mlquantify 0.1.7__py3-none-any.whl → 0.1.9__py3-none-any.whl

mlquantify/meta/_classes.py

@@ -0,0 +1,761 @@
+import numpy as np
+import pandas as pd
+from copy import deepcopy
+from tqdm import tqdm
+from sklearn.linear_model import LogisticRegression
+from sklearn.model_selection import GridSearchCV, cross_val_predict, train_test_split
+
+from mlquantify.base import BaseQuantifier, MetaquantifierMixin
+from mlquantify.metrics._slq import MSE
+from mlquantify.mixture._classes import SORD, DyS
+from mlquantify.mixture._utils import getHist, hellinger
+from mlquantify.utils import Options, Interval
+from mlquantify.utils import _fit_context
+from mlquantify.confidence import (
+    construct_confidence_region
+)
+from mlquantify.base_aggregative import (
+    _get_learner_function, 
+    is_aggregative_quantifier,
+    get_aggregation_requirements,
+    uses_soft_predictions
+)
+from mlquantify.utils._sampling import (
+    bootstrap_sample_indices
+)
+from mlquantify.model_selection import APP, NPP, UPP
+from mlquantify.utils._validation import validate_data, validate_prevalences
+from mlquantify.utils.prevalence import get_prev_from_labels
+
+
+
+def get_protocol_sampler(protocol_name, batch_size, n_prevalences, min_prev, max_prev, n_classes):
+    """ Returns a prevalence sampler function based on the specified protocol name.
+    
+    Parameters
+    ----------
+    protocol_name : str
+        The name of the protocol ('app', 'npp', 'upp', 'upp-k').
+    batch_size : int
+        The size of each batch.
+    n_prevalences : int
+        The number of prevalences to sample.
+    min_prev : float
+        The minimum prevalence value.
+    max_prev : float
+        The maximum prevalence value.
+    n_classes : int
+        The number of classes.
+        
+    Returns
+    -------
+    callable
+        A function that generates prevalence samples according to the specified protocol.
+    """
+    
+    if protocol_name == 'artificial':
+        protocol = APP(batch_size=batch_size,
+                           n_prevalences=n_prevalences,
+                           min_prev=min_prev,
+                           max_prev=max_prev)
+
+    elif protocol_name == 'natural':
+        protocol = NPP(batch_size=batch_size,
+                           n_samples=n_prevalences)
+
+    elif protocol_name == 'uniform':
+            protocol = UPP(batch_size=batch_size,
+                           n_prevalences=n_prevalences,
+                           algorithm='uniform',
+                           min_prev=min_prev,
+                           max_prev=max_prev)
+    elif protocol_name == 'kraemer':
+        protocol = UPP(batch_size=batch_size,
+                           n_prevalences=n_prevalences,
+                           algorithm='kraemer',
+                           min_prev=min_prev,
+                           max_prev=max_prev)
+    else:
+        raise ValueError(f"Unknown protocol: {protocol_name}")
+    return protocol
+
+class EnsembleQ(MetaquantifierMixin, BaseQuantifier):
+    """
+    Ensemble-based Quantifier combining multiple models trained on varied data samples 
+    with controlled prevalence distributions to improve robustness and accuracy.
+
+    This quantifier constructs an ensemble of quantification models using batches of training 
+    data sampled according to an evaluation protocol (e.g. 'artificial', 'natural', 'uniform', 'kraemer') 
+    with specified prevalence constraints. Diverse models are trained on these subsamples, 
+    and their prevalence estimates aggregated using various selection metrics and aggregation methods.
+
+    Parameters
+    ----------
+    quantifier : BaseQuantifier
+        The quantifier model class to be used for ensemble members.
+    size : int, default=50
+        Number of ensemble members (sub-models) to train.
+    min_prop, max_prop : float, default=(0.1, 1.0)
+        Minimum and maximum class prevalence proportions for generating training batches.
