=Paper=
{{Paper
|id=Vol-2030/HAICTA_2017_paper94
|storemode=property
|title=Modeling of Posting Behavior in Social Media
|pdfUrl=https://ceur-ws.org/Vol-2030/HAICTA_2017_paper94.pdf
|volume=Vol-2030
|authors=Kateryna Kononova,Anton Dek,Maksym Shpakovych
|dblpUrl=https://dblp.org/rec/conf/haicta/KononovaDS17
}}
==Modeling of Posting Behavior in Social Media==
https://ceur-ws.org/Vol-2030/HAICTA_2017_paper94.pdf
Modeling of Posting Behavior in Social Media
1 2 3
Kateryna Kononova , Anton Dek , Maksym Shpakovych
1
Department of Economic Cybernetics and Applied Economics, V. N. Karazin Kharkiv
National University, e-mail: kateryna.kononova@karazin.ua
2
Department of Economic Cybernetics and Applied Economics, V. N. Karazin Kharkiv
National University, e-mail: dektox@gmail.com
3
Department of Applied Mathematics, Kharkiv National University of Radioelectronics,
maksym.shpakovych@gmail.com
Abstract. The aim of the research is a modeling of online social networks
(OSNs) users posting behavior. The research is based on the data retrieved
from two most popular Russian-speaking OSNs – Vkontakte and
Odnoklassniki. The article explores the distribution of the users, their posts,
friends, and groups. To check the common hypothesis that content creators are
also the main channel of information propagation, we have applied Kohonen
maps. Users clustering allowed identifying the types of their posting behavior;
in particular, we distinguished “writers” from “propagators” and “readers”.
The research has shown that the cluster of “writers” was the smallest in both
OSNs; however, these users generated the main content. Next in number was a
cluster of “propagators”, contained users with the largest number of reposts.
Among those who are actively interested on a given topic, the most numerous
group was “readers”, however, the absolute majority of users stayed
“indifferent”; they could be considered as a potential audience. Obtained
results can be used for Green Technologies promoting in society enforcing
their wide implementing.
Keywords: users posting behavior, online social networks, generation and
dissemination of information, clustering, neural networks, Kohonen maps.
1 Introduction
Despite high expectations and a huge amount of publications on environmental
safety, which is one of the main aims of National Agendas in many countries, it
should be admitted that the general attention to green technologies decreases in
recent years (Figure 1).
The previous study (Kononova et al., 2016) showed that there was no significant
correlation between indices of Green Technologies and Environmental Sustainability
at the macro level. Nevertheless, it was observed at the micro level, emphasizing the
importance of Green Technologies promotion not only on the state level but also in
society. Taking into account the fact that more than a quarter of the world population
are users of several online social networks (OSNs), they offer one of the most
837
effective channels of opinion forming. However, to find the optimal (shortest and
cheapest) way to promote any idea via OSN, its structure and users posting behavior
should be studied in details.
Fig. 1. Google trends on “Environmental Technology” and “Green Technology” search
queries, 2004-2017 (Google, 2017)
Nowadays numerous researchers deal with agents behavior, studying its
properties, motivation and significant factors. However, despite the growing number
of scientific publications in this field, relevant research is fragmented and there is not
yet a holistic understanding of agents behavior. At the same time, big social data
accumulation in recent years has formed a separate interdisciplinary research
direction – Social Media Mining, which studies social networks structures, users’
profiles and their behavior. Content, generated by users of social networks, is an
important source of information, which could be effectively used to identify implicit
patterns of users posting behavior.
The research includes the data about users’ profiles, their posts, and comments
retrieved from two the most popular Russian-speaking social networks Vkontakte
(VK) and Odnoklassniki (OK). The purpose of the study is modeling of users posting
behavior with the data collected from the OSNs on a given topic (which is not
environmental issues, but more popular one). To achieve it, the following tasks were
set:
− to download the users’ profiles residing in the target regions;
− to download the posts from users’ pages;
− to recognize the posts on a given topic using preliminary developed
dictionaries;
− to identify active on a given topic users;
− to identify and describe the types of their posting behavior.
This paper is organized as follows: in the next Section we explain the previous
studies on social streams; in Section 3, we present the methodology of data collection
and preparation; in Section 4 you can find frequency analysis results; the posting
behavior of VK and OK users is described in Sections 5 and 6; and finally, we
discuss the results and conclude the paper in Section 7.
