Psychological Analysis for Depression Detection from Social Networking Sites.

Rapid technological advancements are altering people's communication styles. With the growth of the Internet, social networks (Twitter, Facebook, Telegram, and Instagram) have become popular forums for people to share their thoughts, psychological behavior, and emotions. Psychological analysis analyzes text and extracts facts, features, and important information from the opinions of users. Researchers working on psychological analysis rely on social networks for the detection of depression-related behavior and activity. Social networks provide innumerable data on mindsets of a person's onset of depression, such as low sociology and activities such as undergoing medical treatment, a primary emphasis on oneself, and a high rate of activity during the day and night. In this paper, we used five machine learning classifiers-decision trees, K-nearest neighbor, support vector machines, logistic regression, and LSTM-for depression detection in tweets. The dataset is collected in two forms-balanced and imbalanced-where the oversampling of techniques is studied technically. The results show that the LSTM classification model outperforms the other baseline models in the depression detection healthcare approach for both balanced and imbalanced data.

1. Introduction

Psychological analysis is a process in which psychological data are extracted from text-based data. To eliminate emotions, opinions, and judgement-forming, text data are used. Having opinions or views toward some products or any topic is human psychology, which defines what one thinks about the products or topic. Nowadays, the way of expressing various emotions and giving opinions has changed drastically with the advancement of social media and Internet technology. People use blogs, product recommendation and review websites, and other social media to give opinions about products, movies, and political parties and on current important topics. Famous social media platforms such as Facebook, Twitter, and Reddit have become the most reliable platforms for sharing opinions and reviews among a new generation of Internet users [1]. Business firms and organizations use people-oriented psychological feedback to increase their products' value and quality.

1.1. Role of Psychology

Human beings usually have a greater sense of emotions and feelings; these feelings, when merged with technology, can be converted into useful tools. Another word used for human feelings is psychology. Psychology is used in various research works, such as identification of mental health issues, product reviewing, customer satisfaction identification, and business advertisements. In business advertisements, psychology helps in decision-making, e.g., to buy a product online, when the product is not available to touch and feel, then product reviews help a customer take decisions on whether the product is good to buy or not. Usually, online reviews are a mix of true and false opinions. Reviews with high polarity toward positive psychology increase the product value [2]. Reviews also hint about defaults and needs in a product that help business firms improve their products and satisfy their customers. Half of the business world depends on customer feedback and reviews. Another application of psychological analysis is market research that includes collecting data through social media and other websites. These data help in understanding the current market trend and advertisements' quality that affect people. The reviews provided by people for a product or a movie follow a collect-and-pass-through recommendation system, which depicts the polarity of psychology that shows whether people like the product or the movie or not. People's opinions and reviews can be analyzed thoroughly, which helps boost business performance and design future services [3].

Second, psychology plays an important role in mental health research, such as identification of anxiety attacks, major depressive disorders, bipolar disorder, and many more. We focus on major depressive disorder in this paper. Depression is also known as major depressive disorder and is commonly found in people with anxiety issues. Every 1 in 5 people have been suffering or have suffered from depression [4]. Of the total world population, 4.5% of the population is suffering from depressive disorder. Depression has some common symptoms such as anxiety attacks, loss of appetite, feeling sad for a long duration (1 month), and losing interest in favorite activities. Usually, during the initial state of depressive disorder, an individual avoids social gatherings, shows lack of energy, makes fewer efforts to communicate with friends and family, and shows a feeling of incompleteness. Due to lack of social interaction and the fear of being judged, depression survivors indulge in social networking to share their thoughts and feelings with people similar to themselves (Munmun et al, 2021). By being hidden and still saying what they had in their mind, this made social media more useful for depression detection research work. Mainly, the Twitter platform comprises many eco-groups where people of the same interest connect with each other through, for example, the “depression group,” “mental health club,” etc. We use the Twitter platform in this work, which provides sufficient data for depression detection and classifying users into depressed and nondepressed categories.

