3.1

Analysis of the research status of machine learning

Machine learning is an important method for data prediction and data decision analysis through feature extraction and feature training. It can extract features of IDT, analyze the weight of multiple factors of IDT, and effectively analyze IDT. As one of the most important methods, DBN training has important application value. The network model is initialized by Restricted Boltzmann Machine (RBM), and then the initialization speed of the model is improved by this method. Using the backpropagation algorithm to study the data supervised. In this process, it optimized the parameters and proposed the expected model^10-13.

In order to get network parameters, we use an unsupervised greedy algorithm to pre-train the network model. First, the first layer RBM1 is trained, a vector is generated in the visual layer of RBM1, and the generated value is transmitted to the hidden layer; Then, Thirusangu et al. reconstructed the input signal of the visual layer¹⁴, and the data of the visual layer is randomly extracted and transmitted to the hidden layer. It trains each layer by using the maximum function learning method and combining deep learning network and training mechanism. So we need less training time.

The weights of IDT are optimized and adjusted¹⁵, after the DBN pre-training is completed, the feature data of IDT extracted layer by layer from the input multi-layer RBM is substituted into the BP neural network. With the help of label data, the back propagation algorithm is used to optimize and adjust the weight extraction of IDT. After training the DBN, the initial weight w₁, w₂,.., w_n of the network is obtained:

In the formula, θ={w₁, w₂,.., w_n, w} and θ are the weights to be fine-tuned, y_i is the label of the sample, and y_i’ is the prediction feature parameters in the model. Input the data sample of IDT with marked information into BP neural network, use BP algorithm to send back the error to the whole network model, fine-tune the parameters of (DBN) in each layer, and after constantly adjusting the parameters, the weight prediction value of IDT is obtained.

3.2

Construction of MFWM

In this paper, we use the MFWM of IDT to analyze the characteristics of IDT. In this paper, we use the deep learning method to carry out the Boltzmann machine variant of the IDT processing unit, and obtain the MFWM of two-way connection IDT^{16, 17}, as shown in Figure 1.

Figure 1.

MFWM of IDT.

This paper reconstructs the data of Internet addiction according to the MFWM of IDT and applies it to unsupervised learning for regression or classification of IDT data. In addition, RBM can be used as a generic model. Through this MFWM of IDT, the following analysis is carried out.

4.1

Deep belief networks

In this paper, the feature mining method is used to extract the data neurons of students’ Internet addiction, and the complexity features of Internet addiction data in each layer are obtained. Multiple feature maps are used to obtain the feature synaptic weight set of Internet addiction data under the constraint of neurons¹⁸. After the local average, the feature map resolution of each convolution layer is obtained, finally, we output the trend characteristics of Internet addiction data after various slight changes, and the specific deep network structure is shown in Figure 2.

Figure 2.

Deep learning network structure of Internet addiction data.

This paper introduces a deep learning model, which can extract a large number of data features of Internet addiction, provide a lot of data for analysis and prediction of Internet addiction, improve the knowledge spectrum of data characteristics of Internet addiction, obtain more accurate data fitting ability of Internet addiction, and obtain the global optimal solution by repeated training, The traditional method can easily generate local optimal problem when predicting the multi factor weight of the tendency of Internet addiction data¹⁹. Without relying on the data label of sample Internet addiction in knowledge spectrum, the accuracy of prediction of the characteristics of Internet addiction data can be improved.

4.2

Convolution feature extraction of internet addiction data

In order to improve the effect of MFWM analysis of IDT, this paper uses a convolution feature to extract Internet addiction data, improves the Convolutional Neural Network (CNN) model through data transformation, and obtains structural variants of this CNN model. The structure of this convolution feature extraction process of Internet addiction data is shown in Figure 3.

Figure 3.

Convolution feature extraction structure of Internet addiction data.

