uber面经safety组

自己还觉得挺顺利的，没想到第二天就接到据信。

1. 小哥讲了一下他们组做的project,目的是预测发生在非洲之一类贫困落后地区司机被打劫的概率， 然后block this request. 请问你要怎么收集feature?
关于从geographic (city, area, lon + lat), time during the day 上面建立tagging.
2. 模型建好了，你要考虑用什么metrics to evaluate?
我说recall，就是尽量completely capture the incidences. 然后选threshold尽量降低type I error, 就算type II error 升高一些也没关系。
3. a/b test来evaluate the model.要怎么split?
4. test 组如果预测有高的概率就block this request, how do we compare test & control?
I said assuming test and control are completely random, we can score both test and control, and then segment by the block threshold.  
Then we can compare the non blocked populate to ensure a fair comparison.

metric应该用precision-recall curve吧，因为imbalance data and true positive is more important than true negative

赞赞赞，谢谢楼主啊！

谢谢楼主分享

请问这个是有没有coding link的？

对的， 不需要电脑， 只是打电话。

谢谢楼主分享！

谢谢楼主分享，这个是不是应该尽量降低type II error啊，metric选择是要选business related metric还是recall这些啊