Analyzing Social Media Discourse

An Approach using Semi-supervised Learning

Álvaro Figueira

and Luciana Oliveira

CRACS/INESC TEC, University of Porto, Rua do Campo Alegre, Porto, Portugal

CICE/ISCAP & INESC TEC, Polytechnic of Porto, Porto, Portugal

Keywords: Social Media Publications, Text Mining, Automatic Categorization, Higher Education Sector,

Strategic Benchmarking.

Abstract: The ability to handle large amounts of unstructured information, to optimize strategic business

opportunities, and to identify fundamental lessons among competitors through benchmarking, are essential

skills of every business sector. Currently, there are dozens of social media analytics’ applications aiming at

providing organizations with informed decision making tools. However, these applications rely on

providing quantitative information, rather than qualitative information that is relevant and intelligible for

managers. In order to address these aspects, we propose a semi-supervised learning procedure that discovers

and compiles information taken from online social media, organizing it in a scheme that can be strategically

relevant. We illustrate our procedure using a case study where we collected and analysed the social media

discourse of 43 organizations operating on the Higher Public Polytechnic Education Sector. During the

analysis we created an “editorial model” that characterizes the posts in the area. We describe in detail the

training and the execution of an ensemble of classifying algorithms. In this study we focus on the techniques

used to increase the accuracy and stability of the classifiers.

1 INTRODUCTION

The undeniable growth of social media

environments has been introducing profound

changes in society and in the communication

management landscape. Though social media

impacts are still subject of research in a wide variety

of fields, in what organizations are concerned, two

main aspects are consistently revealed throughout

literature: the newly empowered role of millions of

social media users, co-creators, active voices and

active influencers, which organizations fail to

understand and engage with, and the fact that

organizations are still struggling with the

development of a social media strategy and budget,

thus mismanaging the potential and barriers

presented by the new consumer and by social

networks in general.

In fact, organizations are rushing into social

media networks following the worldwide trend to

create a social presence in multiple channels,

reaching for and aiming at mediatization, without

previously defining a clear strategic approach, which

should, for instance, be built upon clear insights on

their target audience and an editorial plan/calendar,

that can foster the achievement of the overall

business objectives. Nevertheless, when adopting

social media, organizations are, in fact, allocating

time, effort, skills, human resources and technology

and this raises the constant need to measure the

return on these investments (ROI) and legitimize

them in the context of organizational development.

However, how can organizations attempt to

measure the efficiency and return on investments on

a social media approach that has not been

strategically designed/aligned and is a set of

unarticulated processes and situational messages?

On top of the absence of a strategic alignment

between social media approaches and organizational

goals/performance, organizations are also lacking

strategically relevant social media monitoring

methods.

The social media analytics provided by the

thousands of free/commercial web based

applications are able to provide some interesting and

valuable insights, but fail to support a relevant and

insightful benchmarking process. Social media

monitoring has been turned into a process where

organizations are on the run to acquire, for instance,

more fans that their competitors’, more likes and, in

188

Figueira, Á. and Oliveira, L.

Analyzing Social Media Discourse - An Approach using Semi-supervised Learning.

In Proceedings of the 12th International Conference on Web Information Systems and Technologies (WEBIST 2016) - Volume 2, pages 188-195

ISBN: 978-989-758-186-1

some cases, more positive feedback, where

sentiment analysis is part of the equation.

However, a well sustained strategic

benchmarking assessment that provides information

on the current implicit/explicit strategy and

knowledge on how to further develop it, it’s still

lacking.

Therefore, the assessment of social media

performance should rather be built upon the most

relevant business areas, as a key priority for

organizations that aim at turning social networks

into true business assets.

As a result, we present a persistent monitoring

methodology built upon benchmarking methods,

which rely heavily on the identification and analysis

of a set of strategically relevant editorial areas that

can foster organizational performance. According to

the proposed methodology, organizations are

propelled to focus not only on the traditional Social

Media key performance indicators, but to

incorporate them on a deeper editorial analysis that

may allow them to gain medium and long term

competitive advantage.

