Pattern Recognition course 笔记- causal inference and learning

仅为个人笔记使用

---

Causality: ubiquitous in science and daily life

• health
• climate
• cosial science
• Neuroscience

if you change something here , there might be high probablity of certain behaviour

progmatic approach

causality = “effective manipulability”

• relevant, measurable influence
• synonyms:control, explanation, responsibility,effect
• group level, not instance level

在这里插入代码片

speed slippery brake -->crash
key contributing factors

who’s regulating who?

“Simpson’s paradox”

• treatment of kidney stones
  

which often occurs in statistical analysis, by controling different parameter you can get different outcome
How to compute the correct effect?

• adjustment formula to compensate for confounding
• do-calculus

Causal effect estimation

key model assumption: causal DAG

• real world consists of networks of causally interacting variables,
• structure corresponds to a directed acyclic graph (DAG)
• subset of these variables observed in experiments
  

Basic graphical model terminology

• a path $\pi =〈X,..,Y〉$ is unblocked given nodes Z if:

• all non-colliders along π are not in Z
• all colliders along π are ancestor of some Z ∈ Z

path 〈A,B,F,G〉 is unblocked given the empty set
path 〈A,B,D,G〉 becomes unblocked given F
path 〈A,B,F,G〉 is blocked given F
path 〈A,B,D,G〉 is again blocked given {D,F}

d-separation

• in a graph G, nodes X and Y are d-separated given Z, iff
  there are no unblocked paths in G between X and Y given Z,
  otherwise they are d-connected