Depression is a major public health issue that affects people psychologically all over the world. It is defined as a collection of mixed impairment symptoms and disturbance in one's cognition and behavior (Orabi et al., 2018). During the 2019–2021 COVID era, there was a rapid increase in mental health issues and suicidal cases (Zulfiker et al., 2021). The World Health Organization states that more than 300 million people worldwide suffer from depression, prompting many researchers to focus on this topic (Priya et al., 2020). Chronic diseases can also be caused by depression. Figure 1 illustrates the symptoms of depression in a detailed manner.

Depression affects men and women differently, in such a way that women experience depression more severely than men (Seney et al., 2018). Because women are more likely to have anxiety disorders. Women are more involved in social gatherings, which exposes them to intimacy and emotional disclosure. Mental health issues can also be exacerbated by modern lifestyles influenced by social media and the pressure to live up to a certain standard that requires the approval of others (Kumar et al., 2020). Furthermore, studies revealed that younger individuals are disproportionately affected by mental health issues such as anxiety, depression, and obsessive-compulsive behavior, which has led many to commit suicide or consider it (Orsolini et al., 2021).

People use social media to share and communicate their ideas, and emotional states of being across many social platforms. Imagery is also a popular method of self-expression on social media platforms such as instagram (Mun and Kim, 2021). More evidence of depression can be found in Instagram photos (Smith and Anderson, 2018). As a result, this method effectively elicits deeper psychological consciousness by allowing emotional expressions that cannot be expressed in writing.

Another popular way to deal with mental health issues is through expressive writing or texting. Users of these social media platforms tend to document and narrate their lives through these platforms making it easy for us to understand their personal lives (Orabi et al., 2018). Twitter and Facebook are also popular social media platforms and important for person-to-person communication where people share their perspectives on real-life events (Ricard et al., 2018; Sood et al., 2018). Thus, depression can be identified through word sentiment analysis (Seabrook et al., 2018). According to Ricard et al. (2018), community-generated content responses can be used to identify levels of depression in people who have similar user-generated content. In retrospect, social media can also be beneficial to people's mental health.

Before this paper can go into what social media mining entails, it is necessary to comprehend what this topic stems from. Social media mining stems from data mining (Babu and Kanaga, 2022), which could also be considered as a stem of machine learning in various fields (Babu and Kanaga, 2022). Data mining can be considered the process of finding useful sets of data from larger pools of accessible data (Jagadishwari et al., 2021; Babu and Kanaga, 2022). In some studies, this is considered an automated process of uncovering knowledge, relationships, and patterns in interrelated sets of data (Babu and Kanaga, 2022). Data mining techniques are used in the process of social media mining for several reasons and several methods could be implemented in this process.

The growth and impact of social media have grown rapidly over the past few years. With this growth, more and more data has become available for both private and commercial use through these platforms. Noticeably, we are all constantly interacting with each other through these social media platforms, and one could say, our whole lives are now documented through this data. From whom we talk to, whom we know, and what we like or dislike, this information is accessible to almost anyone through these platforms, with this, the analysis of people or their social cues can be done through this data. Social media mining can be considered a process of gathering interrelated data from social media platforms to identify patterns and relationships in the data (Babu and Kanaga, 2022). There are several forms in which this data could be found. Text, image, and voice are three common forms of data that are mined during the social media mining process (Jagadishwari et al., 2021).

This is a growing topic that is used to understand people's sentiments about their everyday lives. This can be defined as a classification of text blocks, traditionally as either neutral, negative, or positive (Babu and Kanaga, 2022). However, this does not only depend on the polarity of the text, but the emotions associated with it, be it happy, sad angry, etc. Many studies have been conducted and commonly, various Natural Language Processing algorithms are used (Babu and Kanaga, 2022). Studies also show that Binary and Ternary classification techniques are regularly used, with multi-class classification providing more accurate results (Babu and Kanaga, 2022). Multi-class classification divides the data into multiple sub-classes and then works on these sub-classes separately based on the class's polarities (Gaikwad et al., 2014; Babu and Kanaga, 2022). Deep learning techniques are also used for the classification process (Gaikwad et al., 2014; Babu and Kanaga, 2022). However, the two common sentiment analysis techniques are rule-based sentiment analysis and machine learning-based sentiment analysis.

Rule-based sentiment analysis makes use of rules and word collections labeled by polarity to identify the opinion or context of the text (Babu and Kanaga, 2022). In this technique, sentiment value is made up of a combination of attributes to understand sarcasm, negation, or dependent clauses (Babu and Kanaga, 2022).

Machine learning-based sentiment analysis is focused on training a machine learning model using a sentiment-labeled training set (AlSagri and Ykhlef, 2020; Babu and Kanaga, 2022), to train the model to understand the polarity of words given in a certain order.