Words convey different psychologies and tell about the current psychological state of a being. For example, consider words such as “not feeling good” or “why people behave like this to me.” Sentences comprise words with both positive and negative polarity, but with the use of “not” and “why,” the very meaning and feelings change. The dictionary software LIWC (Linguistic Inquiry and Word Count) is usually used for analyzing words and extracting their meaning in terms of psychology [5-7]. This dictionary software comprises multiple categories according to the types of human psychology, which helps in psychological analysis using textual data only. Multiple factors are considered when we perform psychological analysis to detect mental health illness, such as temporal factors, emotional factors (positive: “excited,” “wonderful,” and “lovely”) (negative: “empty,” “lonely,” and “forgetful”), and use of personal pronouns (“I,” “me,” and “them”) (Verma et.al, 2020).

1.2. Psychological Analysis Using Social Media

Social media platforms such as Facebook, Twitter, Instagram, and Reddit consist of rich amounts of data to perform psychological classification tasks. Psychological analysis is all about extracting psychological data from textual data. Twitter is a famous social media platform used by millions of people worldwide. People share their opinions about current hot topics or any political chaos and discuss famous incidents. The opinions or views in Twitter are in the forms of tweets, which have a maximum limit of 250 words approximately. This limit made Twitter data important. As tweets are meant to have a certain word limit, users use specific words to express their views and emotions. Twitter, nowadays known as a hub of political and government people— mainly politicians and government officials, is used for various announcements about international and national projects [8]. Users discuss the views given by various political parties. During elections, this kind of discussion will help in determining results of elections if psychological analysis is carried out using tweets collect from the Twitter platform. Another famous social networking platform is Facebook. As the Internet expanded, Facebook became famous in 2007, and now, in 2021, Facebook currently has 5 billion users around the world. People who use the platform provide useful insights into businesses, movies, politics, and trends. Facebook generates a vast number of data every year, which comprise images, text, and links. Facebook comprises multiple online communities where people of likeable interest connect with each other and share their psychological views and opinions. Online communities, such as political groups that discuss their favorite and hated political personalities, talk about parties that have different opinions. These types of data are useful in understanding what kind of psychology people have for their political parties. Another group is mental health groups where people discuss their mental health—how they are fighting every day to live and undergoing treatment [9]. Data from these types of groups provide sufficient text data to understand the psychology of people in these groups. Identifying psychology through text data is much easier than identifying psychology through image-based data as images require use of deep learning classifiers and high-quality image-based data. However, with text-based data, psychological analysis is easily performed on these giant social networking platforms. A lot of research has been carried out in this field, and it uses artificial intelligence, which in combination with LIWC provides satisfactory results. Facebook contains ads for various companies and data about companies that we search on the Google platform. Usually, these ads sometimes use surveys to understand user psychology as to what kind of image they have about the business organization. As shown in Figure 1, there are some basic human emotions that are usually conveyed through specific words. Facebook provides some reaction emoticons in the “Posts” section that contain emoticons similar to these basic human emotions. Figure 2 shows various methods for psychological classification, which include dictionary-based and machine learning–based methods. In dictionary-based methods, the sets of words are classified using dictionary values, and machine learning–based methods include supervised and unsupervised learning methods, such as neural networks, that provide some of the best results. There are hybrid-based methods that combine dictionary-based methods with machine learning–based methods.

Figure 1

Commonly used human emotions.

Figure 2

Various methods for psychological classification.

In this paper, we have selected and trained the tweet data. We have performed data preprocessing on datasets and removal of the raw data from the dataset. After value extraction from the datasets, training the data of tweets and cross-validating the training dataset were carried out. We used 5 machine learning classifiers—support vector machines, decision trees, logistic regression, K-nearest neighbor, and LSTM—for depression detection in tweets. The results show that the LSTM classification model outperforms the other baseline models in the depression detection approach. To handle the imbalanced dataset, the oversampling and undersampling of class imbalance approaches are implemented and analyzed.

