Utilizing K-Means Clustering for the Detection of Cyberbullying Within Instagram Comments

The advent of social media, a product of technological advancement, has instigated a paradigm shift in people's worldviews, lifestyles, and cultural practices. As more individuals are drawn towards online networks in daily life, certain activities, inextricably linked with the internet, are now deemed necessities by a majority of groups [1]. Social media, by virtue of its ability to forge broader connections, exerts a significant influence on people's lives [2].

The global penetration of social media is evident from the fact that Indonesia, with a penetration rate of 45 percent, ranks as the third-highest country in terms of social media usage. A total of 175.4 million individuals engage with social media platforms via mobile devices such as smartphones and tablets, accounting for 37% of the total usage in a week [3]. As per the Ministry of Communication and Information, internet access is available to 175.5 million people in Indonesia as of this year.

While social media can yield positive effects, the potential for misuse and criminal activities cannot be ignored [4]. According to a study, 95 percent of internet users regularly engage with social media platforms like Facebook, Twitter, and Instagram, with Instagram attracting a larger millennial audience than Twitter [5, 6]. Instagram's features, ranging from text messages to the transmission of images, videos, and documents, have streamlined long-distance communication. However, the publication of private information by many users opens the door to cybercrimes, including cyberbullying, which ranks third-highest globally in Indonesia [7].

The frequency of cyberbullying incidents on Instagram underlines the challenges in identifying victims and perpetrators due to the public nature of these actions and the lack of a solid reference for evidence [8]. This necessitates expertise in digital forensics to develop techniques for detecting initial acts of cyberbullying and non-cyberbullying in Instagram comments.

According to the anti-bullying organization Ditch the Label, cyberbullying constitutes defamatory personal messages, disparaging remarks on posts, and making fun of others [9]. The online harassment, threats, insults, and other adverse treatments that constitute cyberbullying often inflict more harm than physical assault [10]. Cyberbullying activities encompass a range of actions, including flame, harassment, cyberstalking, defamation, exclusion, trolling, impersonation, and dishonesty [11]. However, detecting verbal or written expressions that incite bullying or hatred can be challenging due to variations in accents and languages, necessitating one-language methodologies and manual data collection [12].

Existing literature reveals the use of methods and algorithms such as naive Bayes Classifier, C45, and K-Nearest Neighbors for cyberbullying analysis, among others like SVM and Naive Bayes Classifier. Although these have not been extensively applied to Instagram, previous studies using Instagram data have successfully classified users based on the appropriateness of specific hashtags through the K-means and tf-idf methods [13]. This paper presents the first instance of employing K-means clustering to detect cyberbullying on Instagram, categorizing cyberbullying behaviors in comments [14, 15].

The K-means algorithm, a partition-type algorithm, groups data into "clusters" based on similarity metrics [16]. It partitions the dataset into non-overlapping groups, ensuring that each data point belongs to a unique group. The cluster value is determined based on the distance between the data and the nearest centroid [17]. The optimization in K-means clustering involves the centered centroid of the cluster and the function used to compute the distance between objects [18, 19]. Therefore, it is hypothesized that bullying-related elements in Instagram comments can be grouped using K-means clustering.

This study aims to detect Instagram cyberbullying by integrating the K-means clustering technique with tf-idf weighting, a method never before applied to Instagram cyberbullying research. The objective is to identify initial acts of cyberbullying and non-cyberbullying in Instagram comments and develop an efficient method of detection. The results, represented by the accuracy, precision, and recall values, will serve as a reference in identifying the initial actions of cyberbullying in the victim's Instagram account and the victim's Instagram comment text. It is hoped that this study will augment the resources available to investigators to identify increasing instances of cyberbullying, particularly in Indonesia.

14-1.png

3.1 Data collection

The data used when using the crawling technique will lead to a less objective assessment of the manual labeling, thus using existing datasets from research sources that have been used. The cyberbullying dataset on Instagram social media used in this study has been used in previous studies using a different method, namely classification. There are several datasets sourced from https://github.com/rizalespe/Dataset-Sentimen-Analisis-Bahasa-Indonesia/blob/master/dataset_komentar_instagram_cyberbullying.csv.

The labeled data set used in this study was collected between 2019 and 2021 and sourced from www.kaggle.com. It is already categorized or labeled with 400 records and 2769 words in the study's open-access dataset. Table 1 shows the dataset has 3 attributes and 1 class attribute.

Table 1. Data attributes for cyberbullying on Instagram

3.2 Preprocessing

This step removes noise from the data to be analyzed to improve clustering results. This step is completed to ensure that the grouping procedure is carried out properly. The preprocessing set, according to the study of Rsa [31], entails a number of processes, including:

(1) Case Folding, the process of converting all letters in the data into lowercase letters. The characters other than letters and numbers are omitted and considered delimiters [32].

(2) Tokenizing is the stage of cutting the string based on each word that composed it. This process breaks sentences into individual words without considering punctuation [33].

(3) Stopword removal Vocabulary that is not a unique word or does not reflect the characteristics of a document will be removed by the stopword list dictionary that already includes the terms in it [34].

(4) Normalization is equating words not by Indonesian language rules, such as abbreviations, slang, and regional languages, which can affect stemming results. Normalization is done automatically using a library and manual methods by looking for samples of inappropriate words to be replaced with appropriate words [35].

