Congratulations to the winners

Just trust your private LB. When I saw the public LB score, I was so determined that I ignored my private LB. I wish I could rely on my CV. I'm excited to see what the top five did to get their scores. Congratulations to the winners.

Discussion 6 answers

harisonmg

Congratulations to the winners. I'd love to hear how those with top scores handled their cross-validation and avoided overfitting as well as feature engineering techniques used.

25 Jul 2022, 10:26

Upvotes 0

J0NNY

Adama science and technology university

Congratulations to all. Thanks to my teammate @Koleshjr. It was really an amazing competition we learned a lot from it and thanks to @zindi for hosting such a competition.

Our approach is

Deal with missing values
Drop unnecessary columns
Drop outliers (rows with the same data but different categories)
Count encoding & new features
MERCHANT CLUSTERING to 10 categories
Aggregation and feature combination
Automatic clustering using KMeans from CountVectrozer
Decomposition of features
Oversampling using SMOTE
StratifiedKFold with 5 split
Voting classification and model stacking

25 Jul 2022, 12:09 (edited 5 minutes later)

Upvotes 4

harisonmg

Did you decompose all features or it's only the outputs from CountVectorizer? If you don't mind, please share the method as well

replied to J0NNY25 Jul 2022, 12:49 (edited 2 minutes later)

Upvotes 0

J0NNY

Adama science and technology university

No we decomposed the aggregate features which were formed from the MERCHANT_NAME. The outputs of CountVectrizer clustered into groups using KMeans clustering.

replied to harisonmg25 Jul 2022, 14:36

Upvotes 1