I am looking for a good pedagogical example use of feature selection for model building. The purpose is to expose students to some very basic methods for feature selection, in the context of boolean classification, as well as the notion of model building -- and to apply it to a real-world data set that someone might care about.
Let me distinguish prediction vs model building. One use of feature selection is for prediction, where the sole goal is to classify (predict the class of instances) in whatever way we can, and we don't care about the model itself; feature selection can be helpful for that by allowing one to focus on relevant attributes. In contrast, in model building, we care about the model as a object of interest in its own right. For instance, perhaps there is some scientific meaning that can be derived from the results of feature selection: maybe knowing that some features are relevant and others aren't might help build a scientific model. I want to teach feature selection and also give an example where they see classification used not just for prediction but also for model building.
Students will have been previously exposed to boolean classification and the $k$-nearest neighbors classifier, but not other more sophisticated classification methods.
I would like to find a concrete example application I can use for teaching feature selection, and that highlights model building as a motivation. Ideally, the application would have a few characteristics:
Involves a boolean classification task.
One where the classification and model building task is well-motivated and potentially of interest to students (e.g., perhaps from some scientific or real-world application), and where it's easy to explain why we might be interested in model building.
There is a data set available, so students can play with it and be exposed to working with real data and see an example of what one can learn from feature selection.
An application where feature selection is effective (i.e., some features are highly relevant and improve prediction, and others are irrelevant)
Ideally, there exists a clean data set without too many complications (e.g., no missing feature values) and of modest size (e.g., at most tens or hundreds of features), and suitable for use with a $k$-nearest neighbors classifier.
Is there an example application that might be suitable for these purposes?