838
2 Related works
Online social networks have attracted considerable scientific attention. Earlier
works (Kumar et al., 2003, Gruhl et al., 2004, Adamic et al., 2005) have focused on
the information propagation behavior in blog-sphere and studied the information
epidemics using classical Diffusion of Innovation model. Kumar studied the
“burstiness” of blogs analyzing the evolving link structure. Gruhl focused on the
propagation of topics from one blog to the next based on the text of the weblog rather
than its hyperlinks. Leskovec (Leskovec et al., 2008) analyzed networks structure of
about 45 thousand blogs and 2.2 million postings and offered a model of network
evolution.
Guo (Guo et al., 2009) analyzing posting behavior has shown strong daily and
weekly patterns; however, for re-posting, the temporal patterns have not been
observed. They also distinguished two groups of users: steadily posting in the
network, and inactively posting (the rest ones posted occasionally). Bamsuk
(Bamsuk, 2012) also investigated temporal characteristics of posting behavior
making comparisons blogosphere vs. Twitter, commercial blogs vs. non-commercial
blogs. Benevenuto (Benevenuto et al., 2010) used clickstream data from a social
network aggregator to compare user behavior across different online social networks.
To understand user behavior, Papagelis, (Papagelis et al., 2011) investigated the
causality between individual behavior and social influence by observing the diffusion
of innovations among social peers. Liu (Liu et al., 2010) predicted user’s interest
based on click behavior. Assuming that user behavior is mainly influenced by three
factors: breaking news, posts from social friends and user’s intrinsic interest, Xu (Xu
et al., 2012) proposed a mixture topic model to analyze users posting behavior.
Roman (Roman et al., 2012) presented a stochastic model based on decision-making
psychology to describe content posting dynamics on OSNs.
However, analyzing posting behavior, most of the studies are based on analytical
models focusing on the ways of users connections, the structure of the networks and
how it evolved over time. Driving by initial assumptions about users posting
behavior they insist on a sensible but doubtable hypothesis that content creators are
the same users who play an important role on information propagation. Our study,
going from the data, offers a bit different view on this issue separating those who
write and those who deliver the information to the majority of readers.
3 Data collection and preparation
The research includes the data retrieved from two most popular Russian OSNs –
Vkontakte and Odnoklassniki concerning a specified topic (political issues in cross-
border regions). VK and OK are relatively similar social networks, both providing
the following data in an accessible form:
− user’s ID;
− geolocation;
− the list of the user’s friends;
839
− the list of the user’s groups;
− user’s posts.
All profiles and posts of those users, who indicated the target geolocation, were
downloaded from these two networks. The sample includes 248k Vkontakte profiles
and 238k Odnoklassniki profiles. First, both datasets were cleaned from
uninformative posts:
− which did not contain textual information (only links, pictures, audio, etc.);
− that language did not match the analyzed one (to identify posts in Russian, the
poliglot library of python was used).
Then, using the tokenizer module of the nltk library of python, we have split the
sentences into separate words and removed extraneous characters (punctuation
marks, emoticons, etc.). With the pymorphy2 library, the words were lemmatized; all
the letters in the words were converted to a lower case. As a result, we have two
datasets (for VK and OK), which meet the following requirements:
− textual format;
− the language of the posts is Russian;
− the words are in a single word form;
− all extraneous characters are deleted;
− all words are written in lower case.
−
4 Frequency characteristics of profiles and posts
After datasets preliminary preparation, we have found and calculated the posts
containing the words from a pre-compiled thematic dictionary (Table 1).
Table 1. A fragment of the VK dataset
Id Num of Num of groups Num of posts Num of
friends friends' posts
195269137 49 3 992 1439
187704426 53 10 949 62
17241807 193 1 925 593
35833523 332 1 789 726
906761 300 3 773 291
23419701 273 4 754 55
51258489 536 10 735 754
Analysis of the datasets frequency characteristics has shown that more than 40%
of VK users and 25% of OK users did not have friends putting under doubts the
relevance of their profiles (Figure 2). However, in general, the distribution density of
friends is described by Gaussian low; it indicates the relevance of the sample.
840
Fig. 2. Distribution of users’ friends
Almost 60% of both OSNs users were not subscribed to the groups specializing on
a given topic (Figure 3).
Fig. 3. Distribution of users by thematic groups
At the same time, about a quarter of a percent of users participated in more than
500 thematic groups; it allows making an assumption about the non-random nature
of their behavior.
The thickening tail of the posts distribution leads to considerations about the
engagement of users who have written more than 500 posts on a given topic (almost
5% of VK users, Figure 4).
841
Fig. 4. Distribution of posts on a given topic
In general, a preliminary study has shown that 10% of users generate 70% of the
content (Figure 5). Taking into account that some users were posting 5-10 times per
day, it could be assumed that posting was kind of work for them.
