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

net/publication/330316652

Application of Artificial Intelligent in the Prediction of Consumer Behavior

from Facebook Posts Analysis

Article · February 2019

DOI: 10.18178/ijmlc.2019.9.1.770

CITATIONS READS

2 162

4 authors, including:

Pornpimon Kachamas Suphamongkol Akkaradamrongrat

Chulalongkorn University Chulalongkorn University
4 PUBLICATIONS   2 CITATIONS    3 PUBLICATIONS   2 CITATIONS   

SEE PROFILE SEE PROFILE

Achara Chandrachai
Chulalongkorn University
47 PUBLICATIONS   143 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Co-author View project

Innovative R&D Evaluation System View project

All content following this page was uploaded by Suphamongkol Akkaradamrongrat on 12 February 2019.

The user has requested enhancement of the downloaded file.

 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

Application of Artificial Intelligent in the Prediction of

Consumer Behavior from Facebook Posts Analysis
Pornpimon Kachamas, Suphamongkol Akkaradamrongrat, Sukree Sinthupinyo, and Achara
Chandrachai

 including sentiment analysis from comment posts. It may be

Abstract—The undeniable fact is that online business today asserted however that we lack research using the capacity of
has increased at a very fast pace everywhere around the globe. Data Mining to conduct an in-depth analysis of consumer
This happens through the widely used of Social Media, behavior probability on Facebook.
especially Facebook which is the most popular platform in the
world. It would be really useful for the digital marketers, if
For the above reason, this research uses Data Mining
there is a certain tool that can predict the intentions of the web technique to analyze consumer behavior probability from
patrons when the brand is posting the message to communicate marketing communication by studying comments and
with their fans or followers. messages posted on Facebook. The data is also used in
The aim of this research is to develop an analytic tool which Machine Learning so as to be an advantage for Facebook
can support online vendors to predict behaviors of the patrons marketers. This research aims to assist online marketers on
according to Dentsu’s AISAS perspective. An Artificial their attempt to communicate to consumers on Facebook in
intelligent model was developed by the results from 75
specialists who evaluated the behavior that will likely occur accordance with desired consumer behavior using integrated
after the comments have been posted. The results, hence, were marketing according to Dentsu’s AISAS model.
collected and prepared for the data modelling process using the
Naïve Bayes probability concept, afterwards, testing for the
model’s accuracy with 10-fold cross validation technique. As the II. BACKGROUND THEORIES
previous study indicated, Naïve Bayes technique gives the best
result for the behavior analysis, which is also true with this A. Machine Learning
study. The predictive model for AISAS behavior from this study Machine learning aims to maximize the efficiency of
can give average accuracy higher than 86 percent.
computer program using sample data or precedent
When bringing the AISAS Model to test with 30 live users
who are online vendors, we can conclude that the overall results experiences. A model is developed with designated
of model have been greatly appreciated and effectively satisfied. parameters in learning to perform computer program
Most vendors also agreed on the ease of use, which creates high operation in order to enhance efficiency of the model using
chances of business opportunities. train data or past experiences. The model is developed to be
capable of predicting or describing knowledge from the data,
Index Terms—AISAS, machine learning, naive-bayes or both [3].
classification, social media, social media analytics.
Machine learning bears a very close relation to statistics [4]
in such a way that the concept of machine learning has
derived from the theory of statistics. Breiman [5] argues that
I. INTRODUCTION
statistics suggests a greater importance of the data, whereas
Text Mining consists of the combination of human machine learning give higher importance to algorithm.
capacity to effectuate natural language processing (NLP) and There are several types of machine learning including
Data Mining [1]. Data Mining is able to extract useful data unsupervised learning, supervised learning and
from diverse resources by means of specific data reinforcement learning. The author of this study resumes
identification and data pattern discovery with an objective to these types as shown in Fig. 1.
manage a great number of data systematically and to make As mentioned above, the objective of this research is to
data ready to be used appropriately [2]. conduct a study of consumer behavior probability from
The application of Data Mining is widely used in business marketing communication on Facebook using machine
environment. One of the applications of such method can be learning. Machine learning effectuates data classification to
observed from data analysis in social media platforms cluster consumer behavior characteristics applying Naï ve
Bayes Classification) [7] from Bayes’ Theorem. This
research also applies Independence Term Assumption [8]. It
Manuscript received August 23, 2018; revised October 24, 2018. is true that several probability equations, including Decision
P. Kachamas is with the School of Technopreneurship and Innovation
Management, Graduate School, Chulalongkorn University, 254 Phayathai Tree, Neutral Networks, Support Vector Machine (SVM),
Rd., Phatumwan Bangkok, 10330 Thailand (e-mail: can be applied to analyze consumer behavior, the best result
pornpimon.kac@student.chula.ac.th). in behavior analysis [9], however, is carried out by Naï ve
S. Akkaradamrongrat and S. Sinthupinyo are with the Department of
Computer Engineering, Chulalongkorn University, 254 Phayathai Rd., Bayesian Classifications [10]. In addition, Naï ve Bayesian
Phatumwan Bangkok, 10330 Thailand (e-mail: Classifications continues to operate to bring about the result
6070327421@student.chula.ac.th, sukree.s@chula.ac.th). even in case of partly lost data making this more efficient that
A. Chandrachai Author is with the Department of Commerce,
Chulalongkorn University, 254 Phayathai Rd., Phatumwan Bangkok, 10330
Support Vector Machine (SVM) [11].
Thailand (e-mail: achandrachai@gmail.com).

