Analysis of news transmission mode based on fuzzy data classification and neural network simulation

Abstract

Online news media websites and mobile news applications can report hot events in real time. With the change of news duration, we urgently need a tool that can automatically extract hot events from massive data and show how events change dynamically with time. In this paper, the authors analyze the news transmission mode based on fuzzy data classification and neural network simulation. Due to the limitation of BP neural network algorithm, there are some problems in the prediction of information transmission. Therefore, we use fuzzy algorithm to optimize BP neural network, which is easy to fall into local minimum and slow convergence speed, so that BP neural network has higher prediction accuracy. The simulation results show that after the introduction of the neural network model, the features of neural network with richer semantic information can be used, and the new event line can be processed at the same time. The training speed of news content processing is much faster than that of probability graph model. It can be seen that under the influence of new media, news communication shows new characteristics, which further affects people’s news reading habits.

Keywords

Data prediction neural network news information intelligent search

1 Introduction

The mass information provided by news is accompanied by problems. Because of its large scale and wide range of coverage, news leads to information dissemination and development law and complexity [1]. In addition, the nodes involved in information dissemination are highly intelligent and self-adaptive human beings [2]. Their response to information will vary with time, place and mood, which brings more to the research of information dissemination Many challenges [3, 4]. How to spread information in news, how to build a suitable model and how to predict the law of information dissemination have become a research hotspot.

In the early days of biology and infectious diseases, it has made important research achievements in the spread of virus within the population, and has been widely used in the field of in-formation dissemination [5]. The application of neural network in news mainly uses neural network to establish prediction model, but the accuracy of neural network algorithm in prediction is not high [6, 7]. From the research status and development trends at home and abroad, the advantages of designing the prediction algorithm of news information dissemination based on neural network are mainly shown in two aspects: first, the research can establish a solid theoretical basis for news information dissemination and solve the actual information dissemination problems, for example, in the case of social network with complex information dissemination mode, using neural network for information dissemination [8]. The characteristics and methods of information dissemination are optimized. Secondly, after studying the model of news information dissemination based on neural network, we can fully understand the law of information dissemination, which is conducive to improving the accuracy of information dissemination prediction in the field of news [9, 10].

Aiming at the above problems, this paper proposes a prediction algorithm of news information dissemination based on neural network. The goal is to build a more perfect news information dissemination model, and use the improved network algorithm to predict information dissemination on the basis of the model, so as to achieve better simulation effect and higher accuracy [11]. The research work of this paper is helpful to understand the information dissemination mechanism, the dissemination rule and the basic evolution process of public opinion in news, to discover the real behavior characteristics of users in news, to reveal the essential phenomenon hidden behind the huge network, and to lay a solid foundation for the prediction of information dissemination. Therefore, the topic of this paper has certain scientific research, innovation, theoretical value and practical significance [12].

2 Related work

BP neural network can establish the mapping from input to output, which is often used to ad-just the weight of multi-layer perceptron. The main difference between BP neural network and perceptron is that there are more layers than single-layer perceptron, and single-layer perceptron includes an input layer, an output layer and an implicit layer. The hidden layer of BPNN can be composed of multiple neurons [13, 14]. There is no connection between the neurons in the same layer in the network, and the neurons in two adjacent layers are all connected, that is, each neuron in the previous layer points to all neurons in the latter layer. When the training set of data is clear, the number of neurons in input layer and output layer is determined. Next, the design of network topology is to find the number of hidden layers and the number of neurons in each hidden layer. It is proved that BP neural network with one hidden layer is enough to map all continuous functions [15]. Therefore, in the design of network, one hidden layer is preferred. When the number of nodes in the hidden layer cannot change the network performance, two hidden layers are considered. At present, there is no unified theory to determine the number of neurons in the hidden layer [16]. It is more based on the combination of experience and many experiments to solve practical problems.

Aiming at the problem of news communication, we choose BP neural network as the basic framework. The most important problem to be solved is the modeling of news communication data. BP neural network chooses three-layer basic structure. The setting of hidden layer can be found out by experiment [17, 18]. After the input and output of this problem are determined, the model framework is basically established.

