MeanThreshold, MedianThreshold, and other threshold support in GenericUnivariateSelect

[charlesbmi]

Describe the workflow you want to enable

I would like to select features by thresholding their mean value (i.e., mean-across-samples), similar to how VarianceThreshold selects features by thresholding their variance-across-samples.

Describe your proposed solution

Two possible options:

• Implement a new Estimator in sklearn.feature_selection, similar to VarianceThreshold. Example: https://github.com/hermidalc/sklearn-extensions/blob/f9296d0f3ed5d71b7f07779b47d8cf71bbcfa51b/feature_selection/_average_threshold.py#L7-L96
• Add a mode='threshold' option to GenericUnivariateSelect (https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.GenericUnivariateSelect.html)
  • This would allow the user to pass in more flexible score_funcs as well

Describe alternatives you've considered, if relevant

Another alternative, although this seems counter to how these functions are designed

• SelectFromModel(DummyRegressor(strategy='mean'), importance_getter='constant_', threshold=min_mean_value)
  • https://scikit-learn.org/stable/modules/generated/sklearn.feature_selection.SelectFromModel.html
  • https://scikit-learn.org/stable/modules/generated/sklearn.dummy.DummyRegressor.html

Additional context

Setting a MeanThreshold would be useful when working with non-negative features, such as pixel intensity in images. For example, we might want to exclude pixels that are regularly saturated in our dataset, as they may be less informative.

Specifically, in my research field of neuroscience (single-neuron recordings), our "features" are the (non-negative) action-potential-counts for each neuron. We often exclude neurons with very-low-firing-rates to minimize discretization error. Here are a few examples of neuroscience papers that set a MeanThreshold per neuron (i.e., feature):

• https://www.cell.com/neuron/pdfExtended/S0896-6273(17)30592-5
  • Units with mean firing rates less than 1.5 Hz were excluded from the analysis.
• https://www.nature.com/articles/s41586-021-04042-9
  • Neurons with firing rates less than 0.5 Hz were excluded.
• https://www.nature.com/articles/s41598-018-37227-w
  • Firing data were Gaussian smoothed and binned in 0.1 s periods and bins with firing rates less than 0.1 Hz (no spikes) were excluded.

[DuarteSJ]

Is anyone working on this issue? Can I?

[adrinjalali]

I think improving GenericUnivariateSelect sounds like the best path forward here.

[glemaitre]

Is this PR solving the issue: #27722