doi: 10.18178/ijmlc.2019.9.1.770 91
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

understanding. To generate motivation of consumers in

high valued product or service, service providers must use
these external variables to stimulate consumer needs, to
make them see the value in the service and use the service
in fine.
2) Perceived Usefulness consists of the perception of an
individual in advantages generated by using technology.
That is to say that technology is perceived to be useful in
development of his or her performance in work.
3) Perceived ease of Use consists of the ease used to
designate individual perception so as to achieve the
objective easily in terms of quantity, or the success in
getting what corresponds to the need and expectation.
4) Attitude toward using and interest in using technology.
5) Behavioral Intention to Use or behavioral interest to use
Fig. 1. Example of machine learning [6]. technology.

B. AISAS Model
Dentsu’s AISAS mode relates to buying process of a
consumer taking data on internet in consideration prior to
buying. The model consists of Attention, Interest, Search,
Action, and Share [12] which are factors taken into account in
Fig. 3. Technology acceptance model (Davis, 1989).
product and services purchasing decision.
According to TAM, the influence of the external variables
is exercised on belief, attitude and interest in using
technology through two perceptions, the perception of
technology usefulness and the perception of technology use
convenience namely, which effect the implementation of the
Fig. 2. The steps of AISAS models [12].
system. The model is widely applied and adopted as a schema
Working process of AISAS model starts when a consumer in successful decision in the prediction of technology
notices or sees a product, a service or an advertisement acceptance. Moreover, the model illustrates the reasons
(Attention). Then, the Interest is stimulated in order to get related to individual technology perception in terms of
more information about such product or service (Search). In usefulness and convenience so as to build behavior which can
this stage, a consumer will seek for the others’ comments, be interpreted as an interest in technology use. These effect in
reviews, posts from internet, comparison websites, formal turn the application and the acceptance of technology.
webpages of an organization or even from a conversation
with family members or friends who have been using the
product or service. After that, the consumer will express III. METHODOLOGY
his/her own opinion considering the others’ opinions on that This research places an emphasis on supervised learning
product or service. This stage leads to the decision of the according to machine learning theory. We enter data
consumer to buy (Action). Finally, the consumer will act as a extracted from Facebook posts. The behavior type is
messenger by word of mouth, conversation, or online posts of identified and clustered to as potential labels by purposive
his or her opinion and impression (Sharing). Dentsu’s AISAS sampling method. Then, 81 Thai language experts
model does not restrict the order of the occurrence of stages, randomized from ten schools around Bangkok and vicinity is
they can be shifted, altered or repeated [12]. invited to participate in questionnaire. After that, the samples
are specified according to the appropriateness of language
C. Technology Acceptance Model (TAM) expertise [15]. The questionnaire is distributed to Thai
Technology Acceptance Model or TAM consists of the language experts in order to classify the posts in accordance
study with a view to understand and reach end users on how with AISAS model. And finally, we arrive to the conclusion
they come to accept and use technology [13]. This theory is indicating the correspondence of the posts with each AISAS
adapted from Theory of Research Action or TRA [13], [14]. so as to develop further prediction model for the system.
Technology Acceptance Model theory is a model to assert the The detail of each step is given as follows:
acceptance of technology developed to display a clear picture 1) The first step consists of the collection of Facebook posts.
of this behavior. The model also provides a clearer The data will be extracted and collected from three
explanation of acceptance process of an individual to adopt Facebook Pages with the total amount of 8,172 posts. The
technology in everyday use. data collection periods of Page A is between 6th
As seen in Fig. 3, TAM shows factors exercising an December 2012 and 8th October 2017, Page B from 30th
influence in technology use decision as follows: September 2013 to 6th October 2017 and Page C from 24th
1) External Variable is an external influence upon individual September 2009 until 8th October 2017. Once all data is
acceptance including belief, experience, knowledge, and completely collected, we filtrate out repeated posts,

