Prediction mechanism of depression tendency among college students under computer intelligent systems

2.1 Reasons for depression tendencies among college students

With the changes in society, social competition has become increasingly fierce, which has had a very negative impact on the mental health of college students. For example, because they are not sure about whether they can find a job with higher income and social status in the future, they have no sense of control over future events, which leads to the increased psychological pressure on some students. This also makes people pay close attention to the mental health of college students, and depression is a very common negative emotion [26,27]. There are many reasons that cause depression among college students. A survey was conducted on 1,000 students from University A, and two main causes of depression were identified (multiple choices are available). The specific investigation is shown in Table 1.

As shown in Table 1, a survey was conducted on 1,000 selected students, and it was found that they have nine main causes of depression among college students. Among them, the number of students who choose personal and psychological reasons is the highest, and most students believe that a person’s depression is mainly due to internal factors. For external factors, students believe that family reasons are the main external cause of depression, as well as interpersonal and academic pressures, which can also lead to depression in students.

2.2 Experimental materials

2.3 Data collection

The purpose of this study is to understand the current state of depression among students at University A and identify relevant influencing factors. This article proposes solutions based on influencing factors to reduce the overall depression rate of University A students, improve their learning quality, and ensure their mental health. Based on the 1,000 students selected in Table 1, this article expands the sample data. It randomly selected 3,750 students of all grades from University A for scale testing. A total of 3,620 questionnaires were collected, with a recovery rate of 96.53%. The specific situation is shown in Table 2.

As shown in Table 2, a total of 3,620 valid questionnaires were selected. Among them, the proportion of boys and girls is very close, ensuring the rationality of the research data in terms of gender and avoiding large-scale differences. In terms of grade, the number of freshmen is the highest, with 1,000 students accounting for 27.62%, while the number of sophomores is the lowest, with only 820 students accounting for 22.65%. For the selected student sources, the number of students from rural areas is slightly higher than that from urban areas. For the majors of students, the largest number is the accounting major, accounting for 28.18%, while the number of art majors and preschool education majors is relatively small, with a total of only 370 people. Regarding whether students are only children, the number of non only child students is 380 more.

The research object of this research is a student of University A. First, for the criteria of depression tendency and health, students can be divided into those with depression tendency and those with health through the established criteria. Based on the statistical results obtained from the school’s mental health survey scale, individuals who meet the requirements can be selected, and the experimental content and needs can be summarized to these individuals. Next, it would inquire about the students’ willingness to participate in the investigation, whether there is a reason to participate. If they are willing, they can sign an informed consent form and basic personal information registration form. At the same time, they would fill out BDI and SDS according to the requirements and finally conduct the experiment. The informed consent form is an important legal document in the medical process. It records the process by which the doctor informs the patient of medical information when the patient receives medical services, and the patient decides whether to receive medical services. Throughout the entire experiment, the subjects were alone in a quiet place using professional writing equipment. If there were any questions, the experimenter could help answer them. The specific framework is shown in Figure 1.

Figure 1

Data collection framework diagram.

As shown in Figure 1, this article collected data from students at University A and screened the learning criteria for inclusion. If it does not meet the criteria, it ends. If it does, it is determined whether the criteria for compliance are below BDI 13. If it is, it is divided into a healthy group; if it is not, it is divided into a depression-prone group. Next, these two groups can be numbered, and eligible students can sign informed consent forms to conduct depression tendency prediction experiments and obtain experimental results.

2.4 Feature extraction

Emotional features can be extracted, and then sentiment analysis models can be used to determine the sentiment of student scale texts. Emotion can be used to study students’ depressive tendencies, which takes into account students’ emotional demands and is not limited to identifying students’ language patterns. Considering that depression patients have negative emotions and symptoms, they are also more inclined to express negative emotions in social settings. Common ones include anxiety and tension, avoidance and withdrawal, low self-esteem and self-doubt, and anger and hostility. Therefore, emotions can be used to identify whether students suffer from depression. Therefore, emotional features are used to extract the depressive emotional features of University A students [31,32]. Using an emotion dictionary can effectively capture users’ emotional characteristics and judge students’ emotions. It is currently widely used in emotion analysis and has achieved good results. Based on the above considerations, sentiment distribution is used to analyze users’ language and text features, and sentiment features are evaluated based on sentiment dictionaries. In the feature quantification stage, emotional features Senta are added. The language structure of users is disassembled and refined, and the user’s emotion value, word frequency weight, and other information are calculated through feature quantification. The calculation formula is as follows:

where n represents the number of negative sentiment Weibo posts detected by the sentiment classification model, and N is the total number of Weibo posts posted by users.

