Kaggle Santander

Nothing yet.

• Cross-Validation issue and Public LB unreliability: https://www.kaggle.com/c/santander-customer-satisfaction/forums/t/19946/who-believe-in-standard-deviation
• Game of randomness: https://www.kaggle.com/c/santander-customer-satisfaction/forums/t/20299/does-it-make-sense-to-loop-through-seed-to-get-best
• Example of feature selection: https://www.kaggle.com/kobakhit/santander-customer-satisfaction/0-84-score-with-36-features-only/discussion
• Public LB unrepresentative of the Private LB: https://www.kaggle.com/c/santander-customer-satisfaction/forums/t/20226/best-way-to-the-top-of-the-leader-board
• Duplicated Rows (will add more myself about the repo): https://www.kaggle.com/c/santander-customer-satisfaction/forums/t/20276/duplicated-rows
• Data Dictionary: https://www.kaggle.com/c/santander-customer-satisfaction/forums/t/19291/data-dictionary

Major issues:

• There are many duplicated rows
• Many variables are linear combinations of other variables: any machine learning algorithm requiring non ill-conditioned matrix will fail hard (ex: Linear Discriminant Analysis)
• There are many outliers: 99999999 values, etc.
• There are variables that are constant
• There are duplicates where the target is different. I'll compile a list later