Mathematical Methods in Engineering 2 (Machine Learning)
Dept. of Control & Instrumentation Engineering, Korea Univ. Jooyoung Park
(Textbook: C. Bishop, Pattern Recognition and Machine Learning, Cambridge Univ. Press, 2006)
Lecture #1 2016.3.15
Graphical models (PGM)
Ch8 Graphical models (PGM)
-diagram instead of algebraic manipulation
-useful properties of PGM:
① visualization (e.q. for factorization)
② insights (e.q. for cond. indep.)
③ computation made easy (e.q. sum-product alg.)
-PGM, 𝐺 = 𝑣 , 𝜀
vertex (node) represents a r.v.
edge (link) represents probabilistic
Graphical models (PGM)
-Three types of PGM
① directed graphs (Bayesian networks)
② undirected graph (Markov random fields)
③ factor graphs
EX)
a a
b c b c
BN MRF FG
Bayesian networks
8.1 Baysesian networks (DG)
EX) 𝑝 𝑎, 𝑏, 𝑐 = 𝑝 𝑎 𝑝 𝑏 𝑎 𝑝 𝑐 𝑎, 𝑏 = 3𝑘=1𝑝(𝑥𝑘|𝑝𝑎 𝑥𝑘 )
Note: 𝑝 𝑥1, ⋯ , 𝑥𝑘 = 𝐾𝑘=1𝑝(𝑥𝑘|𝑝𝑎 𝑥𝑘 ) joint pdf a factorization
a
b c
𝑥1
𝑥2 𝑥3
Bayesian networks
EX) Lin. regress prob (Bayesian approach)
(A Bayesian approach for the lin. reg. prob.) trn data D= 𝑥𝑛, 𝑡𝑛 𝑛=1𝑁
𝑡𝑛 = 𝑤𝑇∅ 𝑥𝑛 + 𝜖𝑛, 𝜖𝑛~𝑁 0, 𝜎2 IID In the Bayesian approach,
𝑤~𝑁 0, 𝛼𝐼
more precisely
∴joint pdf 𝑝 𝑡, 𝑤 𝑥, 𝛼, 𝜎2 = 𝑝 𝑤 𝛼 𝑝 𝑡 𝑤, 𝑥, 𝜎2 = 𝑝 𝑤 𝛼 𝑝 𝑡𝑛 𝑤, 𝑥𝑛, 𝜎2 Joint dist. : 𝑝(𝑡, 𝑤) = 𝑝(𝑤) 𝑝(𝑡|𝑤) = 𝑝(𝑤) 𝑁𝑛=1𝑝(𝑡𝑛|𝑤)
hyper parameter
b
𝑡1 … 𝑡𝑛 𝑡𝑛 N
w
Conditional indep.
8.2 Conditional indep.
Def. 𝑎 and 𝑏 are conditionally indep. Given 𝑐,
if 𝑝 𝑎, 𝑏 𝑐 = 𝑝 𝑎 𝑐 𝑝(𝑏|𝑐) (where a, 𝑏, 𝑐 are r.v.)
Notation : 𝑎 ⊥ 𝑏
Note : Assume 𝑝(𝑏|𝑐) ≠ 0 𝑎 ⊥ 𝑏|𝑐, we have 𝑝 𝑎 𝑐 = 𝑝(𝑎|𝑏, 𝑐)
∵ 𝑝 𝑎 𝑏, 𝑐 = 𝑝 𝑎,𝑏 𝑐𝑝 𝑏 𝑐 = 𝑝(𝑎|𝑐)
Conditional indep.
8.2.2 d-separation property
Consider a directed graph, where A, B, and C are arbitrary non intersecting sets of nodes.
Def. Consider all possible (undirected) paths from any node in A to any node in B.
We say that any such path, P, is blocked if at least one of the following hold:
a. P contains either HT or TT node, and the node is in the set C.
(즉, 관찰된 non-consider 노드는 이 노드를 지나는 path 를 block 함) b. P contains a collider(HH node), and neither the node nor any of its
descendants belongs to the set C.
(즉, 관찰되지 않은(latent) HH노드는 이 노드를 지나는 path를 block 가능) 단, 이 collider의 후손이 관찰되면 이 bock 상황이 해제됨.
Means “directed”
Conditional indep.
If all the possible paths are blocked, then A is d-separated from B given C.
(또는 A and B are d-separated C)
d-separation Thm. :
Let A, B and C can separated sets of node in a directed graph.
If A and B are d-separated by C, then we have 𝑎 ⊥ 𝑏|𝑐