When x_i ∈ R^k, x_i represents the weight vector of multi factor k -dimension of IDT, which is related to the k -th eigenvector in the data of Internet addiction. At this time, for the feature vector with n, the length of the characteristic data of the Internet addiction data is expressed as:

In equation (2), ⊕ is the connection operator of IDT data. X_i : i + j is used to show the connection of X_i, X_i+1,…, X_i+j data of addiction tendency. w ∈ R^hk represents the convolution operation of IDT data. At this time, the convolution kernel can extract the weights of X_i : i + h –1 features in IDT data at the same time²⁰. For example, the feature ci of IDT data can be obtained from the feature vector window h. The calculation formula of the feature of IDT data is as follows:

In equation (3), b ∈ R, b is the bias value of IDT data, and f is the nonlinear function of IDT weight. In this paper, convolution kernel is used to realize feature mapping set on feature security margin [X₁ : h, X₂ : h + 1,…, Х_n–h+1 : n] of IDT data:

Among them, c ∈ R^n–h+1. The feature mapping of Internet addiction data is carried out by maximum pooling, and the maximum weight max{c} of addiction data is selected as the best correlation feature to obtain the important weight features in the feature mapping of IDT.

In order to improve the convolution feature extraction rate of addiction data, the Gated Recurrent Unit (GRU) unit is used to determine the memory direction of Internet addiction information flow. The unit structure of Internet addiction data characteristic GRU is shown in Figure 4.

Figure 4.

GRU unit structure.

t represents the time activation value of Internet addiction data extraction, represents the candidate activation value of Internet addiction data, represents the previous activation value of Internet addiction data, and represents the previous candidate activation value of Internet addiction data:

The update gate determines how far to update its activation value or content. The calculation formula of the update gate is shown in equation (6):

This process uses a linear sum between the existing state and the new calculated state. The calculation method of this candidate state is similar to the traditional RNN cycle unit, and the calculation method is shown in equation (7):

In the formula, r_t is the reset gate component and x_t is the multiplication of elements. When is close to 0, the reset gate can effectively make the unit read the first symbol of the input sequence and allow it to forget the previous state.

The calculation method of the reset gate is similar to that of the update gate, and the calculation method is shown in equation (8):

In this paper, a two-layer recurrent neural network is used to process the multi factor characteristics of IDT. Each unit will have 128 GRU units, among which the size of the hidden IDT state data is 128. By concatenating the output features of IDT factors, a 100 * 128 matrix of IDT factors is obtained. At this time, this 100 * 128 matrix is the result of convolution feature extraction of IDT factors.

5.1

Experimental design

In order to verify the effectiveness of different methods, this paper adopts ReHo based analysis method of College Students’ Internet addiction (Method 1)⁷, VMHC based analysis method of adolescent Internet addiction (Method 2)⁸, improved Analytic Hierarchy Process (AHP) based analysis method of College Students’ Internet addiction behavior (Method 3)⁹ and this method to predict the accuracy of IDT, the effectiveness of different methods is verified by the prediction time and recall rate of IDT, the experimental analysis is carried out in Matlab environment, and the running system is windows8. Table 1 show the experimental parameter settings:

Table 1.

hardware and software environment configuration.

According to the configuration parameters of software and hardware environment in Table 1, the specific experimental results are as follows:

5.2

Experimental result

(1) Prediction accuracy of IDT

Methods 1-3 and this paper are used to test the prediction accuracy of IDT, and Table 2 show prediction accuracy.

Table 2.

Prediction accuracy of IDT under different methods.

There are differences in the prediction accuracy of IDT in Table 2. When the value of iterations is 15, the prediction accuracy of Method 1 is 38%, Method 2 is 69%, Method 3 is 66%, and this method is 89%. When the iteration is 30, the prediction accuracy of Method 1 is 60%, that of Method 2 is 56%, that of Method 3 is 55%, and that of this paper is 96%. The prediction accuracy of this new method is much better.

(2) Prediction time of IDT

The time-consuming test experiment of prediction of IDT is carried out based on Methods 1-3, and the proposed method. Table 3 shows the results.

Table 3.

Prediction time of IDT by different method (unit: s).

