added model selection for quantification

quapy/data/base.py

@@ -4,11 +4,19 @@ from sklearn.model_selection import train_test_split
 from quapy.functional import artificial_prevalence_sampling
 from scipy.sparse import vstack
 
+from util import temp_seed
+
 
 class LabelledCollection:
 
     def __init__(self, instances, labels, n_classes=None):
-        self.instances = instances if issparse(instances) else np.asarray(instances)
+        if issparse(instances):
+            self.instances = instances
+        elif isinstance(instances, list) and len(instances)>0 and isinstance(instances[0], str):
+            # lists of strings occupy too much as ndarrays (although python-objects add a heavy overload)
+            self.instances = np.asarray(instances, dtype=object)
+        else:
+            self.instances = np.asarray(instances)
         self.labels = np.asarray(labels, dtype=int)
         n_docs = len(self)
         if n_classes is None:
@@ -87,6 +95,7 @@ class LabelledCollection:
         return LabelledCollection(documents, labels, n_classes=self.n_classes)
 
     def split_stratified(self, train_prop=0.6):
+        # with temp_seed(42):
         tr_docs, te_docs, tr_labels, te_labels = \
             train_test_split(self.instances, self.labels, train_size=train_prop, stratify=self.labels)
         return LabelledCollection(tr_docs, tr_labels), LabelledCollection(te_docs, te_labels)
@@ -110,14 +119,16 @@ class LabelledCollection:
             yield self.uniform_sampling_index(sample_size)
 
     def __add__(self, other):
-        if issparse(self.instances) and issparse(other.documents):
+        if issparse(self.instances) and issparse(other.instances):
-            docs = vstack([self.instances, other.documents])
+            join_instances = vstack([self.instances, other.instances])
-        elif isinstance(self.instances, list) and isinstance(other.documents, list):
+        elif isinstance(self.instances, list) and isinstance(other.instances, list):
-            docs = self.instances + other.documents
+            join_instances = self.instances + other.instances
+        elif isinstance(self.instances, np.ndarray) and isinstance(other.instances, np.ndarray):
+            join_instances = np.concatenate([self.instances, other.instances])
         else:
             raise NotImplementedError('unsupported operation for collection types')
         labels = np.concatenate([self.labels, other.labels])
-        return LabelledCollection(docs, labels)
+        return LabelledCollection(join_instances, labels)
 
 
 

quapy/data/datasets.py

@@ -1,5 +1,5 @@
 import zipfile
-from util import download_file_if_not_exists, download_file, get_quapy_home
+from util import download_file_if_not_exists, download_file, get_quapy_home, pickled_resource
 import os
 from os.path import join
 from data.base import Dataset
@@ -12,7 +12,21 @@ TWITTER_SENTIMENT_DATASETS = ['gasp', 'hcr', 'omd', 'sanders', 'semeval13', 'sem
                               'sst', 'wa', 'wb']
 
 
-def fetch_reviews(dataset_name, tfidf=False, min_df=None, data_home=None):
+def fetch_reviews(dataset_name, tfidf=False, min_df=None, data_home=None, pickle=False):
+    """
+    Load a Reviews dataset as a Dataset instance, as used in:
+    Esuli, A., Moreo, A., and Sebastiani, F. "A recurrent neural network for sentiment quantification."
+    Proceedings of the 27th ACM International Conference on Information and Knowledge Management. 2018.
+    :param dataset_name: the name of the dataset: valid ones are 'hp', 'kindle', 'imdb'
+    :param tfidf: set to True to transform the raw documents into tfidf weighted matrices
+    :param min_df: minimun number of documents that should contain a term in order for the term to be
+    kept (ignored if tfidf==False)
+    :param data_home: specify the quapy home directory where collections will be dumped (leave empty to use the default
+    ~/quay_data/ directory)
+    :param pickle: set to True to pickle the Dataset object the first time it is generated, in order to allow for
+    faster subsequent invokations
+    :return: a Dataset instance
+    """
     assert dataset_name in REVIEWS_SENTIMENT_DATASETS, \
         f'Name {dataset_name} does not match any known dataset for sentiment reviews. ' \
         f'Valid ones are {REVIEWS_SENTIMENT_DATASETS}'
@@ -27,18 +41,37 @@ def fetch_reviews(dataset_name, tfidf=False, min_df=None, data_home=None):
     download_file_if_not_exists(URL_TRAIN, train_path)
     download_file_if_not_exists(URL_TEST, test_path)
 
