Optimizing TF-IDF schemes

An example of optimizing TF-IDF weighting schemes using 5 fold cross-validation

from __future__ import print_function

import os
from itertools import product

import numpy as np
from sklearn.datasets import fetch_20newsgroups
from sklearn.linear_model import LogisticRegression
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.pipeline import Pipeline

from freediscovery.feature_weighting import SmartTfidfTransformer

rng = np.random.RandomState(34)

We load and vectorize 2 classes from the 20 newsgroup dataset,

newsgroups = fetch_20newsgroups(subset='train',
                                categories=['sci.space', 'comp.graphics'])
vectorizer = CountVectorizer(stop_words='english')
X = vectorizer.fit_transform(newsgroups.data)

then compute baseline categorization performance using Logistic Regression and the TF-IDF transfomer from scikit-learn

X_tfidf = TfidfTransformer().fit_transform(X)

X_train, X_test, y_train, y_test = train_test_split(X, newsgroups.target,
                                                    random_state=rng)

pipe = Pipeline(steps=[('tfidf', TfidfTransformer()),
                       ('logisticregression', LogisticRegression())])

pipe.fit(X_train, y_train)
print('Baseline TF-IDF categorization accuracy: {:.3f}'
      .format(pipe.score(X_test, y_test)))

Out:

Baseline TF-IDF categorization accuracy: 0.973

Next, we search, using 5 fold cross-validation, for the best TF-IDF weighting scheme among the 80+ combinations supported by SmartTfidfTransformer. Two hyper-parameters are worth optimizing in this case,

• weighting parameter that defines the TF-IDF weighting (see the TF-IDF schemes user manual section for more details)
• norm_alpha is the α parameter in the pivoted normalization when weighting=="???p".

To reduce the search parameter space in this example, we also can exclude the case when either the term weighting, feature weighing or normalization is not used as it expected to yield worse than baseline performance. We also exclude the non smoothed IDF weightings (?t?, ?p?) since thay return NaNs when some of the document frequency is 0 (which will be the case during cross-validation). Finally, by noticing that the case xxxp with norm_alpha=1.0 corresponds to the weighing xxx (i.e. with pivoted normalization disabled) we can reduce the search space even further.

pipe = Pipeline(steps=[('tfidf', SmartTfidfTransformer()),
                       ('logisticregression', LogisticRegression())])

param_grid = {'tfidf__weighting': ["".join(el) + 'p'
                                   for el in product('labLd', 'sd',
                                                     "clu")],
              'tfidf__norm_alpha': np.linspace(0, 1, 10)}

pipe_cv = GridSearchCV(pipe,
                       param_grid=param_grid,
                       verbose=1,
                       n_jobs=(1 if os.name == 'nt' else -1),
                       cv=5)
pipe_cv.fit(X_train, y_train)
print('Best CV params: weighting={weighting}, norm_alpha={norm_alpha:.3f} '
      .format(**pipe_cv.best_estimator_.steps[0][1].get_params()))
print('Best TF-IDF categorization accuracy: {:.3f}'
      .format(pipe_cv.score(X_test, y_test)))

Out:

Fitting 5 folds for each of 300 candidates, totalling 1500 fits
Best CV params: weighting=lsup, norm_alpha=0.778
Best TF-IDF categorization accuracy: 0.990

In this example, by tuning TF-IDF weighting scheme with pivoted normalization, we obtain a categorization accuracy score of 0.99 as compared to a baseline TF-IDF score of 0.973. It is also interesting to notice that the best weighting hyper-parameter in this case is lnup which corresponds to the “unique pivoted normalization” case proposed by Singhal et al. (1996), although with a different α value.

Total running time of the script: ( 0 minutes 48.113 seconds)