=Paper=
{{Paper
|id=Vol-2616/paper26
|storemode=property
|title=The Intelligent Monitoring of Messages on Social Networks
|pdfUrl=https://ceur-ws.org/Vol-2616/paper26.pdf
|volume=Vol-2616
|authors= Serhii Holub, Natalіia Khymytsia, Maria Holub, Solomiia Fedushko
|dblpUrl=https://dblp.org/rec/conf/coapsn/HolubKHF20
}}
==The Intelligent Monitoring of Messages on Social Networks==
https://ceur-ws.org/Vol-2616/paper26.pdf
The Intelligent Monitoring of Messages on Social
Networks
Serhii Holub1 [0000-0002-5523-6120] and Natalіia Khymytsia2 [0000-0003-4076-3830]
and Maria Holub1 [0000-0002-0553-8163], and Solomiia Fedushko2 [0000-0001-7548-5856]
1
Cherkasy State Technological University, Shevchenko Boulevard 460, Cherkasy, Ukraine.
2
Lviv Polytechnic National University, Lviv 79013, Ukraine
s.holub@chdtu.edu.ua, nhymytsa@gmail.com,
sashasokolovsksa92@gmail.com, solomiia.s.fedushko@lpnu.ua
Abstract. The results of research on the processes of use of information
technology of multilevel intellectual monitoring for classification of text
internet-messages in social networks are presented. The selection of messages
was carried out by experts. Two textbook classes were formed. The first class
included texts of malefactors. The second class was formed on the basis of
posts that, in the opinion of the authors, are not malefactors. The processes of
decomposition of text messages of social networks, the adaptive formation of
dictionaries of signs, use of the agent approach to the construction of the
classifier, and its use for recognition of malefactors are investigated in the
work. Experimentally confirmed hypotheses about the existence in Ukrainian-
speaking social communities of participants who have a common style of
presenting messages. This may be the result of their joint special training for
this type of activity. The effectiveness of classifying agents was assessed by the
number of correctly classified text messages. Conclusions about the belonging
of messages to a certain class were made based on the results of the
classification of observation points that describe individual texts. The number
of correctly classified messages prepared by the agent was over 92%. The
hypothesis about the possibility of using the method of adaptive formation of
the dictionary of features in the technology of classification of text messages,
with a volume of more than 100 characters, has been experimentally confirmed.
This allows you to automate the processes of protection of the information
space of Ukrainian-language Internet content of social networks.
Keywords: social networks, web communities, content, information, sources,
monitoring, content analysis, intellectual agents, dictionary of features.
1 Introduction
Intelligent monitoring is an information technology that provides knowledge to
decision-making processes by organizing continuous observations and processing
their results. Information is obtained in the form of information about the properties
Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons
License Attribution 4.0 International (CC BY 4.0). COAPSN-2020: International Workshop on
Control, Optimisation and Analytical Processing of Social Networks
of the object of monitoring, knowledge is obtained in the form of identified patterns
and trends. Monitoring intelligent systems (MIS) is a software implementation of
information technology of multilevel intelligent monitoring. Features of application of
this technology demand efficiency in the formation of new knowledge and their
adaptability to change of properties of the environment. Therefore, MIS organizes
continuous observations and implements their processing to build a model knowledge
base (MKB) - as a means of obtaining information and storing knowledge after the
transformation of the results of observations of the objects of monitoring. Building
MKB using agent approaches based on cloud technologies allows quickly respond to
changing monitoring tasks and adapt to changing properties of monitoring objects.
Information monitoring is one of the priorities in a hybrid war. Depending on the
media, the task of converting the results of observations to a typical form of input data
is performed. The most common media in social networks are text messages. The use
of intelligent monitoring of text messages in social networks has certain features.
These messages are short and presented in a special style. Therefore, conventional
methods of analysis and previously constructed models for their development are not
effective. In particular, the previously solved attribution problems, which allow
classifying the authors of printed works, will not allow classifying the authors of
messages on social networks.