92
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

picture posts, video posts which do not contain message. 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠𝑖
𝑅𝑒𝑐𝑎𝑙𝑙𝑖 = (4)
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠𝑖 + 𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒𝑠𝑖
Finally, for each page, we arrived at 1,000 posts
representing the highest rate of interaction of page 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠𝑖
members, these posts is then analyzed accordingly. 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖 = (5)
𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠𝑖 +𝐹𝑎𝑙𝑠𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒𝑠𝑖
2) The next step is to select messages by ordering them
following the ranking of reaction amount together with TABLE I: CLASSIFICATION DECISION
the number of comments and shares from highest to Value classified by Value of Classi (Ci)
classifier Yes No
lowest. The selected message must be in Thai language, Yes TP FP
clear and legible filtrated from three pages with the total No FN TN
of 3,000 posts: 1,000 posts equally from each page.
3) The third step consists of AISAS classification on 1) A user interface is designed in this step together with the
messages. The step starts with creating labels indicating development of analysis system of Facebook posts so as
Attention, Interest, Search, Action and share namely. to predict whether that post affects consumers
During this process, the researcher uses survey method to considering AISAS pattern. The natural language
indicate the label for learning system via questionnaire. processing is applied to cluster and analyze extensively so
Then, the questionnaire is presented to Thai language that the owners of the pages and the vendors enter the
experts from schools and universities in order to classify posts they wanted to communicate. This aims to test
each post to corresponding Dentsu’s AISAS integrated whether the probability is in line with the objectives of
marketing pattern. The obtained data is considered as the communication to consumers. The result is carried out in
set of training data for machine learning applying Naï ve form of visualization. If it corresponds to the objective,
Bayes Classification [7]. the message will be posted directly from Web application
4) The fourth step aims to verify and validate model quality. without signing in Facebook to post. In the other hand, in
We call on K-Fold Cross Validation which is the model to case that it does not correspond to the objective, the
detect errors in posts classification into AISAS pattern on system will suggest a group of words to business owners.
training dataset. The random begins with clustering data 2) Technology acceptance model test is situated in this step.
into equal K-part; one K-part for testing dataset, the We select testers to test and use and send model system
others K-parts for training dataset. A model will be then prototype of analysis system of messages posted on
created to test AISAS classification of each post. After Facebook pages influencing upon consumer behavior
that, testing dataset is shuffled until K [16] time is according to Dentsu’s AISAS model to marketers who
completed. In this test, the training dataset consists of possess Facebook account to test. The messages posted
3,000 posts used to develop a training model applying through Facebook pages are then collected so as to
Naï ve Bayes algorithm in AISAS classification of each perform a study of online consumer behavior according to
post. We limit the number of dataset at 10 datasets (k=10) Dentsu’s AISAA model theory. This group is comprised
in cross validation. of 152,583 pages [17] of businesses using Facebook to
communicate or operate the business in Thailand. The
Page A sample in this research is selected using Purposive
Sampling method among 31 marketers present on
Page B
Data verification AISAS Cross-Validatio Facebook, and who use Facebook as public relations tool
and selection
Page C
classification n to promote their business. [18]. It can be stated that the
amount of 30 marketers is the minimum to have normal
distribution. Normal distribution provides average value
TAM and standard deviation which shows the least and
acceptable errors. In addition, normal distribution
provides an explanation of the fact fairly even though
Design and develop there is no need of normal distribution in statistics.
Fig. 4. Research methodology.
The criteria of score value obtained by the
questionnaire in opinion section follows the model Sig
Accuracy defines the value of how accurate the program Sense [19], which divide result following Class interval, and
can predict. can be interpreted as follows:
Average value between 4.21 and 5.00 means very good.
(𝑇𝑃+𝑇𝑁) Average value between 3.41 and 4.20 means good.
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = (2)
(TP+TN+FP+FN)
Average value between 2.61 and 3.40 means average.
Average value between 1.81 and 2.60 means fair.
Recall = the number of all related documents which have
Average value between 1.00 and 1.80 means poor and need
been extracted divided by all related documents.
to be improved.
Precision = the number of related documents which have
been extracted divided by the number of extracted document,
as illustrated in Table I and equation (3), (4), (5) respectively.
IV. RESULTS
𝑅𝑒𝑐𝑎𝑙𝑙𝑖 × 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖
𝐹𝑀𝑒𝑎𝑠𝑢𝑟𝑒𝑖 = 2 × (3) A. Results from AISAS Classification on Messages with
𝑅𝑒𝑐𝑎𝑙𝑙𝑖 + 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛𝑖
Naïve Bayes Method

