DISCOVERING USERS TOPIC OF INTEREST FROM TWEET

International Journal of Computer Science & Information Technology (IJCSIT) Vol 10, No 1, February 2018
DOI:10.5121/ijcsit.2018.10108 95
DISCOVERING USERS TOPIC OF INTEREST
FROM TWEET
Muhammad Kamal Hossen1
, Md. Ali Faiad1
, Md. Shahnur Azad Chowdhury2
and
Md. Sajjatul Islam3
1
Department of Computer Science and Engineering, CUET, Chittagong, Bangladesh
2
Department of Business Administration, IIUC, Chittagong, Bangladesh
3
Department of Computer Science and Engineering, CIU, Chittagong, Bangladesh
ABSTRACT
Nowadays social media has become one of the largest gatherings of people in online. There are many ways
for the industries to promote their products to the public through advertising. The variety of advertisement
is increasing dramatically. Businessmen are so much dependent on the advertisement that significantly it
really brought out success in the market and hence practiced by major industries. Thus, companies are
trying hard to draw the attention of customers on social networks through online advertisement. One of the
most popular social media is Twitter which is popular for short text sharing named ‘Tweet'. People here
create their profile with basic information. To ensure the advertisements are shown to relative people,
Twitter targets people based on language, gender, interest, follower, device, behaviour, tailored audiences,
keyword, and geography targeting. Twitter generates interest sets based on their activities on Twitter. What
our framework does is that it determines the topic of interest from a given list of Tweets if it has any. This
process is called Entity Intersect Categorizing Value (EICV). Each category topic generates a set of words
or phrases related to that topic. An entity set is created from processing tweets by keyword generation and
Twitters data using Twitter API. Value of entities is matched with the set of categories. If they cross a
threshold value, it results in the category which matched the desired interest category. For smaller amounts
of data sizes, the results show that our framework performs with higher accuracy rate.
KEYWORDS
Tweet analyse, topic generation, social media interest generator, topic of tweet, short text analyse, predict
topic from tweet.
1. INTRODUCTION
General people now gather in popular social media, such as Facebook, Twitter, Instagram,
LinkedIn, WhatsApp, etc. The number of people gathered here is big and can be compared to a
big market for industries. This is an important fact but also is very tough to know the taste and
interests of each people for advertising. User-generated content is the lifeblood of social media
[1].They stores information about users related to their age, gender, interests, location, something
they shared, etc. This stored information allows advertisers to target specifically a particular
group to show their advertisements. The advertisers can thereby analyse the stored information to
categorize the interests of that individual group. This is an advantage for the advertisers and that
specific groups of people will only be shown the advertisements according to their taste. Social
media like Twitter also follow this method. This is why Twitter is so famous for advertising in
online.To advertise on Twitter, there are promoter tweets, trends, promoted accounts and other
things that show up on users’ newsfeeds [2].Most of the Twitter users use English language and