2.5 Data

This study used a simple random sampling method for questionnaire distribution and reviewed the collected questionnaires to eliminate those that were not effectively utilized. The valid questionnaires obtained can be used for statistical calculations using Statistical Product and Service Solution 17.0 (SPSS17.0). At the same time, a dataset can be created and transferred to SPSS21.0 software for preliminary data preprocessing and calculation. The Cronbach’s alpha method was used to study the consistency of questionnaire data. In this study, the relevant techniques included in different levels of depression were clearly described through frequency and percentage.

The methods used in data analysis can be used to study the detection of depression symptoms among students of different grades and majors. It mainly uses BDI and SDS for detection, and these two measurement scales are used for depression monitoring. Scores greater than 50 indicate obvious depressive symptoms, scores between 30 and 49 indicate probable depressive symptoms, while scores below 29 indicate no depressive symptoms, as shown in Tables 3 and 4.

As shown in Table 3, the survey on depression among students of different grades should be divided into three types. It can be found that the number and proportion of students without depressive symptoms are the highest in their freshman year, while the number and proportion of students with obvious depressive symptoms are the lowest. As freshmen, they have just started school and are full of freshness in everything. At the same time, their academic pressure is relatively low, and they do not have to face financial pressure. Everyone has just started school, and interpersonal communication is also relatively less stressful. The number of students without depressive symptoms in their senior year is the lowest, and the proportion is also the lowest. At the same time, the number of students with obvious depressive symptoms is the highest, and the proportion is also the highest. Because the senior year is the graduation season, students need to face internships, enter society, and begin to bear pressure from society. At the same time, their academic focus on papers may also exacerbate stress and lead to depression among students. The proportion of sophomores who may have depressive symptoms is the highest, with 30.49%, while the number of people who may have depressive symptoms in their third year is the highest, with 265 people.

As shown in Table 4, a study on the depression situation of students in different majors reveals that the proportion of medical students without depression symptoms is low, only 10.71%. The proportion of people with obvious depressive symptoms is as high as 58.93%, which is also the highest. This is because medical students face high academic pressure and also face various experiments, which can easily lead to depression. The proportion of students majoring in business administration without depressive symptoms is the highest, with 52.08%. The proportion of civil engineering students with obvious depressive symptoms is the lowest, only 17.18%. The proportion of art students who may have depressive symptoms is the highest, with 50%. Students majoring in accounting may have the highest number of depression symptoms, with 376 people, while students majoring in education during the semester have the lowest number of obvious depression symptoms, with only 40 people.

3.1 Model design

Many methods are involved in computer intelligent systems, and Natural Language Processing (NLP) technology is chosen to study the prediction of depression tendency in college learning [33,34]. By using NLP to analyze students’ emotions and text content mining, students’ current emotional state and thinking mode can be clearly understood. This information can help predict whether students are depressed and providing interventions for school teachers to help avoid increased depressive symptoms. NLP technology can be used to explore important features in student language and improve the effectiveness of stress assessment. First, pre-trained text can be used to create word embedding vectors, and then the vector representation of words can be modified based on semantic content. Then, based on the concept of words and the characteristics of phonetics, the information of words can be embedded into an emotion dictionary to obtain fine-grained feature representations of words. This can detect the feature information with the highest correlation to the model, thereby obtaining the weights of words and phrases, thereby improving the accuracy of search results. The effectiveness of this method can be demonstrated through a large number of experiments in the end.

The first step is to convert the text into a sequence of words. Assuming H represents a text H = { v 1 , v 2 , v 3 , ⋯ , v m } in the data, where v i ( 1 ≤ i ≤ m ) represents the i th word in the text. For sequence H , for each word v i , its corresponding word vector is represented as w i v :

The input text data can be transformed into vector representations of words.

The foundation of NLP technology is to use contextual information to predict the appearance of target words. The formula is

where M t represents the word to be predicted, and j represents the window size.