(5) Stemming, namely the process of obtaining essential words by removing prefixes, suffixes, inserts, and combinations of prefixes and suffixes or confixes [36].

3.3 Term weighting

The noise reduction stage generates a series of keywords or words. The next step is word weighting, which entails giving a term or phrase a weight or value that demonstrates its importance to the document [37]. Each phrase or word in each paragraph is weighted to ensure consistency and coherence in the text [38]. The more times a term appears in the document collection, the more valuable or significant it becomes. The next stage after the weighting is clustering. The method used for weighting is the tf-idf method [39, 40].

The weight of the text is determined using the phrase "frequency," which counts the number of times a term appears in the text. When a term or word appears more frequently in a document [41]. The Inverse Document Frequency (IDF) method concentrates on how frequently a term appears throughout the text collection in comments [42]. In Idf, terms are valued higher because they only sometimes exist in the full-term collection. The Eq. (1) is used to determine Inverse Document Frequency (IDF) [43].

$I d f=\log \left(\frac{\text { number of document }}{\text { the number of documents containing the term }}\quad \quad \right)$                  (1)

The calculation from tf-idf is the term frequency value multiplied by the inverse document frequency. It combines the TF and IDF formulas.

3.4 Clustering

Clustering is a technique used when groups of objects in the same group are combined to create a unique cluster of objects in other clusters. Currently, the following algorithm is used to classify cyberbullying behaviors in comments that have both positive and negative traits [44]:

(1) The weighting process's resulting vector object is allocated after a random selection of the centroid is made.

(2) Euclidian distance is used to determine how far an object or phrase is from its centroid. To compute it using a Eq. (2). Formula for Euclidean Distance:

$\sum_{k=1}^n=\left(x_{i k}-x_{j k}\right)^2$              (2)

With:

dij=level of distinction

n=quantity of vectors

x_ik=input vector image

x_jk=comparison/output image vector

(3) In the event that the centroid shifts once more, step 3 of the process is repeated in order to locate the new centroid using Eq. (3).

$v=\frac{\sum_{i=1}^n \,\, x_i}{n} 1,2,3, \mathrm{n}$             (3)

With:

v=center of the group

x_i=item

n=the total number of objects and the total number of cluster members

(4) If the position of the nearest centroid does not change, the clustering procedure is finished, and the result is the grouping of objects into certain categories depending on the centroid.

3.5 Evaluation

In the evaluation stage, the confusion matrix is employed to assess the effectiveness and accuracy of the model [45]. The confusion matrix provides information on true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN), which aids in evaluating the clustering results [46]. To assess the model or algorithm's performance, this study will utilize 10-fold cross-validation to divide the data. A higher value in the confusion matrix indicates a better model and signifies greater accuracy in the clustering process. Three benchmarks will be used to assess the model's quality, namely Accuracy, Precision, and Recall. These metrics can be represented numerically using percentages (ranging from 1 to 100%) or values between 0 and 1. A recommendation system is considered good if it achieves high Accuracy, Precision, and Recall values.

Eq. (4) is an equation for calculating the accuracy value, while Eq. (5) is an equation for calculating the precision value, and the equation for the recall value available on Eq. (6) from clustering is as follows:

Accuracy $=\frac{T P+T N}{T P+F P+F N+T N} \,\, \times 100 \%$              (4)

Precision $=\frac{T P}{T P+F P} \times 100 \%$              (5)

Recall $=\frac{T P}{T P+F N} \times 100 \%$               (6)

Explanation: 1). True Positive (TP): Predict existing categories and system categories of the same comment that there is a match and an accurate match. 2). True Negative (TN): Predict existing categories and system categories from the same comment that there are no matches and accuracy that there are no matches. 3). False Positive (FP): Predicts existing category and system category of the same comment as a match and turns out to be false to no match. 4). False Negative (FN): Predicting existing categories and system categories from the same comment that there is no match, and it turns out that there is no match.

3.6 Case simulation

In Figure 2, you can observe a case simulation. In this case simulation, it is anticipated that the victim will report the incident to the investigator, who will conduct a cyber investigation.

14-2.png

Figure 2. Simulation of a case cyberbullying

In the case simulation in Figure 2, the victim experienced cyberbullying through comments on the Instagram social media platform. Cyber investigation steps will be carried out after reporting the incident to the investigator. Investigators will start by identifying comments that are considered cyberbullying. Furthermore, the investigator will collect digital evidence through these comments. Relevant information such as the sender's account name, date and time of delivery, and the contents of the comments will be recorded in full. The investigator will save the evidence obtained and make a detailed investigative report. This report will be evidence that can be used in court to prove the initial act of cyberbullying that occurred. At trial, digital evidence prepared by investigators will be included as evidence for initial action that can support lawsuits against perpetrators of cyberbullying. In addition, the results of this investigation can also help provide justice for victims and prevent similar cases from happening in the future.

3.7 Research tools

This study will use tools, including Jupiter Notebook, Microsoft Excel, and Rapidminer Studio. The tools used are shown in Table 2.

Table 2. Research tools

Thus, combining Jupyter Notebook, Microsoft Excel, and RapidMiner Studio will provide advantages and flexibility in data processing and analysis.