This article is organized as follows: in the next

section (2) we describe our methodology, focussing

on the development of an editorial model that will

allow us to categorize social media discourse. In

section 3 we present a case study where we applied

the methodology described. In this section we

elaborate on the type of data to retrieve from social

networks, on the ensemble of chosen algorithms, and

how to improve their accuracy as a whole. We finish

the section with a discussion on the techniques used.

In the last section (5) we present our conclusions.

2 METHODOLOGY

Our proposed approach is based on a 5-stage

method.

On a first stage we verify which are the social

networks being used by the organization under scope

(main agent) and by its direct competitors, on the

Higher Public Polytechnic Education Sector

(HPPES). In order to obtain a comprehensive

analysis, all social networks should be included and

a “representative” time frame should also be chosen.

The time frame that we believe to be the most

suitable is a one-year time frame, because it is in

itself cyclic and more likely to encompass a full

cycle of communication and product/service events.

After deciding on the social networks and on the

ideal time frame, all of the agent’s and agents’

competitors messages must be collected (posts,

tweets, etc.).

Hence, the second stage of the methodology

consists of gathering all the information about all the

sector’s agents’ activities on social media on:

messages’ content, the sector audience (type of

audience, number of fans, followers, etc.) and on the

corresponding responsiveness (likes, shares,

comments, retweets, etc.).

The third stage of the methodology consists of

identifying which are the most relevant areas of

intervention on social media, which we designate by

editorial areas. When this is not previously set by the

agent or is not very clear (i.e. a lack of a formal

content strategy is visible) a first human analysis and

classification of social media messages is required.

In this case, a small sample of messages manually

classified by a communication professional that

determines, in terms of editorial areas, which is the

purpose of each message. When doing so, a set of

guiding principles is considered for the HPPES in

particular, which lead to the construction of the

editorial model presented on section 3.2: (a) the

organizations/enterprise institutional (brand) needs

towards the diversity of stakeholders in its social

media networks; (b) the specifics of its

product(s)/service(s); (c) the need to balance

between institutional and transactional needs in

order to maintain reputation and ensure economic

survival; (d) a multi-channel wide holistic approach

to communication management which facilitates

integrated messages to take the most advantage of

each of the social networks being used; and the

dialogical nature that is intrinsically linked to social

media environments.

These principles should be considered in the

definition of an editorial model for every economy

sector. For instance, for common secondary sector

organizations the main editorial areas could focus

on: (a) core product/service advertisement; (b)

availability of additional services and/or available

customer support; (c) brand reputation (maintenance

/ reinstatement / re-branding, etc.) and (d)

relationship essential in every social media channel,

regardless of the economy / business sector.

The fourth stage of the methodology consists,

then, on the categorization of all the messages

retrieved from all the social networks against the

categories’ of the devised model. Although this is a

complexity linear problem (n × m), it might be too

demanding to be done by hand. In fact, a

categorization of n posts into m categories is a far

too heavy endeavour to be performed by humans if

we are consider thousands of posts and more than

two categories (the common situation). Therefore,

Analyzing Social Media Discourse - An Approach using Semi-supervised Learning

189

for this stage we use an ensemble of tuned

algorithms which are trained to classify posts in a

first moment, and that are presented the full sample

to classify, on a second moment. To mitigate the

occurrence of disputes in classifying a post

according to several possibilities, performed by

different algorithms, we take the majority of the

classifications and we leave unlabelled the posts for

which it was not possible to reach a majority.

Finally, the fifth stage of the methodology

consists on delivering a sectorial performance and

strategic benchmarking, aimed at social media

organizational success.

We begin by building a performance

benchmarking analysis based on social media key

performance indicators (KPI), such as ‘likes’,

‘shares’, ‘comments’, ‘retweets’, etc., using a

weighted scale, in order to measure audience

response to messages. We then measure each

market’s agent’s audience size (i.e. the number of

‘fans’, ‘followers’, etc.) and total communication

efforts (i.e. number of messages sent to social

networks).