-    data = Dataset.load(train_path, test_path, from_text)
+    pickle_path = None
+    if pickle:
+        pickle_path = join(data_home, 'reviews', 'pickle', f'{dataset_name}.pkl')
+    data = pickled_resource(pickle_path, Dataset.load, train_path, test_path, from_text)
 
     if tfidf:
         text2tfidf(data, inplace=True)
-
+        if min_df is not None:
-    if min_df is not None:
+            reduce_columns(data, min_df=min_df, inplace=True)
-        reduce_columns(data, min_df=min_df, inplace=True)
 
     return data
 
 
-def fetch_twitter(dataset_name, model_selection=False, min_df=None, data_home=None):
+def fetch_twitter(dataset_name, for_model_selection=False, min_df=None, data_home=None, pickle=False):
+    """
+    Load a Twitter dataset as a Dataset instance, as used in:
+    Gao, W., Sebastiani, F.: From classification to quantification in tweet sentiment analysis.
+    Social Network Analysis and Mining6(19), 1–22 (2016)
+
+    :param dataset_name: the name of the dataset: valid ones are 'gasp', 'hcr', 'omd', 'sanders', 'semeval13',
+    'semeval14', 'semeval15', 'semeval16', 'sst', 'wa', 'wb'
+    :param for_model_selection: if True, then returns the train split as the training set and the devel split
+    as the test set; if False, then returns the train+devel split as the training set and the test set as the
+    test set
+    :param min_df: minimun number of documents that should contain a term in order for the term to be kept
+    :param data_home: specify the quapy home directory where collections will be dumped (leave empty to use the default
+    ~/quay_data/ directory)
+    :param pickle: set to True to pickle the Dataset object the first time it is generated, in order to allow for
+    faster subsequent invokations
+    :return: a Dataset instance
+    """
     assert dataset_name in TWITTER_SENTIMENT_DATASETS, \
         f'Name {dataset_name} does not match any known dataset for sentiment twitter. ' \
         f'Valid ones are {TWITTER_SENTIMENT_DATASETS}'
@@ -56,23 +89,27 @@ def fetch_twitter(dataset_name, model_selection=False, min_df=None, data_home=No
 
     if dataset_name in {'semeval13', 'semeval14', 'semeval15'}:
         trainset_name = 'semeval'
-        testset_name  = 'semeval' if model_selection else dataset_name
+        testset_name  = 'semeval' if for_model_selection else dataset_name
         print(f"the training and development sets for datasets 'semeval13', 'semeval14', 'semeval15' are common "
               f"(called 'semeval'); returning trainin-set='{trainset_name}' and test-set={testset_name}")
     else:
         trainset_name = testset_name = dataset_name
 
-    if model_selection:
+    if for_model_selection:
         train = join(unzipped_path, 'train', f'{trainset_name}.train.feature.txt')
         test  = join(unzipped_path, 'test', f'{testset_name}.dev.feature.txt')
     else:
         train = join(unzipped_path, 'train', f'{trainset_name}.train+dev.feature.txt')
-        if dataset_name == 'semeval16':
+        if dataset_name == 'semeval16':  # there is a different test name in the case of semeval16 only
             test = join(unzipped_path, 'test', f'{testset_name}.dev-test.feature.txt')
         else:
             test = join(unzipped_path, 'test', f'{testset_name}.test.feature.txt')
 
-    data = Dataset.load(train, test, from_sparse)
+    pickle_path = None
+    if pickle:
+        mode = "train-dev" if for_model_selection else "train+dev-test"
+        pickle_path = join(unzipped_path, 'pickle', f'{testset_name}.{mode}.pkl')
+    data = pickled_resource(pickle_path, Dataset.load, train, test, from_sparse)
 
     if min_df is not None:
         reduce_columns(data, min_df=min_df, inplace=True)