There are two types of data used in the prediction of news dissemination. One is to cluster the attributes of news after analyzing and defining the topological structure of the network, including the direction of news dissemination (directed or undirected), the time of information dissemination (weekend or workday), the length of information, whether the URL is included, whether the picture is included, the number of fans of the publisher, the number of concerns of the publisher, and the transfer The number of fans of the sender, the number of concerns of the forwarder and other attributes [19]. According to the prediction problem, some attributes of news are selected as the input of BP neural network, and the discussion degree of information is taken as the output of BP neural network, so as to predict the discussion degree that information may bring. The other is time series data. The number of related topic discussions in consecutive days is used to predict the degree of topic discussions in the future [20]. That is to say, the number of topic discussions in consecutive days is used as the input of neural network, and the number of topic discussions in the future is used as the output of neural network to establish the information dissemination prediction model.

There are advantages and disadvantages in each of the above two ways. The first way is to select the key attributes. At present, there is no research result that the topic discussion degree is related to which attributes of the information are specific and which attributes are interference items, so the accuracy of prediction model construction cannot be guaranteed. The second way is to use BP neural network as the basic model for topic prediction [21]. The discussion of the problem focuses on the number of neurons in the input layer, that is, how many days of topic discussion in a row to predict the degree of topic discussion in the future. It is not difficult to get this data by trial and error method according to different prediction data, while the output layer generally chooses the next day’s words Number of questions discussed [22, 23]. To select BP neural network as the framework of news communication model, first of all, we need to model specific issues. We can use two different data sets to build the model. One is to build a mapping model between information attributes and information discussion number. The other is to build a mapping model between the information discussion number of the previous days and the information discussion number of the future day according to time series [24]. Both of these mapping models focus on prediction. Then we need to establish BP neural net-work prediction method.

Based on the mathematical model of specific news communication problems, this paper explores the appropriate topology of BPNN, including determining how many hidden layers should be included and the number of nodes in each hidden layer, determining these parameters, so we think that the BP neural network framework is set up, and then carry out the learn-ing and training of BPNN [25, 26]. The specific steps are described in the above section. Finally, the use of BPNN, of course, needs to use the test set to test the performance of each parameter of the trained network in advance, and it can be used to predict after reaching the expectation. Of course, it is undeniable that the prediction accuracy of the unknown depends to some extent on the samples selected before.

3 Theoretical analysis

3.1 The propose IDS system

Because the process of detection is also the process of clustering, the original data category in the cluster is more complete, so that the new data events can be then automatically trained and clusters updated, enabling the system to detect new intrusions.

$\begin{matrix} d_{c} (x_{i}, x_{l}) \equiv \sum_{j = 1}^{p} δ (x_{ij}, x_{lj}) \\ (1 ⩽ i ⩽ n, 1 ⩽ l ⩽ n, i \neq l) \end{matrix}$ (1)

$δ (x_{ij}, x_{lj}) = {\begin{matrix} 0 & if & x_{ij} = x_{lj} \\ 0 & if & x_{ij} \neq x_{lj} \end{matrix}$ (2)

$d_{log} (x_{i}, x_{l}) = \frac{log ((b - 1) (d_{c} (x_{i}, x_{l}) / p) + 1)}{log (b)}$ (3)

In the formula 1∼3, we define the model to be later considered, where the δ (x_ij, x_lj) is the set of clusters, the δ (x_ij, x_lj) is the judgement on the cluster forms and the d_log (x_i, x_l) represent the IDS model to be considered. In this paper, a distributed intrusion detection method based on neural network ensemble is proposed for distributed intrusion detection. First, as each independent detection agent collects data to train two RBF neural networks to form integration. Then, a two-level integrated algorithm is used to perform distributed detection in the process of system operation. That is, the core single Agent classifies suspicious data packets by integration with the general references to the following items.

$min_{W, Z} F (W, Z) = \sum_{i = 1}^{n} \sum_{m = 1}^{c} w_{mi}^{α} d_{log} (Z_{m}, x_{i})$ (4)

$w_{mi} = {\begin{matrix} 1 \\ 0 \\ \frac{1}{{\sum_{s = 1}^{c} [\frac{1}{d_{log} (Z_{s}, x_{i})}]}^{[\frac{1}{α - 1}]}} \end{matrix}$ (5) Where the $w_{mi}^{α} d_{log} (Z_{m}, x_{i})$ is the term waiting for the systematic optimization and the w_mi is the parameter for the further processing. Bagging and the Boosting difference lies in bagging the training regulations choice is stochastic, between each turn training regulations mutually independent, but Boosting training regulations choice is not independent, each turn training regulations choice and front each turn study finally related.

