What is it all above How to use

What is it all above ? • How to use graphs to represent probability distributions over thousands of random variables ? • How to encode conditional independence in directed and undirected graphs ? • How to use such representations for efficient computations of the probability of events of interest ? • How to learn such models from data ? 2

Course Information Meetings: · Lecture: Wednesdays 10: 30 – 12: 30 · Tutorial: Wednesdays 16: 30 – 17: 30 Grade: u 30% in 5 question sets. These questions sets are obligatory. Each contains mostly theoretical problems. Submit in pairs before due time (three weeks). u 40% one or two hours lecture (Priority to graduate students). u 10% Attending at least 12 lectures and recitation classes for a **passing grade**. (In very special circumstances 2% per missing item). u 20% Checking HMWs and presentation of HMW solutions u u Prerequisites: Data structure 1 (cs 234218) Algorithms 1 (cs 234247) Probability (any course) Information and handouts: u http: //webcourse. cs. technion. ac. il/236372 u http: //www. cs. technion. ac. il/~dang/course. BN/ (Only lecture slides) 3

Relations to Some Other Courses. אמור לי מי חבריך ואומר לך מי אתה u Introduction to Artificial Intelligence (cs 236501) u Introduction to Machine Learning (cs 236756) u Introduction to Neural Networks (cs 236950) u Algorithms in Computational Biology (cs 236522) u Error correcting codes u Data mining 4

= Student lectures (8) = TENTATIVE Student lectures (7) 5

u Mathematical Foundations (4 weeks including students’ lectures, based on Pearl’s Chapter 3 + papers). 1. Properties of Conditional Independence (Soundness and completeness of marginal independence, graphoid axioms and their interpretation as “irrelevance”, incompleteness of conditional independence, no disjunctive axioms possible. ) 2. Properties of graph separation (Paz and Pearl 85, Theorem 3), soundness and completeness of saturated independence statements. Undirected Graphs as I-maps of probability distributions. Markov-Blankets, Pairwise independence basis. Representation theorems (Pearl and Paz, from each basis to I-maps). Markov networks, HC representation theorem, Completeness theorem. Markov chains 3. Bayesian Networks, d-separation, Soundness, Completeness. 4. Chordal Graphs as the intersection of BN and Markov networks. Equivalence of their 4 definitions. u Combinatorial Optimiziation of Exact Inference in Graphical models (3 weeks including students lectures). 1. HMMs 2. Exact inference and their combinatorial optimization. 3. Clique tree algorithm. Conditioning. 4. Tree-width. Feedback Vertex Set. u Learning (5 weeks including students lectures). 1. Introduction to Bayesian statistics 2. Learning Bayesian networks 3. Chow and Liu’s algorithm; the TAN model. 4. Structural EM 5. Searching for Bayesian networks u Applications (2 weeks including student lectures). 6

Homeworks • HMW #1. Read Chapter 3. 1 & 3. 2. 1. Answer Questions 3. 1, 3. 2, Prove Eq 3. 5 b, and fully expand/fix the proof of Theorem 2. Submit in pairs no later than 28/10/09 (Two weeks from now). • HMW #2. Read Chapter 2 in Pearl’s book and answer 6 questions of choice at the end. Submit in pairs no later than 11/11/09. • HMW #3. Read fully Chapter 3. Answer additional 5 questions of choice at the end. Submit in pairs no later than 2/12/09. • HMW #4. Submit in pairs 23/12/09 • HMW #5. Submit in pairs 6/1/10 Pearl’s book contains all the notations that I happen not to define in these slides – consult it often – it is also a very unique and interesting classic text book. 7

The Traditional View of Probability in Text Books • Probability theory provides the impression that we need to literally represent a joint distribution explicitly as P(x 1, …, xn) on all propositions and their combinations. It is consistent and exhaustive. • This representation stands in sharp contrast to human reasoning: It requires exponential computations to compute marginal probabilities like P(x 1) or conditionals like P(x 1|x 2). • Humans judge pairwise conditionals swiftly while conjunctions are judged hesitantly. Numerical calculations do not reflect simple reasoning tasks. 8

The Traditional View of Probability in Text Books Given ? Computed ? Estimated or Computed ? P(e | h) Given ? . כחום גבוה e כמחלה ויראלית נדירה ועל h חישבו על 9

The Qualitative Notion of Dependence • Marginal independence is defined numerically as P(x, y)=P(x) P(y). The truth of this equation is hard to judge by humans, while judging whether X and Y are dependent is often easy. • “Burglary within a day” and “nuclear within five years” • Likewise, people tend to judge x Y Z 10

