See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/266141433

News Articles Classification Using Random Forests and Weighted Multimodal

Features

Conference Paper · November 2014

DOI: 10.1007/978-3-319-12979-2_6

CITATIONS READS

63 1,522

5 authors, including:

Dimitris Liparas Yaakov HaCohen-Kerner

Intrasoft International Jerusalem College of Technology
25 PUBLICATIONS 420 CITATIONS 108 PUBLICATIONS 1,086 CITATIONS

SEE PROFILE SEE PROFILE

Anastasia Moumtzidou Stefanos Vrochidis

The Centre for Research and Technology, Hellas The Centre for Research and Technology, Hellas
92 PUBLICATIONS 764 CITATIONS 389 PUBLICATIONS 3,309 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Yaakov HaCohen-Kerner on 29 October 2014.

The user has requested enhancement of the downloaded file.

This is a draft version. The final version is available at
http://link.springer.com/chapter/10.1007/978-3-319-12979-2_6

News articles classification using Random Forests and

weighted multimodal features

Dimitris Liparas1, Yaakov HaCohen-Kerner2, Anastasia Moumtzidou1,

Stefanos Vrochidis1, Ioannis Kompatsiaris1
1
Information Technologies Institute, Centre for Research and Technology Hellas, Thermi-
Thessaloniki, Greece
{dliparas, moumtzid, stefanos, ikom}@iti.gr
2
Dept. of Computer Science, Jerusalem College of Technology – Lev Academic Center,
21 Havaad Haleumi St., P.O.B. 16031, 9116001 Jerusalem, Israel
[email protected]

Abstract. This research investigates the problem of news articles classification.

The classification is performed using N-gram textual features extracted from
text and visual features generated from one representative image. The applica-
tion domain is news articles written in English that belong to four categories:
Business-Finance, Lifestyle-Leisure, Science-Technology and Sports down-
loaded from three well-known news web-sites (BBC, Reuters, and TheGuardi-
an). Various classification experiments have been performed with the Random
Forests machine learning method using N-gram textual features and visual fea-
tures from a representative image. Using the N-gram textual features alone led
to much better accuracy results (84.4%) than using the visual features alone
(53%). However, the use of both N-gram textual features and visual features led
to slightly better accuracy results (86.2%). The main contribution of this work is
the introduction of a news article classification framework based on Random
Forests and multimodal features (textual and visual), as well as the late fusion
strategy that makes use of Random Forests operational capabilities.

Keywords: Document classification, Supervised learning, Multimodal, News

articles, N-gram features, Random Forests, Visual features, Fusion

1 Introduction
The number of news articles published on various websites in general and news
websites in particular had a dramatic increase over the last years. These articles
contain multimodal information including textual and visual (image and video)
descriptions. A visual example of such articles is illustrated in Fig. 1. Nowadays both
journalists and media monitoring companies face the problem of mastering large
amounts of articles in order to identify important topics and events all around the
world. Therefore there is an important need for accurate and rapid clustering and
classification of news articles into a set of categories in order to support journalism
and media monitoring tasks. Despite of the multimodal nature of the articles posted in
the web nowadays, most of the approaches consider only textual data in order to
achieve classification (e.g. [1, 2]). Therefore there is an interesting challenge to
investigate whether the combination of use of visual features in addition to the textual
features will improve the classification accuracy.