+    selection_metric : {'all', 'ptr', 'ds'}, default='all'
+        Metric used to select or weight ensemble members during aggregation:
+        - 'all': uses all models equally,
+        - 'ptr': selects models with prevalences closest to initial test prevalence estimates,
+        - 'ds': selects models with score distributions similar to test data.
+    p_metric : float, default=0.25
+        Proportion of ensemble members to select according to the selection metric.
+    protocol : {'artificial', 'natural', 'uniform', 'kraemer'}, default='uniform'
+        Sampling protocol used to generate training data for ensemble models.
+    return_type : {'mean', 'median'}, default='mean'
+        Aggregation method for ensemble predictions.
+    max_sample_size : int or None, optional
+        Maximum number of samples per training batch; defaults to dataset size if None.
+    max_trials : int, default=100
+        Maximum number of trials for sampling.
+    n_jobs : int, default=1
+        Number of parallel jobs for training ensemble members.
+    verbose : bool, default=False
+        Enable verbose output.
+
+    Attributes
+    ----------
+    models : list
+        List of fitted quantifier ensemble members.
+    train_prevalences : list
+        List of training prevalences corresponding to ensemble members.
+    train_distributions : list
+        List of historical training posterior histograms (used when selection_metric='ds').
+    posteriors_generator : callable or None
+        Function to generate posterior probabilities for new samples.
+
+    Methods
+    -------
+    fit(X, y)
+        Fits all ensemble member quantifiers on sampled training batches.
+    predict(X)
+        Aggregates ensemble member predictions into final prevalence estimates.
+    ptr_selection_metric(prevalences, train_prevalences)
+        Implements PTR-based selection metric on prevalence estimates.
+    ds_get_posteriors(X, y)
+        Computes posterior probabilities for training data with cross-validated logistic regression.
+    ds_selection_metric(X, prevalences, train_distributions, posteriors_generator)
+        Implements DS-based selection metric comparing posterior distributions.
+
+    Notes
+    -----
+    - Ensemble diversity is controlled by sampling prevalences from the specified protocol.
+    - The 'ds' selection metric requires probabilistic quantifiers and computes distribution similarity.
+    - Uses sklearn's LogisticRegression and GridSearchCV internally for posterior computation within 'ds'.
+
+    Examples
+    --------
+    >>> ensemble = EnsembleQ(quantifier=SomeQuantifier(), size=30, protocol='kraemer', selection_metric='ptr')
+    >>> ensemble.fit(X_train, y_train)
+    >>> prevalence_estimates = ensemble.predict(X_test)
+    """  
+      
+    _parameter_constraints = {
+        "quantifier": [BaseQuantifier],
+        "size": [Interval(left=1, right=None, discrete=True)],
+        "min_prop": [Interval(left=0.0, right=1.0, inclusive_left=True, inclusive_right=True)],
+        "max_prop": [Interval(left=0.0, right=1.0, inclusive_left=True, inclusive_right=True)],
+        "selection_metric": [Options(['all', 'ptr', 'ds'])],
+        "p_metric": [Interval(left=0.0, right=1.0, inclusive_left=True, inclusive_right=True)],
+        "protocol": [Options(['artificial', 'natural', 'uniform', 'kraemer'])],
+        "return_type": [Options(['mean', 'median'])],
+        "max_sample_size": [Options([Interval(left=1, right=None, discrete=True), None])],
+        "max_trials": [Interval(left=1, right=None, discrete=True)],
+        "n_jobs": [Interval(left=1, right=None, discrete=True)],
+        "verbose": [bool],
+    }
+
+    def __init__(self,
+                 quantifier,
+                 size=50,
+                 min_prop=0.1,
+                 max_prop=1,
+                 selection_metric='all',
+                 protocol="uniform",
+                 p_metric=0.25,
+                 return_type="mean",
+                 max_sample_size=None,
+                 max_trials=100,
+                 n_jobs=1,
+                 verbose=False):
+        
+        self.quantifier = quantifier
+        self.size = size
+        self.min_prop = min_prop
+        self.max_prop = max_prop
+        self.p_metric = p_metric
+        self.protocol = protocol
+        self.selection_metric = selection_metric
+        self.return_type = return_type
+        self.n_jobs = n_jobs
+        self.verbose = verbose
+        self.max_sample_size = max_sample_size
+        self.max_trials = max_trials
+
+    def sout(self, msg):
+        """Prints a message if verbose is True."""