These performance benchmarks allow us to build

a performance perceptual map of the sector, in

which we relate audience size, audience response

and the agents’ efforts.

The strategic benchmarking analysis is, then,

built on top of the previous analysis, adding the

social media strategies per agent (i.e. the different

combinations of intensities of editorial areas), thus

allowing to identify and examine the most efficient

communication strategies, which are enabling high

performing agents to be successful in social media.

3 CASE STUDY

The present case study was conducted on the

Portuguese Higher Public Polytechnic Education

Sector (HPPES), using the previously presented

methodology.

A total of 137 agents were considered, which

included polytechnic schools integrated into

polytechnic institutes and polytechnic schools

integrated into universities. The number of agents

was then reduced to 94 in order to include only the

schools providing educational services, disregarding

the polytechnic institutes and universities (managing

entities).

During the data collection phase we measured

the HPPEI’s social media networks adoption rates,

in order to include the most relevant communication

channels (social media), and only those that had

been in use at least since the 1st of September 2013,

with the intent of extending the analysis to a full

school year (up to September 2014). In order to

support the research in reliable sources, the study

considered only the social media websites

mentioned on the HPPEI’s official websites. This

method aimed to ensure that the social media

websites under analysis were actually managed by

the HPPEI, instead of other internal or external

stakeholders, such as students, employees

(administrative or faculty) or alumni on their own.

According to these criteria, 43 agents were included

in the study. Facebook proved to be the most

representative social media website, with an

adoption rate of 64% among all agents, as illustrated

in figure 1.

Figure 1: Social media adoption rate by agents.

In this article we focus our analysis only on

Facebook, once it is the most relevant network in the

sector and results adequate enough to provide

evidence on the implementation of the proposed

methodology.

3.1 Retrieval of Facebook Posts

The following stages consisted of retrieving and

classifying all messages posted by HPPEI on

Facebook. We used two methods: an in-house made

system, specially built for the purpose using the

available Facebook API and a third-party software

for collecting information from Social Networks.

From an initial list of the relevant agent Page Id’s,

the two systems accessed the posts retrieving the

following fields:

List 1: Fields collected from Facebook posts.

1) PostId

2) Message

3) Link

4) Name

5) Description

6) Caption

7) #Likes

8) #Comments

9) #Shares

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

190

The two systems retrieved the same number of

posts (15.444), during the entire school year, which

consolidated our confidence about the validity of the

returning set.

3.2 Editorial Model for the HPPES

As previously mentioned, if the identification of the

editorial areas for a specific agent/market is not yet

set, is incomplete or is only implicit, a small sample

of social media messages should be ran through a

communication professional, so an efficient and

comprehensive editorial model can be built.

Specifically concerning the HPPES, some editorial

areas are straight forward and some were added after a

manual classification of a small sample of messages.

In any case, the previously identified guiding

principles for the design of a social media editorial

model where applied and tuned to the HPPES and we

considered the following: (a) the heavy HEI’s mission

towards society and the great diversity of

organizational stakeholders (students, faculty, staff,

employers, partners, research centres, etc.); (b) the

specifics of the educational service (a co-produced

service); (c) a multi-channel wide holistic approach to

communication management; (d) the need to balance

between organizations’ institutional and transactional

needs in order to ensure their competiveness and

financial survival; and (e) the dialogical nature that is

intrinsically linked to social media environments.

The editorial model designed for this specific

case includes the seven main editorial areas that we

believe to have the highest impact on organizational

performance:

Education: messages are aimed at promoting or

providing information about the educational offer,

mainly higher education courses, but also include

complementary internal or external training;

Research: messages are aimed at promoting internal

research results, mainly obtained by faculty

members as inputs to the organization’s areas of

expertise, but also include research results from

other sources with impacts on those same areas.

Also includes information and call for participation

on congresses, seminars and other scientific

meetings held internally or externally;

Society: this category builds upon the, so called,

“third mission” of HEI, which is aimed at engaging

with industry and other cultural and social groups,

encompassing exchanges with society at large.