Fig. 1. Web-based news article from BBC entitled: 2013: The year we all went “mobile”1
News article classification is considered a Document classification (DC) problem.
DC means labeling a document with predefined categories. This can be achieved as
the supervised learning task of assigning documents to one or more predefined
categories [3]. Using machine learning (ML), the goal is to learn classifiers from
examples which perform the category classifications automatically. DC is applied in
many tasks, such as: clustering, document indexing, document filtering, information
retrieval, information extraction and word sense disambiguation. Current-day DC for
news articles poses several research challenges, due to the large number of
multimodal features present in the document set and their dependencies.
In this research, we investigate the task of category-based classification of news ar-
ticles using a combination of visual features and textual N-gram features. The textual
features are extracted from the textual part of the news article, while the visual fea-
tures are generated from the biggest image of the article. Specifically we “learn” two
Random Forest classifiers with textual and visual features and their results are fused
using a late fusion strategy. The main contribution of this work is the usage of visual
features in news article classification in order to leverage the text-based results, as
well as the late fusion strategy that makes use of Random Forests’ operational capa-
bilities (i.e. out-of-bag (OOB) error estimate and proximity ratios).
The rest of this paper is organized as follows: Section 2 presents the theoretical
background and related work. Section 3 describes the textual and visual feature ex-
traction procedure, while section 4 introduces the proposed classification framework.
Section 5 presents the results of the experiments and section 6 concludes the paper.

1
http://www.bbc.com/news/business-25445906
2 Related work and theoretical background
Since in this study we are dealing with supervised machine learning in Document
Classification (DC) in general and with news articles classification in particular, we
report previous work related to these two fields. Furthermore, since Random Forests
(RF) is the machine learning method we use for our proposed classification frame-
work, we provide the theoretical background and related work for this method.

2.1 Document and news articles classification

Several machine learning algorithms have been applied to DC, e.g.: Naïve Bayes [1],
Rocchio [2], Logistic regression and Log-linear models [4], SVM [5] and Random
Forests [6]. All these studies make use of textual features. The Reuters-21578 dataset,
which contains news articles, was for a long time the standard resource dataset for
text classification. Various text classification studies on the Reuters-21578 dataset
make use of N-grams (N>1) in addition to unigrams. For example, [7] investigate the
usefulness of bigrams for document indexing in text categorization (TC), inde-
pendently of any specific learning algorithm. Their experiments reveal that if the fea-
ture evaluation function being used gives rise to a too high bigram penetration level,
effectiveness may decrease. This finding is probably due to the elimination of in-
formative unigrams on the part of bigrams that partly duplicate the information car-
ried by existing unigrams. Moreover, [8] investigate text classification using semi-
supervised ML methods for unlabeled documents. They apply two semi-supervised
algorithms on several text datasets. Their results do not indicate improvement by
combining unigrams and bigrams.
Some DC-related studies deal specifically with documents in web page format. For
instance, [9] employ Neural Networks and Principal Component Analysis for web
page feature selection and classification. Furthermore, in [10] a Random Forests clas-
sifier is employed for multi-category web page classification. Again, it is important to
mention that the aforementioned studies deal with textual features, while our ap-
proach leverages the performance of textual features by using visual features from
representative images.
In addition to the aforementioned works, several research attempts dealt specifical-
ly with news article classification. In this context, [11] have applied a maximum en-
tropy classifier on unigram features to detect emotional charge around links to news
articles in posts from political weblogs. In another work, [12] apply both regression
and classification algorithms that make use of social interaction features to the task of
predicting the popularity of news items on the social web. They demonstrate that one
of the most important predictors of popularity is the source of the article. Finally, [13]
utilize news articles streams and Bayesian text classification in order to classify con-
tents of interest related to geographic communities. These works rely mainly upon
textual data for classification, while the proposed approach leverages the results of
text-based classification by considering visual features.
2.2 Random Forests
Random Forests (RF) is an ensemble learning method for classification and regression
[14]. The basic notion of the methodology is the construction of a group of decision
trees. RF employs two sources of randomness in its operational procedures:
1. Each decision tree is grown on a different bootstrap sample drawn randomly
from the training data.
2. At each node split during the construction of a decision tree, a random subset
of m variables is selected from the original variable set and the best split
based on these m variables is used.
For an unknown case, the predictions of the trees that are constructed by the RF are
aggregated (majority voting for classification / averaging for regression). For a RF
consisting of N trees, the following equation is used for predicting the class label l of
a case y through majority voting:
𝑁

𝑙(𝑦) = 𝑎𝑟𝑔𝑚𝑎𝑥𝑐 (∑ 𝐼ℎ𝑛 (𝑦)=𝑐 ) (1)