96
so, nearly half of the tweets are in the English language. This is why the English language plays a
significant role in analysing tweets. Our framework can analyse only English tweets.
As tweet does not necessarily maintain the grammatical rules, direct language processing
algorithms that maintain grammatical rules are not implemented in this framework. Furthermore,
discovering interest set is one very much useful for advertising on Twitter. Every advertiser has a
list of Twitter users who have been on their advertisement's list and bought a product. This list
contains their usernames and advertisers follow them. From this given username, we can retrieve
their tweets that can be very much advantageous for knowing their interests in different topics for
retargeting. All these motivated us to design a system that is not only grammar independent
analyser but also useful for the advertisers.
Our work is done using some free API (Application Program Interface). For the time limit for a
given request, we could only process one tweet at a time and discover interested topics.
Moreover, we have some 'false-positive' issues. Other than that, it has greater accuracy rate and is
implemented for the testing condition.
The remainder of this paper is as follows, section 2possesses some related works. Section 3
formally presents the problem. Section 4 describes the design of the framework. Section 5 shows
the experimental results and observations where accuracy is calculated. Lastly, we concluded the
paper in section 6 where we also suggested some future recommendations.
2. RELATED WORKS
Before starting our work, we followed some other works that led us to this project. These works
show the targeted advertisings' effectiveness that advertises leading to more profits for the
products and how free API is used for social data collection. Chan Mei Lee et al. mentioned that
since using the traditional survey method will limit the coverage of different demographic people
and increase the cost, hence the future researcher is highly encouraged to perform an online
survey as the targeted respondents of this research are internet users [3].
The study of Billy Bai et al. attempted to explore the marketing effectiveness of two different
social media sites (Facebook and Twitter). This study also indicated that different social media
sites demonstrate the same marketing effectiveness [4].Gupta et al. promoted posts that provide
businesses with the ability to push a post out to not only fans but likers of fans, increasing the
reach dramatically. Offers allow businesses to present advertised offers which spread virally as
people accept the offer acting as brand advocates for one’s brand by pushing this to their friends.
This is really only the tip of the Iceberg. Social media is projected to ramp up its revenue-
generating activities [5].
A study was conducted by B. Pikas et al. to test Americans perception of online advertising on the
popular web pages such as Facebook, YouTube, and Twitter [6].The framework proposed by
Kenny Ho et al. helps to identify the most effective and efficient spammer detection features,
evaluate the impact of using different numbers of recent tweets, and therefore, obtaining a faster
and more accurate classifier model [7].
According to Waring et al., given an advertising category, it can be estimated the most promising
areas to be selected for the placement of an advertisement that can maximize its targeted
effectiveness [8].Uchiya et al. specified that when using Twitter, it is difficult for users to find
appropriate users to follow among the vastly numerous users [9]. To solve this problem, we
propose a follow-user recommendation system on Twitter based on an interest domain. Moreover,
we demonstrate the effectiveness of the proposed method through experimentation.

97
Liu et al. proposed a framework through which important research themes and concepts are
synthesized to provide IS researchers with an overview of this research area and to identify those
topics where much research has already been done and those topics where more research is
needed [10]. For click-through prediction, X. He et al. stated that the task of click-through
prediction for online advertising at Twitter is related to the practice in the social media industry,
including online advertising on Facebook, LinkedIn, etc. The most relevant work in literature is a
recent study of predicting clicks on Facebook ads [11].McCorkle et al.presented details for
implementing a Twitter project in a digital marketing course as a means of addressing these gaps.
Student’s feedback on this project indicated that their skills in social media marketing, digital
marketing, and personal branding were improved. Suggestions for use of Twitter in other courses
are also presented [12].
3. PROBLEM DEFINITION
EICV stands for Entity Intersect Categorized Value. It mainly finds the interested topic from
tweet for a given set of interests. The equations for EICV are described below. Here,
Wn = set of words,
Rn = set of redundant words and phrases,
Kn = set of retrieved words from knowledge source for Entities,
Tn = set of topics,
Ttag[n] = set of tags for each topic Tn.
If an element Wn is related to a topic in Tn, then it creates,
= − (1)
where, Ftwitter is a function that generates entities from knowledge source (for our case, it is
Twitter search API). The equation for intersection value is
= ∩ [ ] (2)
For each topic from Tn, it generates an intersection value. We will set a threshold value Tv. The
intersection values greater than the threshold value are our desired topics.
4. EICV - SYSTEM ARCHITECTURE
The architecture of the system comprises four basic modules. These modules are:
Topic and tag input module
Tag generator module
Tweet input module
Interest generator module
The first module takes the advertisement topics and the tags for each advertisement topic as input.
The second module generates related words and phrases. This words and phrases are normally
ambiguous. So, after detecting the ambiguous words and phrases, we remove as much ambiguity
as possible. In the third module, Twitter data can be inputted. We can input Tweets directly into
the system or enter a username to retrieve the last tweet of that account. Lastly (in the interest
generator module), the matched topics to the tweet are generated.