93
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

TABLE II: HOLISTIC VIEW OF AISAS CLASSIFICATION BY THAI LANGUAGE Precision value is the number of related documents which
EXPERTS
have been extracted divided by the number of extracted
Classification A I S Ac Sh document, the calculation is illustrated by equation (5).
Yes 1,148 2,510 1,563 705 1,372 As shown in Table IV, accuracy result of the model with
Precision value, the highest value of Attention (96.38%),
No 7,852 6,490 7,437 8,295 7,628
Interest (82.59%), Search (91.96) is on Page B, whereas
Total 9,000 9,000 9,000 9,000 9,000 Action (98.37%) is on Page C, and Share (95.25%) is on Page
A.
%y 12.76% 27.89% 17.37% 7.83% 15.24%
The lowest Precision value of Attention (95.97%), Interest
%n 87.24% 72.11% 82.63% 92.17% 84.76% (78.68) is on Page C, Search (91.16%), Action (91.16%) is on
Page A, and Share (91.74%) is on Page B.
The labels designated by Thai language experts following
TABLE IV: ACCURACY RESULT OF THE MODEL WITH PRECISION VALUE
a set of questionnaire as shown in Table II is brought to
process with RapidMiner program which calculate the Precision PAGE A,B,C PAGE A Page B PAGE C
probability of each word resulting in Yes or No values in Attention 96.20% 96.20% 96.38% 95.97%
each pattern in AISAS model on applied on each word from
Interest 81.32% 80.95% 82.59% 78.68%
the pages.
Search 91.86% 91.16% 91.96% 91.86%
B. Results from Quality Verification of the Model of
Message Classification Prediction Action 98.18% 97.94% 98.23% 98.37%
Data classification will draw results comparison during Share 93.43% 95.25% 91.74% 92.88%
simultaneous processing of the three pages. Then, the results
in each page will be compared taking into consideration 1)
Precision
accuracy value of the model in all classes, 2) precision value
of data with precision value, and 3) accuracy value of the 100.00%
model with recall value in separated class. 90.00%
The results of accuracy comparison of classification model
80.00% Page A,B,C
with Naï ve Bayes Classifications principle, and with Cross
validation technique as illustrated in Table III asserts that the 70.00% Page A
highest accuracy value of Attention (93.69%) Interest 60.00% Page B
(69.56%) Search (85.20%) Action (97.62%) is on Page B and 50.00% Page C
Share (86.58%) is on Page A.
On the other hand, the lowest accuracy value of Attention
(85.41%) is on Page B and on Page C, Interest (57.50%),
Search (73.47%), Share (77.77%) is on the three pages when Fig. 6. Comparison of accuracy result of the model with Precision value.
we process them together, and Action (93.33%) is on Page C.
Recall value the number of all related documents which
TABLE III: ACCURACY RESULT OF THE MODEL WITH ACCURACY VALUE have been extracted divided by all related documents as
Accuracy PAGE A,B,C PAGE A Page B PAGE C shown in equation (4).
Table V, the result of accuracy of the model with Recall
Attention value shows that the highest value of Attention (97.10),
85.47% 88.79% 93.69% 85.41%
Interest (79.67%), Search (91.86%), Action (91.86%), and
Interest
57.50% 59.66% 69.56% 61.84% Share (91.43%) is on Page B.
Search 73.47% 77.58% 85.20% 78.36% TABLE V: ACCURACY RESULT OF THE MODEL WITH RECALL VALUE
Action 93.83% 95.40% 97.62% 93.33% Recall PAGE A,B,C PAGE A Page B PAGE C
Share 77.77% 86.58% 84.57% 81.07% Attention 88.40% 92.00% 97.10% 88.53%
Interest 61.37% 65.38% 79.67% 71.10%
Accuracy Search 77.97% 83.57% 91.86% 83.96%
100.00% Action 95.49% 97.34% 99.37% 94.80%
90.00% Share 81.90% 90.45% 91.43% 86.07%
80.00% Page A,B,C

