在项目的这一部分，我们首先将数据导入阿里云MaxCompute平台，然后我依托于阿里云MaxCompute平台及PAI平台，对人力资源流动数据进行数据清洗和数据探查。
在MaxCompute中建表代码如下：
DROP TABLE IF EXISTS hr_analytics ;
CREATE TABLE IF NOT EXISTS hr_analytics
(
satisfaction_level DOUBLE
,last_evaluation DOUBLE
,number_project BIGINT
,average_montly_hours BIGINT
,time_spend_company BIGINT
,Work_accident BIGINT
,left1 BIGINT
,promotion_last_5years BIGINT
,position STRING
,salary_level STRING
)
;
然后，进行数据的缺失值处理，检查各个变量的空值数量：
SELECT sum(CASE WHEN satisfaction_level IS NULL THEN 1 ELSE 0 END) AS satisfaction_level
,sum(CASE WHEN last_evaluation IS NULL THEN 1 ELSE 0 END) AS last_evaluation
,sum(CASE WHEN number_project IS NULL THEN 1 ELSE 0 END) AS number_project
,sum(CASE WHEN average_montly_hours IS NULL THEN 1 ELSE 0 END) AS average_montly_hours
,sum(CASE WHEN time_spend_company IS NULL THEN 1 ELSE 0 END) AS time_spend_company
,sum(CASE WHEN work_accident IS NULL THEN 1 ELSE 0 END) AS work_accident
,sum(CASE WHEN left1 IS NULL THEN 1 ELSE 0 END) AS left1
,sum(CASE WHEN promotion_last_5years IS NULL THEN 1 ELSE 0 END) AS promotion_last_5years
,sum(CASE WHEN position IS NULL THEN 1 ELSE 0 END) AS position
,sum(CASE WHEN salary_level IS NULL THEN 1 ELSE 0 END) AS salary_level
FROM hr_analytics
;
检查发现，各项缺失值均为0，即无缺失值存在。
再检查数据的离群值。计算满意度，最近一次评价，参与项目数量，平均工作时间，在职时间，工伤次数的均值及标准差：
SELECT
AVG(satisfaction_level ) AS avg_satisfaction_level
,STDDEV(satisfaction_level ) AS std_satisfaction_level
,AVG(last_evaluation ) AS avg_last_evaluation
,STDDEV(last_evaluation ) AS std_last_evaluation
,AVG(number_project ) AS avg_number_project
,STDDEV(number_project ) AS std_number_project
,AVG(average_montly_hours)AS avg_average_montly_hours
,STDDEV(average_montly_hours ) AS std_average_montly_hours
,AVG(time_spend_company )AS avg_time_spend_company
,STDDEV(time_spend_company ) AS std_time_spend_company
,AVG(work_accident)AS avg_work_accident
,STDDEV(work_accident ) AS std_work_accident
FROM hr_analytics
;
然后利用3σ准则计算离群区间：(μ-3σ,μ+3σ)，发现只有time_spend_company 含有离群值，离群值个数为376，离群率为0.4%。根据实际情况，公司员工在职时间的离群数量不多。可能因为突发情况，导致的在时间过长，本实验的目的为预测离职情况，而在职时间过长或过短更有可能影响离职，故不考虑剔除或替换。
然后利用PAI平台进行线性关系检测，得到决定系数R2为0.01，极为接近0，拟合优度很低。再绘制残差图，得到图形如下：
如图所示，图内数值随机分布 说明数据是线性的，即满意度和最后一次评价间有线性关系。