98
4.1. Topic and Tag Input Module
4.1.1. Topic Input
Firstly, the topics to be advertised are defined. There is a list of advertisement topics in the
Twitter advertisement section. In our system, we input the list of topics in the csv file. Algorithm
1 states the process of inputting topics and tags.
Algorithm 1. Input topics and tags
i. Start
ii. Input topics into the csv file
iii. Create a csv file for each topic
iv. Input tags into the correspondent csv files
v. End
For only simplicity of the research purposes, we took 10 topics that are randomly selected. But
the system is designed to be extendable. That means, we can take as many topics as we like.
These topics can be deleted too. This is the set of topics, Tn.
4.1.2. Tag Input
After topics being set, a csv (comma separated values) file is created for each topic. This type of
file contains different values separated by a comma. Tags are words or phrases that are related to
this topic. For each topic, we can enter tags in the corresponding csv file. For example, if we have
a topic named “Football”, a file will be created named “football.csv” and we can put the tags on
there.
4.2. Tag Generator Module
Algorithm 2 states the process of generating tags using datamuse API and elimination of
ambiguous tags.
Algorithm 2. Generating tags using datamuse API
i. Start
ii. Take topic names from the csv file
iii. Generate tags for each topic into the correspondent csv file
iv. Find the ambiguous words and phrases
v. Delete the ambiguous words and phrases
vi. End
4.2.1 Datamuse API
As tags selected by human are not enough for this research, we took help of API (Application
Program Interface). These are the tools that enable us to use any external application through
interface either on free of any service cost or on paid terms. We used a free API for this purpose
named datamuse API. It provides a lot of features including finding tags for any topic. This is
important to have as many tags as we can get since it results in higher accuracy.

99
4.2.2 Ambiguity
As being free API, this does not generate tags which denote that topic only. Moreover, there are a
lot of words generated that do not have the quality to be tagged. These words are called
ambiguous words. We made a way to find these ambiguous words and show them to the trainer.
A list of these ambiguous words is shown for each topic.
4.2.3 Eliminating Ambiguity
To be able to make this system working, we have to train it first. There is a list of ambiguous
words for each topic. The tags that have ambiguity are selected and removed. But, there are a lot
of tags that still remain in more than one topic's tag list. For example, “player” is a tag that is
generated by both “Football” and “Cricket”. This is not ambiguous as the tag “player” signifies
for both the topic. After eliminating all ambiguities, this module is ready.
4.3 Twitter Input Module
In this work, we will only analyse the tweets. Tweets are generally 140 characters long (from
November 5, 2017). To check the desired interest on topics by analysing tweets we have shared, it
is needed to enter tweets into the system. A tweet can be entered into the system by two ways.
These are:
i. Enter tweet directly into the system,
ii. Enter Twitter's username to retrieve the last tweet of that user.
Algorithm 3 states how tweets can be inputted into the system.
Algorithm 3. Twitter data input
i. Start
ii. Input tweets directly into the system and go to (v)
iii. Otherwise, take Twitter's username as input
iv. Collect the last tweet using the given username
v. End
4.3.1 Direct Tweet
The tweet that we want to analyse can be entered directly into the system.
4.3.2 Twitter's Username
We can also enter any valid username. Luckily, Twitter has their own API named Twitter API,
free of any charge that allows us to collect all the tweets with details in different formats. Twitter
APIis used to get only the last tweet of a user only for simplification of the procedure. This
system can be extended to work with all the tweets of a particular user.
In both ways, the system gets the tweet which is ready to be analysed.

100
4.4 Interest Generator Module
Interest generator module is the key part of this system. It takes tweets and tags as input and
returns set of interests as output. The whole procedure is done in several stages. These stages sum
up the core functionality of the system. The interest generator module can be divided into six
stages to be exact.
They are:
Redundant words remove
Keyword generation
Gather knowledge according to keywords
Ranking
Find intersect values
Generate topics
Algorithm 4 states the process of generating tags using datamuse API and elimination of
ambiguous tags.
Algorithm 4. Generating topics’ list to be advertised
i. Start
ii. Eliminate unnecessary words and phrases
iii. Generate keywords
iv. Search Twitter with keywords
v. Retrieve results
vi. Rank frequent words
vii. Match most frequent words to the tags
viii. Find tags that are most common with the topics
ix. Display the topics
x. Display error rates
xi. End
4.4.1 Redundant Words Remove
Tweets do not necessarily maintain the grammatical rules as they are really hard to be analysed
by natural language processing. Instead, this method divides the tweets into words that are called
entities. This system is only being tested for the English language with the purpose of testing. The
entities are mostly full of redundant words, such as “lol”, “goooaaal”, “suppperb”, URLs, etc.
Now, we have to find Kn that defines the set of words extracted from the knowledge source. The
redundant words, for this reason, have to be deducted. So, initially, all the redundant words have
been removed.