𝑛=1
where I the indicator function and hn the nth tree of the RF.
RF has an internal mechanism that provides an estimation of its generalization er-
ror, called out-of-bag (OOB) error estimate. For the construction of each tree, only
2/3 of the original data’s cases are used in that particular bootstrap sample. The rest
1/3 of the instances (OOB data) are classified by the constructed tree and therefore,
used for testing its performance. The OOB error estimate is the averaged prediction
error for each training case y, using only the trees that do not include y in their boot-
strap sample. Furthermore, when a RF is constructed, all the training cases are put
down each tree and a proximity matrix between the cases is computed, based on
whether pairs of cases end up in the same terminal node of a tree.
Successful applications of RF to an extensive range of disciplines (apart from DC
and web page classification) can be found in the relevant literature. Among others,
image classification [15], network intrusion detection [16] and neuroimaging [17] can
be listed.
In this study, we investigate the application of RF for news articles classification.
Although the RF have been successfully applied to several classification problems (as
discussed above), to the best of our knowledge they haven’t been applied to news
article classification problems. Moreover, an important motivation for using RF was
the application of late fusion strategies based on the RF operational capabilities.

3 Feature Extraction from News Articles Documents

In this work, we assume that the article has two main parts: a) the textual description
and b) the images (e.g. Fig. 1). First we extract N-grams (globally and not per catego-
ry) from the textual description. N-grams were chosen as our textual features because
they were found as relatively easy to compute and effective for various classification
tasks (e.g. [7, 8, 18]). Then, we select the biggest image of the article and extract vis-
ual features. In this case we assume that the biggest image was the representative one.
3.1 N-gram Textual Features
For the extraction of the textual features from a news article web document, the fol-
lowing procedure is applied:
1. All appearances of 421 stopwords for general texts in English are deleted
[19].
2. All possible continuous N-gram words (for N =1, 2, 3, 4) are created, pro-
vided that the all the words in a certain N-gram are in the same sentence.
3. The frequency of each N-gram feature in the corpora is counted.
4. The unigram, bigram, trigram and fourgram (each group alone) features are
sorted in descending order.
To avoid unnecessarily large number of N-grams, only a subset of the most fre-
quent features from each group is selected. More specifically, in our study 195 of the
most frequent N-gram features are selected as follows: a) 100 most frequent uni-
grams; b) 50 most frequent bigrams; c) 30 most frequent trigrams; d) 15 most fre-
quent fourgrams. The motivation for these numbers is as follows: The larger the value
of N is, the smaller the number of relatively frequent N-grams in the corpus is. The
reduction factor was determined to be approximately 2.

