IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 11, No. 4, December 2022, pp. 1448~1459

ISSN: 2252-8938, DOI: 10.11591/ijai.v11.i4.pp1448-1459  1448

Detection of traffic congestion based on twitter using

convolutional neural network model

Rifqi Ramadhani Almassar, Abba Suganda Girsang

Department of Computer Science, Binus Graduate Program-Master of Computer Science, Bina Nusantara University, Jakarta, Indonesia

Article Info ABSTRACT

Article history: Microblogging is a form of communication between users to socialize by
describing the state of events in real-time. Twitter is a platform for
Received Nov 8, 2021 microblogging. Indonesia is one of the countries with the largest Twitter
Revised May 12, 2022 users, people can share information about traffic jams. This research aims to
Accepted Jun 10, 2022 detect traffic jams by extracting tweets in the form of vectors and then
inserting them into the Convolution neural network (CNN) model and
getting the best model from CNN+Word2Vec, CNN+FastText, and support
Keywords: vector machine (SVM). Data retrieval was conducted using the Rapidminer
application. Then, the context of the tweets was checked so that there were
Convolutional neural network 2777 data consisting of 1426 congestion road data and 1351 smooth road
Text classification data. The data was taken from certain coordinate points in around Jakarta,
Traffic congestion Indonesia. Then, preprocessing and changes to vector form were carried out
Twitter using the Word2Vec and FastText methods, then inserted into the CNN
Word embeddings model. The results of CNN+Word2Vec and CNN+FastText were compared
to the SVM method. The evaluation was done manually using the actual
traffic conditions. The highest result obtained using test data by the
CNN+FastText method are 86.33% while CNN+Word2Vec is 85.79% and
SVM is 67.62%.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Rifqi Ramadhani Almassar
Department of Computer Science, Binus Graduate Program-Master of Computer Science, Bina Nusantara
University
Jakarta, Indonesia, 11480
Email: [email protected]

1. INTRODUCTION
Information technology is growing rapidly including social media. Nowadays, it easy for people to
access social media in order to interact or seek entertainment through social media. Twitter is one of social
media for microblogging. Microblogging is a form of communication among internet users to socialize by
describing the state of events through the web [1]. Users can send and read messages on Twitter which are
known as a tweet, but twitter limits the characters in the message.
In October 2020, Twitter users from Indonesia reached 13.2 million, making Indonesia the 5th
largest after Turkey [2]. The data shows that Twitter is widely used by people in Indonesia. Twitter users can
share information and connect with other users in real-time. Users can also provide opinions about what is
happening in real-time, for example providing information and the state of the traffic they are passing.
Indonesia is one of the developing countries that have such problems as high population growth,
inadequate infrastructure, and traffic congestion. High population growth results in high levels of
urbanization and transportation needs, causing traffic jams. Congestion can cause socio-economic problems.
Tweet data can be used to analyze congestion conditions on a road. The main advantage of congestion
analysis using tweet data is that real-time data is obtained directly from users based on tweet time so that they

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  1449

can view traffic history according to the tweet date and if applied to a cellphone application, it will save
battery because it will not run continuously like Global positioning system (GPS).
In [3], image with foreground estimation and cascade classifier were used to detect congestion, but
this research used text from Twitter to detect traffic congestion. Based on research [4], doing preprocessing
by eliminating URLs, or special words before classification, while in [5] hashtag symbols were also removed,
and in [6] deleting unrelated words. Furthermore, to be included in a model according to research by [7],
feature representation can be carried out using the Bag-of-words (BOW) technique to be numerical, while
in [8] feature representation is carried out using the Term frequency-inverse document (TF-IDF) method so
that it becomes a matrix that can be implemented in the Long short-term memory (LSTM) method. In this
study, we use the Word2Vec method, which combines Skip-gram and Continuous bag-of-Words (CBOW)
and FastText. It is the development of Word2Vec which has the advantage of being able to process words
that are not in the vocabulary. Word2Vec and FastText can be used in conjunction with the Convolutional
neural network (CNN) method to do the classification.
Research on tweet-based traffic jam detection has been widely conducted using various methods
such as taking words from each tweet and then classifying the words with the Support vector machine (SVM)
model [9]. Alhumoud [10] obtained Twitter data, entered the data into AsterixDb and then classified it using
Spark. According to research [11], preprocessing stemming and tokenization were applied to the data. Dabiri
and Heaslip [12] carried out mapping tweets into vectors to measure the relationship between words and then
enter them into the CNN and Recurrent neural network (RNN) models to get traffic incidents. CNN is used to
classify by extracting features and word representation of each tweet. By using the Word2Vec model to
distribute words and then enter them into CNN, we can get better extraction results [13]. Currently, the best
results in tweet-based congestion detection research are obtained using the SVM method and the results reach
96.24% [9]. In [12] research predicting traffic incidents, it can reach 98.6% by using the CNN method.
Therefore, this study predicts traffic congestion using Twitter data by using the CNN method. So, this
research aims to congestion detection by classifying tweets and considering the frequency of tweets, by going
through the preprocessing process, extracting features through Word2Vec and FastText, and then entering
them into the CNN model through the Tensorflow framework with Python programming language and then
comparing the two with data testing.