quapy/evaluation.py

@@ -14,7 +14,9 @@ def artificial_sampling_prediction(
         n_prevpoints=210,
         n_repetitions=1,
         n_jobs=-1,
-        random_seed=42):
+        random_seed=42,
+        verbose=True
+):
     """
     Performs the predictions for all samples generated according to the artificial sampling protocol.
     :param model: the model in charge of generating the class prevalence estimations
@@ -25,6 +27,7 @@ def artificial_sampling_prediction(
     :param n_jobs: number of jobs to be run in parallel
     :param random_seed: allows to replicate the samplings. The seed is local to the method and does not affect
     any other random process.
+    :param verbose: if True, shows a progress bar
     :return: two ndarrays of [m,n] with m the number of samples (n_prevpoints*n_repetitions) and n the
      number of classes. The first one contains the true prevalences for the samples generated while the second one
      containing the the prevalences estimations
@@ -36,15 +39,12 @@ def artificial_sampling_prediction(
     if isinstance(model, AggregativeQuantifier):
         quantification_func = model.aggregate
         if isinstance(model, AggregativeProbabilisticQuantifier):
-            print('\tpreclassifying with soft')
             preclassified_instances = model.posterior_probabilities(test.instances)
         else:
-            print('\tpreclassifying with hard')
             preclassified_instances = model.classify(test.instances)
         test = LabelledCollection(preclassified_instances, test.labels)
     else:
         quantification_func = model.quantify
-        print('not an aggregative')
 
     def _predict_prevalences(index):
         sample = test.sampling_from_index(index)
@@ -52,8 +52,9 @@ def artificial_sampling_prediction(
         estim_prevalence = quantification_func(sample.instances)
         return true_prevalence, estim_prevalence
 
+    pbar = tqdm(indexes, desc='[artificial sampling protocol] predicting') if verbose else indexes
     results = Parallel(n_jobs=n_jobs)(
-        delayed(_predict_prevalences)(index) for index in tqdm(indexes, desc='[artificial sampling protocol] predicting')
+        delayed(_predict_prevalences)(index) for index in pbar
     )
 
     true_prevalences, estim_prevalences = zip(*results)

quapy/method/aggregative.py

@@ -3,12 +3,13 @@ from copy import deepcopy
 import functional as F
 import error
 from method.base import BaseQuantifier
-from quapy.classification.svmperf import SVMperf
+from classification.svmperf import SVMperf
-from quapy.data import LabelledCollection
+from data import LabelledCollection
 from sklearn.metrics import confusion_matrix
 from sklearn.calibration import CalibratedClassifierCV
 from joblib import Parallel, delayed
 from abc import abstractmethod
+from typing import Union
 
 
 # Abstract classes
@@ -77,13 +78,19 @@ class AggregativeProbabilisticQuantifier(AggregativeQuantifier):
         self.learner.set_params(**parameters)
 
 
+class BinaryQuantifier(BaseQuantifier):
+    def _check_binary(self, data : LabelledCollection, quantifier_name):
+        assert data.binary, f'{quantifier_name} works only on problems of binary classification. ' \
+                            f'Use the class OneVsAll to enable {quantifier_name} work on single-label data.'
+
+
 # Helper
 # ------------------------------------
 def training_helper(learner,
                     data: LabelledCollection,
                     fit_learner: bool = True,
                     ensure_probabilistic=False,
-                    train_val_split=None):
+                    val_split:Union[LabelledCollection, float]=None):
     """
     Training procedure common to all Aggregative Quantifiers.
     :param learner: the learner to be fit
@@ -91,7 +98,9 @@ def training_helper(learner,
     :param fit_learner: whether or not to fit the learner (if False, then bypasses any action)
     :param ensure_probabilistic: if True, guarantees that the resulting classifier implements predict_proba (if the
     learner is not probabilistic, then a CalibratedCV instance of it is trained)
-    :param train_val_split: if specified, indicates the proportion of training instances on which to fit the learner
+    :param val_split: if specified as a float, indicates the proportion of training instances that will define the
+    validation split (e.g., 0.3 for using 30% of the training set as validation data); if specified as a
+    LabelledCollection, represents the validation split itself
     :return: the learner trained on the training set, and the unused data (a _LabelledCollection_ if train_val_split>0
     or None otherwise) to be used as a validation set for any subsequent parameter fitting
     """
@@ -101,10 +110,17 @@ def training_helper(learner,
                 print(f'The learner {learner.__class__.__name__} does not seem to be probabilistic. '
                       f'The learner will be calibrated.')
                 learner = CalibratedClassifierCV(learner, cv=5)
-        if train_val_split is not None:
+        if val_split is not None:
-            if not (0 < train_val_split < 1):
+            if isinstance(val_split, float):
-                raise ValueError(f'train/val split {train_val_split} out of range, must be in (0,1)')
+                if not (0 < val_split < 1):
-            train, unused = data.split_stratified(train_prop=train_val_split)
+                    raise ValueError(f'train/val split {val_split} out of range, must be in (0,1)')
+                train, unused = data.split_stratified(train_prop=1-val_split)
+            elif isinstance(val_split, LabelledCollection):
+                train = data
+                unused = val_split
+            else:
+                raise ValueError('train_val_split not understood; use either a float indicating the split proportion, '
+                                 'or a LabelledCollection indicating the validation split')
         else:
             train, unused = data, None
         learner.fit(train.instances, train.labels)
@@ -148,8 +164,17 @@ class AdjustedClassifyAndCount(AggregativeQuantifier):
     def __init__(self, learner):
         self.learner = learner
 
