Basis

Discriminative & Generative Model

  • 判别模型 Discriminative Model,又可以称为条件模型,或条件概率模型。估计的是条件概率分布(conditional distribution)
  • 生成模型 Generative Model,又叫产生式模型。估计的是联合概率分布(joint probability distribution)

对于这两个模型,我们用一个例子就能很好的解释:假设我们有一些数据 $(x,y)$,数据只有四组:$(1,0), (1,0), (2,0), (2,1)$

Generative Model :

p(x,y) y=0 y=1
x=0 1/2 0
x=1 1/4 1/4

Discriminative Model :

p(x\ y) y=0 y=1
x=0 1 0
x=1 1/2 1/2

PGM

概率图有这么几种推断方式

概率图的任务通俗的讲有两个

  • Querying:给定或者不给定条件概率下,计算特定变量的概率称为:Inference
  • Estimation:当模型某部分是未知的时候,通过数据 $D$ 来推断出未知的模型,这个称为:learning

Inference

Querying 任务通常来说有三种

Likelihood : Likelihood is calculated to get the conditional probability of a different subset of variables conditioned based on Evidence $\mathbf{E} = \{ X_{k+1},…,X_n \}$, Evidence is the unknown variables, so we need eliminate the unsure variables to get the specific lieklihood.

Conditional Probability : The conditional probability distribution of some query nodes conditioned on an
evidence.

Let $\mathbf{Y}$ be a subset of all domain variables $\mathbf{X} = \{ \mathbf{Y},\mathbf{Z} \}$, $\mathbf{Z}$ is the set of variables under elimination.

Most Probable Assignment : In this query, we are interested in finding only one set of values for the query variables that maximize the given conditional probability instead of finding the entire distribution.

Inference 算法可以分为两种

  • Exact inference(准确推断)

    • Elimination
    • Message-passing
      • sum-product
      • belief propagation
    • Junction Tree
  • Approximate inference(近似推断)a

    • Variational algorithms
      • Loopy belief propagation
      • Mean field approximation

    一般 Inference 最简单的方法是 Elimination 和 brute force,但是由于概率图模型往往有一些比较特别的结构,我们可以用 Message Passing 的算法来做


Learning

根据观测数据,来推理模型中的参数

在做 inference 的时候,我们希望能将有向图和无向图结合起来,比如,原先的子节点的父节点们,我们希望将他们