The Bagging each forecast function does not have the weight, but Boosting has the weight and the Bagging each forecast function may the parallel production, but the Boosting each forecast function only can the smooth production. Like this basic consumes likely regarding the neural network when core study method Bagging may save the massive time expenses through parallel training. Once get from the training data classifier, it can be used for its invasion of real-time detection while because the number of these classes is very limited and need a very small amount of calculation enough to meet requirements of real-time detection. For our system to detect intrusion actually is also the process of data classification. Data classification process as show in Fig. 1.

Fig. 1

Data classification process.

3.2 Neural algorithm improvement

BP network is one of the most widely used neural network models, and it is a typical multi-layer feedforward network. We assume that the BP neural network has n layers, the first layer is the input layer, and the last layer is the output layer. The intermediate layer has a single layer or multiple layers. Moreover, because they have no direct relationship with the outside world, they are also called hidden layers. The input layer acts on the left and right sides of the buffer memory, and the data source is added to the network. Therefore, the input-output relationship of the neurons of the input layer is generally a linear function. In contrast, the input-output relationship of individual neurons in the hidden layer is generally a nonlinear function. Although the hidden layer is not connected to the outside world, their state can affect the relationship be-tween input and output.

We assume that the input layer has n number of neurons and the output layer has p_n number of neurons. When the BP neural network input data X = [x_l, x₂, … x_p1] ^T passes from the input layer to each hidden layer node, the output data Y = [y₁ⁿ, y₂ⁿ, … y_pnⁿ] ^T can be obtained. Therefore, the BP neural network can be thought of as a nonlinear mapping from input to output. At the same time, it is also a supervised Learning.

Next, the standard BP algorithm is derived. x_I represents the neural network input, y_I represents the neural network output; d_I represents the network expected output; w_ijk represents the connection weight of the k-th neuron from the i-th layer to the j-th neuron in the i + 1-th layer. Moreover, the single neuron output of the i-layer of the neural network is denoted by o_ij, the neuron threshold is denoted by o_ij, net_ij denotes the total output of the j-th neuron of the i-th layer, and N_i denotes the number of nodes of the i-th layer of neurons [20].

(1) Calculating the forward propagation of BP networks

$ne t_{ij} = \sum_{k = 1}^{N_{i - 1}} O_{(i - 1) k} K_{(i - 1) kj}$ (6)

$O_{ij} = f_{s} (ne t_{ij}) = \frac{1}{1 + exp [- (ne t_{ij} - θ_{ij})]}$ (7)

(2) Derivation of error back propagation algorithm.

First, the error is defined:

$e_{j} = d_{j} - y_{j}$ (8) The weight is corrected by the error between the actual output of the network and the desired output. Objective function E:

$E = \frac{1}{2} \sum_{J} {(d_{j} - y_{j})}^{2}$ (9) The weight of the network is corrected. Among them, the learning efficiency is represented by η, which is between 0 and 1.

$Δ W_{ijk} = - η \frac{\partial E}{\partial W_{ijk}}$ (10) The recursive relationship between ΔW_ijk and neuron output:

$Δ W_{ijk} = - η \frac{\partial E}{\partial W_{ijk}} = - η \frac{\partial E}{\partial ne t_{(i + 1) k}} \cdot \frac{\partial ne t_{(i + 1) k}}{\partial W_{ijk}} = η δ_{ik} \cdot \frac{\partial ne t_{(i + 1) k}}{\partial W_{ijk}}$ (11) Among them, $δ_{ik} = \frac{\partial E}{\partial ne t_{(i + 1) k}}$

$\frac{\partial ne t_{(i + 1) k}}{\partial W_{ijk}} = \frac{\partial net}{\partial W_{ijk}} (\sum_{h = 1}^{N_{i}} o_{ih} W_{ihk}) = O_{ij}$ (12)

