Online learning is a network learning method with connectivity, accessibility, and flexibility [7]. For some, it offers the potential for learning for new audiences; for others, it offers the opportunity to change the dynamics of learning delivery and competition fundamentally [8]. With the development of online education, students, parents, and educators about online teaching satisfaction have been the focus of scholars around the world.

Before research into neural networks became hot, many scholars studied the factors affecting teaching satisfaction by various statistical methods and traditional machine learning methods. Some typical studies are as follows: Through regression analysis, Kuo et al. [9] explored the contribution of independent variables to student satisfaction and the effect of student background on dependent variables. The results showed that learner-instructor interaction, learner-content interaction, and Internet self-efficacy were good predictors of student satisfaction. Cole et al. [10] used three years to conduct an e-learning satisfaction survey on 553 students, and the results showed no statistically significant difference in satisfaction based on gender, age, or study level. “Convenience” and “Lack of interaction” are the two most influential factors on student satisfaction. Bolliger and Wasilik [11] developed and validated a tool that can be used to perceive teacher satisfaction in online learning. Eom et al. [12] used structural equation modeling to study the determinants of student satisfaction in the context of online courses, and the structural model results reveal that instructor feedback and learning style are significant predictors of learning outcomes. Gunawardena et al. [13] designed a hybrid method including qualitative and quantitative methods to explore the factors affecting employee satisfaction in a multinational company’s online education project. Lee [14] used factor analysis, structural equation modeling, independent sample t-test, and other techniques to investigate the potential differences in online learning acceptance and satisfaction between Korean and American students. Simpson [15] studied a sample of 157 students participating in online education and found that teaching content that passed the teacher review can improve student satisfaction. Roach et al. [16] believe that educators should always pay attention to students’ progress when conducting online education. Eom [17] analyzes the causes and consequences of teacher-student interaction in asynchronous learning courses and finds that curriculum structure, students’ self-motivation, and learning style all affect teacher-student interaction.

Wang et al. [18] used the structural equation model to explore the relationship between student characteristics, self-regulated learning, self-efficacy, and course results in an online learning environment. The model results found that students with previous online learning experience tend to have more effective learning strategies when taking online courses. In order to reduce the sense of isolation that online learners may experience, which in turn affects their satisfaction with online learning, McInnerney and Roberts [19] point out, the authors offer several suggestions:

1) more use of synchronous rather than asynchronous communication facilities.

2) adding a “Warm-up” phase to the curriculum structure.

3) focusing on communication in the teaching process.

With the maturity of neural network technology, various neural network models have been applied to academic research in education. Guo [6] uses the three-layer multilayer perceptron (MLP) models to build a prediction model of student’s curriculum satisfaction that is more accurate than linear regression. The study shows that the number of students enrolled in courses and the high distinction rate in the final exam is the two most important factors affecting course satisfaction. Kardan et al. [20] developed a neural network model for student satisfaction and course selection, and the results show that the model is superior to three famous machine learning algorithms and two previous naive algorithms.

Aydogdu [21] used a neural network model to study the effects of gender, course time, and homework completion on students’ final course scores. The prediction accuracy was 80.47%. Agaoglu [22] uses artificial neural networks to mine teacher performance data. Yukselturk et al. [23] use a genetic algorithm-based feature selection algorithm and a neural network to predict who will drop out of school during online education.

Internationally, online education is seen as an opportunity to improve primary education in rural areas. A large body of literature discusses the contribution of different online education models to narrowing the socio-economic gap between developing and developed regions.

A survey by Jinqiu et al. [24] shows that communities in northwest China have benefited enormously from Internet and communications technology programs. Zhang et al. [25] designed and used scaffold strategies to conduct a cross-regional collaborative learning activity among secondary schools. The study found that systematic scaffold strategies were essential for students to have compelling cross-regional online collaborative learning experiences. Wang and Zhao [26] summed up the achievements of fundamental education reform in rural China, analyzed the problems existing in the reform, and put forward countermeasures to solve these problems. Research by Wei et al. [27] has shown a significant positive relationship between education and income in rural areas.

Several departments designed the questionnaire for teachers. First, three university educators and three provincial education department staff designed the first draft of the questionnaire. This first draft was then distributed to 30 teachers from different schools. Feedback and suggestions were received from these thirty teachers so that the questionnaire could be revised and adjusted. As shown in Table 1, the final teacher questionnaire consisted of twenty questions, which can be grouped into four categories as follows.

