Attrition Classification
Implement any classifier using library functions to predict whether an employee will leave the company or not. Data: Test and Train data are given with 33 and 34 columns respectively and 1028 rows. The column vector consists of: Age, Attrition, BusinessTravel, DailyRate, Department, DistanceFromHome, Education, EducationField, EmployeeCount, EmployeeNumber, EnvironmentSatisfaction, Gender, HourlyRate, JobInvolvement, JobLevel, JobRole, JobSatisfaction, MaritalStatus, MonthlyIncome, MonthlyRate, NumCompaniesWorked, OverTime, PercentSalaryHike, PerformanceRating, RelationshipSatisfaction, StockOptionLevel, TotalWorkingYears, TrainingTimesLastYear, WorkLifeBalance, YearsAtCompany, YearsInCurrentRole, YearsSinceLastPromotion, YearsWithCurrManager, ID
Data Processing:
• Encoding
The first step taken was to encode the data such that we are left with numerical values only for the categorical data. Label Encoding converts categorical labels into numerical values. For the categorical data which have more than two unique values are encoded using One Hot Encoding. Features like, BusinessTravel, Department, OverTime, etc are converted based on their unique column values. One hot encoding is used to avoid any bias with large values assigned to data having more than two unique values.
• Scaling
The classification algorithms works best for similarly scaled features. Thus feautre scaling is apllied using MinMaxScaler to range values between 0 to 1.
• Removing Target and redundant features
The target feature : Attrition is removed from train data for model training. Also the features like- EmployeeNumber and ID which are unique for every row are removed as they will not help in any training. The feature called EmployeeCount which has same value all over is also removed.
Lastly the feautures like, MonthlyRate, HourlyRate and DailyRate seems correlated and removed except one for reducing irrelevant features. Thus reducing efforts on model training.
After above data processing, the columns for test and train data are, Age, DailyRate, DistanceFromHome, Education, EnvironmentSatisfaction, Gender, JobInvolvement, JobLevel, JobSatisfaction, MonthlyIncome, NumCompaniesWorked, OverTime, PercentSalaryHike, PerformanceRating, RelationshipSatisfaction, StockOptionLevel, TotalWorkingYears, TrainingTimesLastYear, WorkLifeBalance, YearsAtCompany, YearsInCurrentRole, YearsSinceLastPromotion, YearsWithCurrManager, BusinessTravel_Travel_Frequently, BusinessTravel_Travel_Rarely, Department_Research & Development, Department_Sales, EducationField_Life Sciences, EducationField_Marketing, EducationField_Medical, EducationField_Other, EducationField_Technical Degree, JobRole_Human Resources, JobRole_Laboratory Technician, JobRole_Manager, JobRole_Manufacturing Director, JobRole_Research Director, JobRole_Research Scientist, JobRole_Sales Executive, JobRole_Sales Representative, MaritalStatus_Married, MaritalStatus_Single
Train-Test data split & Cross validation
Using train_test_split helper function from scikit-learn, one can evaluate hyper paramaters for their classifier. However, there is a risk of overfitting on the test set because the hyper-parameters can be tuned for optimal values. By spliting the available data into two sets, the learning model can be deprived of the sufficient number of training samples. Also, a particluar choice of training and validation data may increase the accuracy of overall result but have poor predictability for the new data.
Cross validationIt helps to solve the above problem where training samples are split into k sets (k-folds) for validation. An average of the k folds can be taken as good measure of accuracy for that model.