[WIP] Instance hardness

[dvro]

Added Instance Hardness Threshold under sampling method.

• unbalanced_dataset/under_sampling/instance_hardness_threshold.py
• example/under-sampling/plot_instance_hardness_threshold.py

The higher the threshold, the more samples of the majority class are removed.

[glemaitre]

I saw some discrepancies with PEP8 standard. Could you make a check for that.

[glemaitre]

I would rename mode by kind_sel. We used this keyword in the other part of the API for almost similar. I think this is in the ENN

[dvro]

done.

[glemaitre]

yep you can push your changes

[glemaitre]

You have some blanks lines. You should probably remove them

[glemaitre]

@dvro Do you think that we could replace the parameter threshold by the parameter ratio. In fact, ratio could determine the probability to apply in order to obtain the desirable ratio between minority and majority classes.

In this case I would also remove the possibility to remove some instances from the minority class.

[fmfn]

Do you think that we could replace the parameter threshold by the parameter ratio.

I came here to say that, can we make the API consistent? ratio is the choice so far, would it make sense to re-name threshold?

[glemaitre]

@fmfn I agree but I would even enforce the same behaviour for this parameter: number of minority samples over number of majority samples.

[fmfn]

@glemaitre Yes! Much better indeed.

[dvro]

I can rename it to ration, but imo ratio = #maj / #min (or the other way around); when I used threshold it means the probability threshold in which samples are removed. A user might think that, in setting ratio=0.5, the output would be 2*X samples of class A and X samples of class B ... when that is not the case. (I don't know if I am being clear about this).

What you guys think?

[fmfn]

I agree tha simply renaming it would cause confusion. I suppose what I mean was: is that a transformation that can be done to threshold though would allow us to translate it to a ratio?

[glemaitre]

For this case I would all the probability level such that the ratio is nearest of what we want. Roughly without trying it:

ratios = np.zeros(100, ) probs = np.linspace(0., 1., 100) for i, p in enumerate(probs): ratios[i] = self.stats_c_[self.min_c_] / np.count_nonzero(np.logical_or(probabilities >= p, y == self.min_c_)) ratios = np.abs(ratios - self.ratio) threshold = probs[ratios.argmin()]