The target group for this study was primary, middle, and high school teachers in a Chinese province, covering all stages of K-12 education. The questionnaire was used to investigate the target group’s participation in online education during the COVID-19 outbreak. A questionnaire was anonymous and voluntary to ensure the confidentiality and reliability of the data and the privacy of the respondents. The questionnaire was released online on 23 March 2020, and 498,481 questionnaires were returned as of 4 April 2020.

In order to reduce the influence of dirty data on the results and to ensure the validity of the experimental results, we performed a clean data filter. The simplest and most effective way is to analyze the set of values for individual features. Assuming that x is a feature with continuous values and that x * is a valid value in that feature, it should satisfy the following relationship.

x ¯ − 3 σ ≤ x * ≤ 3 σ + x ¯ ,   x ¯ = ∑ i = 1 n x i n ,   σ = ∑ i = 1 n ( x i − x ¯ ) 2 n . (1)

In addition, we believe that the span of grades taught by the same teacher should not exceed 6. If a teacher teaches both Year 1 and Year 7 students, we would consider this figure unreasonable. Similarly, we believe that the total number of subjects taught by the same teacher should not exceed three. If a teacher teaches four subjects simultaneously—language, mathematics, English, and physics—we do not consider the data reasonable.

After data cleaning, the statistics showed that the number of valid questionnaires in the teachers’ questionnaire was 493,747, with an effective rate of 99.05%.

As shown in Table 2, in three age ranges of 29 and below, 40 to 49, and 60 and above, the proportions of rural and urban teachers are similar. In contrast, in the age range of 30 to 39, the proportion of teachers in urban areas is higher than in rural areas, and in the age range of 50 to 59, the proportion of urban teachers is lower than the overall proportion, while the proportion of rural teachers is higher than the overall proportion.

One of the most critical factors hindering the development of primary education in Chinese townships is the lack of quality teachers [28]. Looking at rural teachers compared to urban teachers in terms of degree and age, the proportion of teachers with a Ph.D. is smaller in both rural and urban areas, but of the proportion of teachers with a master’s degree, the proportion of rural teachers is only one-third that of urban teachers. In addition, the proportion of teachers

with a bachelor’s degree in townships is nearly ten percentage points lower than that of urban teachers.

In urban schools, nearly 90% of teachers teach only one course during their employment, while this proportion is only 76.91% of teachers in rural areas. Moreover, we can see from the table that the proportion of rural teachers who teach two or three courses simultaneously is about double that of urban teachers.

In township schools, the shortage of teachers in traditional subjects (such as Chinese and mathematics) is not apparent, but there is an increasing lack of talents to teach modern subjects, such as English, computer science, music, and physical education [28]. As a result, a teacher often has to teach more than one course, even if he has not mastered some of them.

From Table 3, both urban and rural teachers are concerned that student behavior is challenging to monitor and teaching is ineffective when teaching online. In addition, rural teachers differ from urban teachers in the proportional distribution of the following four options: students are not equipped to teach online, parents do not support and cooperate with online teaching, insufficient capacity to teach online, and immature conditions for teaching online.

The Chi-square test showed significant association between school districts and teachers’ worries, χ 2 ( N = 493743 ) = 201.003 , p < 0.01 . The Cramer’s V is 0.08 which suggests a small effect [29].

The implementation of online education has put forward higher requirements for teachers who can master and apply modern educational technology. However, many teachers in villages and towns are old teachers, most of them can’t use electronic devices to teach, which seriously restricts the further development of

online education in villages and towns [30].

Interestingly, there was not much difference in the challenges encountered by urban and rural teachers in online classes during the epidemic, as can be seen in Table 4. χ 2 ( N = 493743 ) = 39.211 , p < 0.01 . The Cramer’s V is 0.009 which suggests a small effect [29].

When teaching online, teachers have a range of problems such as difficulty monitoring students’ learning behavior, distracted students, inconvenient teacher-student communication, and less effective teaching than in a physical classroom. These problems may be related to the medium of learning used by students. The most portable internet tool available to us in our daily lives is the smartphone. The ubiquity, versatility and connectivity of smartphones provide a new and potentially powerful online learning environment [31]. For younger students, the games on the smartphone may be more attractive than the teacher’s teaching content, thereby distracting them. In addition, the small screen of a smartphone makes it more suitable for reading short texts quickly, rather than lengthy materials [32]. As a result, students’ prolonged use of small-screen mobile phones for learning can lead to eye strain and poor concentration, affecting the effectiveness of online classes.

