讨论一道常见的DS面试题

被问到好几次了，答案对方似乎没有特别满意，发上来和大家一起讨论下
某金融机构给潜在客户发信宣传开户offer，记录respond to offer or not, Y非常unbalanced，例如30million里有10000个1剩下都是0，这种时候：
1. logistic的问题在哪里
2. 用tree model的话 用single tree可能出现什么问题 one good thing and one bad thing
3. 用什么measure去evaluate model performance and do model selection
4. influential factors

3. AUC, F measure, confusion matrix, precision, recall. 4. Undersampling, overzampling, SMOTE.

1. Logistic 会predict everything as 0
2. Single tree will give low bias but high variance
3. Prediction & Recall - > F measures, ROC, AUC, etc.

ccqueeny 发表于 2019-03-28 11:40:01
1. Logistic 会predict everything as 0
2. Single tree will give low bias but high

能解释下2么？

Single tree too simple, will undetfitting--->high bias, low variance

That is why we need forest.

single tree may overfit - memorize the noise in the data which cannot generalize to new data. High variance low bias.