Data Preprocessing

What you'll see here

Every cleanup, encoding, and split decision — with the business rationale and the impact on the downstream model.

Rows in

Rows out

No row drops

Features in

Raw columns

Features used

ID + ZIPCode + target excluded

Preprocessing pipeline

Step	Action taken	Reason	Business / ML impact
1. Duplicate check	Detected 0 duplicate IDs.	Customer ID must be unique for honest evaluation.	None — dataset is clean.
2. Missing value check	0 missing values across all columns.	Tree models tolerate some missingness, but cleanliness is preferred for reproducibility.	No imputation needed.
3. Anomalous value treatment · Experience	0 negative Experience values → replaced with 0 in experience_clean.	Negative professional experience is not logically valid. Replacing with 0 preserves rows.	Removes a source of spurious splits; keeps sample size intact.
4. Drop identifier	Drop ID before modeling.	ID has no predictive signal — including it risks overfitting to row order.	Cleaner feature space.
5. Drop ZIPCode (baseline)	Exclude ZIPCode from baseline features.	High-cardinality geographic identifier. Treated as categorical; encoding it would explode the feature space and likely overfit.	Simpler, more generalizable baseline tree. Region engineering left as future work.
6. Binary columns kept as-is	Securities, CD, Online, CreditCard stay as 0/1.	Already model-ready; one-hot encoding binary columns is wasteful.	No feature inflation.
7. Education & Family	Kept as ordered integers (Education 1–3, Family 1–4).	Decision trees split on thresholds — ordinal encoding preserves order and stays interpretable.	Cleaner splits like `Education > 1.5`.
8. Feature engineering	Added mortgage_flag, high_income_flag, high_ccavg_flag, income_band, ccavg_band.	Surfaces business-meaningful thresholds for both EDA and recommendation tiers.	Strengthens analysis and aligns model with how marketing thinks about segments.
9. Stratified train/test split	—	Target class is ~9.6% positives — random splits could starve the test set of positives.	Stable, comparable metrics across hyperparameter sweeps.

Final modeling feature set

Age

Experience (cleaned)

Income

Family

CCAvg

Education

Mortgage

Securities

CD Account

Online

CreditCard

AI: Preprocessing quality review

Click Generate to produce an evidence-based observation, insight, recommendation, and business benefit for this section.