2. RELATED WORKS
Several studies have focused on detecting traffic using tweet data with various approaches.
D’Andrea et al. do classified tweets or in this study called Status update message (SUM) to detect traffic
using several methods [14]. The authors collected Italian-language tweets and then conducted preprocessing
by tokenization and cleansing. The last preprocessing performed feature representation with TF-IDF in
numerical vector form. The study obtained SVM with the highest accuracy of 95.75%.
In research conducted by Yang et al. [1], performs traffic detection with deleted tweets if they are
not related to traffic and performs symbol encoding. Based on the results of the analysis on each tweet, the
authors argue that there are elements of statements that determine the traffic situation, and there are hashtags
that do not always have meaning. So, it is possible to conduct traffic analysis prediction research based on
tweets.
Gu et al. [15] in detecting traffic after collecting Twitter data then removed symbols and synonyms
were extracted from the WordNet database and the authors performed tokenization. Tweet classification was
done using the supervised Latent dirichlet allocation (sLDA) model and the Semi-naïve bayes method. The
study obtained accuracy of 90.5%.
Zulfikar and Suharjito [9] used Twitter data labelled using the X-Means Clustering algorithm. The
authors used the TF-IDF method and performed cross-validation. Furthermore, the authors used the k-NN
and Naïve bayes methods, and then compared the results with the SVM method. The results of the
performance evaluation using the confusion matrix sigmoid SVM method were 96.24%.
Dabiri and Heaslip built a traffic incident detection model based on Twitter using deep learning
architecture [12]. Then the classification was carried out in 3 ways, namely only CNN, only LSTM, and a
combination of CNN and LSTM. The results of this study indicate that using only CNN and converting
vectors into 2 classes with Word2Vec or FastText models and the highest accuracy, which is 98.6%.
Alhumoud explained that an Intelligent transportation system (ITS) is a solution for monitoring traffic
and travel efficiently [10]. The data stored using AsterixDB is as much as possible. Then, the data is filtered,
and data is obtained which is the result of research. Using the spark data processing engine connected to
AsterixDb provides 81%, 89%, and 92% precision for accidents, weather conditions, and road incidents.
Wongcharoen and Senivongse [16] predicted traffic jams in Thailand using data from several
Twitter accounts. The author labelled the data as high (H), medium (M), and low (L). Next, they trained the
Decision tree (DT) using C4.5 with the Weka application. The results of the study with evaluation using 10-
Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)
1450  ISSN: 2252-8938

fold cross-validation obtained precision 0.89, recall 0.898, F-measure 0.892, and receiver operating
characteristic (ROC) Area 0.894.
Gong et al. [17] explained that tweets often include geospatial data such as the user's profile,
latitude/longitude, and time of the tweet. Next, clustering is done using the Density-based spatial clustering
of applications with noise (DBSCAN) algorithm and the authors also does clustering using the ELKI Data
mining framework to get two distances between tweets. The clustering results are then stored in Apache
CouchDB and visualized on a map using the website.
Septianto et al. [18] collected tweet data from Twitter then did preprocessing. Then, Naïve bayes
was done to determine the label. Then, the name of the place was taken by dividing between the words
"direction" and "towards" so that traffic is taken from the 2 place names. The results are then visualized on
Google Maps and repeated every 5 minutes. The results of this study obtained an accuracy of 61.66%.
Subhan et al. [4] collected tweet data from Twitter application programming interface (API) then
change the words in the tweet into standard words via kateglo.com. Next, perform a semantic analysis using
the Generative lexicon method, the results obtained are the location information from and the destination was
separated by the words "menuju", "ke", or "sampai". The results of this study were visualized with Google
Maps to validate the results the authors compared with the actual conditions.
Ankit and Saleena conducted a sentiment analysis of tweets whether they contained positive or
negative sentiments [7]. The authors performed feature representation using the BOW technique, which
represents tweets as numeric. Classification uses several techniques, namely Naïve bayes, random forest
(RF), SVM, Logistic regression (LR). Then, the ensemble classifier is carried out by combining the
classification techniques that have been carried out to improve accuracy, using several data sets, the highest
accuracy results are obtained using the ensemble classifier as much as 85.83%.
Ansari et al. [8] conducted a sentiment analysis of the 2019 presidential election in India. The
authors performed feature extraction based on Term-Frequency to convert data into vectors. Then, it
classified machine learning with the SVM algorithm, DT, LR, RF, and LSTM. The highest result was
obtained by LSTM with 74% accuracy.
Khan et al. introduced Twitter opinion mining (TOM) framework for classified positive, negative,
and neutral sentiments [19]. The Authors performed preprocessing by detecting slang words that will be
interpreted based on the WordNet, Netlingo, and SMS dictionary libraries. The classification algorithm in the
framework used 3 basics, namely Enhanced emoticon classifier (EEC), Improved Polarity Classifier (IPC),
and SentiWordNet classifier (SWNC). Based on the results of this study using the TOM framework, an
average accuracy of 85.7% was obtained with 85.3% precision and 82.2% recall.
Ruz et al. [6] conducted sentiment analysis during natural disasters or political transitions. Using the
Twitter dataset, the first dataset was on the Chile earthquake and the second dataset was taken during the
Catalan independence referendum. The authors performed feature representation using the BOW technique,
to overcome unbalanced sentiment analysis, then apply the Synthetic minority over-sampling technique
(SMOTE). Based on this study, the highest results were found in the first dataset is SVM with an accuracy of
81.2%, and the second dataset is RF with an accuracy of 85.8%.
Lal et al. performed tweet data analysis of tweets data to identify criminal events that require police
attention [5]. Before being analyzed the data was preprocessed by dividing the word into several segments in
1 sentence, then the hashtag word was deleted. The authors perform the TF-IDF conversion and by using the
Waikato environment for knowledge analysis (WEKA) application. The results of the classification using the
RF algorithm get the highest accuracy, which is 98.1%.
Yao and Qian [20] aims to check traffic before 5 a.m. using the previous day's traffic predictions
taken from Twitter. In that study, the tweet data was augmented to reduce noise data based on the tweet clock
and geocode. After that, the data is extracted to get the traffic the next day. By using the tweet2traffic
clustering model for congestion, the accuracy is 79%.
Ahmed et al. explained congestion and traffic information can be identified using the Advanced
traveler information system (ATIS), but these tools are expensive and cannot be installed throughout
the city [11]. The authors performed a Geo-filter to get the location of the tweet in question. The data is then
converted into vectors using TF-IDF and clustering using K-Means, then classified using the SVM model to
get the highest precision of 82%.
Alomari et al. [21] proposed congestion detection and incident detection using Arabic-language
tweets. By using systems, applications, and products in data processing high-performance analytic appliance
(SAP HANA), which is connected to the Tweet API, then the tweet is processed first with the help of SAP
HANA through the tokenization, normalization, and entity extraction stages. Furthermore, sentiment analysis
was carried out using SAP HANA and assisted by SAP Lumira to produce visualizations.
Zhang et al. detect traffic incidents using the deep belief network (DBN) and LSTM methods [22].
So, the authors did preprocess by labeling and extracting each word manually through tokenization and