-    def fit(self, data: LabelledCollection, fit_learner=True, train_val_split=0.6):
+    def fit(self, data: LabelledCollection, fit_learner=True, val_split:Union[float, LabelledCollection]=0.3):
-        self.learner, validation = training_helper(self.learner, data, fit_learner, train_val_split=train_val_split)
+        """
+        Trains a ACC quantifier
+        :param data: the training set
+        :param fit_learner: set to False to bypass the training (the learner is assumed to be already fit)
+        :param val_split: either a float in (0,1) indicating the proportion of training instances to use for
+         validation (e.g., 0.3 for using 30% of the training set as validation data), or a LabelledCollection
+         indicating the validation set itself
+        :return: self
+        """
+        self.learner, validation = training_helper(self.learner, data, fit_learner, val_split=val_split)
         self.cc = ClassifyAndCount(self.learner)
         y_ = self.classify(validation.instances)
         y  = validation.labels
@@ -196,9 +221,18 @@ class ProbabilisticAdjustedClassifyAndCount(AggregativeProbabilisticQuantifier):
     def __init__(self, learner):
         self.learner = learner
 
-    def fit(self, data: LabelledCollection, fit_learner=True, train_val_split=0.6):
+    def fit(self, data: LabelledCollection, fit_learner=True, val_split:Union[float, LabelledCollection]=0.3):
+        """
+        Trains a PACC quantifier
+        :param data: the training set
+        :param fit_learner: set to False to bypass the training (the learner is assumed to be already fit)
+        :param val_split: either a float in (0,1) indicating the proportion of training instances to use for
+         validation (e.g., 0.3 for using 30% of the training set as validation data), or a LabelledCollection
+         indicating the validation set itself
+        :return: self
+        """
         self.learner, validation = training_helper(
-            self.learner, data, fit_learner, ensure_probabilistic=True, train_val_split=train_val_split
+            self.learner, data, fit_learner, ensure_probabilistic=True, val_split=val_split
         )
         self.pcc = ProbabilisticClassifyAndCount(self.learner)
         y_ = self.classify(validation.instances)
@@ -262,7 +296,7 @@ class ExpectationMaximizationQuantifier(AggregativeProbabilisticQuantifier):
         return qs
 
 
-class HellingerDistanceY(AggregativeProbabilisticQuantifier):
+class HellingerDistanceY(AggregativeProbabilisticQuantifier, BinaryQuantifier):
     """
     Implementation of the method based on the Hellinger Distance y (HDy) proposed by
     González-Castro, V., Alaiz-Rodrı́guez, R., and Alegre, E. (2013). Class distribution
@@ -272,11 +306,19 @@ class HellingerDistanceY(AggregativeProbabilisticQuantifier):
     def __init__(self, learner):
         self.learner = learner
 
