Predictive model for rare occurrences

Casper72Casper72 Member Posts: 17 Contributor II
Hi fellow RapidMiners,

What kind of model would you suggest I should look into when trying to predict a binary outcome with a very high class imbalance (97/3)? The problem at hand is medical readmission within 30 days for surgery. Any suggestions would be appreciated. Currently I am planning to test the k-NN algorithm looping through different k-values.

Best regards.

Best Answers

  • lionelderkrikorlionelderkrikor Moderator, RapidMiner Certified Analyst, Member Posts: 1,195 Unicorn
    Solution Accepted
    Hi @Casper72,

    In your case, I advice you to preprocess your data by upsampling your dataset before modelling.
    For that, you can use the SMOTE Upsampling operator from the Operator Toolbox extension available for free in the MarketPlace.

    Regards,


    Lionel

  • Telcontar120Telcontar120 Moderator, RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 1,635 Unicorn
    Solution Accepted
    You can try weighting instead to balance the classes (although not all ML algorithms support weighting).  This might give better results than upsampling with such a small minority class.
    Brian T.
    Lindon Ventures 
    Data Science Consulting from Certified RapidMiner Experts
  • Telcontar120Telcontar120 Moderator, RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 1,635 Unicorn
    Solution Accepted
    Take a look at the tutorial for the Generate Weight (Stratification) operator, that should be the one that you would use.
    Brian T.
    Lindon Ventures 
    Data Science Consulting from Certified RapidMiner Experts

Answers

  • Casper72Casper72 Member Posts: 17 Contributor II
    Thank you Lionel,

    I will try using SMOTE. Have used it before with success, although with less imbalanced datasets (typically in the range of 30/70) 
  • Casper72Casper72 Member Posts: 17 Contributor II
    Tellcontar120: Great idea! I will have to read up upon weighting in RM though. Thank you for your suggestion.
Sign In or Register to comment.