Monitoring of internet-messages is ordered to provide information and knowledge
of various decision-making processes. The most common are classification tasks. But
the results of the tasks of identification and forecasting are useful. And the peculiarity
is that to perform these tasks it is necessary to process text, video, and audio
messages, which in their original state require the transformation into a typical form
of an array of numerical characteristics. And often there is a need to combine
information obtained from different messages of one author. Therefore, special
importance is given to methods of converting messages from the primary form to a
typical matrix of an array of numerical characteristics of informative features and the
processes of consolidation of simulation results. The informativeness of this array
must exceed the limit of informative sufficiency for the available means of model
synthesis. The results of the consolidation of information from different messages
should allow obtaining knowledge in the form of an emergent effect from the
coordination of interactions of diverse models. Therefore, the construction of a
dictionary of features for text, audio, or video messages should be considered as a
solution to one of the local tasks in the process of fulfilling the global task, the
consolidation of the interactions of models of monitoring objects in social networks.
This paper presents the results of research on the processes of intellectual analysis
of text messages from social networks. The process of solving the problem of
monitoring text messages is an element of information technology for intelligent
monitoring of Internet content.
This technology is designed to identify people who have undergone the same type
of training, who use social networks to influence the opinion of readers of the
Ukrainian-language information environment; and can be classified using the
Intelligent Monitoring System (IMS) [1].
2 Related Works
At the beginning of the XXI century, the issue of developing various methods of
analyzing Internet content and developing new approaches to solving this problem is
becoming the subject of interest of even more authoritative scientists.
In particular, practical approaches to the purification, indexing, and extraction of
textual information are published in a collection of scientific papers edited by M.
W.Berry [2]. Valuable for our study were primarily articles that address various
aspects of algorithmic progress in discriminant analysis, investigate spectral
clustering, describe trends in identifying trends and removing synonyms, describe
case studies in web-mining, and analyze customer support logs to extract relevant
topics and query characteristics. Another fundamental edition, prepared by Aggarwal,
Charu C., Zhai, ChengXiang, focuses on text embedded in heterogeneous and
multimedia data, making the extraction process much more complex; methods such as
transfer training and multilingual mining are described [3].
It is important for a comprehensive study of the issues of intellectual monitoring is
the analysis of different approaches to the practical application of content analysis
techniques. In our opinion, it is necessary to note the work of O. Ivanov, which
proposed a technical concept of the content of the analytical program - independent of
the text being analyzed, able to identify significant links between words, determine
the structure of analyzed texts [4]. No less relevant is the approach proposed by
researchers I. Khomytska, V.Teslyuk, O. Morushko, A. Holovatyy. These scientists,
in particular, present the results of the applied methods, models and software, which
confirm that the author's attribution of the text at the phonological level is more
effective [5].
Studies by A. Peleshchyshyn, V.Vus, O. Markovets, S. Albota analyze a special
system of user activity indicators. The authors also explore the linguistic methods of
influence that are purposefully used in online communication [6-8]. The work of O.
Markovets, R. Pazderska, N. Dumanskyi, I. Dronyuk presents methods that will help
determine the level of trust in the author, such as monitoring, attestation, organization
and personification [9]. Y. Syerov, N. Shakhovska and S. Fedushko in the context of
research on artificial intelligence, proposed a new method for determining the
adequacy of data of personal medical profiles [10]. S. Fedushko and E. Benova's work
on computer and language content analysis, authentication of personal data of user
accounts, consolidation, and analysis of user information tracks and their behavioral
models of communication in social areas of the Internet, proposed a comprehensive
approach to identifying dangerous threats, which have a negative impact on Internet
users and are the basis for creating countermeasures against information and
communication threats [11]. Within the research work of N.Khymytsia, T.
Ustyianovych, and I. Dronyuk several series of stages of search and processing of
messages from web forums containing historiographical information by means of web
scraping, data analysis, and big data analysis have been developed [12].
Among the numerous theoretical, methodological and practical studies that
highlight the problems of building monitoring systems, in the context of our topic, it
is important to use multilevel information-analytical monitoring, which is associated
with information analysis, model synthesis, expert evaluation, diagnosis and
forecasting and organization of continuous observations of signs coming from
different sources. The aim is to regularly provide information support for decision-
making processes in various subject areas. Such approaches using intelligent
monitoring are considered in the study of S. Kunytska, S. Holub where it is proposed
to implement the methodology of creating information systems of multilevel
intelligent monitoring by using an agent approach involving cloud technologies [1].