70.00% Page A The model developed to find the probability of word

60.00% Page B groups which affect online consumer behavior in diverse
aspects will be retested to find accuracy with Cross
50.00% Page C
Validation technique for the total number of ten times by
Naïve Bayes technique. The result carried out from the test
demonstrates that we observe a greater accuracy rate of
database classification of the model building when we use
Fig. 5. Comparison of accuracy result of the model with Accuracy value.
separated industry data than that of the overall industry data.

94
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

Moreover, the results of probability of words according to the system and overall satisfaction of the system. The results
AISAS behavior patterns, shown by positive or negative of the acceptance of technology of innovative prototype for
value in each pattern, can be used to evaluate the overall analyzing consumer behavior probability in marketing
probability of sentences which are required to be verified in communication illustrate that users are satisfied with
order to find out the probability. It is worth noted that each innovative prototype at very good rate (average at 4.23).
sentence could fall into more than one behavior pattern. For When we look in detail to the others aspects, the results assert
example, Attention and Interest while the highest probability that users are satisfied (1) with efficiency and usefulness of
rate is on Share. the system as described as very good (average 4.28). The
majority of the sub items in this aspect are at very good result
Recall of satisfaction, except credibility of analysis which obtains an
average score of 4.10 (Good). The highest score at very good
100.00% result falls on suitability of the number of steps (4.45),
90.00% followed by overall satisfaction of service use and Quick
Page A,B,C
response of the system (in each step) (4.32 both) and finally
80.00%
Suitability of menu, quick access to webpage (4.23). An
70.00% Page A
average result of ease of use aspect is situated at 4.20 which
60.00% Page B mean good. The results report that font size and font pattern,
50.00% Page C beautiful and easy to read and data display in clear category
receive an average result at 4.23 which means very good. In
the same time, Menu pattern, easy to read and use at an
average of 4.19 and Well-Design, modern, interesting system
Fig. 7. Comparison of accuracy result of the model with Recall value. at an average of 4.16 receive good result respectively. (3) The
aspect of the intention to use the system shows a good result
C. Result of the Design and Development of Application at an average score of 4.15. The results also show that users
Prototype intend to suggest the others to use (4.19) and to use the
The following steps are applied in system designing. system continually (4.10).
Step 1 – Selection the product; selection of business group TABLE VI: AVERAGE VALUE OF USER SATISFACTION TOWARDS
with whom you wish to communicate on social media. PROTOTYPE INNOVATION
Step 2 – Enter the message which you wish to make it Item Mean S.D. Result
public; key the data which you wish to communicate on
1. Efficiency and Usefulness of the
social media, then click “analyze” to process or click “delete” system
in case that you wish to delete all message and retype new Suitability of menu, quick access Very
messages. 1.1 4.23 0.62 good
to webpage
Step 3 – Select the channel of public communication; you Quick response of the system (in Very
1.2 4.32 0.60
can share information on social media channel which you each step) good
wish to communicate our message to public such as Very
1.3 Suitability of the number of steps 4.45 0.68
good
Facebook, Twitter or Instagram.
1.4 Credibility of analysis 4.10 0.75 Good
Very
1.5 Overall satisfaction of service use 4.32 0.65
Enter the message which you wish to good
Selection the product
make it public Overall satisfaction of efficiency and Very
4.28 0.44
usefulness of the system good

2. Ease of use

Well-Design, modern, interesting

2.1 4.16 0.64 Good
system
2.2 Menu pattern, easy to read and use 4.19 0.70 Good
Font size, font pattern, beautiful Very
2.3 4.23 0.76
and easy to read good
Very
2.4 Data display in clear category 4.23 0.62
good
Overall satisfaction of ease of use 4.20 0.47 Good
Very
Top 3 Results 3. Attitude of system use 4.23 0.62
Select the channel of public good
communication 4. Intention to use the system
4.1 User intends to use continually 4.10 0.65 Good
Fig. 8. Application prototype.
User intends to suggest the others
4.2 4.19 0.70 Good
to use
D. Results of Use and Acceptance of Application Overall satisfaction of intention to apply 4.15 0.46 Good
Prototype
The application prototype was tested by business or
organization which uses Facebook for their public relations. V. CONCLUSION AND DISCUSSION
Four aspects are taken into consideration; efficiency and
Facebook comprises of two main functions which are the
usefulness of the system, ease of use, intention of users to use