The preliminary research has allowed concluding that the distributions of posts
and thematic groups were non-random, initiating a more detailed study of users
posting behavior.
Fig. 5. Summary statistics for VK and OK users
5 Posting behavior of VK users
To identify and describe the types of posting behavior we have decided to use a
neural network approach, in particular, Kohonen maps, which allow revealing hidden
regularities in the data. Maps are very easy to interpret, because
- each map is a visualization of one of the user's parameters;
- each hexagonal cell contains a certain, in general not the same, number of users;
842
- the users location is the same at the all maps;
- the color of the cell corresponds to the value of clustering parameter (the scale is
indicated at the bottom of each map).
After a series of experiments with VK sub-sample (formed from users who had
more than 200 posts), the following set of maps was obtained (Figure 6).
Fig. 6. Set of Kohonen maps, which are describing the behavior of VK users
The analysis of VK clusters cores (Table 2) allowed describing the following
types of users posting behavior: “writers”, “propagators”, “readers” and
“indifferent”.
Table 2. Characteristics of VK clusters
Clusters Num of Num of Num of Num of
friends groups posts friends'
posts
Writers (81 users, 6.9%) 509 10 500 84
Propagators (83 users, 7.1%) 160 70 272 160
Readers (113 users, 9.7%) 436 20 208 1455
Indifferent (890 users, 76.3%) 130 10 107 119
Let us consider the features of the clusters. Seven percent of users got to the
cluster of “writers”. Although this cluster is the smallest, its users have generated the
main content on the given topic. In addition, they more often then other left
comments when reposting. It is interesting that these users have the largest number of
843
friends, which could be explained by opposite reasons – either their writings are
supported by the community or they expand the channels of information diffusion
themselves.
Next in number is a cluster of “propagators” (also about 7%); these are users with
the largest number of reposts from thematic groups. Unlike other users, they have
few friends focusing on the concentrated content collection from thematic groups and
its further reposting.
Among the people who are actively interested on a given topic the most numerous
group is “readers” (about 10%) whose news feed is full of friends’ posts. Unlike
“propagators”, they are focused on the consumption of information, rather than on it
distribution.
However, the majority of VK users (more than 76%) were indifferent to the
subject and did not show any activity at all.
Graphically identified types of behavior are shown in Figure 7.
1600
1400
Num of friends
1200 Num of groups
1000 Num of posts
Num of friends' posts
800
600
400
200
0
Writers (81 users, 6.9%) Propagators (83 users, 7.1%) Readers (113 users, 9.7%) Indifferent (890 users, 76.3%)
Fig. 7. Average VK clusters characteristics
6 Posting behavior of OK users
The neural network of the same architecture testing on OK dataset has shown
comparable results. There are also clusters of “writers”, “propagators”, “readers” and
“indifferent” users. This indicates the stability of the identified types of posting
behavior. A set of Kohonen maps built on the OK dataset is shown in Figure 8.
844
Fig. 8. Set of Kohonen maps, which are describing the behavior of OK users
The analysis of OK cluster cores (Table 3) shows that in general, the
characteristics of the detected clusters are similar on both OSNs. The only significant
difference was observed in the number of user’s friends: in VK the biggest number
of friends had “writers” and “readers”, then in OK this indicator took the maximum
for “propagators”.
Table 3. Characteristics of OK clusters
Clusters Num of Num of Num of Num of
friends groups posts friends'
posts
Writers (8 users, 2.9%) 121 2 505 31
Propagators (17 users, 6.2%) 455 26 101 60
Readers (28 users, 10.1%) 83 1 225 990
Indifferent (223 users, 80.8%) 73 1 127 20
Graphically identified types of OK users posting behavior are shown in Figure 9.
845
1200
1000 Num of friends
Num of groups
800 Num of posts
Num of friends' posts
600
400
200
0
Writers (8 users, 2.9 %) Propag ators (17 users, 6.2 %) Readers (28 users, 10.1% ) Indifferent (22 3 users, 80.8%)
Fig. 9. Average OK clusters characteristics
7 Conclusions
Analysis of frequency characteristics of 248k VK profiles and 238k OK profiles
has shown that more than 40% of VK users and 25% of OK users did not have
friends; about a quarter of a percent of users took part in more than 500 thematic
groups. In addition, the distribution of posts had the fat tail, identifying the users who
had written more than 500 posts on a given topic. These facts actualized the need for
a more detailed study.
Experimenting with different architectures has allowed creating Kohonen neural
networks of the same structure (Figure 10) to identify users posting behavior on both
OSNs (VK and OK).
