Optimization of support vector machine with cubic kernel function to detect cyberbullying in social networks

ABSTRACT


INTRODUCTION
Social networking is a forum where human beings can disseminate information, discuss, exchange ideas, pour out their hearts and share activities carried out in cyberspace by being delivered via the internet network [1], [2].Various social networks have their respective advantages, such as uploading photos, text, videos and voice recording that are able to interact completely and personally [3].All ages and groups can conduct discussions on social networks without differences in human degrees [4].The government can use the advantages and benefits of social networking to interact with the community in conveying information and receiving complaints from the public [5].
As a result of freedom of opinion in social networks, every government in various countries makes policies so that there are no unpleasant actions for social network users [6].In addition, the government in making policies for a country can be broadcast and disseminated through social networks so that delivery can be done quickly [7].With this, we can reap the pros and cons of policies made by the government because With the pros and cons of government policies made, social networking cannot resist the accelerated rate of giving opinions made on social networks [8].So that arises encouragement from various groups and individuals on social networking [9].This accelerates the delivery of bladder to become cyberbullying, where cyberbullying is an abuse of the use of the internet network by harassing, humiliating, threatening, and insulting others with digital trace recording [10].
Cyberbullying in various literacy which is embodied by LaFrancis and Putnam [11] explained that cyberbullying in the teenage group, the role of parents is needed to avoid raw failure in the real world.Meanwhile, according to Jiang et al. [12] it is necessary to have a government policy in determining the regulations for using cyberbullying.From the literacy, cyberbullying needs to be limited and detected using computer science methods.In accordance with the science of computer science, there are methods for classifying and clustering with results to detect certain events [13]- [15].This knowledge is contained in data mining in combination with artificial neural networks [16], [17].
The speed of increasing data on cyberbullying on social networks cannot be stopped manually [18], [19].So, the need for a method to be able to detect the outpouring of cyberbullying on social networks [20].Potha and Maragoudakis [21] performs cyberbullying to detect and search for cyber predators using the time series method where each predator question is manually explained in terms of severity using a numeric label and applies the dynamic time warping algorithm, so that similarities in the signals are evident, providing a direct indicator for the severity of cyber bullying in this particular dialogue.Meanwhile, Nahar et al. [22] with a data mining approach in detecting cyberbullying with stochastic data and explaining that support vector machine (SVM) is an algorithm with supervised learning where the process must first have a label with fuzzy so that it is continued with the SVM algorithm so that the data becomes complex and multi-dimensional and this method is called fuzzy SVM.
Andriansyah et al. [23] classified the comment column on Instagram on artist accounts which are often referred to as celebrities in detecting cyberbullying where there were 1053 comments and 34 inspection documents containing cyberbullying of 79.412%.Meanwhile, Noviantho et al. [24] identify cyberbullying conversations with a combination of SVM and Naïve Bayes with an accuracy of 92.81% but the algorithm is modified with a poly kernel and achieves optimization at an accuracy of 97.11%.From various studies before doing cyberbullying, they often use SVM so that in this paper a study is designed to detect cyberbullying with data that is tested for cyberbullying on government policies using SVM which is focused on the method, namely using the kernel function to get optimal results using the cubic kernel function.

MATERIAL AND METHOD 2.1. Dataset
The dataset in this paper uses trending data on twitter on government policies regarding the determination of "cipta kerja" which uses the pros and cons so that cyberbullying is identified in every tweet posted by the Twitter social network user.Through keywords using "cipta kerja", the activities are sorted so that the data is crawling and then cyberbullying is detected in government policy.The amount of data achieved is 2400 tweets.

General architecture
This paper is the result of research that focuses on the SVM algorithm that was developed.Where the kernel used is a linear kernel but compared to the use of other kernels such as cubic kernels.So, the contribution in this paper is to produce optimal accuracy values in detecting.So, a general architecture was formed in this research so that it does not spread.The general architecture in this paper can be illustrated in Figure 1.The explanation in Figure 1

Optimization of SVM
SVM is an algorithm that is relatively new and is able to solve regression and classification problems on very large amounts of data [25], [26].In carrying out learning, of course SVM uses supervised learning techniques with supervision and then enters the testing stage [27], [28].In the process, SVM also performs a linear projection from the feature space to the kernel and creates separate and linear classes [29], [30].
The SVM algorithm also has a very significant advantage in optimization, namely by achieving a minimum value in global and local problems.So that SVM is often used in cases to classify, detect, estimate, and predict [27].In its history, SVM was discovered by Vapnik in 1963 which was used in doing classification problems [31], [32].SVM can also basically classify data with high dimensional non-linearity [33].With strong learning, SVM is widely applied and applied to the real world with accurate computational techniques [29], [34].SSVM uses supervised learning techniques which are often found in biophotonics, image detection, classification, estimation, and prediction problems [35].SVM is calculated based on the hyperplane as in the (1) [36], [37]: Where  is the normal value of the hyperplane, () is a function of the input vector, and  is the bias value.
Then from these calculations an optimization is carried out in order to get the minimum value based on (2) [35]: Where   is the input vector value,   is the  class correspondence value, () is the lagrange multiplier value, and  is the linear kernel and cubic kernel function.In the combination of parameters must be determined in advance in getting it.Thus, SVM can be grouped into linear and nonlinear.As seen in ( 3): Where   is the value on the label   ,  is the value to be multiplied by lagrange and  0 is the bias value of 0. In non-linear problems the equation is changed based on (4) [38].
Where the value of  is the value of the number of input vectors.

