Classification for Machine Learning purposes¶
The classification criteria explained in this page complement the more general ones introduced here.
On many occasions, terms "Features", "Fingerprints", "Targets" are used for materials informatics purposes. For example, when constructing a machine learning model, a dataset containing information about multiple materials is used in order to find regular patterns and inter-dependencies. Such dataset is usually split into properties that represent the known data, or features, and the properties to be predicted, or targets.
We clarify this terminology as follows.
- Feature: any property of a material, eg. density or electronic band gap.
- Fingerprint: property of a material used as an input for a (statistical modeling) Workflow, equivalent to the concept of Descriptive property by definition
NOTE: by default, we only store the minimal amount of information required to identify a material enough to reveal a set of its Fingerprints. Such minimal set of properties is called Identity Fingerprints, and the rest - Derived Fingerprints.
- Target: property of a material obtained as output of a (statistical modeling) Workflow, equivalent to the notion of Characteristic property by definition.
Thus, a property-descriptive-characteristic triad is equivalent to feature-fingerprint-target.