Feature Engineering

Feature engineering is turning raw data into inputs a model can learn from. It covers cleaning, transforming, encoding and combining the columns you already have.

It matters because models see numbers, not meaning. A raw timestamp tells a model very little. Day of week and hour of day, pulled out of that same timestamp, can tell it a great deal.

Think of it like this. Think of prepping ingredients before cooking. The recipe assumes chopped onions. Handing over an unpeeled one is technically the same ingredient and produces a much worse dinner.

How it works

• Clean. Handle missing values, fix types, remove duplicates.
• Transform. Scale numbers to comparable ranges, take logs of skewed values.
• Encode. Turn categories into numbers the model can use.
• Derive. Pull day of week out of a date, or a ratio out of two columns.
• Select. Drop features that add noise rather than signal.

Common mistakes

• "Deep learning made this obsolete." It reduced it on images and text. On tabular data, engineered features remain the difference between a good model and a mediocre one.
• "More features is better." Irrelevant features add noise and make overfitting easier.
• "Fit the scaler on everything first." That leaks test data into training. Fit on the training set only, then apply.
• "Any column is fair game." A column that would not exist at prediction time will make your model look brilliant and useless.