeys the glol irete Mrkov property with respet to G, i.e., A B S in `G An(A B S) m X A X B X S. (DL) P oeys the lol irete Mrkov property with respet to G, i.e., v V X v X n(v) X p(v). (DO) P oeys the orere irete Mrkov property with respet to G, i.e., with respet to ny well-orering, X v X pr(v) X p(v). We verifie (DF) (DG) (DL) n (DO) (DF). It is lso strightforwr to show (DL) (DO) (exerise). Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 8 / 22

Other Direte Mrkov Properties n Equivlene Exmples from Erlier Letures Revisite In Le. 2, we verifie the following sttements using equtions. We n now verify them using only grphs. If (X 1,..., X n) is Mrkov hin, then (X n1, X n2,..., X nk ) is lso Mrkov hin, where 1 < n 1 <... < n k n: (X n,..., X 1) is lso Mrkov hin: 1 2 n 1 n 2 n k-1 n k 1 2 n-1 n 1 2 n 1 n 2 n k-1 n k 1 2 n-1 n X j X j (X j 1, X j+1 ), where X j enotes the olletion of ll the vriles ut X j : 1 2 j-1 j j+1 n 1 2 j-1 j j+1 n In n HMM, the oservtion vriles Y 1,..., Y n re generlly fully epenent: Y1 Y2 Yn-1 Yn Y1 Y2 Y n-1 Yn Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 9 / 22

DAG, Unirete Grphs, n Deomposle Grphs Outline Direte Mrkov Properties n Mrkov Equivlent DAGs Other Direte Mrkov Properties n Equivlene DAG, Unirete Grphs, n Deomposle Grphs Mrkov Equivlene of DAGs Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 10 / 22

DAG, Unirete Grphs, n Deomposle Grphs Some Oservtions Rell tht if P ftorizes oring to G m, P oes not neessrily ftorize reursively oring to G: p(x) = Y φ C (x C ) p(x) = Y p`x v x p(v). C C(G m ) v V Exmple: G α β α G m β P(X α, X β, X γ) with p(x) = exp{x αx γ + x β x γ} ftorizes oring to G m, ut γ γ p(x) p(x α)p(x β )p(x γ x α, x β ) euse X α X β. Conversely, unirete grphs n lso represent onitionl inepenene reltions tht nnot e represente y DAGs: e.g., {} {} {, } {} {} {, } Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 11 / 22

DAG, Unirete Grphs, n Deomposle Grphs Some Oservtions But for hins n more generlly, roote trees, the irete n unirete version seems to e equivlent: ftoriztion w.r.t. one implies ftoriztion w.r.t. the other. 1 2 n-1 n 1 2 n-1 n Y1 Y2 Yn-1 Yn Y1 Y2 Y n-1 Yn e e f f Are there other types of grphs for whih this is true? Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 12 / 22

DAG, Unirete Grphs, n Deomposle Grphs Deomposle Grphs: Definition A triple (A, B, C) of isjoint susets of the vertex set V of n unirete grph G is si to eompose G, if V = A B C, n the following two onitions hol: 1. C seprtes A from B; 2. C is omplete suset of V. If oth A n B re non-empty, we sy tht the eomposition is proper. We sy tht n unirete grph G is eomposle if either: (i) it is omplete, or (ii) it possesses proper eomposition (A, B, C) suh tht oth sugrphs G A C n G B C re eomposle. Note tht this efinition is reursive. Exmples: hins, trees, n e f Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 13 / 22

DAG, Unirete Grphs, n Deomposle Grphs Deomposle Grphs: Ftoriztion of Distriutions Proposition 1: Assume tht (A, B, S) eomposes n unirete grph G = (V, E). Then istriution P ftorizes oring to G if n only if oth P A S n P B S ftorize oring to G A S n G B S respetively n P stisfies p(x) = p A S(x A S ) p B S (x B S ). p S (x S ) Implition: If the grph is eomposle, then the mrginl istriutions P A S n P B S mit further ftoriztion in the ove form, n P ftorizes reursively in terms of these mrginls. Exmple: For Mrkov hin X = (X 1,..., X n), P(X ) n e expresse s p(x) = p(x1, x2)p(x2, x3) p(xn 1, xn). p(x 2)p(x 3) p(x n 1) Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 14 / 22

DAG, Unirete Grphs, n Deomposle Grphs Proof of Proposition 1 Let C(G) enote the set of liques of G. Suppose P ftorizes oring to G, i.e.,, p(x) = Y φ C (x C ), for some funtions φ C. C C(G) Sine (A, B, S) eomposes G, ll the liques tht re not susets of A S must e susets of B S, so tht p(x) = Y C C(G A S ) φ C (x C ) Y C C(G B S ) φ C (x C ) = h(x A S ) k(x B S ), for some funtions h, k. It follows then p(x A S ) = X h(x A S ) k(x B S ) = h(x A S ) k(x S ), x B where k(x S ) = X k(x B S ); x B p(x B S ) = X x A h(x A S ) k(x B S ) = h(x S ) k(x B S ), where h(x S ) = X x A h(x A S ); So p(x S ) = X x A X p A S (x A S ) p B S (x B S ) p S (x S ) The onverse is evient. x B h(x A S ) k(x B S ) = h(x S ) k(x S ). = h(x A S ) k(x S ) h(x S ) k(x B S ) h(x S ) k(x S ) = h(x A S ) k(x B S ) = p(x). Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 15 / 22