3.2 Visual Features

The low-level visual features that are extracted in order to capture the characteristics
of images are the MPEG-7 visual descriptors. The MPEG-7 standard specifies a set of
descriptors, each defining the syntax and the semantics of an elementary visual low-
level feature. Each descriptor aims at capturing different aspects of human perception
(i.e., color, texture and shape). In this work, five MPEG-7 visual descriptors capturing
color and texture aspects of human perception are extracted [20]:
1. Color Layout Descriptor: captures the spatial distribution of color or an ar-
bitrary-shaped region.
2. Color Structure Descriptor: is based on color histograms, but aims at iden-
tifying localized color distributions.
3. Scalable Color Descriptor: is a Haar-transform based encoding scheme that
measures color distribution over an entire image.
4. Edge Histogram Descriptor: captures the spatial distribution of edges and
it involves division of image into 16 non-overlapping blocks. Edge infor-
mation is then calculated for each block.
5. Homogenous Texture Descriptor: is based on a filter bank approach em-
ploying scale and orientation sensitive filters.
Then, we apply an early fusion approach, which involves the concatenation of all
the aforementioned descriptors into a single feature vector. In this study, 320 visual
features are extracted in total. The number of features/dimensions that are created
from each descriptor are the following: a) Color Layout Descriptor: 18 fea-
tures/dimensions; b) Color Structure Descriptor: 32 features/dimensions; c) Scalable
Color Descriptor: 128 features/dimensions; d) Edge Histogram Descriptor: 80 fea-
tures/dimensions; e) Homogeneous Texture Descriptor: 62 features/dimensions.
4 Proposed Classification Framework
The flowchart of the proposed classification framework (training phase) is depicted in
Fig. 2. Next, the different steps of the framework are described in detail.
First, all the necessary data is collected in the form of text from news article web
pages, as well as images associated to each web page. In the following step, the pro-
cedures described in Section 3 are applied to the raw data (Parsing / Extraction). In
this way, the visual and textual features are generated. We note that given the fact
that the majority of the web pages contain several images including banners and ad-
vertisement logos, it was decided to keep only the biggest image of each site which
would most probably be the main image of the article. One other important thing to
note is that in this study the features of each modality are treated independently.
Hence, two different feature vectors (one for each modality) are formulated.
In the training phase, the feature vectors from each modality are used as input for
the construction of a RF. From the two constructed RFs (one for the textual and one
for the visual features), we compute the weights for each modality, in order to ap-
ply a late fusion strategy and formulate the final RF predictions. In this study, two
different approaches for the computation of the modality weights are followed:
1. From the OOB error estimate of each modality’s RF, the corresponding
OOB accuracy values are computed. These values are computed for each
class separately. Then, the values are normalized (by dividing them by their
sum) and serve as weights for the two modalities.
2. For the second weighting strategy, the same procedure as in 1. is applied.
However, instead of employing the OOB accuracy values from each RF, the
ratio values between the inner-class and the intra-class proximities (for each
class) are used [21]. First, for each RF the proximity matrix between all
pairs of data cases P={pij,i,j =1, …,w} (w=number of data cases) is con-
structed and then, the aforementioned ratio values are computed as in the
following equation:
𝑃𝑖𝑛𝑛𝑒𝑟 (2)
𝑅=
𝑃𝑖𝑛𝑡𝑟𝑎
where
𝑤
(3)
𝑃𝑖𝑛𝑛𝑒𝑟 = ∑ 𝑝𝑖𝑗 (𝑖𝑓 𝑙𝑖 = 𝑙𝑗 )
𝑖,𝑗=1
𝑤

𝑃𝑖𝑛𝑡𝑟𝑎 = ∑ 𝑝𝑖𝑗 (𝑖𝑓 𝑙𝑖 ≠ 𝑙𝑗 ) (4)

𝑖,𝑗=1
and li, lj the class labels of cases i and j, respectively.
During the testing phase, when the RF predicts a case, it outputs probability esti-
mates per class for that case. The probability outputs Pt and Pv from the textual and
visual RFs respectively are multiplied by their corresponding modality weights Wt and
Wv and summed to produce the final RF predictions as in the following equation:
𝑃𝑓𝑢𝑠𝑒𝑑 = 𝑊𝑡 𝑃𝑡 + 𝑊𝑣 𝑃𝑣 (5)
Fig. 2. Flowchart of proposed classification framework

5 Experimental Results
5.1 Dataset Description
The experiments are realized on a dataset that contains web pages from three well
known News Web Sites, namely BBC, The Guardian and Reuter. Overall, 651, 556
and 360 web pages have been retrieved from each site, respectively. At this point it
should be noted that the manual annotation of the web pages was necessary, regard-
less of the fact that in the three News Web Sites descriptions about the topic of each
web page are provided, since in many cases the descriptions are inconsistent with the
content of the web pages. The manual annotation was realized for a subset of the top-
ics recognized by the IPTC news codes taxonomy 2, which is the global standards
body of the news media. Specifically, we selected the most important topics with the
guidance of media monitoring experts and journalists. Table 1 contains a detailed
description of the final dataset3 and the topics considered.
Table 1. Details of dataset
Topics Num. of
Business, Lifestyle, Science,
Sports documents
finance leisure technology
News Sites per site
BBC 102 68 75 202 447
The Guardian 67 59 116 96 338
Reuter 165 7 29 57 258
Num. of documents per
334 134 220 355 1043
topic