+        if self.verbose:
+            print('[Ensemble]' + msg)
+
+    @_fit_context(prefer_skip_nested_validation=True)
+    def fit(self, X, y):
+        """ Fits the ensemble model to the given training data.
+        
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The input data.
+        y : array-like of shape (n_samples,)
+            The target values.
+            
+        Returns
+        -------
+        self : Ensemble
+            The fitted ensemble model.
+        """
+        self.sout('Fit')
+        
+        self.models = []
+        self.train_prevalences = []
+        self.train_distributions = []
+        self.posteriors_generator = []
+        
+        self.classes = np.unique(y)
+        X, y = validate_data(self, X, y)
+        
+        if self.selection_metric == 'ds' and not len(self.classes) == 2:
+            raise ValueError(f'ds selection_metric is only defined for binary quantification, but this dataset is not binary')
+        # randomly chooses the prevalences for each member of the ensemble (preventing classes with less than
+        # min_pos positive examples)
+        sample_size = len(y) if self.max_sample_size is None else min(self.max_sample_size, len(y))
+        
+        protocol = get_protocol_sampler(
+            protocol_name=self.protocol,
+            batch_size=sample_size, 
+            n_prevalences=self.size, 
+            min_prev=self.min_prop, 
+            max_prev=self.max_prop,
+            n_classes=len(self.classes)
+        )
+
+        posteriors = None
+        if self.selection_metric == 'ds':
+            # precompute the training posterior probabilities
+            posteriors, self.posteriors_generator = self.ds_get_posteriors(X, y)
+            
+        for idx in protocol.split(X, y):
+            X_batch, y_batch = X[idx], y[idx]
+            model = deepcopy(self.quantifier)
+            
+            model.fit(X_batch, y_batch)
+            tr_prev = get_prev_from_labels(y_batch)
+            
+            if self.selection_metric == 'ds':
+                self.train_distributions.append(getHist(posteriors[idx], 8))
+            
+            self.train_prevalences.append(tr_prev)
+            self.models.append(model)
+
+        self.sout('Fit [Done]')
+        return self
+
+    def predict(self, X):
+        """ Predicts the class prevalences for the given test data.
+        
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The input data.
+        
+        Returns
+        -------
+        prevalences : array-like of shape (n_samples, n_classes)
+            The predicted class prevalences.
+        """
+        self.sout('Predict')
+        
+        test_prevalences = []
+        
+        for model in tqdm(self.models, disable=not self.verbose):
+            pred = np.asarray(list(model.predict(X).values()))
+            test_prevalences.append(pred)
+        
+        test_prevalences = np.asarray(test_prevalences)
+
+        if self.selection_metric == 'ptr':
+            test_prevalences = self.ptr_selection_metric(test_prevalences, self.train_prevalences)
+        elif self.selection_metric == 'ds':
+            test_prevalences = self.ds_selection_metric(X, 
+                                                   test_prevalences, 
+                                                   self.train_distributions,
+                                                   self.posteriors_generator)
+
+        if self.return_type == "median":
+            prevalences = np.median(test_prevalences, axis=0)
+        else:      
+            prevalences = np.mean(test_prevalences, axis=0)
+            
+        
+        self.sout('Predict [Done]')
+        prevalences = validate_prevalences(self, prevalences, self.classes)
+        return prevalences
+
+
+    def ptr_selection_metric(self, prevalences, train_prevalences):
+        """
+        Selects the prevalence estimates from models trained on samples whose prevalence is most similar
+        to an initial approximation of the test prevalence as estimated by all models in the ensemble.
+
+        Parameters
+        ----------
+        prevalences : numpy.ndarray
+            An array of prevalence estimates provided by each model in the ensemble.
+
+        Returns
+        -------
+        numpy.ndarray
+            The selected prevalence estimates after applying the PTR selection metric.
