DESCN结构:
包括一个输出倾向分的网络,一个输出tr和cr的网络(论文中用的x-network),输入先通过一个shared network,再分别进入这两个网络。
本质上借鉴了x-learner方法:.
We can see that our X-network is similar to X-learner in that
both try to directly learn the counterfactual treatment effect. In
X-learner, ITE is learned based on the results from base learners,
and its performance is heavily subject to that of the base models.
By contrast, in X-network, ITE is learned together with the base
learners in an integrated