101
In this stage, the value of (Wn-Rn) has been determined where Wn denotes set of words or entities
of the tweet and Rn denotes set of redundant words.
All the URLs have been removed as the tweets are only taken as input. They are removed through
Regular Expression (RE). Then, a set of words is created in a csvfile named “redword.csv”. These
words are the redundant words which are not necessary for the analysis that can be denoted by
Rnr. So, deleting all the redundant words, it remains (Wn - Rnr) words where,
= + " (3)
Here, Rnk are the words we could not delete.
4.4.2 Keyword Generation
Since the numbers of words which are not necessary to be the keywords are numerous, we took
help of a free API supported by “cortical.io”. After deleting all the redundant words, the system
sends the (Wn-Rnr) words to the website. It deletes all the remaining unnecessary words Rnk and
we got back,
#$ = − (4)
Since this is a time-consuming process, a process is utilized to add words, Rnk to the Rnr so that
after a long period of time, our own Rnr set will be enough for finding the keywords. The
keywords extraction API will no more be needed.
After this stage, we have the set of keywords (Wn-Rn) which are prepared to be inputted into the
next stage.
4.4.3 Gather Knowledge According to Keywords
After having the keywords, we will search into the Twitter with those keywords. Twitter
supporting free API has an option to search tweets on the basis of keywords matching to them.
There are two categories of tweets we can find. They are:
Popular tweets
Recent tweets
Twitter restricted the users to have popular tweets upto 20 tweets and recent tweets up to 100
tweets. We retrieved all the tweets and made up an array of all the words of these tweets.
4.4.4 Ranking
After collecting all the words of the retrieved tweets, the array words' frequency is counted. Then,
the words are sorted in descending order depending on their frequency of occurrence. Gathering
knowledge according to keywords and ranking, together it denotes the function Ftwitter(). This
function takes the argument of entities (Wn-Rn). After frequency counting is done, we choose
those words whose frequency is more than three at least and take a certain number of words. This
set of words is Kn.
4.4.5 Find Intersect value
Now, we have Kn and Ttag[n]. Both of these are set of words or phrases. Then the system finds a
number for each topic. This number is determined by the number of words in the set Kn matches
with the number of words or phrase of Ttag[n].

102
Suppose, we have a topic “Football” that has 200tags denoted by Ttag[football]. For an input tweet,
supposeit is found 20 words denoted by Ktweet. If 10 words are matched, then the intersection
value,
%&& ' (( = %&& ' (( ∩ [%&& ' ((] = 10 (5)
So, for each topic Tn, we will get a value of Vn. Finally, we have the intersection values for each
topic.
4.4.6 Generate Topics
From the intersection value, we will first find the maximum value Tmax. This is the primary topic
of advertisement for that particular tweet. Since each tweet can lead us to more than one topic
also, we set another process to get other interest topics. We set a threshold value,
+ = 3 ∗ . //4 (6)
This threshold value is changeable. Initially, we set it on this which leads us least error rate. All
the topics which maintain Vn> Tv for each topic are our topics of advertisement. If the value of
. / is very low, the number of matched topics Vn become high. So, we added an extra feature
that the tweet will only generate topics when,
. / > 34 4 56 (7)
By default, this Minimum Value is also changeable according to practices. Finally, this system
returns the set of topics to be advertised for a given set of topics to be advertised and a given
tweet.
5. EVALUATION OF RESULTS
The framework has been tested for over 1378 tweets or usernames (till December 15, 2017). As
for the experimental review, we specified our system with only ten particular topics which are:
football, cricket, golf, baseball, movie, book, food, politics, drink, and science. We have also
tested by increasing the number of topics to 25 topics which really did not affect the accuracy of
the system. Some of the test cases are shown in Table 1.
Table 1. The generated interest of topics for the proposed system.
Tweets Matched Topic
I like Ronaldo Football
I like Harry Potter Movie
How good Shahrukh Khan is!!1! Movie
This time I am not going to miss pizza Food
I am frustrated No topic matched
Sachin is my idol Cricket
The tweet “I am frustrated” did not result in any tweet because it did not match with any of the
advertisement topics which are true. This clarifies that TweetAd (our proposed system) generates
all the interests of topics correctly from only tweets for this set of results.