In this paper, the study is carried out on two datasets that involve the imbalance and the balance set, and different techniques, such as SMOTE and RUS, are used for oversampling and undersampling to work with the dataset. From the research gap, we are trying to conclude that the performance of LSTM is better than that of other machine learning classifiers.

The methodology that has been followed in this paper is described in Figure 3.

Figure 3

Methodology.

The paper is organized as follows:

Section 2 comprises a literature review. Section 3 comprises the prerequisites of depression detection. Section 4 describes the proposed method, and Section 5 includes the experiment setting, which includes a subsection of data information, modeling that involves the study of five machine learning classifiers used in this research, data preprocessing, and performance measurement. Section 6 includes discussion and results, as well as various measurement comparisons in the form of a graph, and finally, the conclusion and future scope are mentioned in Section 7.

2. Literature Review

Psychological analysis can be carried out using various methods, as explained in Figure 2, with the text-based dataset. In this section, we discuss the previous works performed using various techniques for psychological analysis and the depression detection task.

Using linguistics in depression detection is quite useful as it shows that words used by nondepressed and depressed people may differ. Depressed individuals mainly focus on themselves. In 2014, Nguyen studied two online discussion groups, namely, “control” and “clinical” groups. The “control” group comprised people with a similar interest and fun-loving people, whereas the “clinical” group comprised people suffering from mental illnesses such bipolar disorder, major depressive disorder, SAD, and anxiety attacks. People in the clinical group discussed their issues freely and took advice for medication and intervention. The author finds a difference in online communities involving people of these two groups. People in the “clinical” group usually use first-person pronouns (“I”, “me,” and “my”) in comparison with “control” group people, who use fewer first-person pronouns and discuss various activities such as dancing, singing, and running. This work reveals that use of language plays an important role in depression detection as words describe what someone's current mental state is.

In 2005, Pennebaker used LIWC (Linguistic Inquiry and Word Count), a piece of dictionary software, for analyzing textual data to obtain meaningful insights. Detecting depression using Facebook comments and using 4 categories from the LIWC software to analyze words in the comments were the author's objectives. These categories are language-based, time-based, or emotions-based or include all features. These factors have all minor characteristics of human language use and conditions such as related emotions, time periods, and use of nouns to find out the very meaning of human speech. Using the KNN (K-nearest neighbour) machine learning algorithm for classification, an accuracy of 65% was achieved.

In 2019, Gaikar used SVMs (support vector machines) and Naïve Bayes classifiers to detect depression-related words and sentences and detected the types of depression from those identified words. The authors trained both the classifiers using the bipolar disorder dataset and depression illness–related dataset. The best accuracy achieved was 85% using a machine learning classifier. In [10], the author used audio and text-based data. These data play an important role in natural language processing, and they range from images, texts to videos, and small clips. With text data, the information is limited, but with the audio of depression survivors, it is easy to understand some facts about the affected person's body language. A total of 142 individuals underwent depression detection tests by being asked some questions. They recorded the answers in audio forms and text forms directly from the 142 people, and feature extraction was carried out using the long short-term memory model. The results show that people with depression use more pauses than nondepressed persons and that using first-person pronouns is common among depression survivors; however, nondepressed persons use fewer pauses while talking and focus less on them.

In [11], the author used the Twitter platform to conduct the task of identification of depression using text-based data as Twitter provides short and useful linguistic phrases that directly shows the current mental state of the user. The data were gathered from the CLPsych 2015 conference in which the latest 3000 public tweets are available. For feature extraction, the author used the bag-of-words approach that is famous for identifying mental health illness using machine learning. The method shows word frequencies and the number associated with them and applies various machine learning classifiers in which Naïve Bayes/1-gram gives a recall value accuracy of 86% with 82%, respectively.

