Misusing identifiers, incorrect information splits, and ignoring uncommon characteristic values
One tremendously fulfilling factor about having been concerned within the area of machine studying for so long as I’ve is the chance to at all times be taught one thing new. That one thing new can both be a brand new instrument or methodology (given the fast improvement within the machine studying panorama, there’s by no means a scarcity of that), however it will also be the invention of inaccurate processes in our work that we merely had by no means been conscious of.
A few of these will be fairly obscure and arduous to identify at first look. If these inaccurate processes do slip into your mannequin improvement, there’s an excellent probability it would damage its predictive energy and thus its reliability, and, in the end, its applicability.
On this article, which is the start of a collection exploring widespread pitfalls in machine studying, we’ll concentrate on three information dealing with errors that may happen each throughout the preprocessing part but additionally throughout the modeling part:
- Utilizing Numerical Identifiers as Options
- Random Partitioning As an alternative of Group Partitioning
- Together with Characteristic Values with Inadequate Observations