-    def fit(self, data: LabelledCollection, fit_learner=True, train_val_split=0.6):
+    def fit(self, data: LabelledCollection, fit_learner=True, val_split:Union[float, LabelledCollection]=0.3):
-        assert data.binary, f'{self.__class__.__name__} works only on problems of binary classification. ' \
+        """
-                            f'Use the class OneVsAll to enable {self.__class__.__name__} work on single-label data.'
+        Trains a HDy quantifier
+        :param data: the training set
+        :param fit_learner: set to False to bypass the training (the learner is assumed to be already fit)
+        :param val_split: either a float in (0,1) indicating the proportion of training instances to use for
+         validation (e.g., 0.3 for using 30% of the training set as validation data), or a LabelledCollection
+         indicating the validation set itself
+        :return: self
+        """
+        self._check_binary(data, self.__class__.__name__)
         self.learner, validation = training_helper(
-            self.learner, data, fit_learner, ensure_probabilistic=True, train_val_split=train_val_split)
+            self.learner, data, fit_learner, ensure_probabilistic=True, val_split=val_split)
         Px = self.posterior_probabilities(validation.instances)
         self.Pxy1 = Px[validation.labels == 1]
         self.Pxy0 = Px[validation.labels == 0]
@@ -312,7 +354,7 @@ class HellingerDistanceY(AggregativeProbabilisticQuantifier):
         return np.sqrt(np.sum((np.sqrt(P) - np.sqrt(Q))**2))
 
 
-class ExplicitLossMinimisation(AggregativeQuantifier):
+class ExplicitLossMinimisation(AggregativeQuantifier, BinaryQuantifier):
 
     def __init__(self, svmperf_base, loss, **kwargs):
         self.svmperf_base = svmperf_base
@@ -320,8 +362,7 @@ class ExplicitLossMinimisation(AggregativeQuantifier):
         self.kwargs = kwargs
 
     def fit(self, data: LabelledCollection, fit_learner=True, *args):
-        assert data.binary, f'{self.__class__.__name__} works only on problems of binary classification' \
+        self._check_binary(data, self.__class__.__name__)
-                            f'Use the class OneVsAll to enable {self.__class__.__name__} work on single-label data.'
         assert fit_learner, 'the method requires that fit_learner=True'
         self.learner = SVMperf(self.svmperf_base, loss=self.loss, **self.kwargs).fit(data.instances, data.labels)
         return self
@@ -386,6 +427,9 @@ class OneVsAll(AggregativeQuantifier):
             f'{self.__class__.__name__} expect non-binary data'
         assert isinstance(self.binary_quantifier, BaseQuantifier), \
             f'{self.binary_quantifier} does not seem to be a Quantifier'
+        if not isinstance(self.binary_quantifier, BinaryQuantifier):
+            raise ValueError(f'{self.binary_quantifier.__class__.__name__} does not seem to be an instance of '
+                             f'{BinaryQuantifier.__class__.__name__}')
         self.dict_binary_quantifiers = {c: deepcopy(self.binary_quantifier) for c in data.classes_}
         self.__parallel(self._delayed_binary_fit, data, **kwargs)
         return self

quapy/util.py

@@ -6,6 +6,7 @@ import numpy as np
 import urllib
 import os
 from pathlib import Path
+import pickle
 
 
 def get_parallel_slices(n_tasks, n_jobs=-1):
@@ -58,4 +59,19 @@ def create_if_not_exist(path):
 
 
 def get_quapy_home():
-    return os.path.join(str(Path.home()), 'quapy_data')
+    home = os.path.join(str(Path.home()), 'quapy_data')
+    os.makedirs(home, exist_ok=True)
+    return home
+
+
+def pickled_resource(pickle_path:str, generation_func:callable, *args):
+    if pickle_path is None:
+        return generation_func(*args)
+    else:
+        if os.path.exists(pickle_path):
+            return pickle.load(open(pickle_path, 'rb'))
+        else:
+            instance = generation_func(*args)
+            os.makedirs(str(Path(pickle_path).parent), exist_ok=True)
+            pickle.dump(instance, open(pickle_path, 'wb'), pickle.HIGHEST_PROTOCOL)
+            return instance

test.py

@@ -3,11 +3,13 @@ from sklearn.svm import LinearSVC
 import quapy as qp
 import quapy.functional as F
 import sys
+import numpy as np
 
 #qp.datasets.fetch_reviews('hp')
 #qp.datasets.fetch_twitter('sst')
 