RESULTS AND DISCUSSION
At this stage, the results of applying sentiment analysis to data originating from social media with the keyword "cipta kerja" will be presented using the SVM algorithm with kernel cubic and linear kernel.This study will use as much as 2400 data originating from social media twitter.The data is obtained from the crawling process on twitter social media then the text preprocessing stage is carried out to change unstructured data into structured data which will then be carried out manually labeling, in data labeling there will be positive, negative and neutral labels which will then be carried out sentiment analysis with an SVM algorithm.Following are the results of the sentiment analysis process using the SVM algorithm with cubic kernels and linear kernels.

Text preprocessing results
The result of the text preprocessing process is tweet data with the keyword "cipta kerja" which has been changed and cleaned of irrelevant characters.In the text preprocessing there are 4 stages including case folding, tokenization, stopword and stemming.However, in this study only case folding, tokenization and stopword stages were carried out because the resulting data does not require a stemming process.The following data is the result of text preprocessing for each stage.a. Case folding Case folding is a text preprocessing process that changes the entire contents of tweets to lowercase with the aim of making text data analysis able.Where this research uses text data originating from Twitter and case folding will be carried out.Following are the results of case folding on the dataset contained in Table 1.

c. Stopword
Stopword is the process of removing words that have no meaning contained in the tweet data.By previously carrying out case folding and tokenization and then doing stop words.Following are the results of stopwords in the dataset contained in Table 3.

Sentiment distribution
Sentiment distribution on data to identify and map tweet data on positive, negative and neutral sentiments, the purpose of the distribution is to see opinions contained in tweet data originating from social media with the keyword "cipta kerja".By distributing sentiment, you can find out the opinion patterns given by the community.The following is the sentiment distribution contained in Figure 2.Where Figure 2 is the distribution sentiment from a data warehouse in the form of data from Twitter.Information in Figure 2 contains data with neutral sentiment labels with a total of 1550, negative sentiment labels totaling 700 data and positive sentiment labels totaling 150 data out of the total data used totaling 2400.From the distributed data, training and testing can then be carried out with the SVM algorithm.

Test data sharing
The division of test data is the process of breaking data into two stages, namely data training and data testing, data training will be used to train the sentiment analysis model using the SVM algorithm while data testing will be used to test the model that has been formed.The following is a visualization of the distribution of test data in Figure 3. Figure 3 explains that in the process of building a sentiment analysis model using the SVM algorithm, there is a division of training and testing data, in the training data the percentage is 70% of all data and 30% for data testing of all data.

The results of the SVM model use the linear kernel
In the results of applying sentiment analysis to the SVM model using a linear kernal, it can be seen that the sentiment analysis model produces performance evaluations such as accuracy, precision, recall and f1 score.Following are the results of the SVM model using the linear kernel which can be seen from the results of the following confusion matrix in Figure 4. Description of Figure 4 shows the distribution of predictive data with a confusion matrix with positive, negative, and neutral values.The following is a Table 4 of manual confusion matrix calculations to find accuracy, recall, precision, and F1 score values.Based on Table 4, the calculation of accuracy, precision, recall and F1 score will be carried out as follows: a. Accuracy Accuracy=

The output of the SVM Model uses a cubic kernel
In the results of applying sentiment analysis to the SVM model using cubic kernel, it can be seen that the sentiment analysis model produces performance evaluations such as accuracy, precision, recall and f1 score.The SVM algorithm process certainly does not go through the calculations available from (1)−(4).The following are the results of the SVM model using cubic kernels which can be seen from the results of the following confusion matrix in Figure 6.The description of Figure 6 shows the distribution of predictive data with a confusion matrix with positive, negative and neutral values.The following is a Table 5 of manual confusion matrix calculations to find accuracy, recall, precision and F1 score values.Based on Table 5, the calculation of accuracy, precision, recall and F1 score will be carried out as follows: a. Accuracy Accuracy=

CONCLUSION
In this paper, cyberbullying has been detected on tweets from the word government policy in tweets.From the example of the word "cipta kerja" which fills comments from twitter users, it causes bullying so that cyberbullying is formed.However, the method used in detecting using SVM has been optimized on kernel cubic.The results show that the accuracy of the linear kernel is 92.3% while the cubic kernel function optimization gets an accuracy of 90%.ham shows optimal because the resulting accuracy is more optimal.


ISSN: 1693-6930 TELKOMNIKA Telecommun Comput El Control, Vol.22, No. 2, April 2024: 329-339 330 they are networked by the social community, they can give opinions without any restrictions by using social networks.

Figure 3 .
Figure 3. Distribution of data sharing

Figure 5 .
Figure 5. Result of the ROC curve SVM model with a linear kernel


Optimization of support vector machine with cubic kernel function to … (Al-Khowarizmi)337The ROC curve aims to see the performance of the sentiment analysis model on tweet data with the keyword job creation.The results of the ROC curve will produce a comparison graph between the positive class and other classes.Here are the results of the ROC curve in Figure7.

Figure 7 .
Figure 7. Result of the ROC curve SVM model with a cubic kernel

Table 1 .
The case folding process process carried out to separate tweet data into separate words or tokens.In this case, the data uses text originating from Twitter and has been casefolded and then tokenized.Following are the results of the tokenization of the dataset contained in Table2.
Optimization of support vector machine with cubic kernel function to … (Al-Khowarizmi)333

Table 2 .
The tokenization process

Table 3 .
The stopword process

Table 4 .
Manual confusion matrix calculation table with linier kernel

Table 5 .
Manual confusion matrix calculation table with cubic kernel