data analyst面试统计知识总结~求大米

目前楼主正在找全职da工作，因此复习的时候整理了一些统计相关的面试点，希望可以帮到大家，一起加油！不定期更新~如有错误烦请指正，谢谢~


Hypothesis Testing

Compare 2 data sets            
Compare 1 data set to an idealized data set

Null hypothesis
Alternative hypothesis

Type 1 error: h0 true, reject h0                           false positive            alpha    significance level    5%
Type 2 error: h0 false, fail to reject h0                 false negative           beta

Confidence level: 1-alpha   95%
                         Prob that a parameter falls within a specific range of values
                         Frequency that CI contains the true value of the unknown population parameter

Confidence interval: interval estimate
                             may contain the true value of an unknown population parameter

Power: the prob of rej h0 when h1 is true
          1-beta
          sensitivity

p-value: when h0 is true, the prob of statistical summary >= actual observed results
            e.g. average weight of all babies last year is 200, random sample of this year is 220.
                  Null hyp: average weight of all babies this year is still 200
                  P-value = 0.004
                  If h0 is right, the prob that avg weight of random sample is 220 is extremely small.
                  H0 is wrong


Steps:
1. state h0, h1, alpha
2. determine test statistic, calculate its value, compare it to critical value
3. draw conclusion about h0

Difference between t-test and z-test



补充内容 (2019-1-28 12:36):
因为无法再次编辑，我就把更新内容发在回复里啦

补充内容 (2019-2-14 05:09):
大家可以“只看该作者”

pixie1994 发表于 2019-3-27 05:02
请问楼主准备这些面试的时候用什么复习书比较合适啊 谢谢～～～～～

我用了国内的统计书，然后还有美国研究生上课老师发的讲义~还有就是网上搜哈哈

Logistic Regression
Binomial classification
Robust

Basic steps:
1. hypothesis function to predict, h
2. cost function to measure the difference between h and observed results, sum or avg, J
3. min step 2 by gradient descent

gradient descent: first-order iterative optimization algorithm to find min of a function

Detailed steps:
1. h
Sigmoid function:
g(z)=1/(1+e^(-z) )    Linear decision boundary
   

2. cost function (min J) -- maximum likelihood estimate

3. min J

Difference between linear & logistic

	Linear	Logistic
Outcome	Continuous, can have infinite    Number of possible values	Limited number of possible    values
response	Continuous, negative values	Categorical
Error minimization technique	Ordinary least squares	Maximum likelihood method
Coefficient interpretation	Straightforward	Depends on family (binomial,    Poisson, etc.) and link (logit, log, etc.)

Linear Regression (STATS 500)
a linear approach to model the relationship between a dependent variable & one or more independent variables

simple LR: 1 explanatory variable
multiple LR: more than 1 explanatory variable

multivariate LR: multiple correlated dependent variables are predicted

estimation methods:
1. linear least squares: ordinary LS (OLS)
                                  Weighted LS
                                  Generalized LS (GLS)
2. maximum-likelihood estimation: ridge regression
                                                  Lasso
                                                  Least absolute deviation (LAD): robust
3. other estimation techniques: principal component regression (PCR)
                                                                 # predictor is large / strong correlations among predictors
                                                                 2 stages: reduce predictors -- principal component analysis
.                                                                                 Use reduced predictors in OLS regression fit
                                                                 Extension: partial LS regression

为什么无法编辑了。。。。

ANOVA
Analyze differences among 3 or more group means in a sample
Categorical/qualitative predictors: factors -- have different levels
Parameters: effects -- fixed-effects models -- the only thing changes is mean among levels

Overall variance (y) = explained var (x) + unexplained var(e)
Somewhat robust to non-constant var
Errors: iid N(0, sigma2)

One-way anova: 1 factor, typically for at least 3 groups -- F test
                              (2 groups -- t test, F=t2)
Two-way anova: 2 factors

Central limit theorem: in some situations, when independent random variables are added, their properly normalized sum tends toward a normal distribution even if the original variables are not normally distributed.

谢谢！！！！！！

谢谢总结

楼主有心了！👍

nice 有心整理

doudou1222 发表于 2019-2-5 07:41
楼主有心了！👍

求加米哈哈