Reference [12] focuses on early detection of depression using neural networks. Depressive disorder has various stages, which include initial, intermediate, major, and severe disorders. The initial level of depressive disorder includes visibility of symptoms appearing in affected people such as low appetite, feeling of suicide, social inferiority, and comparing themselves with others. The author suggests identifying depression at its initial stage where it is less harmful is easily curable. Using CNNs (convolutional neural networks), behavioral characteristics from text-based data (CLPsych) are extracted, and some improvements in the early detection parameter ERDE are introduced.

In 2018, Lang used speech data to detect depression, proposing a deep convolution neural network to process speech-based data. Authors use secondary datasets such as AVEC2013 and AVEC2014 depression datasets that comprise 340 short videos of 292 selected people. These data were collected through human beings and computer system interactions using webcams and microphones. The model gives an RMSE value of 9.0001 and an MAE of 7.4211. In Ref. [13], the author analyzed Facebook comments data using various machine learning algorithms. Facebook contains plenty of data that comprise videos, images, and text-based data. Using only text-based data from comments of various publicly available pages, the author analyzed them using machine learning classifiers such as decision trees, KNN, support vector machines, and ensemble classifiers, in which decision trees achieved the best accuracy of 73%.

In Ref. [9], the authors proposed a lexicon-based approach for detection of depression. They constructed a lexicon using Word2Vec, a semantic relationship graph, and the label propagation algorithm. The authors based it on 111,052 Weibo microblogs from 1868 users who were depressed or nondepressed. They have used five classification methods and considered six features to predict depression. The results show that the lexicon generated proved to be better for classification.

In Ref. [14], the researchers focused on mood analysis of human beings with the help of machine learning and deep learning tools of artificial intelligence. In this paper, authors also focused on limitations of artificial intelligence to detect depression. In Ref. [15], the author implemented a machine learning model with preprocessed data for automatic depression classification. Kinect captured a skeletal model used for data extraction and preprocessing. The model achieved 96.47% accuracy in old-age group and 53.85% accuracy in young-age group. In Ref. [16], the authors carried out prediction of depression by noticing the human behavior. For prediction, authors used smart phone datasets. The model achieved 96.44%–98.14% accuracy. In Ref. [17], the researchers implemented an automatic depression detection [AUTO DEP] model by using facial expressions of human beings. The authors used a linear binary pattern descriptor model for feature extraction. The performance of the automatic depression detection model is the same as that of usual previous models. The evaluation of the model was performed on MATLAB and the linear binary pattern descriptor FPGA using Xilinx VIVADO 16.4. Table 1 shows the consolidated literature survey performed by us before research.

Table 1

Literature survey.

Author	Dataset	Model	Result
Nguyen et al. [18]	Clinical group and control group	Lasso	Accuracy 93
Pennebaker et al. [6]	Facebook	Knn	Accuracy 65
Gaikar et al. [19]	Bipolar disorder and depression illness	SVM	Accuracy 85%
Hanai et al. [10]	Audio and text based	LSTM	F1: 0.44 and precision: 0.59%
Lang and Cao [20]	Twitter	Naive Bayes	Accuracy 86%
Trotzek et al. [12]	Reddit message	CNN	Accuracy 87%
Lang and Cao [20]	Speech data	Deep CNN	RMSE: 9.0001 and MAE: 7.4211
Islam et al. [8]	Facebook comments	KNN	Accuracy 73%
Li et al. [9]	Weibo microblogs	Logistic regression	Accuracy: 77% and precision: 77%
Li et al. [21]	Shandong mental health center	Kinetic captured skeleton	Accuracy 96.47%
Asare Kennedy et al. [16]	Smart phone dataset	SVM	Accuracy 96.44–98.14%
Tadalagi and Joshi [17]	Facial expression	SVM	Accuracy 72.8%

After the survey, we have identified the following research questions:

RQ1: Will solving the class imbalance problem improve the performance of the psychological analysis model?

RQ2: Which technique of class imbalance (oversampling/undersampling) gives better results for depression detection?