95
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

posting of message brands wish to communicate to brans in communication with customers more efficiently.
consumers, and opinions and comments of consumers 2) Development of a dictionary used in posts for
responding to brand communication. At present, the research advertisement, direct public relations. One word can vary
or social media analysis tools tends to carry out the studies in meaning and in function. Words may not be related to
and analysis of consumer opinion with data collected from advertisement, public relations. Informal words are more
comment. The results measure consumer sentiment towards likely to be used in posts and should be added to the
the product or service in form of positive, negative and dictionary such as “แซ่บ” (delish) “บูสเตอร์” (booster)
neutral opinion. This research therefore establish the
“อร๊าย” (WTF). This will result in increasing efficiency
objective to analyze message posted by brands
communicating to consumers so as to predict consumer of public relations.
behavior and to determine whether it follows online 3) Add more industries in classification so as to cluster
consumer behavior theory with Dentsu’s AISAS model. according to AISAS model. The system can suggest
The study demonstrates that it would be rather to analyze suitable words for business to communicate accordingly.
and process page by page for more efficiency. This is due to 4) Study of the results of AISAS and real result in business
the fact of the differences in business operation pattern and operation such as sale volume.
communication methods of each page. For example, Page A 5) Study others languages with diversified data.
puts focus on interaction with customers by organizing 6) Develop picture post data used in advertisement, online
activities that customers can join. On the other hand, Page B public relations on social media so as that this data can be
emphasizes on awareness rising to promote product to be then analyzed and classified according to AISAS model.
known and understood, whereas Page C stimulates customers 7) Develop a system to resume message on posts used in
to make a decision to buy the products. advertisement, public relations using Naï ve Bayes
This study corresponds to the concept of theory related to technique which cannot identify the relation of word like
online consumer behavior in buying decision following “ครีมหน้าขาว” (Whitening cream), the system can
Dentsu’s AISAS theory. After verifying posted messages, the indicate that “ครีม” (cream) should be followed by
research found that consumers seek for information from
various sources prior to buy products and services, and they “หน้า” (face) and “ขาว” (white). Those who are
evaluate and reflect their satisfaction after buying products interested can study others techniques which take this
and services. Online social media exercise greater influence issue into account.
so as that consumers can search for all concerned information, 8) Further research can collect the data from online pages
gather data from existing customers as much as possible about consumable goods. This actual research focuses
before making a decision to buy. They do not need to meet solely on non-consumable goods.
the others and ask for their opinion towards product and
services. In addition, they can leave their opinion, impression
and share information to others as well. This illustrates VIII. RESTRICTION
perfectly online consumer behavior in decision making Due to the fact that the Facebook data collection periods
process to buy. are situated between September 2009 and October 2017, the
accuracy therefore is valid in a certain period of time.
However, database is continually updated and modified and
VI. SUGGESTION OF APPLICATION Facebook restricted the rights to access some information
Users of the application should pay attention to spelling such as opinion posted after each post. In case those users set
before posting messages. This is because the database does right to access the opinion posts only for page owners,
not cover misspelling words and this will lead to the researchers cannot collect such data to test or analyze.
prediction. Even though the application can predict without
selection of industry, specific data will result in higher rate of REFERENCES
accuracy. In addition, the application cannot be used to [1] A. Kao and S. R. Poteet, Natural Language Processing and Text
analyze every other aspect including sentiment or interaction Mining, Springer Science & Business Media, 2007.
[2] R. Feldman and J. Sanger, Text Mining Handbook: Advanced
of users in the posts. In the future however, the application Approaches in Analyzing Unstructured Data, p. 13, Cambridge
can be developed and improved to be used to analyze University Press, 2006.
sentiment as all-in-one analysis to facilitate application users. [3] E. Alpaydin, Introduction to Machine Learning: MIT Press, 2014.
Users should also buy advertisement from Facebook. This [4] M. I. Jordan and T. M. Mitchell, “Machine learning: Trends,
perspectives, and prospects,” Science, vol. 349, no. 6245, pp. 255-260,
is due to the fact that Facebook restrict visibility of the posts 2015.
without advertisement of Facebook. This will lead to [5] L. Breiman, “Statistical modeling: The two cultures,” Statistical
invisibility of the products or services on Facebook. Science, vol. 16, no. 3, pp. 199-215, 2001.
[6] A. Harsha, “What is machine learning? Machine learning for
beginners,” Big Data Analytics, March, Edureka, 2018.
[7] W. L. Buntine, “Operations for learning with graphical models,”
VII. FUTURE RESEARCH Journal of Artificial Intelligence Research, vol. 2, pp. 159-225, 1994.
[8] M. A. Ghazanfar, S. A. Alahmari, Y. F. Aldhafiri et al., “Using
From the conclusion and restriction of this research, we machine learning classifiers to predict stock exchange index,”
can formulate a guideline for further research for those who International Journal of Machine Learning and Computing, vol. 7, no.
2, pp. 24-29, 2017.
are interested. [9] A. Stuart and K. Ord, Kendall's Advanced Theory of Statistics, Volume
1) Research on sentence, message or phrase analysis for 1: Distribution Theory: Wiley, 2009.