 #sys.exit()
+from model_selection import GridSearchQ
 
 SAMPLE_SIZE=500
 binary = False
@@ -17,27 +19,31 @@ if binary:
     dataset = qp.datasets.fetch_reviews('kindle', tfidf=True, min_df=5)
 
 else:
-    dataset = qp.datasets.fetch_twitter('semeval13', model_selection=False, min_df=10)
+    dataset = qp.datasets.fetch_twitter('hcr', for_model_selection=False, min_df=10, pickle=True)
-    dataset.training = dataset.training.sampling(SAMPLE_SIZE, 0.2, 0.5, 0.3)
+    # dataset.training = dataset.training.sampling(SAMPLE_SIZE, 0.2, 0.5, 0.3)
 
 print('dataset loaded')
 
 # training a quantifier
-learner = LogisticRegression()
+learner = LogisticRegression(max_iter=1000)
 # model = qp.method.aggregative.ClassifyAndCount(learner)
-# model = qp.method.aggregative.AdjustedClassifyAndCount(learner)
+model = qp.method.aggregative.AdjustedClassifyAndCount(learner)
 # model = qp.method.aggregative.ProbabilisticClassifyAndCount(learner)
 # model = qp.method.aggregative.ProbabilisticAdjustedClassifyAndCount(learner)
 # model = qp.method.aggregative.ExpectationMaximizationQuantifier(learner)
 # model = qp.method.aggregative.ExplicitLossMinimisationBinary(svmperf_home, loss='q', C=100)
-model = qp.method.aggregative.SVMQ(svmperf_home, C=1)
+# model = qp.method.aggregative.SVMQ(svmperf_home, C=1)
 
-if not binary:
+if not binary and isinstance(model, qp.method.aggregative.BinaryQuantifier):
     model = qp.method.aggregative.OneVsAll(model)
 
-print('fitting model')
-model.fit(dataset.training)
 
+# Model fit and Evaluation on the test data
+# ----------------------------------------------------------------------------
+
+print(f'fitting model {model.__class__.__name__}')
+train, val = dataset.training.split_stratified(0.6)
+model.fit(train, val_split=val)
 
 # estimating class prevalences
 print('quantifying')
@@ -47,14 +53,15 @@ prevalences_true  = dataset.test.prevalence()
 # evaluation (one single prediction)
 error = qp.error.mae(prevalences_true, prevalences_estim)
 
-print(f'method {model.__class__.__name__}')
-
 print(f'Evaluation in test (1 eval)')
 print(f'true prevalence {F.strprev(prevalences_true)}')
 print(f'estim prevalence {F.strprev(prevalences_estim)}')
 print(f'mae={error:.3f}')
 
 
+# Model fit and Evaluation according to the artificial sampling protocol
+# ----------------------------------------------------------------------------
+
 max_evaluations = 5000
 n_prevpoints = F.get_nprevpoints_approximation(combinations_budget=max_evaluations, n_classes=dataset.n_classes)
 n_evaluations = F.num_prevalence_combinations(n_prevpoints, dataset.n_classes)
@@ -70,3 +77,30 @@ for error in qp.error.QUANTIFICATION_ERROR:
     score = error(true_prev, estim_prev)
     print(f'{error.__name__}={score:.5f}')
 
+
+# Model selection and Evaluation according to the artificial sampling protocol
+# ----------------------------------------------------------------------------
+
+param_grid = {'C': np.logspace(-3,3,7), 'class_weight': ['balanced', None]}
+
+model_selection = GridSearchQ(model,
+                              param_grid=param_grid,
+                              sample_size=SAMPLE_SIZE,
+                              eval_budget=max_evaluations//10,
+                              error='mae',
+                              refit=True,
+                              verbose=True)
+
+# model = model_selection.fit(dataset.training, validation=0.3)
+model = model_selection.fit(train, validation=val)
+print(f'Model selection: best_params = {model_selection.best_params_}')
+print(f'param scores:')
+for params, score in model_selection.param_scores_.items():
+    print(f'\t{params}: {score:.5f}')
+
+true_prev, estim_prev = qp.evaluation.artificial_sampling_prediction(model, dataset.test, SAMPLE_SIZE, n_prevpoints)
+
+print(f'After model selection: Evaluation according to the artificial sampling protocol ({len(true_prev)} evals)')
+for error in qp.error.QUANTIFICATION_ERROR:
+    score = error(true_prev, estim_prev)
+    print(f'{error.__name__}={score:.5f}')