RQ3: Which classification model outperforms the other baseline models in the depression detection approach?

The above-mentioned research questions have been answered in the subsequent sections.

3. Prerequisites for Depression Detection

Mental health conditions are the common problem for people. According to the WHO, 20% of people suffer from this problem. Adults and children with mental illnesses, such as depression, memory loss, hypertension, and anxiety attacks, were the primarily affected ones. Depression is the fastest growing health disorder; it depends on the mood, which comprises components of motivational and emotional conditions. Traditionally, mental health experts have mainly used clinical examination procedures depending on self-reporting of emotional, behavioral, and cognitive dimensions. To diagnose and calculate the predictions on the condition's evolution, machine learning algorithms have been used since 1998. There are various types of approaches used to detect depression as follows:

3.1. Machine Learning Approach

SVMs, decision trees, random forests, and Naïve Bayes are examples of supervised and unsupervised machine learning algorithms. These algorithms conduct longitudinal, temporal, and sequential analysis on the mood of human beings, and they also analyze social media and social networks. These algorithms detect depression by plotting graphs. The longitudinal analysis plots the user activities with time, and the results show less social interaction, more negative things, strongly clustered ego networks, increased interrelation, and medicinal issues.

3.2. Lexicon-Based Approach

The lexicon-based psychological analysis algorithm uses psychological normalization and evidence-based combined functions. In this, they used Rtexttools of machine learning for comparison of lexicon psychology. The lexicon analysis types in a corpus are adverb lexicons, network word lexicons, and negative word lexicons.

In this approach, the psychological calculation using the number of positive and negative words in a document can be carried out by using the following formula [9]:whereIn the above-mentioned equation, Npos shows the total number of positive words in the current tweet-based data. Npos is the sum of the positive document frequency of each word in the document. Nneg shows the total number of negative words in the tweet-based dataset.

The two methods of lexicon are as follows:

3.2.1. Dictionary-Based Method

The dictionary-based approach is a part of a statistical model that is used to encode the symbols from the datasets. This approach does not encode the single symbols as a variable length of bit vectors. It encodes the variable length, staring into a single token.

The token starts with the dictionary index. If tokens are smaller, then they are replaced with bit vectors. The dictionary-based approach is an easily understandable approach.

3.2.2. Corpus-Based Method

The corpus-based approach is based on study of language, which is based on the language of the text corpus. It is the reliable analysis of languages, psychologies, and disorders. The corpus-based approach collects the natural context of language and performs psychological analysis with minimum experimental interference. The text corpus has been used for linguistic search and to compile dictionaries since 1985, and this method is used for psychological analysis. A phrase's semantic orientation is calculated, where the orientation is determined by mutual information.

Semantic Orientation (phrase) = PMI (phrase, “Amazing”) PMI (phrase, “Destitute”)

The words “Amazing” and “Destitute” are used for calculating the SO of phrases.

3.3. Artificial Neural Network

The most powerful neural network is the ANN, which is a part of machine learning. It is employed in a variety of fields, including computer vision, digital image processing, psychological analysis, and word sequence prediction, and it produces low-error results. ANN algorithms are modeled as the human brain. ANNs work just like how a human brain uses neurons and nerves and learns from the past data. Similarly, the ANN can learn from the past data, predictions, and classifications.

3.4. Class Imbalance Approach

The two types of the class imbalance approach are

SMOTE-Smote balances the class distribution and handles the imbalance problem in the dataset. It calculates the Euclidean distance between the minority class and other instances to find out the K-nearest neighbors.

RUS-It is an undersampling technique of the size of the class in no higher instances. The majority is reduced from the source dataset. It is a simple and easy method of undersampling. A subset of the majority class is selected and merged with the minority class samples of the dataset.