+        """
+        test_prev_estim = prevalences.mean(axis=0)
+        ptr_differences = [MSE(test_prev_estim, ptr_i) for ptr_i in train_prevalences]
+        order = np.argsort(ptr_differences)
+        return _select_k(prevalences, order, k=self.p_metric)
+
+    def ds_get_posteriors(self, X, y):
+        """ 
+        Generate posterior probabilities using cross-validated logistic regression.
+        This method computes posterior probabilities for the training data via cross-validation,
+        using a logistic regression classifier with hyperparameters optimized through grid search.
+        It also returns a function to generate posterior probabilities for new data.
+        
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The feature matrix representing the training data.
+        y : array-like of shape (n_samples,)
+            The target vector representing class labels for the training data.
+            
+        Returns
+        -------
+        posteriors : ndarray of shape (n_samples, n_classes)
+            Posterior probabilities for the training data obtained through cross-validation.
+        posteriors_generator : callable
+            A function that computes posterior probabilities for new input data.
+            
+        Notes
+        -----
+        - In scenarios where the quantifier is not based on a probabilistic classifier, it's necessary
+            to train a separate probabilistic model to obtain posterior probabilities.
+        - Using cross-validation ensures that the posterior probabilities for the training data are unbiased,
+            as each data point is evaluated by a model not trained on that point.
+        - Hyperparameters for the logistic regression classifier are optimized using a grid search with
+            cross-validation to improve the model's performance.
+        """
+        lr_base = LogisticRegression(class_weight='balanced', max_iter=1000)
+
+        optim = GridSearchCV(
+            lr_base, param_grid={'C': np.logspace(-4, 4, 9)}, cv=5, n_jobs=self.n_jobs, refit=True
+        ).fit(X, y)
+
+        posteriors = cross_val_predict(
+            optim.best_estimator_, X, y, cv=5, n_jobs=self.n_jobs, method='predict_proba'
+        )
+        posteriors_generator = optim.best_estimator_.predict_proba
+
+        return posteriors, posteriors_generator
+
+
+    def ds_selection_metric(self, X, prevalences, train_distributions, posteriors_generator):
+        """
+        Selects the prevalence estimates from models trained on samples whose distribution of posterior
+        probabilities is most similar to the distribution of posterior probabilities for the test data.
+        
+        Parameters
+        ----------
+        prevalences : numpy.ndarray
+            An array of prevalence estimates provided by each model in the ensemble.
+        test : array-like of shape (n_samples, n_features)
+            The feature matrix representing the test data.
+        
+        Returns
+        -------
+        numpy.ndarray
+            The selected prevalence estimates after applying the DS selection metric.
+        """
+        test_posteriors = posteriors_generator(X)
+        test_distribution = getHist(test_posteriors, 8)
+        dist = [hellinger(tr_dist_i, test_distribution) for tr_dist_i in train_distributions]
+        order = np.argsort(dist)
+        return _select_k(prevalences, order, k=self.p_metric)
+
+def _select_k(elements, order, k):
+    """
+    Selects the k elements from the list of elements based on the order.
+    If the list is empty, it returns the original list.
+    
+    Parameters
+    ----------
+    elements : array-like
+        The array of elements to be selected from.
+    order : array-like
+        The order of the elements.
+    k : int
+        The number of elements to be selected.
+    
+    Returns
+    -------
+    array-like
+        The selected elements.
+    """
+    elements_k = [elements[idx] for idx in order[:k]]
+    if elements_k:
+        return elements_k
+    print(f"Unable to take {k} for elements with size {len(elements)}")
+    return elements
+
+
+
+
+
+class AggregativeBootstrap(MetaquantifierMixin, BaseQuantifier):
+    """
+    Aggregative Bootstrap Quantifier to compute prevalence confidence regions.
+
+    This metaquantifier applies bootstrapping to both training and test data predictions 
+    to generate multiple bootstrap prevalence estimates. These bootstrapped estimates 
+    are used to construct confidence intervals or elliptical confidence regions for 
+    prevalence predictions, improving uncertainty quantification.
