Please provide legacy support for FeatureDatum input Tuples

[Schmed]

There's legacy code out there that expects to call the classify() method with FeatureDatum Tuples, but this method now supports only TermsDatum Tuples. Why not have a legacy getFeatures() method that takes a term map used by a legacy classify() method that takes a TermsDatum, etc.?

[kkrugler]

So this getFeatures() method would be part of what class?

[Schmed]

I'm referring to the following method within com.scaleunlimited.classify.model.HashedFeaturesLibLinearModel.java:

private Feature[] getFeatures(Map<String, Integer> terms)

I'm suggesting that a second method in the same class be written with the old signature:

private Feature[] getOldFeatures(Map<String, Double> terms)

[kkrugler]

HashedFeaturesLibLinearModel extends BaseLibLinearModel, which extends BaseModel<TermsDatum>. So this means the API for addTrainingTerms(TermsDatum) and classify(TermsDatum) have to stay as such. I think you're proposing adding alternative versions to those two methods in BaseLibLinearModel, which take a FeatureDatum argument.

The tricky part is that BaseLibLinearModel maintains the list of TermsDatum documents for training. So I'd either need to have a parallel array (and verify you were only using one or the other), or I could convert the FeatureDatum to a TermsDatum by say multiplying each feature weight by some large fixed constant. That would impose a constraint on the max weight you could provide, but probably OK. Thoughts?

[kkrugler]

Or I guess I could convert TermDatum to FeatureDatum in addTrainingTerms, and store that in the list, and bail on ever supporting normalization that works across all training documents. Which is probably OK, as I already pulled out the TfIdfNormalizer.