2
http://www.iptc.org/site/Home/
3
The dataset is publicly available at: http://mklab.iti.gr/files/ArticlesNewsSitesData.7z
5.2 Experimental Setup
We randomly split our dataset into training and test sets in order to conduct the exper-
iments. Approximately 2/3 of the cases are kept for training purposes, whereas the
rest (1/3) are used as test set, in order to estimate the classification scheme’s perfor-
mance.
As for the RF parameters that we use in the experiments, we opted to apply the fol-
lowing setting: We set the number of trees for the construction of each RF based on
the OOB error estimate. After several experiments with different numbers of trees, we
noticed that the OOB error estimate was stabilized after using 1000 trees and no long-
er improved. Hence, the number of trees is set to N=1000. For each node split during
the growing of a tree, the number of the subset of variables used to determine the best
split is set to 𝑚 = √k (according to [14]), where k is the total number of features of
the dataset.
Finally, for the evaluation of the performance of the proposed methodology, we
compute the precision, recall and F-score measures for each category, along with their
corresponding macro-averaged values, as well as the accuracy on the entire test set
(all categories included).

5.3 Results
The test set results from the application of RF to each modality separately are summa-
rized in Table 2. We are mainly interested in the values of F-score, since it considers
both precision and recall. We notice that the textual modality outperforms the visual
in all measures, both regarding each topic and the macro-averaged scores. This indi-
cates that textual data is a more reliable and solid source of information, in compari-
son to the visual data. More specifically:
 The RF trained with the textual data achieves a macro-averaged F-score val-
ue of 83.2%, compared to 45.5% for the visual modality
 The accuracy for the textual modality RF is 84.4%, while the visual modality
RF achieves only 53%
 The worst results for the visual data RF are attained for the topics “Lifestyle-
Leisure” (recall 12% and F-score 20.7%) and “Science-Technology” (preci-
sion 45.3%, recall 38.7% and F-score 41.7%). However, the results regarding
the topic “Sports” are considered satisfactory. A possible explanation for this
is the fact that the images from the “Lifestyle-Leisure” web pages depict di-
verse topics and therefore their visual appearance strongly varies. On the
other hand, the images regarding the topic “Sports” contain rather specific
information such as football stadiums (a characteristic example is depicted in
Fig. 3).
 Table 2. Test set results from the application of RF to each modality

Modality Textual Visual

Topics Prec. Rec. F-score Prec. Rec. F-score

Business-
80.0% 87.3% 83.5% 56.3% 57.3% 56.8%
Finance

Lifestyle-
86.7% 78.0% 82.1% 75% 12% 20.7%
Leisure
Science-
79.1% 70.7% 74.6% 45.3% 38.7% 41.7%
Technology
Sports 91.3% 93.8% 92.5% 52.8% 76.8% 62.6%
Macro-
84.3% 82.5% 83.2% 57.4% 46.2% 45.5%
average
Accuracy 84.4% 53.0%

Fig. 3. Characteristic image from a “Sports” web page (left) 4, along with an image regarding a
web page from the “Lifestyle-Leisure” topic (right)5

In Table 3 we provide the test set results from the application of the late fusion
strategy to RF, using the two different weighting methods described in Section 4
(OOB error/ Proximity ratio). The weighting method regarding the proximity ratio
yields better performance results than the corresponding method for the OOB error.
More specifically:
 The accuracy of Textual + Visual (Proximity ratio) is slightly better than the
corresponding accuracy of Textual + Visual (OOB error) (86.2% compared
to 85.9%)
 The two weighted RFs achieve almost equal macro-averaged precision val-
ues (86.8% for Proximity ratio and 86.9% for OOB error), while regarding
the macro-averaged recall and F-score results, Textual + Visual (Proximity