Also, in the works of S. Holyb, N. Khymytsya proposed new aspects of the practical
application of monitoring information systems to identify similar historical periods in
the economy [13, 14].
The results of information retrieval allow us to conclude that the information
technology of intelligent monitoring can be used to solve the problem of classifying
text messages on social networks according to the properties of their authors. The
methods of text mining used by this technology were used to classify messages larger
than 500 characters. The classification of much shorter messages requires the use of
new approaches to the construction of classifiers, in particular the agent approach [1].
Thus, the purpose of this work is to study the processes of classification of texts of
100 characters on the example of messages from one of the social networks. The
results of research should provide an opportunity to automate the processes of
protection of the information space of Ukrainian-language content on the Internet and
create conditions for the development of measures to combat the influence of
attackers.
3 The research of the process of classification of text messages
in social networks
To achieve this goal, several hypotheses were formulated.
1. Ukrainian-language social networks are influenced by persons who have undergone
special training and can be classified according to a similar style of text messages;
2. The use of an adaptive formed dictionary of features based on the results of the deep
decomposition of the text [15] will provide sufficient information on the array of
numerical characteristics of text messages in social networks to perform the task of
classifying their authors.
An experiment was performed to test these hypotheses. The problem of classifying
the authors of text messages on the social network "Facebook" was solved.
46 messages were selected by experts, among which 34 messages formed a finite set
T:
T = {t1, t2, …, t34}, (1)
representing a training sample. Expertly, two classes of texts of the set K were formed:
K = {k1, k2} (2)
The class of texts, the authors of which were identified as generators of unfair
influences, was called "Bots" (Class 1). Class 2 included the original messages of the
authors. Class 2 is called "Original".
It was necessary to construct a classifier f to ensure the display of elements of the set
С = {t35, t36, …, t46}, (3)
that is, new messages in "Facebook", on the elements of the set K:
f : С K. (4)
Figure 1 describes the process of machine learning of the model-classifier of text
Internet-messages.
Fig. 1. Functional diagram of the construction of the classifier of Internet-messages
According to the method described in [15], text messages were divided into sections of
100 characters and converted into a vector form of sequences of observation points in
the multidimensional feature space of the input data array (IDA). 159 vectors of
numerical characteristics of text messages were formed.
Observation points (vectors of numerical characteristics), which are the result of
transformations of texts from the class "Bots", were marked as "Our". Other vectors
were designated as "Alien". 3 sequences of observation points were identified.
Sequence "A" contained 73 observation points and was used to train models. Sequence
"B" was used to implement the external criterion of model quality and contained 72
observation points.
The sequence "C" contained 14 points, was used to test the results of the classification
of texts and did not participate in the synthesis of the model.
Table 1. shows a fragment of the array of input data.
Table 1. An array of input data of the Monitoring intelligent system.
Features (simulation variables)
Class
Observation point а б в г д е є …
Y x1 x2 x3 x4 x5 x6 x7 …
Bot 1_ukr (0) [100] 100 6 3 3 0 2 1 2 …
Bot 2_ukr (0) [100] 100 3 2 3 3 2 5 0 …
Bot 2_ukr (1) [100] 100 3 0 3 3 1 5 0 …
Bot 2_ukr (2) [100] 100 6 3 3 1 1 3 1 …
Bot 3_ukr (0) [100] 100 9 1 7 0 3 1 1 …
Bot 3_ukr (1) [100] 100 2 1 4 0 3 8 1 …
Bot 3_ukr (2) [100] 100 4 2 2 1 1 1 2 …
… … … … … … … … … …
Original 1_ukr (0) [100] -100 1 3 6 0 2 4 0 …
Original 1_ukr (1) [100] -100 5 5 0 1 3 4 0 …
Original 2_ukr (0) [100] -100 5 1 3 2 2 6 1 …
Original 2_ukr (1) [100] -100 9 1 6 0 2 4 1 …
Original 3_ukr (0) [100] -100 8 1 5 3 3 4 0 …
Original 3_ukr (1) [100] -100 3 5 4 0 3 5 0 …
Original 4_ukr (0) [100] -100 9 0 4 0 2 3 1
Original 4_ukr (1) [100] -100 5 0 3 0 4 6 1
Original 4_ukr (2) [100] -100 6 3 3 0 2 4 2
Original 4_ukr (3) [100] -100 12 2 5 2 2 9 0
… … … … … … … … … …
Unlike typical content analysis, the adaptive formation of the dictionary of features
involves the decomposition of texts of one class to the level of individual characters
and their combinations, the choice of criteria for informative features, and setting the
limit of informative sufficiency for the IDA. According to these parameters, the
classification features of messages from social networks are selected. The dictionary of
features is used to convert text elements (windows) into a vector form of an array of
numerical characteristics.
