skbold.estimators package¶
The classifiers subpackage provides two ensemble-type classifiers that
aim at aggregating multivoxel information from multiple local sources in the
brain. They do so by allowing to fit a model on different brain areas, which
predictions are subsequently combined using either a stacked (meta) model
(i.e. the RoiStackingClassifier) or using a voting-strategy (i.e. the
RoiVotingClassifier). The structure and API of these classifiers adhere to
the scikit-learn estimator object.
-
class
RoiStackingClassifier(mvp, preproc_pipe='default', base_clf=None, meta_clf=None, mask_type='unilateral', proba=True, folds=10, meta_fs='univar', meta_gs=None, n_cores=1)[source]¶ Bases:
sklearn.base.BaseEstimator,sklearn.base.ClassifierMixinThis scikit-learn-style classifier implements a stacking classifier that fits a base-classifier on multiple brain-regions separately and subsequently trains a meta-classifier on the outputs of the base- classifiers on the separate brain-regions.
Parameters: - mvp (mvp-object) – An custom object from the skbold package containing data (X, y) and corresponding meta-data (e.g. mask info)
- preproc_pipe (object) – A scikit-learn Pipeline object with desired preprocessing steps
(e.g. scaling, additional feature selection). Defaults to only
scaling and univariate-feature-selection by means of highest
mean-euclidean differences (see
skbold.transformers.mean_euclidean). - base_clf (object) – A scikit-learn style classifier (implementing fit(), predict(), and predict_proba()), that is able to be used in Pipelines.
- meta_clf (object) – A scikit-learn style classifier.
- mask_type (str) – Can be ‘unilateral’ or ‘bilateral’, which will use all masks from the corresponding Harvard-Oxford Cortical (lateralized) atlas. Alternatively, it may be an absolute path to a directory containing a custom set of masks as nifti-files (default: ‘unilateral’).
- meta_gs (list or ndarray) – Optional parameter-grid over which to perform gridsearch.
- n_cores (int) – Number of CPU-cores on which to perform the fitting procedure (here, outer-folds are parallelized).
Variables: - train_scores (ndarray) – Accuracy-scores per brain region (averaged over outer-folds) on the training (fit) phase.
- test_scores (ndarray) – Accuracy-scores per brain region (averaged over outer- and inner-folds) on the test phase.
- masks (list of str) – List of absolute paths to found masks.
- stck_train (ndarray) – Array with outputs from base-classifiers fit on train-set.
- stck_test (ndarray) – Array with outputs from base-classifiers generalized to test-set.
-
fit(X, y)[source]¶ Fits RoiStackingClassfier.
Parameters: - X (ndarray) – Array of shape = [n_samples, n_features].
- y (list or ndarray of int or float) – List or ndarray with floats/ints corresponding to labels.
Returns: self – RoiStackingClassifier instance with fitted parameters.
Return type: object
-
class
RoiVotingClassifier(mvp, preproc_pipeline=None, clf=None, mask_type='unilateral', voting='soft', weights=None)[source]¶ Bases:
sklearn.base.BaseEstimator,sklearn.base.ClassifierMixinThis classifier fits a base-estimator (by default a linear SVM) on different feature sets (i.e. voxels) from different regions of interest (which are drawn from the Harvard-Oxford Cortical atlas), and subsequently the final prediction is derived through a max-voting rule, which can be either ‘soft’ (argmax of mean class probability) or ‘hard’ (max of class prediction).
Notes
This classifier has not been tested!
Parameters: - mvp (mvp-object) – An custom object from the skbold package containing data (X, y) and corresponding meta-data (e.g. mask info)
- preproc_pipeline (object) – A scikit-learn Pipeline object with desired preprocessing steps (e.g. scaling, additional feature selection)
- clf (object) – A scikit-learn style classifier (implementing fit(), predict(), and predict_proba()), that is able to be used in Pipelines.
- mask_type (str) – Can be ‘unilateral’ or ‘bilateral’, which will use all masks from the corresponding Harvard-Oxford Cortical (lateralized) atlas. Alternatively, it may be an absolute path to a directory containing a custom set of masks as nifti-files (default: ‘unilateral’).
- voting (str) – Either ‘hard’ or ‘soft’ (default: ‘soft’).
- weights (list (or ndarray)) – List/array of shape [n_rois] with a relative weighting factor to be used in the voting procedure.
-
fit(X=None, y=None)[source]¶ Fits RoiVotingClassifier.
Parameters: - X (ndarray) – Array of shape = [n_samples, n_features].
- y (list or ndarray of int or float) – List or ndarray with floats/ints corresponding to labels.
Returns: self – RoiStackingClassifier instance with fitted parameters.
Return type: object
-
class
MultimodalVotingClassifier(mvp, clf=None, voting='soft', weights=None)[source]¶ Bases:
sklearn.base.BaseEstimator,sklearn.base.ClassifierMixinThis classifier fits a base-estimator (by default a linear SVM) on different feature sets of different modalities (i.e. VBM, TBSS, BOLD, etc), and subsequently the final prediction is derived through a max-voting rule, which can be either ‘soft’ (argmax of mean class probability) or ‘hard’ (max of class prediction).
Notes
This classifier has not been tested!
Parameters: - mvp (mvp-object) – An custom object from the skbold package containing data (X, y) and corresponding meta-data (e.g. mask info)
- preproc_pipeline (object) – A scikit-learn Pipeline object with desired preprocessing steps (e.g. scaling, additional feature selection)
- clf (object) – A scikit-learn style classifier (implementing fit(), predict(), and predict_proba()), that is able to be used in Pipelines.
- voting (str) – Either ‘hard’ or ‘soft’ (default: ‘soft’).
- weights (list (or ndarray)) – List/array of shape [n_rois] with a relative weighting factor to be used in the voting procedure.
-
fit(X=None, y=None, iterations=1)[source]¶ Fits RoiVotingClassifier.
Parameters: - X (ndarray) – Array of shape = [n_samples, n_features].
- y (list or ndarray of int or float) – List or ndarray with floats/ints corresponding to labels.
Returns: self – RoiStackingClassifier instance with fitted parameters.
Return type: object