Messages in this category include the promotion of

and/or information on: knowledge and technology

transfer, patents, organizational partnerships and

contracts, demonstrations, exhibitions and

showcases conducted by faculty members or

students, and also messages promoting

employability through streaming placement offers

and career opportunities;

Identity/Brand: mainly aimed at the construction,

development and maintenance of the organizational

image and reputation, fostering distinctiveness and

the development of a corporate persona. Messages

consists mainly on promoting and/or informing

about the corporate persona character and include

references to CSR initiatives, institutional events

(such as celebrations, awards, tributes and

graduation ceremonies), students, faculty and staff

honorable mentions and representation activities in

external fairs and exhibitions;

Administration: is aimed at partially extending the

internal administrative communication with internal

publics into social media, but also attending

administrative needs towards external stakeholders.

It informs about deadlines and administrative

processes, procedures and admissions, but it also

promotes and informs on organizational support

services (goals, contacts, working hours, etc.);

Relationship: this category builds upon the

previously mentioned dialogical nature that is

intrinsically linked to social media environments and

aims to foster conversation, boost emotional

connection the organization and its stakeholders,

requiring opinions, introducing current internal,

external, societal or academic issues with which

publics can relate to. Messages in this category tend

to present lower levels of formality in order to

propel interactions and may introduce greetings,

humor, sympathy and motivation.

Information: messages in this category are aimed at

enhancing HEI’s role in fostering citizenship, mainly

among students, thus streaming external social,

economic, political and cultural relevant information,

news, regulations and events that may or may not be

close to the schools adjoining scientific areas.

3.3 Automatic Categorization

Our next step was to perform the classification of the

15444 posts according to our editorial model, listed in

the previous section. This is a time demanding task to

be done by hand and impossible to be undertaken on

the-fly if done exclusively by humans. Therefore, we

propose an automatic method to categorize the posts

based on the conjunction of several of the most recent

and promising algorithms for text classification or

categorization.

Although several text classifiers have been

proposed over the last decades, nowadays this

Analyzing Social Media Discourse - An Approach using Semi-supervised Learning

191

research topic is again gaining a lot of interest from

the research community. The main reason is that

much research is being focused on social networks

due to the abundance of interesting data to work

with. In particular, texts posted in social networks

have special properties that haven’t been considered

in previous research and in the devised algorithms

(e.g. very short texts, abundance of smileys,

inclusion of links, many punctuation signs). These

special characteristics pose new problems and make

text classification, again, a very difficult task, and

prone to failure. Nevertheless, research has been

incorporating these new features and consistently

creating better classification models, especially for

classification under supervised training.

For this step we decided to use six of the most

promising, and prominent, classifiers:

Support Vector Machines. Linear SVMs are a

machine learning algorithm (Cortes, 1995) based on

a geometric method that tries to separate two classes

through an hyperplane, picking the one that

maximizes the margin between the two classes.

More recently, this method was evolved (Crammer,

2002) to deal with a multiple number of classes. We

used the Multi-class SVM lib for this analysis.

Random Forests. RFs were created (Breiman,

2001) to overcome the overfitting effect of the

decision trees. Within this method multiple decision

trees are created during training time, and the mode

of the resulting class is the presented output.

LogiBoost (Friedman, 2000). This algorithm

belongs to a larger category of boosting algorithms

which comprehend AdaBoost, LPBoost and some

others, all based on a common framework called

AnyBoost (Mason, 2000). Generically, the boosting

algorithms try to reduce variance and pre-training

effects in supervised learning by re-weighting a set

of classifiers according to the rule: weak classifiers

should gain weight and strong classifiers should lose

weight. The LogiBoost is implemented in several

regression and classification packages. We used the

one implemented in “caTools” for R.

K-Nearest Neighbours (Altman, 1992). Although

being one of the simplest, machine learning

algorithm, it is still very useful because of it wide

range of applicability. The algorithm relies on the

previous classification of the neighbors to each

training data, classifying according to the majority

up to the defined k elements. The training data is

presented in a vector space model and all trained

examples are vectors in that multidimensional space.