$\begin{matrix} Δ W_{ijk} = - η δ_{ik} O_{ij} \end{matrix}$

$δ_{ik} = - \frac{\partial E}{\partial O_{(i + 1) k}} \frac{\partial O_{(i + 1) k}}{\partial ne t_{(i + 1) k}}$ (13) $\begin{matrix} \frac{\partial O_{(i + 1) k}}{\partial ne t_{(i + 1) k}} = f^{'} (ne t_{(i + 1) k}) \end{matrix}$ $\begin{matrix} f^{'} (x) = \frac{e^{- (x - θ)}}{{[1 + e^{- (x - θ)}]}^{2}} = f (x) [1 - f (x)] \end{matrix}$

$\frac{\partial O_{(i + 1) k}}{\partial ne t_{(i + 1) k}} = f^{'} (ne t_{(i + 1) k}) = f (ne t_{(i + 1) k}) (1 - f) = o_{(i + 1) k} (1 - o_{(i + 1) k})$ (14) $\begin{matrix} \frac{\partial E}{\partial O_{(i + 1) k}} \end{matrix}$ The output node is o_(i+1)k.

$\frac{\partial E}{\partial O_{(i + 1) k}} = y_{k} - d_{k}$ (15)

$δ_{ik} = - \frac{\partial E}{\partial ne t_{(i + 1) k}} = - \frac{\partial E}{\partial O_{(i + 1) k}} = (d_{k} - y_{k}) o_{(i + 1) k} (1 - o_{(i + 1) k})$ (16) Hidden layer node is o_(i+1)k

$\frac{\partial E}{\partial O_{(i + 1) k}} = \sum_{h = 1}^{N_{i + 2}} \frac{\partial E}{\partial ne t_{(i + 2) h}} \cdot \frac{\partial ne t_{(i + 2) h}}{\partial O_{(i + 1) k}} = - \sum_{h = 1}^{N_{i + 2}} δ_{(i + 1) h} w_{(i + 1) kh}$ (17)

$ne t_{ij} = - \sum_{k = 1}^{N_{i - 1}} o_{(i - 1) k} \cdot w_{(i - 1) kj}$ (18) When the hidden layer node is o_(i+1)k, it can only know that the total error has a certain correlation with the hidden layer output, but the correct output result cannot be known in advance. The output of the hidden layer node affects the output of the next hidden layer.

$\begin{matrix} δ_{ik} = O_{(i + 1) k} (1 - O_{(i + 1) k}) \sum_{h = 1}^{N_{i + 2}} δ_{(i + 1) h} w_{(i + 1) kh} \end{matrix}$ (19) The weight formula of the BP algorithm is adjusted to:

$\begin{matrix} Δ W_{ijk} = 〈 \begin{matrix} η (d_{k} - y_{k}) y_{k} (1 - y_{k}) o_{ij} \\ η o_{(i + 1) k} (1 - o_{(i + 1) k}) \cdot \sum_{h = 1}^{N_{i + 2}} δ_{(i + 1) h} w_{(i + 1) kh} o_{ij} \end{matrix} \end{matrix}$ (20)

i-th layer neuron is

$δ_{ik} = 〈 \begin{matrix} (d_{k} - y_{k}) y_{k} (1 - y_{k}) \\ o_{(i + 1) k} (1 - o_{(i + 1) k}) \cdot \sum_{h = 1}^{N_{i + 2}} δ_{(i + 1) h} w_{(i + 1) kh} \end{matrix}$ (21)

In view of the above problems, many effective improvement algorithms have been proposed at home and abroad. The following are some commonly used improved algorithms for training BP networks. Secondly, to solve the training problem of neural network, we must find skills from the learning algorithm, neural network mode as shown in Fig. 2.

Fig. 2

Neural network.

4 Evaluation model optimization

Prediction is mainly to evaluate the development trend and direction of things. It achieves the goal by using the classification of data or estimating the specific value of data. To achieve the goal of information prediction, first we need to get the model according to the obtained data, and then we can predict the unknown variables through the model. However, the current level of people using social network data for prediction is not optimistic. The most important factor is that the data in social network belongs to the massive level, and the analysis algorithm of massive text has not reached the ideal accuracy. It is very important to obtain data from the perspective of social network prediction by using the effective data sampling method of population science. Because of using the existing data in social network to predict, there is always a deviation between the prediction results and the final results.

