COMPUTER SCIENCE EN
Inference_for_Graphs_and_Networks.pdf

# If we substitute the joint distribution pr x x z z θ

• 99

This preview shows pages 85–87. Sign up to view the full content.

If we substitute the joint distribution Pr( X = x , Z = z | θ ), along with γ and ξ , into (3.5), we obtain Q ( θ , θ old ) = γ ( z 10 ) log(1 π ) + γ ( z 11 ) log π + T t =2 1 j =0 1 k =0 ξ ( z t 1 ,j , z tk ) log A jk + T t =1 γ ( z t 0 ) I ( x t = 0) + γ ( z t 1 ) log Pr( X t = x t | Z t = 1 , φ ) . (3.6) Next, we seek an eﬃcient procedure for evaluating the quantities γ ( z tk ) and ξ ( z t 1 ,j , z tk ). The forward–backward algorithm (Baum and Eagon, 1967; Baum and Sell, 1968) is used to accomplish this. First, we define the forward variable as α ( z t,k ) = Pr( X 1 = x 1 , . . . , X t = x T , Z t = k | θ ) k = 0 , 1 . α can be solved for inductively: (1) Initialization: α ( z 1 , 0 ) = 1 π α ( z 1 , 1 ) = π Pr( X 1 = x 1 | Z 1 = 1 , φ ). (2) Induction: For k = 0 , 1 and 1 t T 1, α ( z t +1 ,k ) = [ α ( z t, 0 ) A 0 k + α ( z t, 1 ) A 1 k ] Pr( X t = x t | Z t = k, φ ) . (3.7) Below, we will use the fact that Pr( X = x | θ ) = α ( z T, 0 ) + α ( z T, 1 ). We next need to define the backward variable , the probability of the partial observation sequence from t + 1 to T : β ( z t ) = Pr( X t +1 = x t +1 , . . . , X T = x T | Z t = z t , θ ) . Copyright © 2014. Imperial College Press. All rights reserved. May not be reproduced in any form without permission from the publisher, except fair uses permitted under U.S. or applicable copyright law. EBSCO Publishing : eBook Collection (EBSCOhost) - printed on 2/16/2016 3:37 AM via CGC-GROUP OF COLLEGES (GHARUAN) AN: 779681 ; Heard, Nicholas, Adams, Niall M..; Data Analysis for Network Cyber-security Account: ns224671

This preview has intentionally blurred sections. Sign up to view the full version.

86 J. Neil, C. Storlie, C. Hash and A. Brugh β ( z t ) can be solved for inductively as follows: (1) Initialization: β ( z T,k ) = 1 k = 0 , 1. (2) Induction: For k = 0 , 1 and t = T 1 , . . . , 1, β ( z t,k ) = A k, 0 Pr( X t +1 = x t +1 | Z t +1 = 0) β ( z t +1 , 0 ) + A k, 1 Pr( X t +1 = x t +1 | Z t +1 = 1 , φ ) β ( z t +1 , 1 ) . (3.8) Finally, γ ( z t ) = α ( z t ) β ( z t ) Pr( X = x | θ ) (3.9) ξ ( z t 1 , z t ) = α ( z t 1 ) Pr( X t = x t | Z t = z t , φ ) Pr( Z t = z t | Z t 1 = z t 1 ) β ( z t ) Pr( X = x | θ ) . (3.10) The M Step. In the M step, we maximize (3.6) with respect to θ . Max- imization with respect to π and A is easily achieved using appropriate Lagrange multipliers. Taking the derivative with respect to µ results in a closed form update as well. ˆ π = γ 11 1 j =0 γ ( z 1 j ) ˆ A jk = T t =2 ξ ( z t 1 ,j , z tk ) 1 l =0 T t =2 ξ ( z t 1 ,j , z tl ) j = 0 , 1 k = 0 , 1 ˆ µ = T t =1 γ ( z t 1 ) x t T t =1 γ ( z t 1 ) . The size parameter update is not closed form. From (3.6), we see that it comes down to maximizing log Pr( X t = x t | Z t = 1 , φ ) with respect to s , which we achieve through a numerical grid optimization routine. Scaling. For moderate lengths of chains, the forward and backward vari- ables quickly get too small for the precision of the machine. One cannot work with logarithms, as is the case for independent and identically dis- tributed (i.i.d) data, since here we have sums of products of small numbers. Therefore a rescaling has been developed, and is described in Bishop (2006). Define a normalized version of α as ˆ α ( z t ) = Pr( Z t = z t | X 1 = x 1 , . . . , X t = x t ) = α ( z t ) Pr( X = x | θ ) .
This is the end of the preview. Sign up to access the rest of the document.
• Spring '12
• Kushal Kanwar
• Graph Theory, Statistical hypothesis testing, Imperial College Press, applicable copyright law

{[ snackBarMessage ]}

### What students are saying

• As a current student on this bumpy collegiate pathway, I stumbled upon Course Hero, where I can find study resources for nearly all my courses, get online help from tutors 24/7, and even share my old projects, papers, and lecture notes with other students.

Kiran Temple University Fox School of Business ‘17, Course Hero Intern

• I cannot even describe how much Course Hero helped me this summer. It’s truly become something I can always rely on and help me. In the end, I was not only able to survive summer classes, but I was able to thrive thanks to Course Hero.

Dana University of Pennsylvania ‘17, Course Hero Intern

• The ability to access any university’s resources through Course Hero proved invaluable in my case. I was behind on Tulane coursework and actually used UCLA’s materials to help me move forward and get everything together on time.

Jill Tulane University ‘16, Course Hero Intern