+
+    Parameters
+    ----------
+    quantifier : BaseQuantifier
+        The base quantifier model, which must be aggregative.
+    n_train_bootstraps : int, default=1
+        Number of bootstrap samples to generate from training predictions.
+    n_test_bootstraps : int, default=1
+        Number of bootstrap samples to generate from test predictions.
+    random_state : int or None, optional
+        Random seed for reproducibility.
+    region_type : {'intervals', 'ellipse', 'ellipse-clr'}, default='intervals'
+        Type of confidence region to construct.
+    confidence_level : float between 0 and 1, default=0.95
+        Confidence level for intervals or regions.
+
+    Methods
+    -------
+    fit(X, y, val_split=None)
+        Fits base quantifier and generates training predictions (optionally splitting data).
+    predict(X)
+        Returns prevalence estimates and confidence regions aggregated from bootstrap samples.
+    aggregate(predictions, train_predictions, train_y_values)
+        Performs bootstrap resampling aggregation to obtain prevalence confidence regions.
+
+    Examples
+    --------
+    >>> agg_boot = AggregativeBootstrap(quantifier=SomeQuantifier, n_train_bootstraps=100, n_test_bootstraps=100)
+    >>> agg_boot.fit(X_train, y_train)
+    >>> prevalence, conf_region = agg_boot.predict(X_test)
+    """
+
+    _parameter_constraints = {
+        "quantifier": [BaseQuantifier],
+        "n_train_bootstraps": [Interval(left=1, right=None, discrete=True)],
+        "n_test_bootstraps": [Interval(left=1, right=None, discrete=True)],
+        "random_state": [Options([None, int])],
+        "region_type": [Options(['intervals', 'ellipse', 'ellipse-clr'])],
+        "confidence_level": [Interval(left=0.0, right=1.0)],
+    }
+
+    def __init__(self, 
+                 quantifier, 
+                 n_train_bootstraps=1, 
+                 n_test_bootstraps=1,
+                 random_state=None,
+                 region_type='intervals',
+                 confidence_level=0.95):
+        self.quantifier = quantifier
+        self.n_train_bootstraps = n_train_bootstraps
+        self.n_test_bootstraps = n_test_bootstraps
+        self.random_state = random_state
+        self.region_type = region_type
+        self.confidence_level = confidence_level
+        
+    def fit(self, X, y, val_split=None):
+        """ Fits the aggregative bootstrap model to the given training data.
+        
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The input data.
+        y : array-like of shape (n_samples,)
+            The target values.
+            
+        Returns
+        -------
+        self : AggregativeBootstrap
+            The fitted aggregative bootstrap model.
+        """
+        X, y = validate_data(self, X, y)
+        self.classes = np.unique(y)
+        
+        if not is_aggregative_quantifier(self.quantifier):
+            raise ValueError(f"The quantifier {self.quantifier.__class__.__name__} is not an aggregative quantifier.")
+        self.quantifier_learner = deepcopy(self.quantifier)
+        
+        learner_function = _get_learner_function(self.quantifier_learner)
+        model = self.quantifier_learner.learner
+        
+        if val_split is None:
+            model.fit(X, y)
+            train_y_values = y
+            train_predictions = getattr(model, learner_function)(X)
+        else:
+            X_fit, y_fit, X_val, y_val = train_test_split(X, y, test_size=val_split, random_state=self.random_state)
+            model.fit(X_fit, y_fit)
+            train_y_values = y_val
+            train_predictions = getattr(model, learner_function)(X_val)
+        self.train_predictions = train_predictions
+        self.train_y_values = train_y_values
+        
+        return self
+    
+    def predict(self, X):
+        """ Predicts the class prevalences for the given test data.
+        
+        Parameters
+        ----------
+        X : array-like of shape (n_samples, n_features)
+            The input data.
+        
+        Returns
+        -------
+        prevalences : array-like of shape (n_samples, n_classes)
+            The predicted class prevalences.