This paper explores the factors that affect the usefulness of news comments, and sorts the importance of these factors, and provides ideas and reference methods for later scholars in the study of the popularity of news comments. The system reads the news and comments input by users. First, the news content is segmented, and the stop words are removed. The pre-trained word2vector model is used to convert the news content into sentence vector. Through cosine similarity, K historical news closest to it are selected from the historical news corpus. Based on the number of comments and votes of these historical news, the number of comments and votes of the input news are generated Total. Then, we segment the comment, remove the stop words and turn them into sentence vectors. In the process of word segmentation, we can get the number of words, sentences and the average word length of the comment. The cosine similarity be-tween the sentence vector of news content and the sentence vector of comment content is calculated as the text similarity, and the emotional features of comment are obtained through the emotional dictionary. Finally, a comment can be expressed as a feature vector, which is sent to the trained classifier, and the predicted classification is obtained. Mobile news spread as shown in Fig. 3.

Fig. 3

Mobile news spread.

The Radial Basis Function (RBF) is a scalar function that is symmetric along the radial direction. It is defined as a monotonic function of the Euclidean distance between a point a to the center point in space. The closer the distance, the larger the value of the function, indicating that the greater the effect. However, the further the point a is from the center point O, the smaller the value of the function, indicating that the effect is smaller. The RBF kernel function enables non-linear mapping and can be used to handle linearly inseparable cases with fewer parameters. Compared with the function parameters with more polynomial kernel functions, the training model established by RBF kernel function is relatively simple, and the effect of training model is relatively more stable. It assumes that the eigenvector of the text is x and the category it belongs to is y. The formula for the RBF kernel function is as follows. $\begin{matrix} k (x, y) = \exp (- γ {∥ x - y ∥}^{2}), (γ > 0) \end{matrix}$ The nonlinear classification problem is achieved by mapping it into a high-dimensional space to achieve linear classification. For example, sample X is mapped to a high-dimensional space by an equation Φ (x), and an optimal hyperplane can be found in a high-dimensional space to classify the mapped samples. This optimal hyperplane formula expression is as follows. $\begin{matrix} w • Φ (x) + b = 0 \end{matrix}$ Among them, w is the normal vector of the hyperplane, and the length of the normal vector of the hyperplane is the maximum plane distance sought. The maximum distance found is the optimal hyperplane that is sought, so the maximum value of the classification interval length = 2/ -∥ w ∥ is required to fund. Finding its maximum value is equivalent to finding the minimum value of ∥w ∥ ²/2.

The sample point that satisfies the condition y (w · x + b) = 1 is used as a support vector to construct a Lagrangian function whose formula is as follows. $\begin{matrix} L (w, b, a) = \frac{1}{2} (w \cdot w) - \sum_{i = 1}^{l} a_{i} [y_{i} (w \cdot x_{i}) + b - 1], \\ a_{i} > = 0 () \end{matrix}$

Then, the kernel function k (x, y) is selected so as to satisfy the following formula k (x, y) = Φ (x) · Φ (y).

5 Model experiment analysis

5.1 Data collection

The discovery of news topic is also a branch of news communication research, which has basically formed a model. The hot topic system of social network takes the text obtained after social data grabbing and preprocessing (removing advertisements, labels, filtering pause words, etc.) as input. These texts mainly include release time, title and preprocessed text. Then, the input text is segmented. After the segmentation, the text is vectorized, and then the clustering algorithm is used to cluster it, Finally, according to the heat calculation method, the heat value of each topic is calculated, and the related information of topics is output in the order of heat value from high to low.

All the comment texts in the basic data set are exported to a frequency text to be counted. The text contains 57827 characters, and the total number of words counted is 2921. After obtaining the word frequency statistics results, we manually select the top 1000 words with the highest word frequency, select the stop words that are not helpful for the text classification, and establish the stop word list. In the end, the stop words list obtained contains 182 stop words. The first 15 stop words are listed, and the stop word sample table is shown in Table 1, and the statistical chart is shown in Fig. 4.