Besides having essential online teaching functions for a well-established online teaching platform, it should also support functions such as assignment posting, in-class questioning, and student management. These functions are implemented on mainstream platforms, so there is not much difference in the choice of rural teachers and urban teachers on this issue. χ 2 ( N = 493743 ) = 277.494 ,

p < 0.01 . The Cramer’s V is 0.024 which suggests a small effect [29].

As illustrated in Table 5, Compared to urban teachers, rural teachers are more concerned with the following factors when choosing an online teaching platform, such as whether the platform has rich course resources, whether the content of the platform’s courses is of higher quality, whether the platform’s courses meet current teaching needs, whether the platform’s teaching methods are flexible, and whether the platform’s teaching resources are more easily accessible. However, urban teachers are more concerned about the platform’s interactivity and whether it can provide a relaxed and interactive environment.

Table 6 above shows the similarities between rural and urban teachers on factors that influence online education. Both urban and rural teachers agree that the two most important factors affecting online education are the students’ ability to learn independently and whether their parents help support them.

Similarly, as shown in Table 4, when asked about the difficulties encountered in online teaching, both urban and rural teachers agreed that one of the imperfect aspects of online education was the inability to monitor students effectively.

In addition, if the government is to implement online education on a large scale, it should have supporting hardware and software facilities, and teachers involved in online education need to have some ability to use multimedia equipment for teaching.

As can be seen from Table 7 below, for the choice of teaching method after the epidemic, more than half of the rural teachers opted for the previous offline teaching mode. Similarly, about 46% of urban teachers chose to revert to their

previous mode of teaching. It shows that the time is not fully ripe for the implementation of online education.

For the option “Combine online and offline teaching”, urban teachers are about seven percentage points more likely than rural teachers. Therefore, we suggest that urban teachers are more open to online education than rural teachers.

Teachers who chose to adopt a fully online teaching mode after the epidemic ended represent only a small proportion of teachers in towns and villages.

The statistical analysis above shows the differences between urban and rural teachers when teaching online, but it would make more sense to use neural networks if we had a large amount of data. Since the emergence of online education, many scholars have researched teachers’ and students’ satisfaction in online education, primarily using traditional machine learning methods such as logistic regression and random forest. Considering the current technological hotspots and the size of our data, we decided to use a neural network model to analyze teacher satisfaction.

Since the creation of Alexnet [33] in 2012, the number of layers in artificial neural networks has been moving more profoundly, and it is intuitively assumed that as the number of layers increases, the network’s ability to fit the features will improve. Nevertheless, it was not to be. It has been found that the accuracy of the model does not constantly improve with increasing network depth.

The residual neural network was proposed by Kaiming He [34] and others in 2015. Compared with traditional neural networks, deep residual neural networks add a jump layer connection structure to the forward propagation, which effectively improves the performance of deep neural networks and enhances the feature fitting ability of deep neural networks. Given the excellent performance of the residual structure, we also decided to use the residual structure when building our neural network.

The left-hand side of Figure 1 shows the bottleneck, a component of the residual neural network, which can be defined as:

y = F ( x , W i ) + x , (2)

where x and y are the input and output vectors of the layers to be considered, W_i is the weight in the weight matrix, and F represents the residual mapping to be learned. For an example that has three layers, its residual mapping function is as follows:

F = W 3 ⋅ g ( W 2 ⋅ g ( W 1 x + b 1 ) + b 2 ) + b 3 , (3)

g ( x ) = max ( x , 0 ) , (4)

where g denotes the Rectified Linear Unit activation function, and W₁, W₂, W₃, b₁, b₂, and b₃ are the weights and biases of the first layer, the second layer and the third layer, respectively.

Deep residual neural networks are generally used to process high-latitude image data, and questionnaire data containing just over 100 features is straightforward compared to image data. Therefore, we choose the classical residual neural network structure with the least number of neural network layers, the 18-layer network structure, as our template, and we will further improve on the 18-layer residual neural network.