Each feature of the dictionary forms an element of the vector of characteristics of
the text. A separate vector depicts the observation point in a multidimensional feature
space. The observation point is a row in the IDA matrix (Table 1). Symbols are column
names. They are modeling variables in the synthesis of classifier models.
The IDA, which is formed by sequences "A" and "B" of observation points, was fed
to the entrance of the MIS. Using the agent approach described in [1], MIS synthesized
an agent-classifier of text messages. In the process of classification of the classifier
involved intelligent agents, the structure of which is based on the algorithms of MGCA
[16], neural networks of several topologies, hybrid methods of model synthesis. The
formation and adaptation of the structure of intellectual agents took place in conditions
of competition - they were given: the same modeled indicator and quality criterion -
the number of correctly classified texts. The highest quality was the agent, MKB which
was based on the multi-row algorithm MGCA. After parametric optimization of the
MKB synthesis process, it turned out that the quality criterion of the models used the
condition of the minimum standard deviation of the simulation results from the actual
values of the simulated indicator on the sequence of points "B":
2
∑72 (𝑦𝑖 −𝑦𝑖∗ )
𝑆𝑦 = √ 𝑖=1 , (5)
72
yi – calculated value of the simulated class index, yi* – the actual value of the class
indicator.
In Fig. 2 shows a fragment of the MKB agent-classifier of text Internet-messages.
Y=
+15,6622093047518+0,343662829679701·y2+0,5
7507159195118·x8-3,07611074952025E-
5·y22+4,38629067311242E-
5·x82+0,0326096880710039·x6-
1,51997544750547E-5·x62+9,71819962121924E-
5·y2·x8-1,98675405733841E-
8·y23+8,76337274466889E-9·y2·x82-
2,70919002926536E-8·x8·y22-
1,04270819793911E-8·x83+2,81175741291694E-
12·y24-3,42114384513513E-
12·y22·x82+4,15458173320306E-12·x84-
6,10546690069651E-
5·x6·x8+1,14363884060127E-
9·x63+8,19778350736378E-
9·x6·x82+2,63076457778262E-8·x8·x62-
Fig. 2. A fragment 2,45690316873895E-13·x
of the model knowledge 4 base of the agent-classifier of social
6 -
community messages 3,55243056988161E-12·x62·x82-
1,96675261636832·x17+0,0826746467539376·x17
2
MKB, a fragment of +0,00400450550370684·x
which is presented in fig. 8·x2,
17-is used as an algorithm to convert
2
0,000156465970537728·x ·x
the results of observations of messages in social 2communities,
8 17 - presented in the form
7,49846257181755E-7·x17·x8 -
of a matrix of numerical characteristics in the conclusions
3 about whether the author of
0,000533805451462615·x17 +2,9324169497901E
the message is a wrecker.
