我第一轮考的非常咨询,问题是company A is a competitor of company B. Company B now gives a bundle sales, should company A give bundle sales as well? A/B test 啥都不好使,完全是consulting case, 但他会问你怎么collect data,我答的也不太好
和楼上用逻辑回归同理,如果这200个feature不满足逻辑回归的assumption(比如highly correlated),是否可以比较199个和200个features分别代入tree based model的performance的变化比较?由此看这个feature improve了多少percent
面试官说ideally A/B Test是最优解,但是没办法randomly select to control group和experiment group, 因为你不知道其他199个features的影响,而且t test无法measure the effectiveness of this anti-fraud solution in percentage