Causal Generative Domain Adaptation Networks[Daily Reading]

Causal Generative Domain Adaptation Networks [Daily Reading]

causal graph

话不多说，先上图，这篇文章的因果图和我之前读过的，都有所不同。As the variable that need to be predicted, Y is on the cause position instead of effect position. In the following we will see how this makes sense.

Framework

In this paper they propose a Generative Domain Adaptation Network (G-DAN) with specific latent variables $\theta$ to capture changes in $P_{X|Y}$. $X$denotes features and $Y$ the class label. $P_{X|Y}$ is implicitly represents by functional model $X=g(Y,E,\theta )$. In this function, latent variables $\theta$ may change across domains. E, the noise

Identifiability theory to support $\theta$

Causality: Testing Identifiability
(The above blog explains how to identify joint distribution under confounding in detail.)
In the 3.1 section , the authors use identifiability to make sure whether the estimated $\hat{\theta }$ can capture all the distribution changes.

Model Learning

$\theta$ is estimated by

CG-DAN(Causal Generative Domain Adaptation Networks)