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1448-1459

Int J Artif Intell ISSN: 2252-8938  1451

stemming. Next, tokenize using DBN and LSTM and then compare it with artificial neural network (ANN)
and sLDA. The final result shows that DBN has the best accuracy of 93%.
Essien et al. [23] make traffic predictions using Twitter data, traffic, and weather. For Twitter data is
taken using the Twitter API based on the @OfficialTfGM and @WazeTrafficMAN accounts, weather data is
taken from the Center for Atmospheric Studies (CAS) and traffic data is taken using detectors. The 3 datasets
are combined into one and then entered into the Bidirectional long short-term memory (BiLSTM) model using
k-fold cross-validation. The results of the evaluation using the (Root mean squared error) RMSE get 6.85%.
Ali et al. [24] detected incidents and traffic conditions using ontologies latent Dirichlet allocation
(OLDA), and BiLSTM. Data is retrieved from Twitter and Facebook using the API. Next, do a positive,
neutral, or negative sentiment analysis to find out the traffic conditions. Then, the data is embedded using
FastText and Word2Vec and finally, training is carried out using BiLSTM with softmax regression to classify
conditions and predict polarity. The results in this study achieved an accuracy of 97%.
Alomari et al. [25] in his research, he created a tool to detect traffic-related events using Twitter data
called Iktishaf. It starts by retrieving data via the Twitter API and putting it into MongoDB. The authors
conducted TF-IDF to determine the important words in the dataset. Using spark machine learning (ML),
tweets classified relevant tweets using Logistics regression, SVM, and Naïve Bayes and delete irrelevant
tweets. The results showed that the highest accuracy obtained using SVM reached 90%.
Utari et al. [26] research on natural language processing used data from Twitter accounts that inform
road and traffic conditions. Using data from Twitter, each tweet was separated by token and each word in the
token is analyzed to generate a syntactic structure. The first approach uses anatomy as a determinant of word
classification, then classification approaches from other words that are close to unknown words and finally
used a notional or contextual approach. The classification results are then entered into a web page. The
accuracy in this study reached 88.57%.
Salazar-Carillo et al. [27] in detected congestion, taking data from Twitter and then standardizing
the words that have been separated based on the n-grams, then identifying traffic flows using the Support
vector regression (SVR) library from the results of the accuracy reaching 95.5%. Next, perform the
geocoding process to get the visualization predicted by the model. From these results, in addition to
congestion, information is also obtained related to events that occur on the road.

3. PROPOSED METHOD
Traffic congestion in Indonesia is a problem that often occurs. Many Indonesian people are stuck in
traffic jams. So, road users need to know the traffic conditions to avoid congestion. With the development of
technology, Twitter is one of the mostly used social media for Indonesian people just looking for entertainment
or interacting and one of microblogging platform to socialize by describing the state of events in real-time [1].
During traffic jams, many people provide information or comments about the roads they pass using Twitter.
Indonesia is included in the list of countries with the biggest users of Twitter after Turkey [2]. With
the number of Twitter users, the tweet data at the time of congestion can be processed to show the state of the
road at the time the tweet was made. Research on congestion detection using data from Twitter has been done
before. Twitter has an API that can be used by application developers to retrieve tweet information, user
profiles, messages, and other information through an endpoint [12].
In conducting research, [9] took data from the Twitter API and then processed it by taking weights
from the text using TF-IDF and then classified it using the SVM [12] research predicted traffic incidents
using the CNN to get high accuracy. In this study, we compared the accuracy results between
CNN+Word2Vec, CNN+FastText, and SVM. This study processed tweet data on Twitter related to traffic so
that it can be used as information on traffic jams on a road using the CNN method.
Based on Figure 1, the first step is to collect tweet data with the Indonesian language keywords
"macet" which means congestion road, and "lancar" which means smooth roads from Twitter using the
Rapidminer application. The results of the tweets data are then labeled manually according to the traffic jam on
Google Maps. Then, the data that is out of context was removed by checking one by one using Microsoft Excel.
Then, preprocessing was done because according to [28] a good preprocessing is needed to get
effective and efficient data. In this research, preprocessing was done by encoding labels so that the label
“macet” becomes 0 and "lancar" becomes 1. After that, duplicate data was removed so that each tweet data
is not the same. Next, cleansing was performed by removing symbols or characters that have no meaning and
are not needed in the classification. Then, transform case was performed, which is changing all letters to
lowercase letters to facilitate the classification process. After doing the transforming case, then the word was
changed to synonyms in Indonesian “macet” to “macet”/”padat”/”mandek”/”tertahan” randomly and the
word “lancar” to “lancar”/”mulus”/”lenggang” randomly to reduce discriminatory features. After that,
tokenization was done to separate the words according to the order in the tweet. Each token was done with a
stopword in Indonesian to remove words that have no meaning so that the word will not be considered in the
Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)
1452  ISSN: 2252-8938

classification. The word on each token is also done stemming so that it changes the word into the core word.
Then because CNN is a classification method that requires input in the form of a vector from an image but in
this research, it is text data, it needs to be converted into a vector form. In this process, we will represent
tweets into vector form using Word2Vec because according to research [29] accuracy with pre-trained
vectors gets higher accuracy so words that have the same meaning will get almost the same weight.
Word2Vec is an algorithm model that represents words into a vector using a neural network framework [30].

