Merkmale aus Datumsangaben und Text erstellen

A timestamp or a free-text note holds rich clues a model cannot read directly. Feature engineering pulls that signal into usable columns. 🔍

Before mining a date, convert the string with to_datetime. Only a real datetime exposes the handy parts you want to extract.

The dt accessor unlocks date parts on a datetime column. Reach for it to grab year, month, day, and more in one line.

Pull dayofweek to capture weekly rhythm, where Monday is 0 and Sunday is 6. Behavior often shifts a lot by weekday.

Turn the weekday into a simple boolean: is this a weekend? Such yes-or-no flags are easy for models to learn from.

Hour 23 sits right next to hour 0, yet the numbers look far apart. Cyclical encoding with sine and cosine fixes that wrap-around.

Start simple with text. The character length of a review or its word count is often a surprisingly strong feature.

Flag whether text holds a key term using str.contains. A column for words like refund can capture clear intent.

To use words themselves, count them. A bag-of-words turns text into one column per term holding how often it appears.

Plain counts overrate common words. TF-IDF rewards terms that are frequent in one row yet rare overall, so the distinctive words stand out.

Never build a feature from information unavailable at prediction time. A leaky column inflates scores in testing, then fails in the real world.

You have a datetime column and want the weekday number. What do you use?

You mined raw fields into signal: parse dates then use dt for parts and flags, and turn text into lengths, keyword flags, or word counts. 🎉