Day #83 - Data Splitting Strategies

• Time based splits
• Validation to mimic train / test pic
• Time based trend - differs significantly, Time based patterns important

Different splitting strategies can differ significantly

• In generated features
• In a way model will rely on that features
• In Some kind of target leak

 Split Categories

•  Random Split (Split randomly by rows, Rows independent of each other), Row wise
• Device special features for dependency cases
• Timewise - Before particular date as training, After date as testing data. Useful features based on target
• Moving window validation
• By Id - (By Clustering pictures, grouping them and then finding features)
• Combined (Split date for each shop independently)

Summary

• In most cases split by Rownumber, Time, Id
• Logic for feature generation depends on data splitting strategy
• Set up your validation to mimic the train / test split of competition

Happy Learning and Coding!!!