MultiLayer Perceptrons. The “perceptron” is an

algorithm, in Machine Learning theory, that is able

to classify an input vector using a linear prediction

function, which combines a set of computed weights

to the vector parameters (Freund, 1999). When it is

needed to solve non-linear problems we need more

than a layer of perceptrons. Typically, multi-layer

perceptrons (MLP), use sigmoide function as an

activation function.

Deep Neural Networks. This type of algorithms

(Collobert, 2008) are based on the concept of pre-

training a multi-layered feedforward neural network,

one layer at a time, treating each layer as an

unsupervised restricted Boltzmann machine, and

then using supervised backpropagation for fine-

tuning the neural net.

Deep learning algorithms are based on an

underlying assumption that observed data is generated

by the interactions of a multitude of different factors

on different levels. Deep learning assumes that these

factors can be organized into multiple different levels

of abstraction. Therefore, varying the number of

layers and of layer sizes can provide the needed

amounts of abstraction (Bengio, 2013).

All the algorithms were used through public and

open source libraries (“caret” and “h2o”), available

for the R programming language.

3.3.1 Training Phase

First, we trained manually 350 posts according to the

derived model. As in the sample set there were blank

posts (with no text message) we considered those to

be included in a special category which we labeled

as 0 (zero). The manual classification produced a

total coverage of the 7 + 1 categories, but not

equally balanced, as illustrated in Figure 2.

Figure 2: Label count for the 350 posts.

We then computed the respective accuracy of the

automatic classification. For this we used a

confusion matrix to report the number of false

positives, false negatives, true positives, and true

negatives. We used a standard formula (1) for

computing the accuracy.

 =

∑

  + 

∑



 

(1)

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

192

As a second step, we gave the classifiers a bigger

set of 512 manually classified posts for retraining, and

recomputed the new accuracy. The new set, similarly to

the first one, has a full coverage of the seven categories

plus one, for blank posts, labelled as zero. We illustrate

in Figure 3 the comparison of the two classifications: of

the 350 and of the 512 posts, according to the number

of labels. In Figure 3 the categorization of the 350 posts

is represented by a dashed line and the 512 posts by a

solid line. Categories are distributed along the x axis.

Figure 3: Label count for the 350 and the 512 posts.

It is easy to see that category 6 is still

problematic due to its reduced number of posts.

Apart from that consideration, we may also observe

a tendency for a proportional increase in the number

of posts in the remaining categories, when

expanding the analysis from 350 to 512 posts.

We then computed the accuracy for the new set.

Every method had an accuracy increase, but the total

average for this metric, in the new training, was an

improvement of only 3%, over the 6 techniques.

Therefore, we didn’t feel in the need to classify

more posts manually.

On the other hand the absolute value for the

accuracy was still low (circa 55%).

Maintaining our view to classify the posts relying

only of the post itself and on the features associated

with each post, we augmented List 1 presenting more

information to the classifiers (information that was

already retrieved during post collection, but not used):

List 2: Extra fields collected from Facebook posts.

10) From

11) Date

12) Hour

13) Type

14) Status type

15) Link

16) Name

17) Story

Therefore, we had now 17 fields, possibly some

of them with no values.

We tried various approaches to use all the text in

the fields to help the classifiers, and combinations of

several texts. For examples, we tried:

1) Message or Description

2) Message, Name, Description, Caption and Story

3) Message, Name Description, Caption, Story and

link domain

4) All the 17 fields described in List 1 and 2, and

using a “link explosion” strategy.

We mean by “link explosion” the separation of

each term in an URL that is joined to another term by a

slash, by punctuation signs, or by the protocol’s name.

3.3.2 Classification

We then executed again the trained algorithms on

the 512 sample data using all the approaches. We

computed the accuracy, and found find out that

approach 4 delivers the best result.