-8·x82Y ·x17has
2 a multilayer structure,
+1,0392191238393E-5·x 4 which includes models of
17 -
lower layers (y1, y2, ...) and numerical characteristics of the features of the IDA,
4,30954058257266E-
presented in fig. 2. The ·x6+2,00462385654205E-
6·y2methodology of building models of model knowledge of
2
is2·xa36 separate
intellectual agents [17]9·y +2,16866521324368E-
area of research and3 is not considered in this
work. 12·y2 ·x8+3,83721404755865E-13·y2·x8 -
1,10890735998383E-
The effectiveness of the classifier was assessed by the number of correctly
16·y25+5,61298653236937E-
classified observation points
16·y23·xand 2 the correctness of conclusions about the belonging
8 +1,15804572644461E-
16·y2·x84+6,59300121609306E-9·y2·x6·x8-
1,47320673792522E-
13·y2·x63+1,22109645357258E-12·y2·x6·x82-
2,77651868263496E-
12·y2·x8·x62+3,13009360206965E-17·y2·x64-
5,14017116238555E-
16·y2·x62·x82+4,37269341600039E-17·x8·y24-
2,32870905666926E-
16*x83*y22+1,29373106997219E-16*x85-
3,12952203821126E-
13*x8*x6^3+9,67722773342104E-
13*x83*x6+7,09356187700353E-17*x8*x64-
4,16173684871375E-
16*x83*x62+6,63059211077389E-10*y2^2*x6-
2 2
of new text messages in the social community to one of the classes. They are
combined in the sequence "C". If the message was described by several observation
points, the conclusion that the message belongs to one of the classes was made by
most of the results of the classification of observation points of the sequence "C",
which describe a single message.
4 Results
Table 2 presents the test results of the classifying agent.
Table 2. The classification of text messages from the test sequence "C"
The value of the
Observation point Class Conclusion Result
classifier
Bot 36_ukr (0) [100] 1 -50,82 Bot 1
Bot 37_ukr (0) [100] 1 -74,67
Bot 1
Bot 37_ukr (1) [100] 1 -34,48
Bot 38_ukr (0) [100] 1 -52,53 Bot 1
Bot 39_ukr (0) [100] 1 -38,19 Bot 1
Bot 40_ukr (0) [100] 1 -67,75 Bot 1
Original 41_ukr (0) [100] 2 15,25 Not a bot 1
Original 42_ukr (0) [100] 2 37,86 Not a bot 1
Original 43_ukr (0) [100] 2 50,90 Not a bot 1
Original 44_ukr (0) [100] 2 22,02 Not a bot 1
Original 45_ukr (0) [100] 2 27,03 Not a bot 1
Original 46_ukr (0) [100] 2 -5,14 Bot 0
Original 47_ukr (0) [100] 2 53,30 Not a bot 1
Among the points of the test sequence "C", the agent singled out among the analyzed
texts those whose authors were characterized by experts as malefactors. The agent
error in test № 12 can be eliminated by adjusting the threshold value of the separating
surface. An increase in the diversity of the classifier can be achieved by increasing the
number of texts in the classroom and increasing the number of texts that are described
by observation points in the examination sequence "C".
The test results of the classifier, shown in table 1, allow us to assert its adequacy.
The number of classified text messages from the social community Facebook was
over 92%.
As it was possible to teach the classifier of texts according to the formed text
classes, it means that the authors of text messages are united by common properties.
Therefore, Hypothesis 1 received its experimental confirmation.
Acceptable results of the classifier test prove the effectiveness of the used
intelligent monitoring technology. Therefore, Hypothesis 2 should be considered
experimentally confirmed.
5 Conclusions
For the first time, the results of the application of information technology of
multilevel intellectual monitoring for classification of texts, the volume of 100
characters are received. This allows to expand the capabilities of monitoring
intelligent systems and use them for intelligent analysis of Internet messages in social
communities. It is possible to automate the processes of protection of the information
space of Ukraine, to identify and analyze examples of information impact on readers
of Ukrainian-language content.
6 References
1. Kunytska, S, Holub, S.: Multi-agent Monitoring Information Systems. In: Palagin A.,
Anisimov A., Morozov A., Shkarlet S. (eds) Mathematical Modeling and Simulation of
Systems. MODS 2019. Advances in Intelligent Systems and Computing, vol 1019. pp 164-
171. Springer, Cham (2019).
2. Survey of Text Mining I: Clustering, Classification, and Retrieval. Ed. by M. W. Berry.
2004. Springer, (2003).
3. Aggarwal, C. C., Zhai, C.: Mining Text Data. 527 p. Springer (2012).
4. Ivanov, O.,V.,: Computer content analysis: problems and perspectives of virishennya.
Methodology, theory and practice of sociological analysis of ordinary sou-spіlstva. pp.
335-340. (2009).