When the times of iterations is 5, the prediction time of IDT of Method 1 is 39s, the prediction time of IDT of Methods 2 is 52s, the prediction time of IDT of Methods 3 is 66s, and the prediction time of IDT of this method is only 2S. When the times is 20, the prediction time of IDT of Methods 1 is 58s, the prediction time of IDT of Methods 2 is 59s, the prediction time of IDT of Methods 3 is 62s, and the prediction time of IDT of this method is only 1s. This method needs far less time than others.

(3) IDT prediction recall rate

The paper uses Methods 1-3 and this paper to get the recall rate of IDT prediction. The results are shown in Figure 5.

Figure 5.

Recall rate of IDT Prediction.

Figure 5 shows different recall rates of IDT prediction. When the experimental population is 400, the recall rate of IDT Prediction of Methods 1 can reach 40%, and 61% for Methods 2, 27% for Methods 3, 93% for this paper. When the experimental population is 600 people, the recall rate of IDT Prediction of Methods 1 can reach 63%, and 69% for Methods 2, 46% for Methods 3, 95% for this paper. So, wo can know this method is better for recall rate.

(4) IDT coverage

Based on the user’s test, it determines the coverage according to the comparison between the number of covered users and the remaining users in the evaluation process, calculates the algorithm coverage, and determines whether the algorithm is effective. The calculation formula of coverage is as follows:

In the formula, H_h represents the recommended list; A_a represents the user set; B_b represents a collection of items. The comparative analysis results under different methods are shown in Figure 6.

Figure 6.

Comparison results of coverage.

According to Figure 6, under the condition of no errors and exceptions, the maximum coverage of Methods 1 is 86%, 75% for of Methods 2, 64% for of Methods 3, 92% for this paper. Through comparative analysis, we can find that the coverage in this paper is the best, indicating that the coverage is smooth and complete.

According to the result below, anyway, this method in this paper has better performance for prediction accuracy, prediction time, recall rate and coverage.

6.1

Strengthening ideological and political education of internet addiction

We should carry out ideological intervention on people’s Internet addiction behavior and give full play to the educational role of Ideological and political education. By Focusing on improving the pertinence of Ideological and political education, and explaining the harm of Internet addiction with facts, it can enable people to bravely face the pressure and difficulties of life, transform fear, anxiety and other emotions into internal conscious power. And it also enable people to correctly view and use the network, enhance their ability to distinguish bad information, and maintain network order and abide by network rules in the scene of no monitoring. On the other hand, we should combine network ideological and political education with socialist education and patriotism education, cultivate people’s conscious network moral responsibility, strengthen the ability to identify network cultural values, timely correct the external factors closely related to Internet addiction, and describe the cases of Internet addicts, so as to enable people to have correct values and network cognition, to regard the Internet as a learning and leisure tool to avoid excessive use of the Internet and affecting their normal life. Relevant educators should pay attention to ideological and political education, build a good interactive relationship with internet addicts, give full play to their initiative, and ensure the long-term effectiveness of ideological and political education for Internet addicts. When carrying out ideological intervention on Internet addicts, we should also deeply explore the psychological problems of Internet Dependence addiction, understand people’s family life, temperament, interests and hobbies and learning life through psychological counseling and humanistic care, exchange ideas with people, dredge their internal feelings, promote the ideological transformation of Internet addicts, and guide them to establish correct three outlooks, So that people can actively face real life, change their original ideas, fully understand the ideological misunderstanding of Internet Dependence and addiction, and help people gradually reduce the number and time of surfing the Internet.