In Table 1 we can see that there is a 3%

improvement from expanding the training set to 512

samples, and that, using all the 17 fields as text

features, results in a significantly better accuracy.

To assess the stability of the method we

performed a 10-fold cross validation, with the input

data averagely distributed, throughout the whole

sample (the 512 posts) and then we computed the

respective accuracy (using formula 1) in each pass,

for the whole ensemble of classifying algorithms.

Table 1: Accuracy during the training phase.

ccurac

350 512 Dif

512-all

SVM 0.3264 0.3553 0.0289 0.4523

RF 0.5347 0.5866 0.0519 0.6862

LB 0.6044 0.6507 0.0463 0.7243

KNN 0.4722 0.4803 0.0081 0.5451

MLP 0.5844 0.606 0.0216 0.7703

H2ODL 0.5781 0.6031 0.025 0.6527

verage 0.5167 0.547 0.0303 0.63848

The above table represents just the first run of

the training. We list below, in Table 2, the results of

the 10-fold cross validation.

Table 2: Accuracy of the aggregation for the six

algorithms in each of the 10 runs.

1 0.63848

2 0.71432

3 0.75420

4 0.65130

5 0.70034

6 0.65470

7 0.65382

8 0.72943

9 0.73342

10 0.62832

verage 0.68583

Analyzing Social Media Discourse - An Approach using Semi-supervised Learning

193

We ended up with an average accuracy above

68%, which seemed a fair base for classifying the

whole set of posts.

We then run the whole set of 15444 posts on our 6

trained classifiers to obtain a predicted category for

each post, by each technique. Finally, we used the

prevailing category of these six techniques as the final

result, i.e., we used the prevailing category of the six-

set as the final predictive category. When there was

no mode, the post was labelled with a zero. The

number of posts per category is illustrated in Figure 4.

Figure 4: Final label count for the 15444 posts.

As we can see category 6 remained with very

few posts (15) and less than 1% were unclassified

posts (labelled as zero). Apart from Logiboost, every

classifier could reach some category for every post.

Logiboost failed to categorize 5055 posts (about 1/3

of the total of posts). All in all, 15376 posts (i.e.,

more than 99%) were successfully given some label.

3.3.3 Strategic Benchmarking for Business

Intelligence

Having the posts categorized in each of the seven

categories defined in our model allows us to build a

statistics model and parameters to assess the effort

and gain with each message. For instance, we can

compare the number of posts in each category for

every competitor and its return in the form of Likes,

Shares and Comments (eventually with different

weights). In a previous study (Oliveira and Figueira,

2015) it was shown that some competitors have

centralized strategies whilst others have

decentralized or hybrid strategies, according to the

amount of effort per editorial area. Such research

outputs and methodology are typically framed in

strategic benchmarking processes that rely heavily

on business intelligence skills. In fact, the ability to

handle large amounts of unstructured data, to help

identify and develop new strategic business

opportunities and the identification of fundamental

lessons among competitors are essential to the

formation of well sustained medium / long term

decision making processes.

3.3.4 Discussion and Notes

The “link explosion” strategy increased the accuracy

of the classifiers around 2%. The use of all collected

features associated with each post leveraged the

accuracy in about 7%. All the text was concatenated

and transformed into a confusion matrix of posts and

terms. The TF and TF-IDF metrics were used to try

to discover relevant words to the classification. This

procedure resulted in a slightly better results for TF

but not relevant, as neither were conclusive about its

discriminatory power.

Before the training phase the more sparse

features derived from the message of each post were

removed, in order to obtain a maximum of 0.999 of

sparsity. A 0.99 of sparsity was allowed for features

derived from the link explosion.

Curiously, LogiBoost had the best accuracy for the

512 trained set, despite not being able to classify all the

posts. The Multi-Layer Perceptron, was the second best

during training, and became first when all features were

used. Moreover, this algorithm was capable of

classifying all the posts. The accuracy assessment used

a 10-fold cross validation, implemented through the

“caret” library. Whenever a classifier was trained using

the caret library, a cross validation was used and, for

each of the 10-fold was computed the accuracy and the

kappa index. When we used the “h2o” library the

accuracy was computed using a manually built, but

standard, implementation of the metric.