846
Num of
friends Writers
Num of Readers
groups
Num of
Propagators
posts
Num of
friends' Indifferent
posts
Fig. 10. Kohonen neural network, which is describing users posting behavior
Comparing the clustering results on both social networks, we note that the
identified types of users behavior, namely: “writers”, “propagators”, “readers” and
“indifferent”, have proved to be stable. In addition, we distinguished “writers” from
“propagators” and “readers” not proving the common hypothesis that content
creators are also the main channel of information propagation.
It was shown that on both OSNs the cluster of “writers” was the smallest,
however, its users had generated the main content on a given topic (the number of
their daily posts allows making an assumption about their bias); in addition, they
more often than others left comments when reposting. The main identifying criterion
of this cluster was the number of posts generated by one user.
Next in number was a cluster of “propagators”, which contained the users with the
largest number of reposts from thematic groups. The main clustering criterion here
was the number of thematic groups subscribed by the user.
Among those who were actively interested on a given topic, the most numerous
group were “readers” whose news feed was full of relevant posts. The main
clustering criterion here was the number of user friends’ posts on the given topic.
However, the absolute majority of both OSNs users were indifferent to the topic;
they could be considered as a potential audience (Figure 11).
847
100
90
Writers
80 Propagators
Readers
70
Indifferent
60
50
40
30
20
10
0
VK OK VK OK VK OK
Num of groups Num of posts Num of friends' posts
Fig. 11. VK and OK clusters characteristic
Experimental data and obtained conclusions can be used both in the theoretical
analysis (substantiation of behavioral axioms and hypotheses) and in research of
applied problems related to the posting behavior of social network agents for the
development of effective mechanisms for the information flows formation. It could
be of help in Green Technologies promoting in society enforcing their wide
implementing.
References
1. Adamic L., and Glance N. (2005). The political blogosphere and the 2004
U.S. election: Divided they blog. Retrieved from:
https://pdfs.semanticscholar.org/1197/
1e428132ade5439f77eea258140302865ad7.pdf
2. Bumsuk L. (2012). A temporal analysis of posting behavior in social media
streams. Retrieved from:
http://www.aaai.org/ocs/index.php/ICWSM/ICWSM12/
paper/viewFile/4741/5094
3. Benevenuto F., Magno G., Rodrigues T., and Almeida V. (2010). Detecting
Spammers on Twitter. Retrieved from: http://www.decom.ufop.br/fabricio/
download/ceas10.pdf
4. Google Trends (2017). Retrieved from:
https://trends.google.com/trends/explore?
date=all,all&geo=,&q=Environmental%20Technology,Green%20Technology
5. Gruhl D., Guha R., Liben-Nowell D., and Tomkins A. (2004). Information
diffusion through blogspace. Retrieved from:
http://people.csail.mit.edu/dln/papers/ blogs/idib.pdf
848
6. Guo L., Tan E., Chen S., Zhang X., and Zhao Y. (2009). Analyzing patterns
of user content generation in online social networks. Retrieved from:
https://cs.gmu.edu/~sqchen/publications/kdd09.pdf
7. Kononova K., Kovpak E. (2016). Green ICTs: impact on the environmental
sustainability. International Journal of Sustainable Agricultural Management
and Informatics. Inderscience, Vol. 2, No. 2/3/4, pp. 95-109.
8. Kumar R., Novak J., Raghavan P., and Tomkins A. (2003). On the bursty
evolution of blogspace. Retrieved from:
http://www.disco.ethz.ch/lectures/fs12/ seminar/paper/Barbara/32.pdf
9. Leskovec J., Backstrom L., Kumar R., and Tomkins A. (2008). Microscopic
evolution of social networks. Retrieved from:
http://citeseerx.ist.psu.edu/viewdoc/
download?doi=10.1.1.141.6919&rep=rep1&type=pdf
10. Liu J., Dolan P., and Pedersen E. R. (2010). Personalized news
recommendation based on click behavior. Retrieved from:
http://cs.northwestern.edu/~jli156/IUI224-liu.pdf
11. Papagelis M., Murdock V., and van Zwol R. (2011). Individual behavior and
social influence in online social systems. Retrieved from:
http://www.cs.toronto.edu/ ~papaggel/docs/papers/all/HT11-Individual-
Behavior-and-Social-Influence-in-Online-Social-Systems.pdf
12. Roman P. E., Gutierrez M.E., and Rios S.A. (2012). A model for content
generation in On-line social network. Retrieved from:
https://www.researchgate.net/ publication/233897861
13. Xu Z., Zhang Y., Wu Y. and Yang Q. (2012). Modeling User Posting
Behavior on Social Media. Retrieved from: http://yaowu.co/docs/sigir12.pdf
849