96
 International Journal of Machine Learning and Computing, Vol. 9, No. 1, February 2019

[10] M. N. Murty and V. S. Devi, Pattern Recognition: An Algorithmic Suphamongkol Akkaradamrongrat was born in
Approach, Springer Science & Business Media, 2011. Bangkok, Thailand in 1993. He received the bachelor’s
[11] H. Shi and Y. Liu, “Naï ve bayes vs. support vector machine: Resilience degree in computer science from Department of
to missing data,” Artificial Intelligence and Computational Intelligence, Computer Science, Thammasat University. He is
pp. 680-687. currently studying the master’s degree in computer
[12] K. Sugiyama and T. Andree, The Dentsu Way: 9 Lessons for science at the Department of Computer Engineering,
Innovation in Marketing from the World's Leading Advertising Agency, Chulalongkorn University. His research interests are
McGraw-Hill, 2011. machine learning, natural language processing, and
[13] F. D. Davis, “Perceived usefulness, perceived ease of use, and user social network analysis.
acceptance of information technology,” MIS Q., vol. 13, no. 3, pp.
319-340, 1989. Sukree Sinthupinyo received his bachelor degree,
[14] I. Ajzen, “The theory of planned behavior,” Organizational Behavior master degree and Ph.D. (computer engineering)
and Human Decision Processes, vol. 50, no. 2, pp. 179-211, 1991. from Chulalongkorn University, Thailand. His main
[15] S. Bjurstrom and M. Plachkinova, Sentiment Analysis Methodology for areas of research are artificial intelligence, machine
Social Web Intelligence, 2015. learning & pattern recognition, and engineering.
[16] P. Refaeilzadeh, L. Tang, and H. Liu, “Cross-validation,”
Encyclopedia of Database Systems, pp. 532-538, Springer, 2009.
[17] Zocialrank. (2018). Thailand Facebook Ranking.[Online]. Available:
http://zocialrank.com/facebook/
[18] W. S. Gosset, “The probable error of a mean,” Biometrika, vol. 6, no. 1,
pp. 1-25, 1908. Achara Chandrachai received a bachelor degree of
[19] J. W. Best, Research in Education, Englewood Cliffs, New Jersey: commerce and accountancy from Chulalongkorn
Prentice-Hall, Inc., 1977. University, master of business administration (2nd
honor roll) from National Institute of Development
Pornpimon Kachamas was born in Bangkok, Administration, Ph.D. (quantitative business
Thailand, 1985. She has a B.Sc. in business analysis) of Arizona State University. She has been a
administration from Assumption University. Her professor emeritus of technopreneurship &
M.Sc. is in business administration (international innovation management program since 2006 and
program) (1st honor roll) from Stamford International manager of IMET (Institute of Management
University. She is a Ph.D. candidate in School of Education Thailand) since 2008. Her research interests are technology
Technopreneurship and Innovation Management, commercialization, business, industry analysis and strategies, business
Graduate School at Chulalongkorn University, innovation.
Thailand. Her research interest is social network
marketing, social network analysis, AISAS model, and machine learning.

97

View publication stats