knn classifier accepts various data structures

[wildart]

Classifier can accept data structures form NearestNeighbors package.

[wildart]

I propose two new abstract types Classifier and Regressor to identify majority vote and averaging approach to results of particular prediction.

abstract Classifier
immutable kNNClassifier{T <: NearestNeighborTree} <: Classifier
    t::T
    y::Vector
end

Surely, Classifier must be interface that would define a method specification which handles results of a particular model prediction through majority vote or as-is. Same approach would apply to Regressor with averaging.

[johnmyleswhite]

The community's spent a lot of time discussing how to define things like Regressor in the past: https://github.com/JuliaStats/Roadmap.jl/issues/4

I'm not sure it's the most fruitful thing to do at the moment.

[johnmyleswhite]

Coming back to this, perhaps we should introduce even finer type distinctions? One might want to use majority voting with a parametric type that stores the predictions in a vector of the same type as the input labels (which could be arbitrary types). In contrast, the averaging/probability case seems to call for a very different data structure as output.

[wildart]

I agree with parametric type for labels in classification case. As for regression case it would always be numerical value even though implementations could output auxiliary data.