Coding Hygiene

• One data frame for each unit of analysis. E.g.
  • Customer
  • Transaction
  • Item
• One model for each considered dependent variable and features' set
  • Use the model that performs and suits the best
  • Leave only the model with the features' space that was finally used
• Use longer pipes to avoid partial data.frames

• It’s hard to track results from the end till the beginning - leave filter statement to the very late end of modeling
• Last filter statement is very useful as the introduction to the reporting stage

• Follow the How we name saved files at Gradient instructions
• Always store raw, main and final results
• For partial results try to store only those that require a lot of time to reproduce

• Always be able to determine the filters applied to the considered data.frame
• Keep track of the number of rows after filtering
• Keep track of the number of rows after applying a join operation
• Consider assertr and assertive packages to add verifications steps for above 2 points or use data verification functions in our g6tr package.

• Don’t merge branches that do not contribute to the final result
• Alone model is not always enough - apply coefficients in convenient way to read with tables and plots that can be used

• See Our Naming Conventions
• Use our standard libraries