• The notions of relevance and dependence are far more basic than the numerical values. In a resonating system it should be asserted once and not be sensitive to numerical changes. • Acquisition of new facts may destroy existing dependencies as well as creating new once. • Learning child’s age Z destroys the dependency between height X and reading ability Y. • Learning symptoms Z of a patient creates dependencies between the diseases X and Y that could account for it. Probability theory provides in principle such a device via P(X | Y, K) = P(X |K) But can we model the dynamics of dependency changes based on logic, without reference to numerical quantities ? 11

Definition of Marginal Independence Definition: IP(X, Y) iff for all x DX and y Dy Pr(X=x, Y=y) = Pr(X=x) Pr(Y=y) Comments: u Each Variable X has a domain DX with value (or state) x in DX. u We often abbreviate via P(x, y) = P(x) P(y). u When Y is the emptyset, we get Pr(X=x) = Pr(X=x). u Alternative notations to IP(X, Y) such as: I(X, Y) or X Y u Next few slides on properties of marginal independence are based on “Axioms and algorithms for inferences involving probabilistic independence. ” 12

Properties of Marginal Independence Trivial Independence: Ip(X, ) Symmetry: Ip(X, Y) Ip(Y, X) Decomposition: Ip(X, YW) Ip(X, Y) Mixing: Ip(X, Y) and Ip(XY, W) Ip(X, YW) Proof (Soundness). Trivial independence and Symmetry follow from the definition. Decomposition: Given P(x, y, w) = P(x) P(y, w), simply sum over w on both sides of the given equation. Mixing: Given: P(x, y) = P(x) P(y) and P(x, y, w) = P(x, y) P(w). Hence, P(x, y, w) = P(x) P(y) P(w) = P(x) P(y, w). 13

Properties of Marginal Independence Are there more such independent properties of independence ? No. There are none ! Horn axioms are of the form 1 & … & n where each statement stands for an independence statement. We use the symbol for a set of independence statements. Namely: is derivable from via these properties if and only if is entailed by (i. e. , holds in all probability distributions that satisfy ). Put differently: For every set and a statement not derivable from , there exists a probability distribution P that satisfies S and not s. 14

Properties of Marginal Independence Can we use these properties to infer a new independence statements from a set of given independence statements in polynomial time ? YES. The “membership algorithm” and completeness proof in Recitation class (Paper P 2). Comment. The question “does entail ” could in principle be undecidable, drops to being decidable via a complete set of axioms, and then drops to polynomial with this claim. 15

Properties of Marginal Independence we check consistency of a set + independence plus a set - of negated independence statements ? u. Can The membership algorithm in previous slide applies only for G that includes one negated statement – simply use the algorithm to check that it is not entailed from +. But another property of independence called “Amstrong Relation” guarantees that consistency is indeed verified by checking separately (in isolation) that each statement in is not entailed from +. 16

Properties of Marginal Independence We define a “product” of two distributions P 1 and P 2 over a set of n variables X 1, …, Xn as follows: P (a 1 1 , …, an n) P 1(a 1 , …, an) P 2( 1 , …, n) Note that the domains of Xi change in this definition ! This definition implies that for every independence statement : hold in P 1 and holds in P 2 if and only if holds in P*. Hence, consistency can be checked statement by statement. 17

Definitions of Conditional Independence Ip(X, Z, Y) if and only if whenever P(Y=y, Z=z) >0 (3. 1) Pr(X=x | Y=y , Z=z) = Pr(X=x |Z=z) 18

Properties of Conditional Independence Same Properties of conditional independence: Symmetry: I(X, Z, Y) I(Y, Z, X) Decomposition: I(X, Z, YW) I(X, Z, Y) Mixing: I(X, Z, Y) and I(XY, Z, W) I(X, Z, YW) BAD NEWS. Are there more properties of independence ? Yes, infinitely many independent Horn axioms. No answer to the membership problem, nor to the consistency problem. 19

Important Properties of Conditional Independence Recall there are some more notations. 20

Dependency Models – abstraction of Probability distributions 21

Graphical Interpretation 22

The Intersection Property Revisited Intersection: I(X, ZW, Y) and I(X, ZY, W) I(X, Z, YW) This property holds for positive distributions. Counter example: X=Y=W, Z is the emptyset. A set of independence statements I(x, y, z) (namely, a dependency model) that satisfies Symmetry, Decomposition, Weak Union, Contraction, and Intersection is called a graphoid, and without intersection is called semi-graphoid. 23