Suggestions

[WhiteChair]

next steps

• [ ] Adding more variables.
• [x] Up sampling! struggeled to make this work myself (caret package)
• [ ] Try to make xgboost work?

[WhiteChair]

I think we can improve the score by:

• Using a different imputation method.
• Adding more variables.
• Looking at the correlation between the features. I did not do that because I was lazy and corrplot was a bit of a nightmare with 40 variables.
• Trying again to make Xgboost work
• Maybe outliers are present and are not being taken care off? aka increase robustness. (I hope daniel can look at this since he took the robust course :) )
• Up sampling! struggeled to make this work myself (caret package).

[WhiteChair]

Suggestion for new variables, calculate ratio's: phonecal/sms offnet / onnet calls

cathegorical:
weekend user or week user? see if week use > weekend use. social, are ONNET >= OFNET connections? (communication to people with the same provider)