4
http://www.bbc.com/sport/0/football/27897075-"_75602744_ochoa.jpg"
5
http://www.bbc.com/travel/feature/20140710-living-in-istanbul-“p022ktsw.jpg”
 ratio) is better (84.2% to 82.9% for the macro-averaged recall and 85.3% to
84.3% for the macro-averaged F-score)
For comparison purposes, we also constructed a fused RF model, where equal
weights were assigned to each modality. We notice that after following this weighting
approach (i.e. with equal weights), the performance of RF diminishes in all aspects.
The superiority of the weighting strategy based on the proximity ratio of each topic is
also evident in Fig. 4, where the macro-averaged F-score values of all 5 RF models
constructed in this study are sorted in ascending order. We observe that Textual +
Visual (Proximity ratio) is the best performing model among all cases.

Table 3. Test set results after the late fusion of RF regarding three different weighting schemes

Weighting Textual + Visual

Textual + Visual Textual + Visual
method (Weighting based on
(Weighting based on (Equal weights per
proximity ratio per
OOB error per topic) topic)
topic)
Topics Prec. Rec. F-score Prec. Rec. F-score Prec. Rec. F-score

Business-
80.3% 92.7% 86.1% 82.4% 89.1% 85.6% 71.1% 91.8% 80.1%
Finance
Lifestyle-
92.5% 74.0% 82.2% 92.9% 78.0% 84.8% 91.4% 64% 75.3%
Leisure
Science-
83.9% 69.3% 75.9% 81.4% 76.0% 78.6% 83.1% 65.3% 73.1%
Technology
Sports 90.7% 95.5% 93% 90.5% 93.8% 92.1% 87.4% 86.6% 87%
Macro-
86.9% 82.9% 84.3% 86.8% 84.2% 85.3% 83.3% 76.9% 78.9%
average
Accuracy 85.9% 86.2% 80.4%

Fig. 4. Macro-averaged F-score values for all RF models sorted in ascending order
6 Summary, Conclusions and Future Work
In this research, we investigate the use of N-gram textual and visual features for clas-
sification of news articles that fall into four categories (Business-Finance, Lifestyle-
Leisure, Science-Technology, and Sports) downloaded from three news web-sites
(BBC, Reuters, and TheGuardian).
Using the N-gram textual features alone led to much better accuracy results
(84.4%) than using the visual features alone (53%). However, the use of both N-gram
textual features and visual features (weighting based on proximity ratio per category)
led to slightly better accuracy results (86.2%).
Future directions for research are: (1) Defining and applying additional various
types of features such as: function words, key-phrases, morphological features (e.g.:
nouns, verbs and adjectives), quantitative features (various averages such as average
number of letters per a word, average number of words per a sentence) and syntactic
features (frequencies and distribution of parts of speech tags, such as: noun, verb,
adjective, adverb), (2) Applying various kinds of classification models based on tex-
tual and visual features for a larger number of documents that belong to more than
four categories in the news articles area, as well as in other areas, applications and
languages, (3) Selecting a representation for images based on visual concepts. In such
a case, the approach could consider more than one image per article. Visual concepts
could be extracted from each image and the average score for each visual concept
could be calculated, in order to represent the article based on multiple images.

Acknowledgments. The authors would like to acknowledge networking support by

the COST Action IC1002: Multilingual and multifaceted interactive information ac-
cess (MUMIA) and the COST Action IC1307: The European Network on Integrating
Vision and Language (iV&L Net). The authors would also like to thank Avi Rosen-
feld, Maor Tzidkani, Asaf Sabag and Daniel Nisim Cohen from Jerusalem College of
Technology, Lev Academic Center, for their assistance to the authors in providing the
software tool to generate the textual features used in this research.