The aim of this study is to identify the individuals in depression on Twitter using tweets, which are short messages created by the individual users on Twitter. We are using 5 machine learning classifiers—SVMs, KNN, decision trees, logistic regression, and LSTM—for the classification job, and feasibility of this classification is examined through evaluation metrics precision, F1-score, and recall. To handle the imbalanced dataset, the oversampling and undersampling of class imbalance approaches are implemented and analyzed.

4. Proposed Methodology

In this paper, we have selected and trained the tweet data. We have performed data preprocessing on datasets and removal of the raw data from the dataset. After value extraction from the datasets, training the data of tweets and cross-validating the training dataset have been carried out. The sampling process is performed over the dataset so that it can sample the data according to the psychology.

The classifier is implemented, which will classify the actual data and apply various learning methods to it. In the classification technique, the data are categorized in two parts: one is imbalanced data, and the other one is balanced data. In imbalanced data, the data which do not contain any missed values can be oversampled. Balanced data are categorized into two types: (1) positive tweets and (2) negative tweets. The flowchart is shown in Figure 4.

Figure 4

Proposed methodology for handling the data of tweets.

5. Experiment Setting

5.1. Data Acquisition

In this study, we considered two datasets: Sentiment_140 and Sentiment_tweets3; these are publicly available on https://www.kaggle.comfor for research and study purposes.

Sentiment_140: This is one of the most famous datasets for sentiment analysis and for natural language processing. The dataset consists of 1.6 million tweets, which is extracted using Twitter API. The tweets have annotation 0 = negative, 2 = neutral, and 4 = positive, but we use different annotations for our study, which is 0 = negative and 1 = positive. There are 3 fields in the dataset:

Id = id of the tweet.

Text = the tweet data.

Label = the polarity of tweets (0 = negative and 1 = positive).

Sentiment_tweets3: This dataset consist of 14000 tweets, which is publicly available for use, from which only 6000 tweets are taken for use. For this study, we combined these two datasets and collectively used them; if the dataset is imbalanced, then the missing values of the dataset are checked, and the data are oversampled.

In the imbalanced dataset, first the number of instances with positive and negative tweets is checked. This will degrade the overall predictive performance of the model. The imbalanced data modeling is carried out with oversampling data techniques.

5.2. Modeling

Modeling is the mathematical expression that represents the data in the context of a problem, often a business problem. The main aim of modeling is to form data for insights. Machine learning algorithms perform processing of data. To process the data, some models are used. The following are the data processing models used in this paper:

Decision tree (DT): A decision tree is a decision support tool that includes the chances of event outcomes, resource cost, and utility. It is used in predictive modeling approaches, statistics, and machine learning.

One of the important parameters in decision trees is information gain, which minimizes the amount of information required to differentiate between two data points for partition[18].

Here, pi shows the probability that a tuple in dataset D may belongs to class ci.

Info (D) is the sum of the mean amount of data required to determine a data object class D to which it belongs. The info gain is calculated as

Support vector machine (SVM): A support vector machine is a model that uses the classification algorithm for two-group classification problems. An SVM is a fast and dependable algorithm. It consists of a line that separates two data objects, known as the decision boundary. This acts as the main separation axis. The equation of the separation axis is

Y = mx + c

where m stands for the slope.

Now, the hyperplane equation separating the data objects are

H: wt(x)+b = 0, where b stands for the bias term.

K-nearest neighbor (KNN): This algorithm is used for classification and regression. The KNN algorithm assumes the similarity between the new case data and available cases and put the new cases into the category that is most similar to the available categories.

The KNN algorithm utilizes the Euclidean distance formula to determine the minimum distance between two data points in the plane with coordinates (x, y) and (a, b), which is represented as

LSTM: It is an artificial recurrent neural network architecture used in the field of deep learning. It can only process single data but can also process an entire sequence of data.

LSTM transition function[9]:

It = σ (Wi. [ht-1, xt] + bi).

Ft = σ (Wf. [ht-1, xt] + bf).

Qt = tanh (Wf. [ht-1, xt] + bq).