Table 1
Sample table of stop words

No. Words Occurrences Frequency

1 of 1888 4.9585

2 Up 517 1.3580

3 Enter 313 0.8224

4 on 218 0.5908

5 Pull 178 0.4821

6 Got 168 0.4542

7 also 166 0.4487

8 and also 158 0.4264

9 Has 136 0.3679

10 All 129 0.3483

11 still is 89 0.2397

12 give 84 0.2286

13 Put 72 0.1951

14 but 68 0.1840

15 Let 63 0.1702

No.	Words	Occurrences	Frequency
1	of	1888	4.9585
2	Up	517	1.3580
3	Enter	313	0.8224
4	on	218	0.5908
5	Pull	178	0.4821
6	Got	168	0.4542
7	also	166	0.4487
8	and also	158	0.4264
9	Has	136	0.3679
10	All	129	0.3483
11	still is	89	0.2397
12	give	84	0.2286
13	Put	72	0.1951
14	but	68	0.1840
15	Let	63	0.1702

Fig. 4

Sample diagram of stop words.

The above stop word list is imported into the basic data set, and the stop words appearing in the stop word list are deleted to obtain the data set of the experiment.

5.2 Experimental design

The experimental design flow of this experiment is: First, when the data is preprocessed, the stop words in the data are deleted, which can reduce the experimental error and improve the accuracy of the experiment. Then, the test set of the entire data set is subjected to negative word analysis processing, and the text whose negative words appear odd times is marked. These texts are classified by the PMI algorithm, and the SVM algorithm is used to train and classify the remaining texts that are not labeled in the test set, and the final classification result is obtained.

The experimental design of the PMI algorithm can be expressed as:

The classification experiment based on PMI algorithm can further improve the classification accuracy of the algorithm after adding the negative word analysis. The judgment of the negative word relies on the negative vocabulary, and this experiment constructs a negative vocabulary, which contains the current common negative words, and the weight of the negative word is set to –1.

For the comment text marked as PMI to be classified, the high-frequency words of three positive categories are selected based on the NTUSD sentiment dictionary: “good”, “good”, “like”. Moreover, the high-frequency sub-words of the three negative classes are selected as: “poor”, “unused”, and “bad” as reference words for text classification.

5.3 Experimental results

The data set of the improved experiment based on the PMI algorithm is the same as the basic data set based on the PMI algorithm experiment. The only difference is that the data set has been further improved, i.e. stop word processing, which has the advantage of reducing the noise of subsequent experiments. After the stop word processing of the data set, the negative word analysis of the test set and the annotations that the number of negative words is an odd number are classified by PMI algorithm, and the remaining unlabeled texts are classified by RBF kernel function using libsvm tool. The experimental results after optimizing the sentiment orientation algorithm of the online course review text are shown in Table 2. Statistical diagram as show in Fig. 5.

Table 2
Experimental results table

Name Positive class Negative class

RecalI 0.837 0.795

Precision 0.820 0.843

F-Measure 0.829 0.818

Name	Positive class	Negative class
RecalI	0.837	0.795
Precision	0.820	0.843
F-Measure	0.829	0.818

Fig. 5

Statistical diagram of the experimental results.

Experiment using “good” and “bad” as reference words is used as experiments, and experiment using high-frequency words in the emotional dictionary as the reference words is used as the experiment 2, and the experiment of the sentiment orientation algorithm based on SVM for online course review text is used as experiment 3, and the improved experiment is used as experiment 4. The recall rate, accuracy, and F-Measure of these four experiments were compared and analyzed. The results of the analysis are shown in Tables 3 –5, and Figs. 6 –8.

Table 3

Comparison table of recall rates

	Positive class	Negative class
Test 1	0.62	0.71
Test 2	0.72	0.83
Trial 3	0.73	0.68
Test 4	0.82	0.78

Table 4

Comparison table of accuracy

	Positive class	Negative class
Test 1	0.63	0.65
Test 2	0.76	0.73
Trial 3	0.7	0.72
Test 4	0.79	0.82

Table 5

Comparison table of F-Measure

	Positive class	Negative class
Test 1	0.63	0.67
Test 2	0.73	0.77
Trial 3	0.71	0.69
Test 4	0.81	0.79

Fig. 6

Comparison chart of recall rates.

