Features are the individual properties of a knowledge component (KC) that determine the retrieval conditions of that KC, that is, when a student uses or thinks of a particular action or idea (e.g., divide both sides of an equation, pick 'a' vs. 'the' as an article).
If the KC is part of the knowledge that we want students to learn, and it makes sense to distinguish contexts where the KC should and should not be applied, then we can also distinguish relevant features of the KC from irrelevant features. Relevant features tend to be present in contexts where the KC should be applied and absent in contexts where the KC should not be applied. Irrelevant features tend to be absent in contexts where the KC should be applied and/or present in contexts where the KC should not be applied. A knowledge component that has just relevant features and no irrelevant features has high feature validity.
Sometimes features are relatively directly perceivable (seen or heard). In the language literature, such features are called cues. Sometimes the relevant features of a knowledge component require more complex inference to be detected by the student. For example, Chi, Feltovich, and Glaser (1981) distinguish between shallow features of physics problems, like pulley system or inclined plane, that are irrelevant to correct problem solving (i.e., KC application) and deep features, like conservation of energy, that are relevant to accessing correct knowledge components.
A number of projects provide some good examples of KC feature analysis including Julie Booth's in Algebra and Amy Ogan's in French. In both, much of the instructional design is focused on helping students to learn the relevant deep features (e.g., a term includes a number and its sign, positive or negative) and distinguish them from irrelevant shallow features (e.g., a number without it's sign).