How Data Preprocessing Works

Real-world data is never clean. Missing values leave gaps, scale differences make features incomparable, and categorical text can't be fed directly to most algorithms. Data preprocessing fixes these problems before training begins.

Dropping rows removes any row with missing values. Simple but wasteful — you lose all other data in that row. Works when few rows are affected.

Without scaling, Income (0-1000k) dominates Score (0-1). The model treats the bigger numbers as more important — even when they shouldn't be.

One-hot encoding creates a binary column for each category. No fake ordering is imposed — Red is not "greater than" Blue. Use this for nominal (unordered) categories.

Training data teaches the model. Validation data tunes hyperparameters. Test data gives the final, unbiased score. If the model ever sees test data during training, your evaluation is meaningless — the model simply memorizes answers.

A preprocessing pipeline chains these steps in order: handle missing values first (so scaling doesn't break), then scale, then encode, and split last so test data stays untouched. In practice, use tools like scikit-learn's Pipeline to keep this reproducible.