Fig. 7

Comparison chart of accuracy.

Fig. 8

Comparison chart of F-Measure.

6 Innovative mode of news communication

(1) Bring self media into news report

It is decided by the interactivity of news communication and the diversity of communication subjects in the era of new media that self media is included in news reports. Therefore, we need to constantly adapt to the trend of the times and bring the self media into the news report. In the news report, we can set up a special self media section, which can be named “personal experience” and so on, to serve as a position for netizens to release news. At the same time, set up corresponding news editors in charge of the section, verify and track the valuable and significant news, and publish it in other corresponding sections for readers who need to read serious news. After the verification of valuable news, the first publisher of news can be rewarded appropriately to encourage more people to join in the news report with a serious and responsible attitude and encourage the vigorous development of self media.

(2) Report in-depth and valuable news

Under the influence of new media, the news content shows the phenomenon of homogeneity. In order to stimulate the readers, some news organizations began to use some sensationalism, exaggeration, extreme views and seriously distorted news. This has seriously affected the healthy development of news. Therefore, in the era of new media, if we want to innovate the mode of news communication, we must do the opposite, insist on reporting the news with depth and value, so as to enable readers to think independently, rather than only be stimulated by some extreme remarks. We should have the ability to find valuable news, insist on reporting news that has a real impact on people’s lives, has practical significance for social development, and has a certain role in national development, and avoid blindly stimulating the curiosity of the audience through all kinds of anecdotes and anecdotes. Secondly, journalists should have a deep coverage of the news. The depth of news is reflected in that the excavation of news can inspire people to think, affect people’s actual behavior, and make people feel the real society. Therefore, journalists should insist on in-depth reporting of news, and in the process of reporting news, they should report news facts, understand the reasons behind news facts, supervise and report the handling of news facts, and promote the improvement of relevant laws, regulations and policies.

(3) Press release should be consistent with readers’ reading habits

Modern society is a fast-paced development society. With the increase of the amount of news release, it brings people a kind of reading burden, which makes it more and more difficult for people to find the needed news. Therefore, when we release news, we must make the news fit with the readers’ reading habits, and try to shorten the time for readers to screen the news. Only in this way can we gain the loyalty of readers. First of all, news publishing needs to be classified. News can be classified according to the types of news events, the time of news occurrence, etc., which is convenient for readers to filter news, and provide readers with classified subscriptions, so as to facilitate readers’ personalized reading. Secondly, press releases should be as clear and concise as possible. In the page of news release, we can use the way of Title listing to show readers all kinds of news. Readers can click the title according to their own interests to read the news content, which can effectively save readers’ reading time

7 Conclusion

In the era of new media, news communication presents new features, which also affect people’s news reading habits all the time. Therefore, we need to fully understand the characteristics of news communication in the new media era, and then fully grasp people’s news reading habits on this basis, and then put forward the innovative mode of news communication in the new media era on the basis of these two points. We believe that only by deeply grasping the characteristics of news communication mode in the new media era and constantly understanding people’s news reading habits, can we put forward innovative news communication mode and greatly improve the efficiency of news communication.

To sum up, under the influence of new media, news communication shows new characteristics, which further affects people’s news reading habits. Therefore, we need to fully grasp the characteristics of news communication and people’s news reading habits. Based on these conditions, we need to create a communication mode that is in line with the future development trend of news media, so as to maximize the efficiency of news communication.

References

, Xu

and Yang

et al., 3D convolutional neural networks for human action recogni-tion, IEEE Transactions on Pattern Analysis and Machine Intelligence35(1) (2013), 221–231.

Chen

, Jafari

and Kehtarnavaz

, Improving human action recognition using fusion of depth camera and inertial sensors, IEEE Transactions on Human-Machine Systems45(1) (2015), 51–61.

Devanne

, Wannous

and Berretti

et al., 3-D human action recognition by shape analy-sis of motion trajectories on riemannian manifold, IEEE Transactions on Cybernetics45(7) (2015), 1340–1352.

Ahn

and Neumann

, Measuring effectiveness: A DEA approach under predetermined tar-gets, International Journal of Business Analytics1(1) (2014), 16–28.