+        """
+        X = validate_data(self, X, None)
+        learner_function = _get_learner_function(self.quantifier_learner)
+        model = self.quantifier_learner.learner
+        
+        predictions = getattr(model, learner_function)(X)
+
+        return self.aggregate(predictions, self.train_predictions, self.train_y_values)
+
+
+    def aggregate(self, predictions, train_predictions, train_y_values):
+        """ Aggregates the predictions using bootstrap resampling.
+        
+        Parameters
+        ----------
+        predictions : array-like of shape (n_samples, n_classes)
+            The input data.
+        train_predictions : array-like of shape (n_samples, n_classes)
+            The training predictions.
+        train_y_values : array-like of shape (n_samples,)
+            The training target values.
+            
+        Returns
+        -------
+        prevalences : array-like of shape (n_samples, n_classes)
+            The predicted class prevalences.
+        """
+        prevalences = []
+        
+        self.classes = np.unique(train_y_values)
+        
+        for train_idx in bootstrap_sample_indices(
+            n_samples=len(train_predictions),
+            n_bootstraps=self.n_train_bootstraps,
+            batch_size=len(train_predictions),
+            random_state=self.random_state
+        ):
+            train_pred_boot = train_predictions[train_idx]
+            train_y_boot = train_y_values[train_idx]
+            
+            for test_idx in bootstrap_sample_indices(
+                n_samples=len(predictions),
+                n_bootstraps=self.n_test_bootstraps,
+                batch_size=len(predictions),
+                random_state=self.random_state
+            ):
+                test_pred_boot = predictions[test_idx]
+
+                requirements = get_aggregation_requirements(self.quantifier)
+                
+                if requirements.requires_train_proba and requirements.requires_train_labels:
+                    prevalences_boot = self.quantifier.aggregate(test_pred_boot, train_pred_boot, train_y_boot)
+                elif requirements.requires_train_labels:
+                    prevalences_boot = self.quantifier.aggregate(test_pred_boot, train_y_boot)
+                else:
+                    prevalences_boot = self.quantifier.aggregate(test_pred_boot)
+
+                prevalences_boot = np.asarray(list(prevalences_boot.values()))
+                prevalences.append(prevalences_boot)
+
+        prevalences = np.asarray(prevalences)
+        confidence_region = construct_confidence_region(
+            prev_estims=prevalences,
+            method=self.region_type,
+            confidence_level=self.confidence_level,
+        )
+
+        prevalence = confidence_region.get_point_estimate()
+        
+        prevalence = validate_prevalences(self, prevalence, self.classes)
+
+        return prevalence
+
+
+
+
+class QuaDapt(MetaquantifierMixin, BaseQuantifier):
+    r"""QuaDapt Metaquantifier: Adaptive quantification using score merging and distance measures.
+
+    This metaquantifier improves prevalence estimation by merging training samples 
+    with different score distributions using a merging factor \( m \). It evaluates 
+    candidate merging factors, chooses the best by minimizing a distribution distance 
+    metric (Hellinger, Topsoe, ProbSymm, or SORD), and aggregates quantification accordingly.
+
+    Parameters
+    ----------
+    quantifier : BaseQuantifier
+        The base quantifier model to adapt.
+    measure : {'hellinger', 'topsoe', 'probsymm', 'sord'}, default='topsoe'
+        The distribution distance metric used to select the best merging factor.
+    merging_factor : array-like
+        Candidate merging factor values to evaluate.
+
+    Methods
+    -------
+    fit(X, y)
+        Fits the base learner on training data.
+    predict(X)
+        Predicts prevalence aggregating via the best merging factor.
+    aggregate(predictions, train_y_values)
+        Performs adaptation and aggregation based on merged score distributions.
+    _get_best_merging_factor(predictions)
+        Evaluates merging factors and selects the best based on minimum distance.
+    _get_best_distance(predictions, pos_scores, neg_scores)
+        Computes the distance metric between predicted and class score distributions.
+
+    Class Methods
+    -------------
+    MoSS(n, alpha, m)
+        Generates merged score samples modeling class conditional distributions 
+        parameterized by mixing proportion alpha and merging factor m.