5. Khomytska, І., Teslyuk, V., Holovatyy, A., Morushko O.: Methods, models and means of
the system for differentiation of phonostatistical structures of engglish functional styles.
Development of methods, models and means of authorship attribution of a text. In:
Eastern-European Journal of Enterprise Technologies. № 3/12 (93). pp. 41–46. (2018).
6. Peleshchyshyn, A., Vus V., Markovets O., Albota S.: Identifying Specific Roles of Users
of Social Networks and Their Influence Methods. In: Proceedings of the 13th International
Scientific and Technical Conference on Computer Sciences and Information Technologies,
CSIT 2018, pp. 39–42. Lviv. DOI: 10.1109/STC-CSIT.2018.8526635. (2018).
7. Peleshchyshyn, A., Vus, V., Albota, S., Markovets, O. A Formal Approach to Modeling
the Characteristics of Users of Social Networks Regarding Information Security Issues.
Advances in Intelligent Systems and Computing. Volume 902, 2020, Pages 485-494. 2nd
International Conference of Artificial Intelligence, Medical Engineering, Education,
AIMEE 2018; Moscow; Russian Federation; 6 October 2018 through 8 October 2018;
Code 226259. (2018).
8. Trach, O., Peleshchyshyn, A.: Development of directions tasks indicators of virtual
community life cycle organization, International Scientific and Technical Conference
"Computer Sciences and Information Technologies, pp. 127-130 (2017).
9. Markovets, O., Pazderska, R., Dumanskyi, N., Dronyuk, I. Analysis of citizens’ appeals in
heterogeneous web services CEUR Workshop Proceedings Volume 2392, 2019, Pages
184-198 1st International Workshop on Control, Optimisation and Analytical Processing
of Social Networks, COAPSN 2019; Lviv; Ukraine; 16 May 2019 through 17 May 2019;
Code 149063
10. Syerov, Y., Shakhovska, N., Fedushko, S.: Method of the Data Adequacy Determination
of Personal Medical Profiles. Proceedings of the International Conference of Artificial
Intelligence, Medical Engineering, Education (AIMEE2018). Advances in Artificial
Systems for Medicine and Education II. Volume 902, 2019. pp. 333-343.
https://doi.org/10.1007/978-3-030-12082-5_31. (2019).
11. Fedushko, S., Benova, E.: Semantic analysis for information and communication threats
detection of online service users. The 10th International Conference on Emerging
Ubiquitous Systems and Pervasive Networks (EUSPN 2019) November 4-7, 2019,
Coimbra, Portugal. Procedia Computer Science, Volume 160, 2019, Pages 254-259.
https://doi.org/10.1016/j.procs.2019.09.465. (2019).
12. Khymytsia, N., Ustyianovych, T., Dronyuk, I. Identification and Modeling of
Historiographic Data in the Content of Web Forums.CEUR Workshop Proceedings
Volume 2392, 2019, Pages 297-308. 1st International Workshop on Control, Optimisation
and Analytical Processing of Social Networks, COAPSN 2019; Lviv; Ukraine; 16 May
2019 through 17 May 2019; Code 149063. (2019).
13. Holyb, S., Khymytsya N.: The use of multi-level modeling in the cliometric studies
process. In: Proceedings of the 13th International Conference “Modern Problems of Radio
Engineering, Telecommunications and Computer Science”, TCSET 2016, pp. 733–735.
Lviv-Slavske, Ukraine (2016).
14. Golub, S., Khymytsia, N.: The Method of Cliodinamik Monitoring. In: Proceedings of the
2nd International Conference on Data Stream Mining and Processing, TDSMP 2018, pp.
223–226. Lviv (2018).
15. Holub, M.,S.: Form of mass input of recent tributes in the classification of texts in the
technology of information monitoring. Mathematical machines and systems. 2018. No. 1.
pp. 59-66. (2018).
16. Ivakhnenko, A.G.:Inductive method of self-organization of models of complex systems.
Kiev, Naukova Dumka, 1981. 296 p. (1981).
17. Zhiryakova I.A., Holub S.V. New approach to conceptual knowledge. Technical science
and technology. 2015, No. 2. pp. 78-82. (2015).