Classification use-case

[wbecker]

This project doesn't currently allow for the predicting the type of an input, as there is no sense of knowing to what type an input value maps.

Normally when using a classifier, there is a two stage process. 1 - fit(X, y), using training input and output data 2 - predict(X), using unknown data, and returning the estimated

It would be good if this project presented a similar interface.

I would suggest creating a class, wmd_classifier, which implements these two models.

fit, which would:

• take in an array of documents and break them down into bows
• create a WMD instance
• cache centroids

predict, which would:

• take in a document
• break it into a bow
• calculate its centroid
• call nearest_neighbours
• calculate the output type, based on the k nearest neighbours, weighted by their closeness

[wbecker]

I'd be happy to contribute something like this!

[vmarkovtsev]

@wbecker This is :+1: sklearn-like interface would be really useful. Feel free to PR.

My only suggestion is to abstract the way a document is transformed into nBOW. E.g. provide a function in __init__ and let the documents be "objects", with nice defaults for spacy/strings.

And let's name it WmdClassifier. I have just stated the contribution guidelines in https://github.com/src-d/wmd-relax/wiki/Contributions