Ot = σ (Wo. [ht-1, xt] + bo).

Ct = ft ® ct-1 + it ® qt.

Ht = ot ® tanh (ct).

The LSTM model contains multiple iterative steps for every time stamp. Here, ht-1 shows the hidden states (old), xt shows the input of the current time stamp, and ft is the forget gate that partially removes data, which are redundant or not useful from the old memory cell, for further processing in LSTM with ot as the output gate that selects an appropriate output for the current time stamp. The sigmoid function is used here that has a range [0, 1], which is shown using the σ symbol, with ® used for multiplication of elements.

5.3. Data Preprocessing

The data collected from the social media platform contain some error or may contain useless text, which causes difficulty in semantic analysis. As the dataset we are using is free from emojis, there is no need for emoji processing. Second, the stop words removal task is performed by using NLTK in Python. We can download the list of stop words, and stemming is used, which ignores stop words and creates systems by removing suffixes or prefixes that are used with the word. In this study, we use a snowball stemmer, which is different from a porter stemmer, as it allows performing multiple language stemmers. Tfidfvectorizer is used to tokenize the given document.

5.4. Performance Measure

The performance measurement of datasets is carried out with the help of a confusion matrix from where we get the results, which calculates the values of accuracy, precision, and recall with the help of positive and negative values of datasets. The formulas used for the calculation of values are

6. Result and Discussion

We have performed the experiments in this proposed work in three folds. Experiment 1 performs depression detection using the imbalanced datasets. In Experiment 2, the SMOTE technique of oversampling has been implemented to handle the class imbalance issue of the training dataset. Experiment 3 implements the RUS technique of undersampling for solving the imbalance issue in the dataset. The performance of the classifiers discussed above is analyzed and compared. The classification algorithms are analyzed using scikit-learn library, and we plotted the graphs using Matplotlib library.

Table 2 tabulates the results of the performance measurement with the imbalanced dataset. Table 3 lists the performance of the various classification models using the balanced dataset of SMOTE. Table 4 tabulates the results of the performance measurement of the learning models using the balanced dataset of RUS.

Table 2

Results of performance measures with imbalanced dataset.

Classifier	Precision	Recall	F1-score	Accuracy
Decision tree	0.73	0.63	0.67	0.03
SVM	0.76	0.67	0.71	0.52
KNN	0.67	0.56	0.61	0.65
LR	0.75	0.68	0.71	0.71
LSTM	0.79	0.72	0.74	0.78
Average	0.74	0.65	0.68	0.53

Table 3

Results of performance measures using SMOTE.

Classifier	Precision	Recall	F1-score	Accuracy
Decision tree	0.76	0.64	0.69	0.68
SVM	0.79	0.69	0.73	0.62
KNN	0.69	0.59	0.63	0.71
LR	0.77	0.72	0.74	0.76
LSTM	0.84	0.75	0.79	0.83
Average	0.77	0.67	0.71	0.72

Table 4

Result of performance measure using RUS.

Classifier	Precision	Recall	F1-score	Accuracy
DT	0.72	0.64	1.80	0.67
SVM	0.77	0.67	0.71	0.59
KNN	0.67	0.57	0.61	0.68
LR	0.75	0.69	0.07	0.72
LSTM	0.82	0.73	0.77	0.80
Average	0.74	0.66	0.79	0.69

From Table 2 it is noted that the average precision, recall, F1 score, and accuracy with the imbalanced dataset are 0.74, 0.65, 0.68, and 0.53, respectively, whereas the average precision, recall, F1 score, and accuracy with the balanced dataset using SMOTE are 0.77, 0.67, 0.71, and 0.72, respectively. There is a percentage increase of 4.05, 3.07, and 35.84 in the values of precision, recall, and accuracy, respectively. The results show that solving the class imbalance problem improves the performance of the psychological analysis model.