However, even if it is the 18-layers residual neural network with the smallest number of layers, each of its layers contains four 3 × 3 convolutional layers to achieve feature extraction. Compared with image data, our questionnaire data has fewer features, and continuous convolution may lead to the loss of essential features. Therefore, we need to modify the traditional 18-layers residual neural network. We used a 1 × 1 convolutional layer and two bottleneck structures to form the building block (Figure 1. Centre) of our neural network. We then implemented our own network (Figure 1. Right) using building blocks instead of the two basic block structures for each layer of the classic 18-layer ResNet. It significantly weakens the feature extraction capability of the traditional 18-layer residual neural network, thereby retaining essential features.

In our neural network, Rectified Linear Unit (ReLU) is used as the activation function. The output neuron of the j_th layer can be expressed as:

O j = ReLU ( S j ) = max ( 0 , W j T ⋅ x j + b ) , (5)

where w and b denote the weight and bias, respectively. Use real numbers as parameters and return real values within [ 0 , + ∞ ) . For the output layer, use the softmax function as the activation function. Similar to the above, the output neuron of the j_th layer can be expressed as:

O j = σ ( S j ) = e S j ∑ k = 1 m e S k , (6)

m is the number of output neurons. In our neural network structure, the value of m is 3. The softmax function takes real numbers as parameters and maps them to real values between 0 and 1, and makes the sum equal to 1. Since the sum of the output is 1, the softmax layer can be regarded as a probability distribution, and the O_j value can be interpreted as the estimated probability of the input classification by the network.

We use the commonly used Adam as the optimizer during model training. When training the model, momentum and adaptive learning rate can be used to speed up the convergence speed. We choose Categorical Cross entropy as the loss function. Cross entropy can be used to evaluate the difference between the probability distribution obtained by the current training and the true distribution. It describes the distance between the actual output and the expected output, that is, the smaller the value of cross entropy, the closer the two probability distributions are.

After the basic neural network framework has been built, the question arises as to how to process the data to meet the input requirements of the neural network. The teacher questionnaire consists of 20 questions, including 8 single-choice and 12 multiple-choice questions. Each option for a single-choice question is represented by a number that has a practical meaning, for example, for question 4: the highest qualification obtained by the teacher. The number 1 represents a doctorate, the number 2 a master’s degree, the number 3 a bachelor’s degree, and the number 4 a college degree. We chose the answer to each multiple-choice question as one of our independent variables. We use 0 and 1 to indicate whether the teacher selected multiple-choice options when completing the questionnaire. If this option was selected, we use 1 to indicate this, and if not, we use 0. We, therefore, use a vector of length the number of multiple-choice options to represent the results of the multiple-choice questions.

It is worth mentioning that in multiple-choice questions, there are some meaningless options, such as “other,” and after removing some of these options, we combine all the independent variables to form a 117 × 1 vector data. To make the data satisfy the input conditions of our neural network, we added four zeros to the end of the original data and adjusted them to an 11 × 11 numerical matrix. We then greyed out the digital matrix to obtain a grey-scale image, the input data for our neural network.

We chose teacher satisfaction with online education as the dependent variable. Teacher satisfaction with online education is a three-category variable with three values. Zero indicates that teachers are satisfied with online education, one indicates that teachers are generally satisfied with online education, and two indicates that teachers are dissatisfied with online education. We use this value as the output data of the neural network.

The 18-layer residual neural network is the classical residual neural network structure. However, it is unknown whether an 18-layer residual neural network with modified essential components of the neural network can still maintain high prediction accuracy. In order to verify the rationality of our neural network structure, we conducted a comparison experiment with the traditional 18-layers residual neural network.

We conducted experiments with a data volume of 30,000, and compared two neural networks in terms of prediction accuracy and training time, and set the number of iterations to 128.

As can be seen from Table 8, our neural network model is more accurate than ResNet-18 in terms of prediction and our neural network requires less training time.

After comparing the impact of different layer neural network structures on our dataset, we compared the performance of our neural network with traditional machine learning algorithms (Logistic Regression, SVM, Naive Bayesian, Decision Tree, Random Forest) on three datasets with different data sizes of 50,000, 150,000 and 450,000.

In addition, we performed a cross-sectional comparison of the models, replicating the artificial neural network model from the Bang Won Seok paper [35] (named ANN) and the backpropagation neural network model from the Tinggui [36] Chen paper (named BP), and experimented with them on our dataset. The experimental results are shown in Table 9.