Figure 1. Research stages

This research also used FastText to convert words into vectors for comparison. FastText has an n-
gram parameter, which is a text breakdown of n words so that words that are not in the corpus can be used as
vectors. In this study, several variations of parameter values were tested and the results are not too different.
The Word2Vec and FastText parameters used in this study can be seen in Table 1. Furthermore, in the
classification stage, the CNN architecture model used will combine the results from the Pooling Layer with
kernel sizes 2, 3, and 4 so that it looks like in Figure 2.
In [8], the classification of tweets is divided into 2, namely “Macet” and “Bukan-Macet” so that in this
study the classification is also carried out in 2 classes “macet” class containing tweets about traffic jams based
on point coordinates and “lancar” class contains tweets about smooth traffic based on point coordinates. CNN is
made with several layers so that it gets a classification from tweets. The description of each parameter on CNN
can be seen in Table 2 and Table 3 is a list of layers and their CNN settings in this study.
In text classification using CNN, the results of 1-dimensional vectors that have been carried out
using Word2Vec and FastText are then embedded so that they are divided into several channels and stored in
the first layer in the vector, then entered into a convolutional layer with kernels 2, 3, and 4. Each
convolutional layer result is continued into the pooling layer. The pooling layer results from each kernel are
then combined and entered into a fully connected layer consisting of dense, dropout, and dense again, then
the activation function is carried out using sigmoid.

Table 1. Word2Vec and FastText parameter

Parameter Description Value
Word2Vec FastText
Size The sum of the dimensions of the vector 100 100
Negative Number of noise words 5 5
Window Distance measure from the data context 5 5
Min_count Minimum number of words available 1 1
Workers The number of threads used to train the model 8 8
Alpha Learning rate value 0.065 0.065
Min_alpha Learning rate drop 0.065 0.065
Word_ngrams The maximum length of the word ngram None 3

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1448-1459

Int J Artif Intell ISSN: 2252-8938  1453

Figure 2. CNN model architecture

Table 2. CNN parameter description

Parameter Description
input_dim Number of vocabulary
ouput_dim Embedding dimensions
input_length Input sequence length
filters Output dimension of convolution layer
padding Vectors added on each side of the matrix
activation Activation function
kernel_size Convolution window length in each step
strides Convolution step length
units Output dimension
rate Part of the input unit to be omitted

Table 3. CNN layer list

Layer Output Parameter Setting
Input Layer (None, 43) 0
Embedding (None, 43, 200) 20000000 input_dim = 100000; ouput_dim = 200;
input_length = 43
Convolution 1 (None, 42, 100) 40100 filters = 100; kernel_size = 2; padding =
“valid”; activation = “relu”; strides = 1
Convolution 2 (None, 41, 100) 60100 filters = 100; kernel_size = 3; padding =
“valid”; activation = “relu”; strides = 1
Convolution 3 (None, 40, 100) 80100 filters = 100; kernel_size = 4
padding = “valid”; activation = “relu”; strides =
1
Average Pooling 1D 1 (None, 100) 0
Average Pooling 1D 2 (None, 100) 0
Average Pooling 1D 3 (None, 100) 0
Concatenate (None, 300) 0
Dense (None, 256) 77056 units = 128; activation = “relu”
Dropout (None, 256) 0 rate = 0.1
Dense (None, 1) 257 units = 1
Activation (None, 1) 0 activation = “sigmoid”

Based on Keras library documentation on Python programming, if the result is less than 5, then it
returns a sigmoid value close to zero and more than 5 then the sigmoid result is close to 1. If the sigmoid result
is close to zero then the tweet is in the “macet” which means congestion category, while if the result is close to 1
then the tweet is in the “lancar” which means smooth category. The results of each tweet will be calculated by
the number of congestion and smooth, the most results are the categories of traffic at these coordinates.
The dataset is divided into three, namely 60% training, 20% test, and 20% validation. Training is
used to train the model so that it can increase its accuracy. For each epoch iteration, validation was carried
Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)
1454  ISSN: 2252-8938

out to determine the increase in the accuracy of the model, and the test is used to perform testing when the
model has completed training.
To calculate the performance of this study, the confusion matrix is used and the result from
CNN+Word2Vec, CNN+FastText, and SVM are compared. The confusion matrix is in the form of a
comparison table which contains the number of test data and the number of prediction results using the model
that has been made true or not. In the confusion matrix, there are several values, namely True positive (TP)
which is the number of data jams that are predicted to be correct, True negative (TN) which is the amount of
current data that is predicted to be correct, False positive (FP) which is the amount of current data that is
predicted to be false and False negative (FN) which is the number of data crashes that are predicted to be
incorrect. This evaluation was divided into 2, namely evaluation of tweet data with testing data and
evaluation of actual congestion manually. In the evaluation of tweet data, each data will be predicted using
data testing and then entered into the confusion matrix. Meanwhile, the evaluation of actual congestion will
be done manually by determining the coordinates and hours to be used for 30 different locations or hours,
then taking tweets with the RapidMiner application according to the coordinates and hours. For each data in 1
data set, congestion classification is carried out and the classification results in 1 set will be taken from the
most classifications. Then the results in the 1 set classification will be compared with traffic conditions on
Google Maps. The results of the actual congestion evaluation manually will also be entered into the
confusion matrix.

4. RESULTS AND DISCUSSION

4.1. Data collection
The data used in this study is tweet data taken using the Rapidminer application with the keywords
“macet” and “lancar” with several coordinate points around Jakarta, Indonesia. The data that was
successfully retrieved were 4087 tweets during the months of February 2021-June 2021. Then the data is
checked manually by comparing the labels on Google Maps according to the hours and coordinates, then
looking at the actual situation according to the keywords “macet” or “lancar”. In the dataset, it can contain
congestion tweet data category but in reality, it is smooth and vice versa. The results were manually checked
again to find out the context of the tweets related to traffic or not so that checking the dataset became 2777
tweets consisting of 1426 data congestion or labeled “macet” and 1351 data smooth or labeled “lancar”. An
example of a dataset can be seen in Table 4.