On comparison of Table 3 and Table 4, there is a percentage increase of 4.05, 1.51, and 4.34 in precision, recall, and accuracy, respectively, on using the SMOTE technique of oversampling in contrast with the RUS approach. The results illustrate that the SMOTE approach to handle class imbalance gives better results for depression detection.

From Tables 2–4, it can also be observed that the values of precision, recall, F1 score, and accuracy are high when LSTM is used as a learning model. The highest recall value of 0.75 is achieved using LSTM. Similar results with the accuracy value as high as 0.83 and the highest precision of 0.84 have been recorded with LSTM as the learning model. The observations mentioned above infer that LSTM outperforms the other baseline models in the depression detection approach.

6.1. Balanced Data

Figures 5–8 show the graphical representation of the performance measurement of precision, recall, F1score, and accuracy using the imbalanced dataset on the various leaning models. The graphs indicate that LSTM outperforms other baseline learning models in performance measurement.

Figure 5

The precision value of imbalanced dataset.

Figures 9–12 show the graphical representation of the performance measurement of precision, recall, F1score, and accuracy using the balanced dataset on the various leaning models after the application of the SMOTE oversampling technique. From the graphs, it can be clearly inferred that on resolving the class imbalance issue of the training dataset, the predictive performance of the models increases. The models are less biased and give more accurate results for depression detection.

Figure 9

The precision value of balanced dataset using SMOTE.

Figures 13–16 show the graphical representation of the performance measurement of precision, recall, F1 score, and accuracy using the balanced dataset on the various leaning models after the application of the RUS undersampling technique. The plotted graphs indicate that although RUS gave better results than the models trained with imbalanced datasets, it did not outperform the oversampling approach of class imbalance.

Figure 13

The precision value of balanced dataset using RUS.

From the above-mentioned statistical analysis and graphical interpretations, the answers to the research questions stated above are clearly provided as follows:

For RQ1: The solution of the class imbalance problem improves the performance of the psychological analysis model.

For RQ2: The SMOTE oversampling technique of class imbalance gives better results for depression detection using psychological analysis.

For RQ3: The LSTM classification model outperforms the other baseline models in the depression detection approach.

7. Conclusion and Future Scope

This study describes various techniques for psychological analysis for depression identification, including machine learning and lexicon (dictionary and corpus)-based approaches, with the goal of detecting depression using machine learning classifiers. We also investigated the classification of both balanced and imbalanced techniques. Machine learning algorithms such as decision trees, support vector machines, and LSTM show good accuracy, but a hybrid approach may provide better results in detecting mental health disorders than individual classifiers. Younger generations are quite active on social media, expressing their thoughts and opinions on a variety of topics and circumstances. Business firms use these platforms for their benefits, such as collecting customer reviews and complaints about services. These opinions help in identifying customer psychology (positive, negative, and neutral) toward products and any activity. The social networking platform has proved to be a great tool for identifying the mental state of social networking users. Because depression is a new problem in the society, examining social network data is helpful in detecting depression. This study discusses various obstacles in psychological analysis, but these issues will be addressed in the future as more psychological analysis research is conducted. The limitation to this research is that it can help in depression detection on social media. A large number of people of the world do not use social media because they find themselves uncomfortable on the social platform, which will remain undiagnosed.

For future work, we may identify the ability to estimate depressive disorder and daily activity, which leads to depression analysis with more accuracy. By merging LSTM and SVMs, we can create a hybrid model that can more accurately identify depression, benefiting people all across the world. As we know, LSTM can handle large datasets, and SVM performance in text classification for psychological analysis has greater accuracy; combining SVMs with LSTM may give better results for the imbalanced dataset as well as the balanced data set. In balanced datasets, SMOTE and RVS are used in the detection of depression. People may suffer from anxiety, depression, or suicide thoughts, and we can predict the same using AI and machine learning techniques. However, what about those who do not use social networking platforms? Developing a system that can identify depressive behavior in persons who do not use social media is a research objective in the future [20-24].