+
+    Examples
+    --------
+    >>> quadapt = QuaDapt(quantifier=SomeQuantifier, merging_factor=[0.1, 0.5, 1.0], measure='sord')
+    >>> quadapt.fit(X_train, y_train)
+    >>> prevalence = quadapt.predict(X_test)
+    """
+    
+    _parameter_constraints = {
+        "quantifier": [BaseQuantifier],
+        "merging_factor": "array-like",
+        "measure": [Options(["hellinger", "topsoe", "probsymm", "sord"])],
+        "random_state": [Options([None, int])],
+    }
+    
+    def __init__(self, 
+                 quantifier,
+                 measure="topsoe", 
+                 merging_factor=(0.1, 1.0, 0.2)):
+        self.quantifier = quantifier
+        self.measure = measure
+        self.merging_factor = merging_factor
+        
+    
+    def fit(self, X, y):
+        X, y = validate_data(self, X, y)
+        self.classes = np.unique(y)
+        
+        if not uses_soft_predictions(self.quantifier):
+            raise ValueError(f"The quantifier {self.quantifier.__class__.__name__} is not a soft (probabilistic) quantifier.")
+        
+        requirements = get_aggregation_requirements(self.quantifier)
+        if not requirements.requires_train_proba:
+            raise ValueError(f"The quantifier {self.quantifier.__class__.__name__} does not use training probabilities, which are required for QuaDapt.")
+        
+        self.quantifier.learner.fit(X, y)
+        self.train_y_values = y
+        
+        return self
+        
+    def predict(self, X):
+
+        X = validate_data(self, X, None)
+        
+        model = self.quantifier.learner
+        
+        predictions = getattr(model, "predict_proba")(X)
+
+        return self.aggregate(predictions, self.train_y_values)
+    
+    
+    def aggregate(self, predictions, train_y_values):
+
+        pos_predictions = predictions[:, 1]
+        m = self._get_best_merging_factor(pos_predictions)
+        
+        self.classes = self.classes if hasattr(self, 'classes') else np.unique(train_y_values)
+
+        moss = QuaDapt.MoSS(1000, 0.5, m)
+
+        moss_scores = moss[:, :2]
+        moss_labels = moss[:, 2]
+
+        prevalences = self.quantifier.aggregate(predictions,
+                                                moss_scores,
+                                                moss_labels)
+        
+        prevalences = {self.classes[i]: v for i, v in enumerate(prevalences.values())}
+        return prevalences
+
+        
+    def _get_best_merging_factor(self, predictions):
+        
+        MF = np.atleast_1d(np.round(self.merging_factor, 2)).astype(float)
+        
+        distances = []
+        
+        for mf in MF:
+            scores = QuaDapt.MoSS(1000, 0.5, mf)
+            pos_scores = scores[scores[:, 2] == 1][:, :2]
+            neg_scores = scores[scores[:, 2] == 0][:, :2]
+            
+            best_distance = self._get_best_distance(predictions, pos_scores, neg_scores)
+            
+            distances.append(best_distance)
+        
+        best_m = MF[np.argmin(distances)]
+        return best_m
+    
+    def _get_best_distance(self, predictions, pos_scores, neg_scores):
+        
+        if self.measure in ["hellinger", "topsoe", "probsymm"]:
+            method = DyS(measure=self.measure)
+        elif self.measure == "sord":
+            method = SORD()
+        
+        best_distance = method.get_best_distance(predictions, pos_scores, neg_scores)
+        return best_distance
+        
+
+    @classmethod
+    def MoSS(cls, n, alpha, m):
+        p_score = np.random.uniform(size=int(n * alpha)) ** m
+        n_score = 1 - (np.random.uniform(size=int(round(n * (1 - alpha), 0))) ** m)
+        scores = np.column_stack(
+            (np.concatenate((p_score, n_score)), 
+             np.concatenate((p_score, n_score)), 
+             np.concatenate((
+                 np.ones(len(p_score)), 
+                 np.full(len(n_score), 0))))
+        )
+        return scores
+