DOC improve documentation of nearmiss #1028
Conversation
|
|
||
| >>> from imblearn.under_sampling import NearMiss | ||
| >>> nm1 = NearMiss(version=1) | ||
| >>> X_resampled_nm1, y_resampled = nm1.fit_resample(X, y) | ||
| >>> print(sorted(Counter(y_resampled).items())) | ||
| [(0, 64), (1, 64), (2, 64)] | ||
|
|
||
| As later stated in the next section, :class:`NearMiss` heuristic rules are |
There was a problem hiding this comment.
I think that describing parameters without having described the logic breaks the flow and doesn't really help the user understand what the parameters do.
I suggest removing this paragraph altogether and expanding on the meaning of the parameters (if necessary) on the docstrings.
| is the largest. | ||
| **NearMiss-3** is a 2-steps algorithm: | ||
|
|
||
| First, for each negative sample, that is, for each observation of the minority class, |
There was a problem hiding this comment.
I refer back to target class and minority because positive and negative are a bit confusing, and at this stage, the reader may have forgotten which is which.
| previous toy example. It can be seen that the decision functions obtained in | ||
| each case are different. | ||
|
|
||
| When under-sampling a specific class, NearMiss-1 can be altered by the presence |
There was a problem hiding this comment.
This is the paragraph that I don't understand.
Are we saying that Nearmiss-1 is sensitive to noise, because the samples from the majority closer to the minority are noise?
That is what I understood and rephrased it. Correct me if I am wrong.
But besides that, is it always the case that observations from the majority closer to the minority are noise? I get it that they will be the hardest to classify, but that does not necessarily make them noise. Noise, as I understand it, is some sort of random variation. Or in this case, it would be samples from the majority class that are not necessarily representative of the majority, or in other words sort of outliers. But all of these, is not in the text. The text as it stands right now just says that if they are close, they are noise, and I don't think that is correct.
Thoughts?
There was a problem hiding this comment.
A noisy observation would be something outside of the sample distribution. So I agree that it does not have anything to do with the decision boundary. Those are just harder to classify if the data overlap but they are not noise.
|
|
||
| NearMiss-3 is probably the least sensitive version to noise due to the first sample | ||
| selection step. | ||
|
|
There was a problem hiding this comment.
More thoughts on NearMiss:
These is just my opinion, I would be happy to be challenged on this one.
To me, it feels that the authors almost used some random logic to select the samples: those with the smallest average distance, those with the largest and then something else in version 3. In their article, they don't really describe how they came to this conclusion, so it seems arbitrary to me. Given this, I am not sure it makes sense to discuss anything further on this user guide. I would remove all the discussion about noise in fact.
In addition, this method was designed and tested on text data, something similar to bag of words / tokens, which is a very particular format. And I am not aware that it has been tested / used somewhere else? So overall, I am not a great fan of this method, but that aside, should we not add a disclaimer somewhere saying that it was designed for text, so caution when implementing in more traditional datasets?
There was a problem hiding this comment.
Actually we don't even implement the right things here :)
#980
|
@glemaitre I am done with the initial changes, but I think it would be worth to discuss this a bit. Thank you! |
related to #853
Separating in smaller PRs as requested.
I had some trouble understanding the documentation for this class, so I hope that when I re-worded it, I captured the message that wanted to be given in the original version. More comments in the code.