Table 4. Dataset examples

Created-At From-User From-User- Text Retweet- Label
Id Count
08/02/2021 Maria_ 60450573 "RT @SonoraFM92: 16.40 #LalinSONORA-Jembatan Dua 2 Macet
16:50 arah Pluit.... Macet
https://t.co/nSNtTslLh8 https://t.co/0sCgHy09mD"
08/02/2021 Ahmad Haris 95354503 jalanan mulai underpass soleh iskandar bogor, 2 minggu 0 Macet
16:34 ini macet mulu yang arah ke cibinong. problemnya cuma
jalan berlubang pas jembatan.
10/03/2021 Radio Sonora 180696019 "07.00 Situasi arus lalu lintas terkini di Jl. RS Fatmawati 1 Lancar
07:33 Jakarta menuju Cipete Raya/Blok M, dan sebaliknya menuju TB
Simatupang/Lebak Bulus terpantau lancar.
https://t.co/jc6K9D5FMF
@TMCPoldaMetro"
10/03/2021 PT. 540625251 07.37 WIB #Tol_JLJ Ulujami-Pd Ranji-Serpong 0 Lancar
07:37 JASAMARGA LANCAR.; Serpong-Pd Ranji-Ulujami LANCAR.

4.2. Preprocessing
Furthermore, to make it easier to classify the label “macet” is changed to 0 and “lancar” is changed
to 1. The next preprocessing in this process includes deleted unused columns, cleansing, transform case,
tokenization, stopword and stemming. This stage is carried out with the literature and NLTK libraries.
Delete unused columns. The deleted columns are Created-At, From-user, From-user-id, To-user, To-
user-id, language, source, Geo-location-latitude, Geo-location-longitude, Retweet-count so that leaves a column
Text and Label. Cleansing. Cleansing is the process of removing symbols or characters that have no meaning
and are not needed in the classification. Example: RT SonoraFM92 16.40 LalinSONORA Jembatan Dua arah
Pluit Macet. Transform Case. Transform Case is the process of changing all letters to lowercase to facilitate
the classification process. Example: rt sonorafm92 16.40 lalinsonora jembatan dua arah pluit macet. Changing

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1448-1459

Int J Artif Intell ISSN: 2252-8938  1455

the words “macet” and “lancar”. Changed into synonyms in Indonesian, the word “macet” becomes
“macet”/”padat”/”mandek”/”tertahan” at random, and the word “lancar” becomes
“lancar”/”mulus”/”lenggang” randomly to reduce discriminatory features. Example: rt sonorafm92 16.40
lalinsonora jembatan dua arah pluit padat. Tokenization. Tokenization is the process of separating words
according to the order in the tweet. Example: [rt, sonorafm92, 1640, lalinsonora, jembatan, dua, arah, pluit,
padat]. Stopword dan Stemming. Stopword is the process of removing words that have no meaning so that the
word will not be considered in the classification and stemming is the process of changing words into their basic
word form. In this study, the stopword and stemming processes become 1 process and the list of stopwords used
is taken from Tala Stopword. Example: rt sonorafm92 1640 lalinsonora jembatan arah pluit padat.

4.3. Convert to vector Word2Vec

In this process, the data has gone through the preprocessing process so that the data can then be used
for word embedding processes. The word embedding process in this study uses the Gensim library. In this
process, input data that has been preprocessed in the form of tokenized text is used and will produce a model.
Each tweet in the dataset is taken with a different word using the TaggedDocument library. Then each word
is entered into Word2Vec.
Word2Vec uses a neural network to get the initial weight values, consisting of 2 architectures are
skip-gram and CBOW. In this study, Word2Vec was trained with the CBOW model with 50 epochs and then
trained with the skip-gram model with 50 epochs. The results of the model with the skip-gram architecture
will be combined with the CBOW model. Table 5 is an example of the final Word2Vec result.

Table 5. Word2Vec final result example

Word Vector
jalan -4.039771, 1.4529599, -0.72466475, 0.14262459, -6.820671, -3.7326174, 0.60550255,
-5.2219214, 1.0444719, -3.082483, … ,
1.0215762, 0.11307827, -0.6540826, 0.07509925, 1.8175708, 0.87743133, -0.5211345, 1.354927, -0.08151489,
0.43929535, …
padat -2.28622794e+00, 7.76517034e-01, -1.54419556e-01, 8.57797921e-01, -3.03995442e+00,
-4.07718992e+00, 3.46807949e-02, …,
-7.54858077e-01, 7.09322989e-01, 1.37110937e+00, 5.55417128e-02, -1.02720189e+00,
-2.23155960e-01, 5.28727114e-01, …
cakung 1.4838842, 0.13154435, -0.10948927, -1.2377591, -0.04943107, -0.55506617, 0.5633306, -0.5293238, 0.52347064,
rorotan 0.5164963, …,
lenggang -0.5595496, 0.45256287, -0.8794082, 0.5641887, 0.1843323, -0.3603339, 0.2183291, 0.17377803, 0.01006237, -
0.42547587, 0.15067908, …

4.4. Convert to vector FastText

In addition to using Word2Vec vector formation also uses FastText which is a development of
Word2Vec by dividing words into n-grams. This process uses the Gensim library, by entering data in the
form of text that has passed tokenization and then goes through training and produces a model. By using
TaggedDocument the words entered in the training will be different as in the vector formation stage using
Word2Vec. FastText also uses a neural network consisting of 2 architectures are skip-gram and CBOW.
Next, training the FastText with the CBOW model with 50 epochs and also training with the skip-gram
model with 50 epochs. The results of the model with the skip-gram architecture are combined with the
CBOW model as shown in Table 6.

