A feature, or signal, is an input variable or measurable attribute supplied to a machine learning model. Raw data often needs preprocessing to create consistent, useful, informative inputs before modeling. section 7.2
A feature engineering plan and pipeline can address data issues and make useful information accessible to a model:
- Check for duplicated information across fields, distribution skews, differences of scale, and information spread across hierarchies. section 7.2
- Encode relevant human insight so the model can use relationships that raw values may obscure. section 7.2
Calendar dates
A model may not infer that January 1 follows December 31 from the numbers 1 and 12 alone; feature engineering can represent that relationship. section 7.2
Figure 7.1 Temperatures plotted by day of the year (data for Suffolk Wattlesham weather station, 2019, from UK Government).
Compare temperatures plotted by day of the year with the alternative view plotted by distance from midsummer. section 7.2
