Documentation for DRFSC

Distributed Randomised Feature Selection for Classification (DRFSC)

__init__(n_vbins=1, n_hbins=1, n_runs=1, redistribute_features=False, feature_sharing='all', k=0, output='single', metric='roc_auc', verbose=False, polynomial=1, preprocess=True, max_processes=None)

Constructor for DRFSC class. Initialises the DRFSC class with the given parameters.

Parameters:

Name	Type	Description	Default
n_vbins	int, optional	Number of vertical partitions to create for the data. Defaults to 1.	1
n_hbins	int, optional	Number of horizontal partitions to create for the data. If output = 'ensemble', each hbin will converge to its own best model. Defaults to 1.	1
n_runs	int, optional	Number of feature-sharing iterations to perform. Larger numbers may yield better results, but also take longer. Defaults to 1.	1
redistribute_features	bool, optional	If True, the base features included in each bin will be shuffled at each feature-sharing iteration. Does not affect feature sharing. Defaults to False.	False
feature_sharing	str, optional	The method used to share features. Defaults to 'all'. Options (str): 'all', 'latest', 'top_k'. If feature_sharing = 'all', the entire history of best features from all sub-processes will be shared. If feature)sharing = 'latest', features from all sub-processes at the current iteration will be shared. If feature_sharing = 'top_k', the best k features will be shared.	'all'
k	int, optional	Number of best features to share. Only used if feature_sharing = 'top_k'. Defaults to 0.	0
output	str, optional	Output type desired. Options (str): 'single', 'ensemble'. If output = 'single', the best model from all sub-processes will be returned. If output = 'ensemble', the best model from each horizontal partition will be returned. If output = 'ensemb;e'. no features between different horizontal partitions will be created. Defaults to 'single'.	'single'
metric	str, optional	Evaluation metric used in the optimisation process. Options (str) : ['acc', 'roc_auc', 'weighted', 'avg_prec', 'f1', 'auprc']. Defaults to 'roc_auc'. For more information on the metrics, see the documentation for the sklearn.metrics module.	'roc_auc'
verbose	bool, optional	if True, prints extra information. Defaults to False.	False
polynomial	int, optional	degree of polynomial features to use. Defaults to 1.	1
preprocess	bool, optional	If True, will scale-data, create dummies for categorical variables, and create polynomial features based on passed polynomial parameter. If False, will only convert data to numpy arrays. Defaults to True.	True
max_processes	int, optional	Enforces maximum number of processes that can be generated. If None, will use all available cores. Defaults to None.	None

set_initial_mu(mu_init, _type='name')

Setter for initial mu values for RFSC.

If type(mu_init) == float, will set all mu values to that float.

If type(mu_init) == dict, will set the mu values for the given features to the given values.

If dict is passed, _type must be set to 'name' or 'index'. If 'name', the keys of the dictionary must be the names of the features. If 'index', the keys of the dictionary must be the indices of the features.

Parameters:

Name	Type	Description	Default
mu_init	Union[dict, float]	Initial mu values for RFSC. Can be a dictionary or float.	required
_type	str, optional	Only required if type(mu_init) == dict. Type of keys in mu_init.	'name'

get_rfsc_params()

Getter for RFSC parameters. Returns a dictionary of the current RFSC parameters

set_rfsc_params(params)

Setter for RFSC parameters. Updates the RFSC parameters with the given dictionary. Dictionary must be in the form of {parameter_name: parameter_value}.

To view available parameters, see help(RFSC_base.init)

load_data(X_train, X_val, Y_train, Y_val, X_test=None, Y_test=None, polynomial=1, preprocess=True)

Preprocesses the data in the required way for the DRFSC model. Can be used to load data into the model if it has not been loaded yet. Scales the data to [0,1] and creates polynomial expansion based on the passed 'polynomial' parameter.

Parameters:

Name	Type	Description	Default
X_train	np.ndarray or pd.DataFrame	Train set data	required
X_val	np.ndarray or pd.DataFrame	Validation set data	required
Y_train	np.ndarray or pd.DataFrame	Train set labels	required
Y_val	np.ndarray or pd.DataFrame	Validation set labels	required
X_test	np.ndarray or pd.DataFrame, optional	Test set data. Only required if postprocessing is required. Defaults to None.	None
Y_test	np.ndarray or pd.DataFrame optional	Test set labels. Only required if postprocessing is required. Defaults to None.	None
polynomial	int, optional	degree of polynomial features to use. Defaults to 1.	1
preprocess	bool, optional	If True, will scale-data, create dummies for categorical variables, and create polynomial features based on passed polynomial parameter. If False, will only convert data to numpy arrays. Defaults to True.	True

Returns

X_train, X_val, Y_train, Y_val, X_test, Y_test : np.ndarray Preprocessed data and labels

fit(X_train, X_val, Y_train, Y_val)

The main function for fitting the model. Returns the a single final model if output == 'single', else returns a model ensemble based on the number of horizontal partitions (n_hbins).

Parameters:

Name	Type	Description	Default
X_train	np.ndarray or pd.DataFrame	Train set data	required
X_val	np.ndarray or pd.DataFrame	Validation set data	required
Y_train	np.ndarray or pd.DataFrame	Train set labels	required
Y_val	np.ndarray or pd.DataFrame	Validation set labels	required

score(X_test, Y_test, metric=None)

Used to evaluate the final model on the test set.

Parameters:

Name	Type	Description	Default
X_test	np.ndarray or pd.DataFrame	Test set data	required
Y_test	np.ndarray or pd.DataFrame or pd.Series	Test set labels	required
metric	str, optional	Metric to use for evaluation. By default uses the metric specified in the constructor. Other options: ('acc', 'roc_auc', 'weighted', 'avg_prec', 'f1', 'auprc').	None

Returns:

Name	Type	Description
evaluation	dict	returns the score of the model based on the metric specified.

feature_importance()

Creates a bar plot of the features of the model and their contribution to the final prediction.

Returns

Returns:

Name	Type	Description
figure	matplotlib figure	hisogram of feature coefficients for features of the final model.

pos_neg_prediction(data_index=0, X_test=None)

Creates a plot of the positive and negative parts of the prediction. Returned figure shows, for a given sample (indexed by data_index), the value of the coefficients and multiplies them by the feature values. These components of the prediction are then plotted.

Parameters:

Name	Type	Description	Default
data_index	int	Index of the data observation to be plotted. If X_test is not provided, then the index is relative to the provided training/validation data. If X_test is provided, then the index is relative to the provided test data. Default is 0.	0
X_test	np.array or pd.DataFrame	Test data to be used for prediction. If provided, then the index is relative to the provided test data. Default is None.	None

Returns:

Name	Type	Description
figure	matplotlib.pyplot.figure	output figure

single_prediction(data_index=0, X_test=None)

Creates a plot of the single prediction of the final model. Figure shows for a given sample (indexed by data_index) the coefficients of the final model, weighted by the feature values for the indexed observation.

Parameters:

Name	Type	Description	Default
data_index	int	Index of the data observation to be plotted. If X_test is not provided, then the index is relative to the provided training/validation data. If X_test is provided, then the index is relative to the provided test data. Default is 0.	0
X_test	np.array or pd.DataFrame	Test data to be used for prediction. If provided, then the index is relative to the provided test data. Default is None.	None

Returns:

Name	Type	Description
figure	matplotlib.pyplot.figure	output figure