Table 6. FastText final result example

Word Vector
jalan -2.50172567e+00, -1.02271485e+00, 3.24512750e-01, 8.21203768e-01, 2.77345330e-01, -2.61418581e-01, -
6.10480928e+00, …,
-1.22443938e+00, -6.80972517e-01, -4.08609390e-01, 1.36039710e+00, -1.08312324e-01, -9.95488167e-01,
1.90991676e+00, …
padat -1.4134253, 0.4434754, -1.1915249, -0.5949095, 0.02905422, 0.16940174, -1.3152181, 0.9656274, -0.10436693,
0.71517915, …,
-0.59788555, -0.48828775, -1.0627398, 0.7557978, 0.19303735, 0.15080197, 1.4271564, -0.8112281, -0.0476714, -
0.12192555, …
cakung -0.17935084, -2.7250762, 1.4563868, -0.11476076, 0.03896635, -0.42993864, 1.867808,
rorotan -0.4077298, 0.98534757, -0.93986416, …,
lenggang 0.3597058, 0.32056263, -0.9387656, 0.47813663, -1.7315956, -1.467028, 1.4627337, 1.0339308, 0.01472377,
0.5048158, 1.7824717, …

Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)

1456  ISSN: 2252-8938

4.5. Classification results

The data that has been converted to vector is divided into 60% training data, 20% validation data, and
20% test data. Then the data is classified using the Keras library using the CNN method. Before doing the
training, the researcher performed CNN hyperparameters to get the best parameters that were tested with
training data. The CNN hyperparameters are carried out using Grid Search Cross-Validation, which is to tests
the combinations one by one and validate for each combination. The parameters tested can be seen in Table 7.

Table 7. CNN Hyperparameter parameters used

Parameter Tested Value
learn_rate 0.1, 0.2, and 0.3
dropout_rate 0.1, 0.2, and 0.3
neurons 16, 32, 64, and 128
activation “relu”
strides 1, 2, and 3

Table 8. Model hyperparameter results

Parameter Model
No Learning Rate Neurons Dropout Rate Strides CNN+Word2Vec CNN+FastText
1 0.1 16 0.1 1 0.8607 0.8673
2 0.2 32 0.1 2 0.8601 0.8697
3 0.3 128 0.1 1 0.8739 0.8373
4 0.1 16 0.2 2 0.8601 0.8643
5 0.2 64 0.2 1 0.8109 0.8775
6 0.3 64 0.2 1 0.8817 0.8493
7 0.3 128 0.2 1 0.8787 0.8439
8 0.1 16 0.3 2 0.8535 0.8577
9 0.3 128 0.3 2 0.8577 0.7832
10 0.3 128 0.3 3 0.8631 0.8571

Based on Table 8 hyperparameters on the CNN+Word2Vec model, the accuracy with training data
is 0.8817 or 88.17% with a learning rate of 0.3 neurons 64, dropout rate 0.2 and strides 1 while
CNN+FastText is 0.8775 or 87.75% with a learning rate parameter of 0.2, neurons 64, dropout rate 0.2 and
strides 1. By using the parameters that have been obtained from the hyperparameter, then training is carried
out for 50 epochs using training data, and then for each epoch is evaluated using validation data, the results
of training and validation with the CNN+Word2Vec method are 89.08% and 87.38% while in the
CNN+FastText method are 87.88% and 87.38%. The validation results are similar from the two models
because the data is less and less varied.

4.6. Evaluation results

In this study, the evaluation phase is carried out by comparing the results of the confusion matrix
between the CNN+Word2Vec, CNN+FastText, and SVM methods using test data. The SVM model used is
taken from the Sklearn library, then the results are entered into a confusion matrix to get accuracy, precision,
recall, and f1-score. Table 9 is confusion matrix evaluation using test data by CNN+Word2Vec,
CNN+FastText, SVM.
Based on Table 10 or Figure 3, the highest accuracy results were obtained by the CNN+FastText
method of 0.8633 or 86.33%, while the lowest value was obtained by SVM of 0.6762 or 67.62%. Then, it
was evaluated by comparing the predicted results with the actual congestion manually. The test was carried
out 30 times consisting of 15 times the data was stuck and 15 times the data was smooth from a different
place or time. Table 11 is the result of the actual evaluation manually.

Table 9. Confusion matrix evaluation using test data

Actual Value
CNN+Word2Vec CNN+FastText SVM
Congestion Smooth Congestion Smooth Congestion Smooth
Predicted Value Congestion 232 59 234 57 180 111
Smooth 20 245 19 246 69 196

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1448-1459

Int J Artif Intell ISSN: 2252-8938  1457

Table 10. Calculation based on confusion matrix with data testing

CNN+Word2Vec CNN+FastText SVM
Accuracy 0.8579 0.8633 0.6762
Precision 0.8059 0.8118 0.6384
Recall 0.9245 0.9283 0.7396
F1-score 0.8611 0.9661 0.6853

100

80

60

40

20

0
CNN+Word2Vec CNN+FastText SVM

Accuracy Precision Recall F1-Score

Figure 3. Performance results with test data

Table 11. Result confusion matrix manual actual evaluation

TP TN FP FN Accuracy Precision Recall F1-Score
CNN+Word2Vec 11 10 5 4 0.7000 0.6875 0.7333 0.7097
CNN+FastText 11 10 5 4 0.7000 0.6875 0.7333 0.7097
SVM 8 11 4 7 0.6333 0.6667 0.5333 0.5926

Based on Table 11 or Figure 4 the highest accuracy value was obtained by CNN+Word2Vec and
CNN+FastText methods which reached 0.7 or 70%, while the lowest accuracy was obtained by SVM of
0.6333 or 63.33%. Based on this research, traffic detection using the CNN method can be applied as
supporting data or congestion detection on roads that are not supported by navigation applications, view
traffic jam history by date, and can also be implemented in applications for devices that do not have GPS.