In order to improve the categorization, we tried

the selection of disjoint sets of words of features for

every class, ie, words that are used in just one

category. We also tested the removal of sparse

features which had high correlation and the removal

of features with high variance. All in all, the

improvement was quite small which probably is due

to the fact that most of the algorithms are based on

Figure 5: Correct assigned labels per algorithm according

to the final label produced by the ensemble of classifiers.

WEBIST 2016 - 12th International Conference on Web Information Systems and Technologies

194

trees, which already do select relevant features during

training. Finally, we must stress that the Multi-Layer

Perceptron was the algorithm that more often categorized

the posts as the majority of the ensemble of classifiers did

(cf. Figure 2).

4 CONCLUSIONS

In this article we presented a model for analysing

social media discourse by categorizing the messages

posted in social networks according to a n-category

model, created during analysis. Posts from an entire

year are collected and a small amount is used to

build the editorial model. Another sample was used

for training. In our case study, the accuracy did not

increase the classifying procedure with a sample

bigger than 3% of the total data. In the case study we

used six of the most well-known algorithms to

perform the categorization. We showed how to

improve their classifying accuracy by using the

features associated with each post, and how to fine

tune the categorization parameters. The resulting

accuracy of the method was increased from 51% up

to 68%. After the ensemble of algorithms were

trained the whole sample was ran through it.

ACKNOWLEDGMENTS

This work is supported by the ERDF – European

Regional Development Fund through the

COMPETE Programme (operational programme for

competitiveness) and by National Funds through the

FCT (Portuguese Foundation for Science and

Technology) within project «Reminds/ UTAP-

ICDT/EEI-CTP/0022/2014.

REFERENCES

Altman, N. S. (1992). "An introduction to kernel and

nearest-neighbor nonparametric regression". The

American Statistician 46 (3): 175–185.

Bengio, Y.; Courville, A.; Vincent, P. (2013).

"Representation Learning: A Review and New

Perspectives". IEEE Transactions on Pattern Analysis

and Machine Intelligence 35 (8): 1798–1828.

Breiman, Leo (2001). "Random Forests". Machine

Learning 45 (1): 5–32.

Cortes, C.; Vapnik, V. (1995). "Support-vector networks".

Machine Learning 20 (3): 273.

Freund, Y.; Schapire, R. E. (1999). "Large margin

classification using the perceptron algorithm" (PDF).

Machine Learning 37 (3): 277–296.

Friedman, J., Hastie, T., and Tibshirani, R. (2000).

Additive logistic regression: a statistical view of

boosting. Annals of Statistics 28(2): 337–407.

Koby Crammer and Yoram Singer. 2002. On the

algorithmic implementation of multiclass kernel-based

vector machines. J. Mach. Learn. Res. 2 (March

2002), 265-292.

Llew Mason, Jonathan Baxter, Peter Bartlett, and Marcus

Frean (2000); Boosting Algorithms as Gradient

Descent, in S. A. Solla, T. K. Leen, and K.-R. Muller,

editors, Advances in Neural Information Processing

Systems 12, pp. 512-518, MIT Press.

Oliveira, L., Figueira, A. (2015). Benchmarking analysis

of social media strategies in the Higher Education

Sector. In Proceedings of Conference on ENTERprise

Information Systems (CENTERIS’16). Vol64:779-

786.

Ronan Collobert and Jason Weston. (2008). A unified

architecture for natural language processing: deep

neural networks with multitask learning. In

Proceedings of the 25th international conference on

Machine learning (ICML '08). ACM, New York, NY,

USA, 160-167.

Rud, Olivia (2009). Business Intelligence Success Factors:

Tools for Aligning Your Business in the Global

Economy. Hoboken, N.J: Wiley & Sons. ISBN 978-0-

470-39240-9.

Analyzing Social Media Discourse - An Approach using Semi-supervised Learning

195