Multiply the initial vector Y with the corresponding matrix M to obtain the result of the hidden vector:

Then multiply the hidden layer f 0 by the hidden layer matrix M ′ to obtain the hidden layer output:

During the training process, the training error is calculated through the loss function, and the matrices M and M ′ are iterated and optimized through backpropagation.

After preprocessing and word segmentation of text data, a depression candidate sentiment word set would be generated. At the same time, it is necessary to remove irrelevant useless words related to depression and stress and only retain useful words related to them. To this end, the term frequency-inverse document frequency (TF-IDF) weighting algorithm was used for word frequency weight analysis, resulting in a depression seed word set [35,36]. The formulas for calculating the frequency index used in TF-IDF are as follows:

In the above formulas, the feature word i belongs to the candidate emotion word set, and tf ( i ) represents the frequency of the feature word i in the total word count d of the text. i d f ( i ) represents the frequency of the inverse document, which is the degree of domain generalization of the word to the text. m represents the total number of text. Overall, the tf − i d f value of the i th word is determined by the frequency of its occurrence in the text and the domain generalization of the text.

By using TF-IDF to process the weighted algorithm of college students’ texts and extending the calculation of cosine similarity between words, the calculation formula for cosine similarity is as follows:

where θ refers to the angle between two words after vectorization; a i is the numerical value of the word i on the dimension a after vectorization representation; the cos value of the angle between two words indicates their similarity.

The seed word set filtered by the TF-IDF weighted algorithm belongs to a set of depression emotion dictionaries. Due to the incomplete generalization ability of subset domains, the seed word set also needs to be expanded. In order to improve the quality and efficiency of the dictionary, this article uses a basic sentiment dictionary to expand and support the depression seed word set on the basis of promises. Figure 2 shows the extension of the depression dictionary:

Figure 2

Expansion process of depression emotion dictionary.

By fusing TF-IDF word frequency features with text sentiment features, a text feature vector is obtained, and the input vector h i of this part is represented as:

where w tfidf represents the TF-IDF word frequency feature, and w rule represents the emotional value feature of the text.

The depression tendency detection module would fuse the W i and H i feature vectors obtained, represented as formulas (11) and (12).

where Mixed Feature represents the fused feature vector, and Fea represents the output vector obtained from the fused feature vector.

Then, input Fea into the Sigmoid layer for classification and obtain the final prediction Result , represented as:

3.2 Model construction and training

In order to evaluate the performance of the model, precision rate, recall rate, and F1 score values are mainly used for evaluation. Among them, the text of patients with depression is the positive sample size, while the normal text is the negative sample size. In general, accuracy and recall affect the overall performance of a model and are positively correlated with the overall performance of the model. The specific formulas are as follows:

where T P represents the number of correctly classified positive samples, F P represents the number of negative samples classified as positive samples, and F M represents the number of positive samples classified as negative samples.

Based on the collected data, a predictive model can be constructed, with depression tendency as the target variable and other features as input variables. Many methods are involved in computer intelligent systems for conducting experiments on extracted sample data using different methods. In addition to NLP, there are also NN, support vector machine (SVM), and other methods that use text to predict depression tendencies, such as recurrent neural network (RNN), long short term memory (LSTM), transformer model, etc. The results of the depression tendency prediction models constructed using these different methods can be compared with the results obtained from the research methods in this article. It is mainly compared from three aspects: precision rate, recall rate, and F1 score value, as shown in Table 5.

As shown in Table 5, regardless of the precision rate, recall rate, or F1 score value, the student depression tendency prediction model constructed based on NLP technology is higher than other student depression tendency prediction models. Among them, the model built based on the LSTM algorithm has the lowest precision rate, while the model based on the Transformer has the lowest recall rate, and the model based on the SVM algorithm has the lowest F1 score value. The model precision rate based on NLP technology is 8.2, 6.99, 7.43, 8.81, and 8.34% higher than that based on NN, SVM, RNN, LSTM, and Transformer, respectively. The recall rate of the model based on NLP technology is 6.16% higher than that of the model constructed based on the NN algorithm. In summary, the model constructed based on NLP technology has more accurate performance results in predicting students’ depressive tendencies through text.