Struggling with normalization/Standardisation for machine learning dataset

Question

Sorry for what is probably a very obvious/rookie question. I'm currently doing a data science module for my degree and making very slow progress with the work.

The case study i'm doing is around HR data for a fictitious organisation to measure the impact that various attributes (Age, Employee satisfaction and Salary) on the employee performance score.

In particular, I don't massively understand standardisation/normalisation and when it should and shouldn't be done.

I need to use the dataset (See attached for sample) to create machine learning models for both Random Forest and Linear Regression. I'll be using R for the machine learning element.

What I have currently done is:

• Pseudonymized the employee ID
• Used R to express an employee date of birth as an age
• Used one-hot encoding to represent the 'Department' and 'Sex' fields as a binary value
• Used label encoding for the 'Performancescore' attribute which I believe is an ordinal relationship and changed the values from ('Exceeds', Fully Meets, Needs improvement and PIP) to 3,2,1,0 respectively.

The part i'm struggling with is: Do the Age and Salary columns need to be Standardised/Normalised? Will they work for Random Forest and Linear Regression models or do I need to do anything else to the data? I've read a lot of conflicting things online and don't have a lot of confidence with the module so any advice would be greatly appreciated :)

Answer 1

Random Forest doesn’t care about the scale of your {age, salary} features, as it will choose appropriate breakpoints no matter what scale they’re on. Might as well feed it Z scores anyway, assuming it’s easy enough to transform them back to the original scale.

Linear Regression, OTOH, is likely to care. Make the features Z scores, zero mean with unit variance, to make them comparable.

This being a learning exercise, take a moment to scale up one individual feature or another by 1e6, and by 1e-6, then score the resulting models to see how they perform. Some modeling techniques are insensitive to affine transformations, while for others, not so much.