很多DS的面试会聊到experimentation,分享几个我觉得非常有帮助的blog

IT公司回答how to measure the success of a feature一些方法的讨论和相关的blog

Experimentation/Stats 是几乎所有data scientist面试必考的方向。下面让我来试着给出一些常用的方法并和reference。

Experimentation想要解决的根本问题就是correlation <> causation。笼统的讲，我们可以把experimentation分成可以做randomization的情况（比如tranditional的AB test）和不能做randomization的情况（比如causalimpact，difference-in-difference）。下面的几个blog会给你各个方法浅显易懂的解释：

1)        Randomized experiment
Why power is important? https://multithreaded.stitchfix. ... ngs-about-power/#f1

Increase power using pre-exp data: https://booking.ai/how-booking-c ... -cuped-995d186fff1d

Bayesian AB testing: https://medium.com/convoy-tech/t ... testing-f859d2219d5

.google  иSelection bias and holdout group: https://medium.com/airbnb-engine ... tation-c3d67795cceb


2)        Non-randomized experiment
Causal impact for pre-post analysis: https://google.github.io/CausalImpact/CausalImpact.html

Difference-in-difference analysis: https://www.mailman.columbia.edu ... fference-estimation

感谢分享！感觉a/b testing 学问很大啊。

感谢分享！刚开始学习diff in diff.