100

80

60

40

20

0
CNN+Word2Vec CNN+FastText SVM

Accuracy Precision Recall F1-Score

Figure 4. Evaluation results with manual testing

5. CONCLUSION
Based on the results of this study, it can be concluded that making a model to detect congestion
using tweets takes several stages such as data collection, preprocessing, and data classification. Before

Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)

1458  ISSN: 2252-8938

training can be added hyperparameters to get the best parameters for the model. Then, in this research, the
CNN+FastText model at the time of evaluation using data testing has the highest level of accuracy compared
to CNN+Word2Vec and the SVM method gets the lowest accuracy, while the actual manual evaluation of
CNN+FastText and CNN+Word2Vec gets the highest accuracy. For further research, it is recommended to
perform a combination of classification methods to increase the accuracy of congestion detection using tweet
data. This study only focuses on the classification of tweet data so that further researchers can implement it in
the form of an application. Users can directly use the congestion detection application. Then, the dataset can
be multiplied to get variations of tweets so that the results are more optimal and in collecting data it is better
to involve more than one person to avoid human errors.

REFERENCES
[1] L. C. Yang, B. Selvaretnam, P. K. Hoong, I. K. T. Tan, E. K. Howg, and L. H. Kar, “Exploration of road traffic Tweets for
congestion monitoring,” Journal of Telecommunication, Electronic and Computer Engineering, vol. 8, no. 2, pp. 141–145, 2016.
[2] Statista, “Twitter: most users by country,” Statista. 2021, [Online]. Available: https://www.statista.com/statistics/242606/number-
of-active-twitter-users-in-selected-countries/.
[3] U. Masud, F. Jeribi, M. Alhameed, A. Tahir, Q. Javaid, and F. Akram, “Traffic congestion avoidance system using foreground
estimation and cascade classifier,” IEEE Access, vol. 8, pp. 178859–178869, 2020, doi: 10.1109/ACCESS.2020.3027715.
[4] Subhan Subhan, E. Sediyono, and Farikhin Farikhin, “The semantic analysis of Twitter data with generative lexicon for the
information of traffic congestion,” Journal of Advances in Information Systems and Technology, vol. 1, no. 1, pp. 45–54, 2019.
[5] S. Lal, L. Tiwari, R. Ranjan, A. Verma, N. Sardana, and R. Mourya, “Analysis and classification of crime Tweets,” Procedia
Computer Science, vol. 167, pp. 1911–1919, 2020, doi: 10.1016/j.procs.2020.03.211.
[6] G. A. Ruz, P. A. Henríquez, and A. Mascareño, “Sentiment analysis of Twitter data during critical events through Bayesian
networks classifiers,” Future Generation Computer Systems, vol. 106, pp. 92–104, May 2020, doi: 10.1016/j.future.2020.01.005.
[7] Ankit and N. Saleena, “An ensemble classification system for Twitter sentiment analysis,” Procedia Computer Science, vol. 132,
pp. 937–946, 2018, doi: 10.1016/j.procs.2018.05.109.
[8] M. Z. Ansari, M. B. Aziz, M. O. Siddiqui, H. Mehra, and K. P. Singh, “Analysis of political sentiment orientations on Twitter,”
Procedia Computer Science, vol. 167, pp. 1821–1828, 2020, doi: 10.1016/j.procs.2020.03.201.
[9] M. T. Zulfikar and Suharjito, “Detection traffic congestion based on twitter data using machine learning,” Procedia Computer
Science, vol. 157, pp. 118–124, 2019, doi: 10.1016/j.procs.2019.08.148.
[10] S. Alhumoud, “Twitter analysis for intelligent transportation,” Computer Journal, vol. 62, no. 11, pp. 1547–1556, Nov. 2019, doi:
10.1093/comjnl/bxy129.
[11] M. F. Ahmed, L. Vanajakshi, and R. Suriyanarayanan, “Real-time traffic congestion information from tweets using supervised
and unsupervised machine learning techniques,” Transportation in Developing Economies, vol. 5, no. 2, p. 20, Oct. 2019, doi:
10.1007/s40890-019-0088-2.
[12] S. Dabiri and K. Heaslip, “Developing a Twitter-based traffic event detection model using deep learning architectures,” Expert
Systems with Applications, vol. 118, pp. 425–439, Mar. 2019, doi: 10.1016/j.eswa.2018.10.017.
[13] A. K. Sharma, S. Chaurasia, and D. K. Srivastava, “Sentimental short sentences classification by using CNN deep learning model
with fine tuned Word2Vec,” Procedia Computer Science, vol. 167, pp. 1139–1147, 2020, doi: 10.1016/j.procs.2020.03.416.
[14] E. D’Andrea, P. Ducange, B. Lazzerini, and F. Marcelloni, “Real-time detection of traffic from twitter stream analysis,” IEEE
Transactions on Intelligent Transportation Systems, vol. 16, no. 4, pp. 2269–2283, Aug. 2015, doi: 10.1109/TITS.2015.2404431.
[15] Y. Gu, Z. Qian, and F. Chen, “From Twitter to detector: real-time traffic incident detection using social media data,”
Transportation Research Part C: Emerging Technologies, vol. 67, pp. 321–342, Jun. 2016, doi: 10.1016/j.trc.2016.02.011.
[16] S. Wongcharoen and T. Senivongse, “Twitter analysis of road traffic congestion severity estimation,” in 2016 13th International
Joint Conference on Computer Science and Software Engineering, JCSSE 2016, Jul. 2016, pp. 1–6, doi:
10.1109/JCSSE.2016.7748850.
[17] Y. Gong, F. Deng, and R. O. Sinnott, “Identification of (near) real-time traffic congestion in the cities of Australia through
twitter,” in UCUI 2015 - Proceedings of the ACM 1st International Workshop on Understanding the City with Urban Informatics,
co-located with CIKM 2015, Oct. 2015, pp. 7–12, doi: 10.1145/2811271.2811276.
[18] G. R. Septianto, F. F. Mukti, M. Nasrun, and A. A. Gozali, “Jakarta congestion mapping and classification from Twitter data
extraction using tokenization and naïve bayes classifier,” in Proceedings - APMediaCast: 2015 Asia Pacific Conference on
Multimedia and Broadcasting, Apr. 2015, pp. 14–19, doi: 10.1109/APMediaCast.2015.7210266.
[19] F. H. Khan, S. Bashir, and U. Qamar, “TOM: Twitter opinion mining framework using hybrid classification scheme,” Decision
Support Systems, vol. 57, no. 1, pp. 245–257, Jan. 2014, doi: 10.1016/j.dss.2013.09.004.
[20] W. Yao and S. Qian, “From Twitter to traffic predictor: next-day morning traffic prediction using social media data,”
Transportation Research Part C: Emerging Technologies, vol. 124, p. 102938, Mar. 2021, doi: 10.1016/j.trc.2020.102938.
[21] E. Alomari, R. Mehmood, and I. Katib, “Sentiment analysis of arabic tweets for road traffic congestion and event detection,” in
EAI/Springer Innovations in Communication and Computing, 2020, pp. 37–54.
[22] Z. Zhang, Q. He, J. Gao, and M. Ni, “A deep learning approach for detecting traffic accidents from social media data,”
Transportation Research Part C: Emerging Technologies, vol. 86, pp. 580–596, Jan. 2018, doi: 10.1016/j.trc.2017.11.027.
[23] A. Essien, I. Petrounias, P. Sampaio, and S. Sampaio, “A deep-learning model for urban traffic flow prediction with traffic events
mined from Twitter,” World Wide Web, vol. 24, no. 4, pp. 1345–1368, Jul. 2021, doi: 10.1007/s11280-020-00800-3.
[24] F. Ali, A. Ali, M. Imran, R. A. Naqvi, M. H. Siddiqi, and K. S. Kwak, “Traffic accident detection and condition analysis based on
social networking data,” Accident Analysis and Prevention, vol. 151, p. 105973, Mar. 2021, doi: 10.1016/j.aap.2021.105973.
[25] E. Alomari, I. Katib, and R. Mehmood, “Iktishaf: a big data road-traffic event detection tool using Twitter and spark machine
learning,” Mobile Networks and Applications, 2020, doi: 10.1007/s11036-020-01635-y.
[26] D. R. Utari, A. Wibowo, and A. A. Sobari, “Natural language processing of Twitter data for presenting road and traffic
information (in Bahasa),” in Senamika, 2021, no. April, pp. 756–765.
[27] J. Salazar‐carrillo, M. Torres‐ruiz, C. A. Davis, R. Quintero, M. Moreno‐ibarra, and G. Guzmán, “Traffic congestion analysis
based on a web‐gis and data mining of traffic events from twitter,” Sensors, vol. 21, no. 9, p. 2964, Apr. 2021, doi:
10.3390/s21092964.

