The core rule

A model can only learn from the information you feed it.

If your data is:

• missing values
• inconsistent formats
• skewed scales
• messy categories
• leaking target information

…the model will learn the wrong thing.

The “modeling is the easy part” reality

In practice, time spent is often:

• Data cleaning + feature engineering: ~60–80%
• Modeling: ~10–20%
• Deployment + monitoring: ~10–20%

Common failures caused by bad preprocessing

1) Data leakage

You accidentally include information that wouldn’t be available at prediction time.

Example:

• predicting customer churn
• but you include “days_since_last_purchase” computed after churn happens

Symptom: extremely high validation scores that collapse in production.

  flowchart LR
  A[Raw Dataset] --> B[Leakage Feature]
  B --> C[Great CV Score]
  C --> D[Bad Production]

2) Train/test contamination

You fit preprocessing on the full dataset (train + test) instead of only train.

Example: scaling with mean/std from all data.

3) Incorrect handling of categories

If category values appear in test data that weren’t in training, encoding can break.

4) Different distributions (dataset shift)

Your training data doesn’t match reality.

• prices change
• user behavior changes
• sensors drift

The correct pattern: pipelines

The safest approach is to use a pipeline that:

• fits preprocessing only on training data
• applies the same transform to validation/test

In scikit-learn, that means:

• PipelinePipeline
• ColumnTransformerColumnTransformer

Mini-checkpoint

Before you train a model, answer:

• What values are missing? Why?
• Do any columns “peek into the future”?
• Are there categories that will grow over time?
• Which features have wildly different scales?