Guo

and Lai

, A survey on still image based human action recognition, Pattern Rec-ognition47(10) (2014), 3343–3361.

Chaaraoui Padilla-López

A.A.

, Ramón

José

, Climent-Pérez , Pau , et al., Evolutionary joint selection to improve human action recognition with RGB-D de-vices, Expert Systems with Applications41(3) (2014), 786–794.

Chen

, Liu

and Kehtarnavaz

, Real-time human action recognition based on depth mo-tion maps, Journal of Real-Time Image Processing12(1) (2013), 155–163.

Barnachon

, Bouakaz

and Boufama

et al., Ongoing human action recognition with mo-tion capture, Pattern Recognition47(1) (2014), 238–247.

Everts

, Van Gemert

J.C.

and Gevers

, Evaluation of color spatio-temporal interest points for human action recognition, IEEE Transactions on Image Processing23(4) (2014), 1569–1580.

10.

Theodorakopoulos

, Kastaniotis

and Economou

et al., Pose-based human action recog-nition via sparse representation in dissimilarity space, Journal of Visual Communication and Image Representation25(1) (2014), 12–23.

11.

Ben Aoun

, Mejdoub

and Ben Amar

, Graph-based approach for human action recogni-tion using spatio-temporal features, Journal of Visual Communication and Image Repre-sentation25(2) (2014), 329–338.

12.

Zhi-Jie

and Chun-Yan

Y.U.

, An improved support vector machines with balanced binary decision tree for multi-class classification, Journal of Chinese Computer Systems35(5) (2014), 1124–1127.

13.

Dongxiao

, Tiannan

and Bingyi

, Power load forecasting by wavelet least squares sup-port vector machine with improved fruit fly optimization algorithm, Journal of Combinato-rial Optimization33(3) (2016), 1–22.

14.

, Hua

and Tang

et al., Modeling of the hot metal silicon content in blast furnace using support vector machine optimized by an improved particle swarm optimizer, Neural Computing and Applications27(6) (2016), 1451–1461.

15.

Dharamkar

, Singh

R.R.

and Dharamkar

et al., Cyber-attack classification using improved ensemble technique based on support vector machine and neural network, International Journal of Computer Applications103(11) (2014), 1–7.

16.

Wang

, Fan

and Hsu

C.H.

et al., A vertical handoff method via self-selection decision tree for internet of vehicles, IEEE Systems Journal10(3) (2017), 1183–1192.

17.

, Ren

and Niu

, Landslide susceptibility assessment using object mapping units, de-cision tree, and support vector machine models in the Three Gorges of China, Environ-mental Earth Sciences71(11) (2014), 4725–4738.

18.

Parvin

, Mirnabibaboli

and Alinejad-Rokny

, Proposing a classifier ensemble frame-work based on classifier selection and decision tree, Engineering Applications of Artificial Intelligence37(37) (2015), 34–42.

19.

Song

Y.Y.

and Ying

L.U.

, Decision tree methods: Applications for classification and predic-tion, Shanghai Archives of Psychiatry, 2015.

20.

Mikučionienė

Rʉta

, Martinaitis

and Keras

, Evaluation of energy efficiency measures sustainability by decision tree method, Energy and Buildings76 (2014), 64–71.

21.

Fernández

Alberto

, Peralta

and Manuel Benítez

José

, et al., E-learning and educational data mining in cloud computing: an overview, International Journal of Learning Technology9(1) (2014), 25–52.

22.

Miller

L.D.

, Soh

L.K.

and Samal

et al., A comparison of educational statistics and data min-ing approaches to identify characteristics that impact online learning, Journal of Educa-tional Data Mining7(3) (2015), 737–741.

23.

Liu

X.H.

, Application of data mining technology in higher education, Applied Mechanics and Materials574 (2014), 5.

24.

Mining library and university data to understand library use patterns, The Electronic Li-brary33(3) (2015), 355–372.

25.

Lopes

, Clement

and Roy

et al., Multi-armed bandits for intelligent tutoring systems, Journal of Educational Data Mining7(4) (2015), A–705.

26.

Reiner

B.I.

, The quality/safety medical index: Implementation and analysis, Journal of Digital Imaging28(1) (2015), 1–6.