103
For calculating accuracy, we used the following equation:
788 589 5: =
;$.' &% <<$ &# < & ∗=>>
;$.' &% ? < ? ?
(8)
It has correctly discovered 1282 times among 1378 tweets or usernames and could not detect
correctly 96 times. So, the accuracy rate is 93.03%.Again to be mentioned, let we have a tweet
with an entity “Ronaldo”. If we test this on the framework, most of the times it displayed only
one topic “Football” which is considered as correct, 2nd
most topic generation was both
“Football” and “Cricket” which is also considered as correct, and 3rd
most frequent result was
“Cricket” and “Football” which is also considered as correct. We considered a topic being
guessed correct if the topic is at most of the 2nd
rank and that did not result in any other topic
rather than it being a game. If one of the topics is “Politics”, we would not consider that as an
accurate topic generation.
6. CONCLUSIONS
Advertising is one of the hottest topics in recent days in social media. Researchers are working
hard for a long time to pinpoint Twitter users' interest in an advertisement topic. The performance
of the existing system is good. Still, a lot of targeting options are missed through the crack of the
system. The proposed system can be used in two ways. In one way, implementing it in Twitter, it
would have more data for analysing which might result in greater accuracy. In another way, it can
be used as an external application for the advertisers who can determine the people interested in
their products with more accuracy alongside with Twitter advertisement system.
As every system has some limitations, this system also has that are stated below:
This system detects wrong topics in a rare case as the resources were limited.
This system will only work for the English language.
There are some cases where topics are not generated, though it surely has a topic to be
generated.
If any tags are given as input, this system can only retrieve the last tweet only
Using keyword generation API supported by cortical.io is not a great solution which is
time-consuming.
As we used free API, the number of requests is limited. It will work unlimited on paid
service.
As this is only one leap toward a development, there are a lot of works can be done to improve
this system or leading to another great system. Future systems can have features for other
languages also. Increasing the data source, the tags generated for each topic will be more
accurate. Future systems can determine interest set for all tweets for a given user. Lastly, our
suggestion will be to make a complete advertisement system along-side with the existing method
which will be perfect for commercial uses.
REFERENCES
[1] J. A. Obar and S. S. Wildman, (2015) “Social media definition and the governance challenge: An
introduction to the special issue”, Telecommunications Policy, vol. 39(9), pp. 745-750.

104
DOI:http://dx.doi.org/10.2139/ssrn.2647377
[2] S. Goyal, (2013) “Advertising on social media”, Scientific Journal of Pure and Applied Sciences,
vol.2, no. 5, pp. 220–223.
[3] M. L. Chan, H. L. Chee, S. P. Chin, and G. X. Sim, (2014)“Consumers' attitude towards online
advertising: the study on information responses”, BIB (Hons.) thesis, UTAR.
http://eprints.utar.edu.my/id/eprint/1348
[4] X. Y. Leung, B. Bai, and K. A. Stahura, (2015) “The marketing effectiveness of social media in the
hotel industry: A comparison of Facebook and Twitter", Journal of Hospitality & Tourism Research,
vol. 39, no. 2, pp. 147-169. DOI: https://doi.org/10.1177/1096348012471381
[5] G. Bakshi, and S. K. Gupta, (2013) “Online advertising and its impact on consumer buying behavior",
International Journal of Research in Finance and Marketing, vol. 3, no. 1, pp. 21-30.
[6] B. Pikas, and G. Sorrentino, (2014) “The effectiveness of online advertising: consumers perceptions
of ads on Facebook, Twitter and Youtube", Journal of Applied Business and Economics, vol. 16, no.
4, p. 70-81.
[7] K. Ho, V. Liesaputra, S. Yongchareon, and M. Mohaghegh,(2017) “A framework for evaluating anti
spammer systems for Twitter", In Panetto H. et al. (eds): OTM Conferences On the Move to
Meaningful Internet Systems. Lecture Notes in Computer Science, vol. 10573, Springer, pp. 648-662.
DOI: https://doi.org/10.1007/978-3-319-69462-7_41
[8] M. E. Waring, K. L. Schneider, B. M. Appelhans, T. A.M. Simas, R. S. Xiao, M. C. Whited, A. M.
Busch, M. M.Evans, and S. L. Pagoto, (2016) “Interest in a Twitter-delivered weight loss program
among women of childbearing age", Translational Behavioral Medicine, vol. 6, no. 2, Springer, pp.
277-284. DOI:10.1007/s13142-015-0382-4
[9] T. Uchiya, Y. Ishida, Y. Kume, and I. Takumi, (2016) “Proposal of follow user recommendation
system on Twitter based on interest domain", In IEEE 5th Global Conference on Consumer
Electronics. DOI: 10.1109/GCCE.2016.7800413
[10] D. Liu and H. R.Weistroffer, (2016) “A look at online targeted advertising in information systems
research”, In Proceedings of the Southern Association for Information Systems Conference (SAIS),
USA. https://aisel.aisnet.org/sais2016/17
[11] X. He, J. Pan, O. Jin, T. Xu, B. Liu, T. Xu, Y. Shi, A. Atallah, R. Herbrich,S. Bowers, et al., (2014)
“Practical lessons from predicting clicks on ads at Facebook”, In Proceedings of the Eighth
International Workshop on Data Mining for Online Advertising, ACM, pp. 1–9. DOI:
10.1145/2648584.2648589
[12] D. E. McCorkle and J. M. Payan, (2017) “Using Twitter in the marketing and advertising classroom
to develop skills for social media marketing and personal branding”, Journal of Advertising
Education, vol.21(1), pp. 33-43.
AUTHORS
Muhammad Kamal Hossen has received his B. Sc. and M. Sc. in Computer Science
& Engineering (CSE) degrees from the department of Computer Science &
Engineering of Chittagong University of Engineering & Technology (CUET),
Bangladesh in 2005 and 2015 respectively. He is now pursuing his Ph. D. degree in
CSE from the same university. Since 2006, he has been serving as a faculty member in
the Department of CSE, CUET. His research interests include digital image
processing, cryptography, steganography, and pattern recognition, data mining.

