Skip to content

Latest commit

 

History

History
106 lines (83 loc) · 4.13 KB

README.md

File metadata and controls

106 lines (83 loc) · 4.13 KB
Raw

feature-specific-contribution

Different methods to compute the specific contribution of a groups of features X = [A, B, C] to predict a target Y.

Objective

We aim at computing the specific contribution of different groups of features contained in X (ex: X = [A, B, C], with A, B, C having different dimensionality) to predict each dimension of a target Y.

(Note that would like to keep track of the contribution for each dimension of Y).

We use the three methods below. All three methods are coded in the src/get_contrib.py file, and are called by the run_experiments() function contained in src/run_experiments.py. The three functions take as inputs: X, y, hiearchy (e.g [0, 0, 0, 1, 1, 2], with 0 (resp. 1 and 2) refering to the indices of A (resp. B and C) in X), the predictive scikit-learn model, and the three parameters scale (whether to scale per group), n_repeats (for permutation importance) and pca (number of components if we apply PCA to each group before fitting).

Method 1: "permute_current"

Function contrib_permute_current() in get_contrib.py

  • (optional) Scale each group of features separately
  • (optional) PCA on each group
  • orthogonalise
  • concatenate features
  • fit on train
  • compute permutation importance, for each feature group separately, on test
  • re-iterate on folds

Method 2: "permute_below"

Function contrib_permute_below() in get_contrib.py

  • (optional) Scale each group of features separately
  • (optional) PCA on each group
  • concatenate features
  • fit on train
  • compute permutation importance, for (+ refers to concatenation)
    • the first group A: importance[A] = r - r_with_A_shuffled
    • the first + second group: importance(A+B) = r - r_with_A_and_B_shuffled
    • the first + second + third group: importance(A+B+C) = r - r_with_A_and_B_and_C_shuffled
    • ...
    • all groups shuffled
  • extract specific contribution for each level
    • contrib_A = importance(A)
    • contrib_B = importance(A+B) - importance of first group (A)
    • ...
  • re-iterate on folds

Method 3: "concat"

Function contrib_concat() in get_contrib.py

  • (optional) Scale each group of features separately
  • (optional) PCA on each group
  • concatenate features
  • in order to predict Y, fit in hiearchical order and compute scores of: (+ refers to concatenation)
    • the first group A: r(A)
    • the first + second group: r(A+B)
    • ...
  • extract specific contribution for each level
    • contrib_A = r(A)
    • contrib_B = r(A+B) - r(A)
    • contrib_C = r(A+B+C) - r(A+B)

Parameters

Each function computes the contribution of feature groups (e.g [A, B, C]), indicated in the parameter hierarchy. Ex: if hierarchy = [0, 0, 0, 1, 1, 2] (n_groups=3), and y is shape (n, 10) (y_dim=10), each method would return an array of shape (n_groups, y_dim) = (3, 10), giving the contribution of each group to the prediction of a particular dimension.

Each function takes as input:

  • X : array of shape (n, x_dim)
    • features
  • y : array of shape (n, y_dim)
    • target
  • hierarchy : array of int of shape (x_dim).
    • Indices of the group for each feature.
    • Ex: [0, 0, 0, 1, 1, 2] refers to 3 groups, the first with dim=3, second with dim=2, last dim=1
  • model : sklearn model, optional, by default RidgeCV(np.logspace(-2, 8, 20)).
    • Predictive model used
      • to predict y given X,
      • if exp="permute_one", the model is also used to orthogonlize features
  • scale : bool, optional.
    • Whether to scale each group of features (independantly) before fitting, by default True
  • n_repeats : int, optional.
    • Number of repeats for exp "permute_xxx", by default 50.
    • Scores are always averaged across repeats.
  • pca : int, optional.
    • If > 0, PCA is applied to each group of features (independantly) before fitting, refers to the number of components to use in the PCA, by default 20.

Returns

Each function outputs:

  • importance : an array of shape (n_groups, y_dim), with n_groups = len(unique(hierarchy)) importance[i, j] gives the contribution of features group i for th prediction of y_j.