DAG, Unirete Grphs, n Deomposle Grphs Deomposle Grphs: Direte or Unirete? Definition: A DAG G = (V, E) is si to e perfet if for ll v V, p(v) is omplete suset of G, in other wors, if its unirete version G = G m. Note: G of perfet DAG is eomposle. (How to see this?) Proposition 2: Let G e perfet DAG n G its unirete version. Then istriution P is irete Mrkov with respet to G if n only if it ftorizes oring to G. Implition: Proilisti Inepenenes/epenenes Deomposle Moels DAG Moels Unirete Moels Chin Grph Moels Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 16 / 22

DAG, Unirete Grphs, n Deomposle Grphs Proof of Proposition 2 Proof: Sine G = G m, if P ftorizes reursively oring to G, then it ftorizes oring to G. Conversely, if P ftorizes oring to G, we show P reursively ftorizes oring to G y inution on the numer of verties V. For V = 1, the onlusion hols trivilly. For V > 1, onsier terminl vertex α. Then, (V \ ({α} p(α)), {α}, p(α)) eomposes D. So y Proposition 1, P n e expresse s p(x) = p(x V \{α}) p(x α, x p(α) ) p(x p(α) ) = p(x V \{α} )p(x α x p(α) ), (2) n P V \{α} ftorizes oring to GV \{α}. By inution, P V \{α} ftorizes reursively oring to G V \{α} : p(x V \{α} ) = Y p(x v x p(v) ). v V \{α} Together with Eq. (2) this shows tht P ftorizes reursively oring to G. Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 17 / 22

Mrkov Equivlene of DAGs Outline Direte Mrkov Properties n Mrkov Equivlent DAGs Other Direte Mrkov Properties n Equivlene DAG, Unirete Grphs, n Deomposle Grphs Mrkov Equivlene of DAGs Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 18 / 22

Mrkov Equivlene of DAGs Mrkov Equivlene of DAGs Terminologies for DAG G: Skeleton of G: G, the unirete version of G. A omplex (lso lle n immorlity): sugrph of G inue y three verties (α, γ, β), of the form shown on the right. I.e., the two prents α, β re not joine in G. α γ β Definition: We sy tht two DAG G n e G re Mrkov equivlent if they represent the sme onitionl inepenene reltions. Theorem: Two DAG G n e G re Mrkov equivlent if n only if they hve the sme skeleton n the sme omplexes. This theorem hs importnt implition for utomti moel seletion: Sine (DG) (DF), the set of istriutions ssoite with G is the sme s tht ssoite with ny DAG Mrkov equivlent to G. When eges o not hve usl interprettions, it woul e in vin to try to istinguish etween equivlent moels, n we nee to ompre equivlent lsses of DAGs inste. (The representtive grph of n equivlent lss is hin grph, interestingly.) Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 19 / 22

Mrkov Equivlene of DAGs Exmples of Mrkov Equivlent DAGs An equivlent lss onsisting of three DAGs n their representtive (from Anersson et l. 1997): Wht hppene to the other 5 possiilities? Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 20 / 22

Mrkov Equivlene of DAGs Proof of the Mrkov Equivlene Theorem for DAG Proof of suffiieny: if G n G e hve the sme skeleton n the sme omplexes, then they hve the sme morl grphs for ll sugrphs inue y the sme set of verties. This mens tht for ny isjoint susets A, B, S of verties, A B S in `G An(A B S) m A B S in `e GAn(A B S) m. So G n e G represent the sme onitionl inepenene reltions. Proof of neessity: we rgue this y using ounterexmples. Sme skeleton: suppose there is n ege etween α n β in e G ut not in G. Then α, β is non-jent in `G An({α,β}) m, so α β An({α, β}) \ {α, β}. Assume x { 1, 1} V. Let p(x) exp{x αx β }. Then P ftorizes oring to e G, ut X α n X β re not inepenent given X An({α,β})\{α,β}, ontrition. Sme omplexes: suppose there is omplex inue y (α, γ, β) in G, ut in eg, the onnetion etween them is ifferent. Agin, α, β is non-jent in `GAn({α,β}) m, so α β An({α, β}) \ {α, β}. Assume x { 1, 1} V. Let p(x) exp{x αx γ + x β x γ}. Then P ftorizes oring to e G, ut X α n X β re not inepenent without X γ given, ontrition. Note: This theorem is shown y Verm n Perl (1991) n inepenently y Fryenerg (1990) (for hin grphs). In the ove we followe the ltter proof, whih uses (DG) inste of -seprtion. Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 21 / 22

Mrkov Equivlene of DAGs Further Reings For irete Mrkov properties n eomposle grphs: 1. Roert G. Cowell et l. Proilisti Networks n Expert Systems, Springer, 2007. Chp. 5.2, 5.3. Huizhen Yu (U.H.) Byesin Networks: Direte Mrkov Properties (Cont ) n Mrkov Equivlent DAGsFe. 11 22 / 22