References
1. Schneider, K. M.: Techniques for improving the performance of naive Bayes for text clas-
sification. In Computational Linguistics and Intelligent Text Processing, pp. 682-693,
Springer Berlin Heidelberg (2005)
2. Zeng, A., & Huang, Y.: A text classification algorithm based on rocchio and hierarchical
clustering. In Advanced Intelligent Computing (pp. 432-439), Springer Berlin Heidelberg
(2012)
3. Sebastiani, F.: Machine learning in automated text categorization. ACM Computing Sur-
veys, 34 (1), pp. 1-47 (2002)
4. Toutanova, K.: Competitive generative models with structure learning for NLP classifica-
tion tasks. In Proceedings of the 2006 Conference on Empirical Methods in Natural Lan-
guage Processing, (pp. 576-584) (2006)
5. Ho, A. K. N., Ragot, N., Ramel, J. Y., Eglin, V., & Sidere, N.: Document Classification in
a Non-stationary Environment: A One-Class SVM Approach. In Proceedings of the 2013
12th International Conference on Document Analysis and Recognition, (ICDAR), (pp.
616-620) (2013)
 6. Klassen, M., & Paturi, N.: Web document classification by keywords using random for-
ests. In Networked Digital Technologies, pp. 256-261, Springer Berlin Heidelberg (2010)
7. Caropreso, M. F., Matwin, S., & Sebastiani, F.: Statistical phrases in automated text cate-
gorization. Centre National de la Recherche Scientifique, Paris, France (2000)‫‏‬
8. Braga, I., Monard, M., & Matsubara, E.: Combining unigrams and bigrams in semi-
supervised text classification. In Proceedings of Progress in Artificial Intelligence, 14th
Portuguese Conference on Artificial Intelligence (EPIA 2009), Aveiro (pp. 489-500),‫‏‬
(2009)
9. Selamat, A., & Omatu, S.: Web page feature selection and classification using neural net-
works. Information Sciences, 158, pp. 69-88 (2004)
10. Aung, W. T., & Hla, K. H. M. S.: Random forest classifier for multi-category classification
of web pages. In Services Computing Conference, 2009. APSCC 2009. IEEE Asia-
Pacific, pp. 372-376, IEEE (2009)
11. Gamon, M., Basu, S., Belenko, D., Fisher, D., Hurst, M., & König, A. C.: BLEWS: Using
Blogs to Provide Context for News Articles. In ICWSM (2008)‫‏‬
12. Bandari, R., Asur, S., & Huberman, B. A.: The Pulse of News in Social Media: Forecast-
ing Popularity. In ICWSM (2012)
13. Swezey, R. M., Sano, H., Shiramatsu, S., Ozono, T., & Shintani, T.: Automatic detection
of news articles of interest to regional communities. IJCSNS, 12(6), 100 (2012)
14. Breiman, L.: Random Forests. In Machine Learning, 45(1), pp. 5-32 (2001)
15. Xu, B., Ye, Y., & Nie, L.: An improved random forest classifier for image classification.
In Information and Automation (ICIA), 2012 International Conference on (pp. 795-800),
IEEE (2012)
16. Li, W., & Meng, Y.: Improving the performance of neural networks with random forest in
detecting network intrusions. In Advances in Neural Networks–ISNN 2013 (pp. 622-
629), Springer Berlin Heidelberg (2013)
17. Gray, K. R., Aljabar, P., Heckemann, R. A., Hammers, A., & Rueckert, D.: Random for-
est-based similarity measures for multi-modal classification of Alzheimer's disease. Neu-
roImage, 65, pp. 167-175 (2013)
18. HaCohen-Kerner, Y., Mughaz, D., Beck, H., & Yehudai, E.: Words as Classifiers of Doc-
uments According to their Historical Period and the Ethnic Origin of their Au-
thors. Cybernetics and Systems: An International Journal, 39(3), pp. 213-228 (2008)
19. Fox, C.: A stop list for general text. ACM SIGIR Forum, 24 (1-2), ACM (1989)
20. Sikora, T.: The MPEG-7 visual standard for content description-an overview. IEEE Trans-
actions on Circuits and Systems for Video Technology, 11(6), pp. 696-702 (2001)
21. Zhou, Q., Hong, W., Luo, L., & Yang, F.: Gene selection using random forest and prox-
imity differences criterion on DNA microarray data. Journal of Convergence Information
Technology, 5(6), pp. 161-170 (2010)

View publication stats