✨ Introduce BalancedBagging #2
Conversation
|
Check out this pull request on See visual diffs & provide feedback on Jupyter Notebooks. Powered by ReviewNB |
.github/workflows/CI.yml
Outdated
| - '1.8' | ||
| - 'nightly' | ||
| os: |
There was a problem hiding this comment.
Use 1 and not 1.0 to get the latest stable release. And include 1.6 as this is the Long Term Stable release.
There was a problem hiding this comment.
@ablaom, for some reason Imbalance had julia = "1.8" when we released it. I changed it to 1.6 but we won't be able to test on 1.6 so long as Imbalance is a dependency or the next release isn't released yet. For now will keep 1.8 and 1 until Imbalance's next release.
src/BalancedBagging.jl
Outdated
| """ | ||
| function group_inds(y::AbstractVector{T}) where {T} | ||
| result = LittleDict{T,AbstractVector{Int}}() | ||
| freeze(result) |
There was a problem hiding this comment.
freeze is not mutating. So this line achieves nothing.
There was a problem hiding this comment.
I assume you mean the first line is not mutating? so freeze achieves nothing?
There was a problem hiding this comment.
No, I mean that you do not use the output of line 8 (you do not assign it to any variable used later) so this line might as well not exist. My impression was you thought that freeze(result) mutated result but it does not. It just returns a new immutable dictionary with faster look-up and smaller allocation than result.
src/BalancedBagging.jl
Outdated
| end | ||
|
|
||
| """ | ||
| $(MMI.doc_header(BalancedBaggingClassifier)) |
There was a problem hiding this comment.
If you are going to explicitly require specification of a hyperparameter, because it has not default (in this case model),then you cannot use the doc_header shortcut, because it says that the zero-argument constructor works: https://alan-turing-institute.github.io/MLJ.jl/dev/adding_models_for_general_use/#MLJModelInterface.doc_header
So you will have to customise this part. Ditto the other wrapper.
There was a problem hiding this comment.
Solved for this wrapper. The other wrapper doesn't have an MLJ-compliant docstring, only the quick start in the README and basic documentation. Would you like me to add one for it?
There was a problem hiding this comment.
No, wrappers don't need to have compliant doc-strings.
src/BalancedBagging.jl
Outdated
| - `X`: any table of input features (e.g., a `DataFrame`) so long as elements in each column | ||
| are subtypes of either the `Finite` or `Infinite` scientific types. | ||
|
|
||
| - `y`: the binary target, which can be any `AbstractVector` where `length(unique(y)) == 2` |
There was a problem hiding this comment.
Again, I don't see this needs to be binary, but again this depends on the atomic classifier being wrapped.
There was a problem hiding this comment.
@ablaom I throw an error when y is not binary. The resampling scheme proposed by the paper assumes binary data.
In this paper, we examine only binary classification problems.
It's not trivial to generalize but I may look into that later.
src/BalancedBagging.jl
Outdated
| - `predict(mach, Xnew)`: return predictions of the target given | ||
| features `Xnew` having the same scitype as `X` above. Predictions | ||
| are probabilistic, but uncalibrated. |
There was a problem hiding this comment.
| - `predict(mach, Xnew)`: return predictions of the target given | |
| features `Xnew` having the same scitype as `X` above. Predictions | |
| are probabilistic, but uncalibrated. | |
| - `predict(mach, Xnew)`: return predictions of the target given | |
| features `Xnew` having the same scitype as `X` above. Predictions | |
| have the same form as that of the classifier, `model` |
There was a problem hiding this comment.
@ablaom model has to be probabilistic. I thought we agreed in the meeting to write that they are uncalibrated.
src/BalancedBagging.jl
Outdated
| - `predict_mode(mach, Xnew)`: instead return the mode of each | ||
| prediction above. |
There was a problem hiding this comment.
| - `predict_mode(mach, Xnew)`: instead return the mode of each | |
| prediction above. | |
| - `predict_mode(mach, Xnew)`: if `model` is a `Probabilistic` model, return the mode of each | |
| prediction above. |
There was a problem hiding this comment.
@ablaom, I constraint that the model has to be probabilistic so I will ignore "if model is a Probabilistic model".
Co-authored-by: Anthony Blaom, PhD <anthony.blaom@gmail.com>
…cing into BalancedBagging
|
Tests aren't failing, Github just reached it's limit with me for today. |
|
Is this a daily limit, then? It seems we don't get a reset until October 1st, and this effects all JuliaAI private repos. I think you can go public to resolve early. |
Good point. Let me make is public immediately. |
Description
This PR adds BalancedModel which allows parallel composition of random undersampling models followed by a probabilistic classification model. In other words, it's the EasyEnsemble algorithm but for a generic probabilistic classifier rather than Adaboost.
@ablaom could you please review the additions when you get the chance.