6.2

Building community monitoring and management mechanism

We should take the community as a unit, build a community monitoring and management mechanism. Community is an important carrier of people’s life, and the environment and people’s development play an important role. Therefore, it is very important to establish a community monitoring and management mechanism. The monitoring and management mechanism needs the support of society and family. We should always pay attention to people’s network psychology and network behavior, and give help, care and emotional support to people who depend on the network, and actively exchange ideas and communicate with them, and timely curb the tendency of Internet Dependence and addiction. To persuade and monitor the Internet addicts, master the subsequent use of the Internet, and timely stifle the tendency of Internet addiction. Community monitoring and management also has strong permeability. The Internet addicts will continue to be influenced by the people around them in the community, and people can communicate with each other. This mechanism is not easy to arouse people’s disgust, so that the intervention effect can be brought into full play. The monitoring and management mechanism should also give play to the monitoring role of the party and league organizations, constantly improve the code of conduct and articles of association of the party and League organizations, make them feedback on the network status of Internet addicts, standardize people’s Internet behavior and avoid them falling into the state of Internet addiction. The party and League organizations should actively carry out Internet themed education activities, and people’s hobbies and ages are relatively close. The degree of support and understanding between organizations is high, and there are no obstacles to communicate with each other. While enriching people’s daily life, it can also divert excessive attention to the network, and strengthen the pertinence and uniqueness of Internet addiction monitoring through the organization’s self-education function.

6.3

Improving the legal system and purify the source of the network

We should improve the construction of relevant laws, provide institutional guarantee for the network information industry and purify the source of network information. And we also fully implement the network real name system, protect the security and privacy of users’ identity information, put an end to bad information arbitrarily released by anonymous means, detect bad information on the network through the semantic understanding of network information, and then implement network filtering and shielding technology to filter and shield bad information on the network. The government should continue to give environmental, policy and financial support. And based on the original technology, they should further develop relevant technologies, and provide filtering and shielding technology to schools, stations, public libraries, communities, Internet cafe operators and other public fields. They also should establish a network bad information reporting platform, create a network information supervision and reporting mechanism to continuously improve the supervision linkage mechanism dominated by network police while applying network filtering and shielding technology. By the end, they should strengthen the organization and coordination ability, actively give play to the supervision role of network self-discipline organizations and network volunteers, and standardize the benign network order. The reporting platform should strengthen the punishment of bad information, ensure the importance of reporting processing efficiency, set up special boxes and columns for reporting network bad information, strictly abide by the reporting acceptance, answering and feedback mechanism, and make the reporting process public, regularly hold special rectification of network bad information, and target individuals and organizations that publish a large number of bad information. So that they cannot establish websites all their life, reduce the negative impact of bad information, at the same time, implement the high tax system of Internet cafes, so that people can only surf the Internet at home or school, restrain people’s irrational demand for Internet cafes, and provide people with a good network environment.

7.1

Conclusion

The network has broadened people’s horizons, enabling them to enjoy the unprecedented convenience and rich life brought by high technology. At the same time, they have also been affected and impacted by many negative aspects such as thinking methods, values, behavior patterns and personal growth. Therefore, how to use the network reasonably and how to surf the Internet healthily is an urgent problem to be studied at present. Internet use is divided into normal Internet use, excessive Internet use, IDT and Internet addiction. Therefore, it is considered that the phenomenon of addiction in the process of Internet use does not occur suddenly, but often has a continuous process. For example, it may be normal to use the Internet at the beginning of use, and then there is excessive Internet use, followed by the tendency of Internet addiction, and finally there is Internet addiction, that is, it is a continuous pedigree process from healthy Internet access to Internet addiction. The normal use of the Internet is called healthy Internet behavior, and the excessive use of the Internet, IDT and Internet addiction is collectively referred to as unhealthy Internet behavior.

This paper proposes a multi factor weight model analysis method of IDT based on the machine learning method. By constructing the MFWM of IDT, The DBN is used to realize the weight sharing of IDT factors, extract the convolution characteristics of IDT factors, and realize the analysis of IDT factors weight model. This study provides guidance, reference and suggestions for people to make rational use of the Internet and intervene Internet addiction behavior from the aspect of psychology and behavior. The results are as follows: The prediction efficiency of IDT of this method is high, which can effectively predict IDT, and the recall rate of IDT prediction is high, and the coverage of this method is closest to the ideal state, which shows that the prediction effect of IDT of this method is better.

7.2

Prospect