How do you handle missing values in Pandas without intr...

The rule that governs everything else here is: whatever number you use to fill a gap, it must have been learned from training data only. In practice that means checking df.isnull().mean() to see how much is missing and whether the pattern looks random or informative — and adding a binary “was this value missing” column before you impute, since the fact of missingness can itself carry signal.

For the imputation itself, use sklearn’s SimpleImputer: call .fit() on the training split so it learns the median or mode from training data only, then call .transform() on both train and test using those learned values. The mistake that quietly inflates your evaluation metrics is computing df['age'].fillna(df['age'].mean()) on the whole dataset before splitting, or fitting an imputer on the full data — both let test-set statistics leak into what the model effectively “knew” during training, so your offline scores look better than what you’ll actually see in production.