Sentiment analysis can be broken down into four major processes, namely, (1) data collection, (2) text preparation (data preprocessing), (3) sentiment detection (feature extraction), and (4) sentiment classification and presentation as output (Babu and Kanaga, 2022). The data collection process is made to allow for the relevant data to be collected, in this case, from various social media platforms. In a paper by Samsari et al. (2022) this data collection process consisted of a dataset made up of tweets collected during the COVID pandemic. Similarly in an article by Lui (2020), the data used in the study was collected from the Facebook and Twitter social media platforms. AlSagri and Ykhlef (2020), also collected data from tweets, and like most studies that made use of the text mining method, emoticons and emojis were removed as part of the preprocessing process (Babu and Kanaga, 2022). The data preparation process is used to clean the data by removing unrelated data and removing noise and words with no analytical significance (Babu and Kanaga, 2022). The third is the sentiment detection process, here the text is analyzed to extract opinions, reviews, and feedback while removing any text related to facts or popular knowledge (Babu and Kanaga, 2022). Thereafter, the classification and presentation process are conducted (Babu and Kanaga, 2022). Samsari et al. (2022) made use of the Naïve Bayes classifier to classify the data as either positive, negative, or neutral. Jagadishwari et al. (2021) made use of the Linear regression model as well as the Support Vector Machine (SMV) model, and these two models generated similar results. Ranganathan and Tzacheva (2019) used the Support Vector Machine LibLinear in an article titled Emotion Mining in Social Media Data. Tiwari et al. (2021) analyzed the performance of five different classification models and the most accurate results were generated by the Decision Tree classifier. The sentiment analysis process is shown in Figure 4 below.

These studies suggest that the two most common sentiment analysis methods used are the Rule-based method and the Machine learning-based method (AlSagri and Ykhlef, 2020). Regardless of the method used, there are four major processes in the sentiment analysis life cycle as mentioned, (1) data collection, (2) text preparation (data preprocessing), (3) sentiment detection (feature extraction), and finally (4) sentiment classification and preparation for output. Naïve Bayes, Decision Trees, SVM's, and Linear regression are some of the most common classification models used. However, studies suggest that Decision Trees have shown more accurate results when trained correctly (AlSagri and Ykhlef, 2020).

Four separate Twitter datasets were collected from Kaggle to narrow it down to three columns namely as one dataset: Tweet texts.

Figure 5 is an illustration of the before and after pre-processing of the datasets merged as one dataset.

The four models we looked at included python XGB Classifier, Random Forest Classifier, Logistic Regression, and Support Vector Machine C-Support Vector Classification Model. The four models are described in the Sections 5.1.1–5.1.4.

Table 1 showcases the results of different models, namely XGB Classifier, Random Forest, Logistic Regression, and SVM/SVC. The table presents their performance based on accuracy scores and computation time in seconds. The accuracy scores range from 95.2 to 96.3%, while the computation time varies significantly across the models, with values ranging from 0.29 to 1,072.32 s. These results provide insights into the models' predictive accuracy and computational efficiency, serving as a basis for further analysis and comparison.

The individual results of these models are presented in Table 1.

The results presented suggest that the SVM model and Logistic Regression model produced the most accurate results, with Logistic Regression slightly outperforming the SVM model, while the logistic regression model computed in the shortest amount of time. A more detailed analysis of the results suggests that the accuracy of the results was relatively similar for all four models. The lowest of the four models was the Random Forest Model with an accuracy of 95.2%, surprising as this is an Ensemble method, and it also had the longest computational time of 1,072.32 s. The second was the XGB Classifier with an accuracy of 96.1% and a computational time of 6.75 s. The third is the SVM Model with an accuracy of 96.2% and a computational time of 29.92 s and again, the fastest was the Logistic Regression model with an accuracy of 96.3% and a computational time of 0.29 s.

When comparing the results with previous studies, several insights emerge. Previous studies, however, suggest that ensemble methods should be more effective in sentiment analysis. In the comparison of the accuracy scores presented in Table 2, the random forest classifier's results are quite alarming, considering the higher expectations for ensemble methods. Jain et al. (2022) conducted a similar study and confirmed the effectiveness of the SVM classifier, which outperformed logistic regression and random forest in three out of the represented categories. This aligns with the findings of Jianqiang and Xiaolin (2017), who also highlighted the superior performance of SVM compared to logistic regression and random forest in their study.

Interestingly, Dave (2023) reported a relatively higher accuracy score for logistic regression compared to other models, reaching 83.62%. In addition, this study presents an even higher accuracy score for logistic regression, at 96.3%, indicating its effectiveness in accurately predicting the presence of depression sentiment in tweets. Hence the results obtained in this study indicate that logistic regression emerged as the most suitable model for the paper.

The SVM model, although not consistently outperforming the other models across previous studies, still demonstrates competitive accuracy scores. For instance, in the study by Aljabri et al. (2022), SVM achieved an accuracy of 88%. Similarly, in the current study, SVM/SVC performed well with an accuracy score of 96.2%.

The XGB Classifier achieved an accuracy of 96.1% in this study, indicating its strong performance in detecting depression sentiment in tweets. Compared to other models in the study, the XGB Classifier had the third highest accuracy score. Additionally, one previous study by Aljabri et al. (2022) reported a high accuracy of 90% for the XGB Classifier, further highlighting its effectiveness. The XGB Classifier's ability to capture complex patterns and interactions in the data likely contributed to its successful performance. Overall, the XGB Classifier shows promise as a reliable model for depression detection in tweets.

On the other hand, the random forest model presents mixed results. While it achieved the highest accuracy score in the study by Aliman et al. (2022) at 88.38%, it obtained a relatively lower score in the current study, with 95.2%. These variations could be attributed to different datasets or other factors.

Overall, considering the consistently high accuracy scores and the specific requirements of the paper, logistic regression emerged as the best model choice. However, the inclusion of other models such as SVM and XGB Classifier allows for a comprehensive comparison and exploration of their performance in sentiment analysis.