International Journal of Computer Science & Information Technology (IJCSIT) Vol 10, No 1, Februa
Md. Ali Faiad received B. Sc. in Computer Science & Engineering degree from
Department of Computer Science and Engineering of CUET), Bangladesh in 2017. His
research interests include data privacy and data mining, distributed systems and cloud
computing, big data management.
Md. Shahnur Azad Chowdhury
Engineering from International Islamic University Chittagong (IIUC), Bangladesh in
2003 and M. Sc. in 2012 from Daffodil Int'l University, Dhaka. He has
the IIUC for the last fourteen years as a lecturer, assistant professor and associate
professor respectively. His areas of research interest are
data mining, IoT.
Md. Sajjatul Islam has received his B. Sc. i
degree from the department of CSE of Chittagong University of Engineering &
Technology (CUET), Bangladesh in 2005 and and Master of Science in Computing
from University Of Wales (UoW), United Kingdom in 2011. He is pu
degree in CSE from CUET. He is now serving as an assistant professor in the
Department of CSE of Chittagong Independent University (CIU), Bangladesh. His
research interests include big data analysis, cloud computing.
.
International Journal of Computer Science & Information Technology (IJCSIT) Vol 10, No 1, Februa
received B. Sc. in Computer Science & Engineering degree from
Department of Computer Science and Engineering of CUET), Bangladesh in 2017. His
research interests include data privacy and data mining, distributed systems and cloud
Md. Shahnur Azad Chowdhury received B. Sc. in Computer Science and
Engineering from International Islamic University Chittagong (IIUC), Bangladesh in
2003 and M. Sc. in 2012 from Daffodil Int'l University, Dhaka. He has been serving
the IIUC for the last fourteen years as a lecturer, assistant professor and associate
professor respectively. His areas of research interest are natural language processing,
has received his B. Sc. in Computer Science & Engineering (CSE)
degree from the department of CSE of Chittagong University of Engineering &
Technology (CUET), Bangladesh in 2005 and and Master of Science in Computing
from University Of Wales (UoW), United Kingdom in 2011. He is pursuing his Ph. D.
degree in CSE from CUET. He is now serving as an assistant professor in the
Department of CSE of Chittagong Independent University (CIU), Bangladesh. His
research interests include big data analysis, cloud computing.
105

DISCOVERING USERS TOPIC OF INTEREST FROM TWEET

More Related Content

What's hot (15)

Similar to DISCOVERING USERS TOPIC OF INTEREST FROM TWEET (20)

Recently uploaded (20)

DISCOVERING USERS TOPIC OF INTEREST FROM TWEET