Int J Artif Intell, Vol. 11, No. 4, December 2022: 1448-1459

Int J Artif Intell ISSN: 2252-8938  1459

[28] E. Utami, A. D. Hartanto, S. Adi, I. Oyong, and S. Raharjo, “Profiling analysis of DISC personality traits based on Twitter posts
in Bahasa Indonesia,” Journal of King Saud University - Computer and Information Sciences, vol. 34, no. 2, pp. 264–269, Feb.
2022, doi: 10.1016/j.jksuci.2019.10.008.
[29] Y. Kim, “Convolutional neural networks for sentence classification,” 2014, doi: 10.3115/v1/d14-1181.
[30] P. F. Muhammad, R. Kusumaningrum, and A. Wibowo, “Sentiment analysis using Word2vec and long short-term memory
(LSTM) for Indonesian hotel reviews,” Procedia Computer Science, vol. 179, pp. 728–735, 2021, doi:
10.1016/j.procs.2021.01.061.

BIOGRAPHIES OF AUTHORS

Rifqi Ramadhani Almassar is currently an employee at Maybank Indonesia,

Jakarta since 2022. He got bachelor’s in informatics from Universitas Internasional Semen
Indonesia, Indonesia, in 2020. He possessed his Master of Informatics (MTI) from University
Bina Nusantara, Indonesia, in 2022. He was a programmer at Sinergi Informasi Semen
Indonesia from 2019 to 2020, and then joined programmer at Global Service Indonesia from
2020 to 2021. His area of work interest is in the field of Machine Learning, Web
Programming, Mobile Programming, Internet of Things, and Artificial Intelligence. He can be
contacted at email: [email protected].

Abba Suganda Girsang is currently a lecturer at Master in Computer Science,

Bina Nusantara University, Jakarta, Indonesia Since 2015. He got Ph.D. degree in 2015 at the
Institute of Computer and Communication Engineering, Department of Electrical Engineering,
National Cheng Kung University, Tainan, Taiwan. He graduated bachelor from the
Department of Electrical Engineering, Gadjah Mada University (UGM), Yogyakarta,
Indonesia, in 2000. He then continued his master’s degree in the Department of Computer
Science at the same university in 2006–2008. He was a staff consultant programmer in
Bethesda Hospital, Yogyakarta, in 2001 and also worked as a web developer in 2002–2003.
He then joined the faculty of the Department of Informatics Engineering in Janabadra
University as a lecturer in 2003-2015. His research interests include Swarm, Intelligence,
Combinatorial Optimization, and Decision Support System. Email: [email protected]

Detection of traffic congestion based on twitter using … (Rifqi Ramadhani Almassar)