True positive (TP), false negative (FN), false positive (FP), and true negative (TN) are the four kinds of outcomes (TN). TP denotes successfully projected positive group samples, FN denotes a positive group that was wrongly forecasted as unfavorable, FP denotes an antagonistic group that was incorrectly predicted as positive, and TN denotes a correctly predicted negative group. We utilize the following assessment criteria to forecast the performance of our approach and compare it to other ways based on these indices. The proportion of the correct number of positive samples in the overall number of positive samples computed by the classifier is referred to as precision.

Precision i = TP i TP i + FP i . (7)

The proportion of the right number of positive samples in the total number of positive samples is referred to as recall.

Recall i = TP i TP i + FN i . (8)

The harmonic mean of accuracy and recall is used to get the F1 score.

F 1 i = 2 ⋅ Precision i ⋅ Recall i Precision i + Recall i . (9)

In our dataset, the output is triple classification rather than binary classification,

so we need to calculate Precision, Recall and F1 separately for each output type and then average them.

The experimental results in the Tableshow that our neural network model has a higher prediction accuracy than the traditional machine learning model for data amounts of 50,000, 150,000, and 450,000. For 150,000 and 450,000, our neural network model performs well in Precision, Recall, and F1-Score, and is far better than other traditional machine learning models.

In the previous experiment, we built a neural network model and used several features to predict teacher satisfaction accurately. The black-box nature of neural networks makes it difficult to analyze which features play a key role in the decision-making process of neural networks. We changed the representation of the data before feeding it into the neural network, which allowed us to label the key features in the decision-making process of the neural network by means of a heat map, via the Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm [37]. For every target concept, Grad-CAM employs gradients of that concept to build a coarse localization map that highlights the locations in the picture where the concept may be predicted.

In order to obtain the location mapping for the classification discriminations of class c, we first compute the gradient of the score for class c, y c (before the

softmax), with respect to feature maps A k of a convolutional layer, i.e. ∂ y c ∂ A k .

Neuron significance weights ϖ k c are derived from these gradients by averaging the global-average-pooled values:

ϖ k c = 1 z ∑ ∑ ∂ y c ∂ A k . (10)

We combine forward activation maps with a ReLU and weight them to get,

L Grad-CAM c = ReLU ( ∑ k ϖ k c ⋅ A i j k ) . (11)

The ReLU function is used to remove the influence of negative values on the feature map on the classification results, and the final classification task is expressed by Equation (11).

We integrate the Grad-CAM algorithm into our overall process to obtain the reasons for the decisions made by the neural network model for each input data. Figure 2 illustrates our complete work, including data collection, filtering, building the neural network model, and using the Grad-CAM algorithm to label the essential features of the neural network model for making decisions.

The Grad-CAM technique makes our model more transparent by generating a visual interpretation of which regions play an essential role in the decision-making of the neural network. We, therefore, selected six individuals from the data of teachers successfully predicted by the neural network model: three teachers who were satisfied, generally satisfied, and dissatisfied with online education in the urban teacher, and three teachers who were satisfied, generally satisfied, and dissatisfied with online education in the rural teacher. The data from the six teacher samples were processed using Grad-CAM technology, and Figure 3 below shows the results of our experiment.

Among the teachers who were satisfied with online education, features 46, 47, 48, 57, 58, and 59 in the urban teachers’ data played an essential role in the decision-making process of the neural network. By reviewing the questionnaire, we found that these features correspond to the usability of the online education platform and the abundance of online course resources, respectively. Features 82, 83, 84 and 93, 94 and 95 in the data on rural teachers play an essential role in the decision-making process of the neural network. The practical implication of these features is that online education improves the ability of students to use digital devices.

Of those teachers who were generally satisfied with online education, features 84, 85, 86 and 95, 96 and 97 in the urban teacher data played a significant role in the decision-making process of the neural network, with the corresponding questions in the questionnaire being the impact of online education on students’ expression and communication respectively. Features 2 and 3 in the township teacher data significantly influenced the decision-making of the neural network, corresponding to the questions in the questionnaire about the teacher’s age, educational background, and the subject taught.

Some of the teachers were dissatisfied with this online education practice. By analyzing the heat map, we can learn that features 92, 103, 111, and 112 in the urban teachers’ data played an essential role in the decision-making process of the neural network, and these features correspond to problems such as poor online education in the questionnaire, respectively. Features 24, 25, 35, 36, 46, and 47 of the rural teacher data play an essential role in the decision-making process of the neural network, and these features correspond to issues such as inadequate

online facilities in the questionnaire.