Uneingeschränkter Zugang

Innovation and Practice of Ideological and Political Education Communication Mode Driven by Artificial Intelligence

,  und   
21. März 2025

Zitieren
COVER HERUNTERLADEN

Introduction

Under the wave of intelligent technology, the communication pattern of ideological and political education in colleges and universities has undergone profound changes, gradually showing the blurred boundaries of “human-machine” and intelligent scenarios. In the era of artificial intelligence, the change of technical means makes the information dissemination channels more and more extensive, and increasingly penetrate into the learning life, profoundly affecting the way of thinking, behavior and values of college students, and inserting “wisdom wings” for the dissemination of ideological and political education in colleges and universities [1-3]. In this context, ideological and political education in colleges and universities must plan, move and act in accordance with the situation, grasp the opportunities for the development of artificial intelligence technology, actively explore communication strategies to enhance the effectiveness of ideological and political education in colleges and universities, and synergistically promote technological and ideological education, so as to achieve the modernization and development of human beings.

Examined from the function of information dissemination, intelligent technology is constantly enriching the form of ideological and political education carriers in colleges and universities, and the intelligent communication mode and its material form have become an important part of the ideological and political education carrier system in colleges and universities, and are increasingly playing the advantages that the previous carriers lacked in the ideological and political education in colleges and universities [4-7]. On the one hand, intelligent technology has impacted the traditional way of ideological and political education in colleges and universities [8]. Emerging intelligent technology in the process of information dissemination shows the attraction, become the majority of young college students are happy to accept all kinds of information in an important way, the values and worldview of college students are deeply affected by the emerging intelligent communication methods [9-10]. On the other hand, intelligent technology also provides innovative opportunities for ideological and political education work in colleges and universities [11]. To improve the effectiveness of ideological and political education work, it is undoubtedly necessary to take technology as a support, make full use of new media technology, intelligent technology clusters, grasp the characteristics of the intelligent era and the way of thinking, as an opportunity to promote the innovation and development of ideological and political education carriers in colleges and universities [12-14].

This paper proposes the innovative application strategy of “artificial intelligence +” Civic Education Communication, and constructs the Civic Content Recommendation Algorithm based on graph neural network. Based on graph neural networks, this paper proposes a graph attention memory social network (GAMSN) recommendation algorithm, which consists of an attention memory module, an attention aggregation module, and a model prediction module. The Audience-Civic Content Attentional Memory Module reflects the audience’s inherent multifaceted preferences for Civic Content, and the Audience-Audience Attentional Memory Module simulates the audience’s multiple influences on them. Together, they complete the construction of information vectors in the information dissemination process of graph neural networks. The algorithm is constructed and then tested for performance on the dataset, and analyzed for its effect on disseminating Civics education.

Innovations in the dissemination of civic education based on graph neural networks
Innovative Application Strategies of “Artificial Intelligence +” Civic and Political Education Communication
Personalized Learning Push Based on Artificial Intelligence

Personalized learning push for large audiences based on big data analysis of artificial intelligence in colleges and universities is one of the innovative application strategies of “wisdom +” Civic and political education communication.

Data Analysis and Audience Portrait

Colleges and universities collect audience learning data and behavioral data to accurately understand individual audience’s learning interest, learning level, learning habits, etc., and then apply machine learning algorithms and data mining technology to analyze all the audience information, so as to establish personalized learning portraits of the audience.

Recommendation of personalized learning resources

Based on the audience’s personalized learning portrait, colleges and universities should recommend suitable learning methods for the audience and provide corresponding personalized learning resources [15].

Adaptive learning path design

Intelligent learning paths are tailored to the current learning progress of each individual audience. Colleges and universities can analyze the audience’s learning performance, test scores, etc. to determine the audience’s knowledge mastery status, and then adjust the learning content, difficulty and progress according to this information.

Real-time feedback and interaction

Colleges and universities use AI technology to provide audiences with real-time learning feedback and ways to interact. This can be realized through chatbots, intelligent Q&A systems, etc., and the audience can ask questions and seek help from the system at any time. The system can also provide relevant answers and guidance based on the audience’s questions and feedback.

Continuous optimization and improvement

In the process of personalized learning delivery based on artificial intelligence, colleges and universities should collect relevant information about the audience’s learning process, receive feedback from individual audiences, and continuously optimize and improve the personalized learning delivery system. Colleges and universities can also provide personalized learning resources according to the individual differences and learning needs of the audience to improve the learning effect.

Applying Virtual Reality Technology to Civic and Political Education

At present, virtual reality technology is becoming increasingly advanced, and colleges and universities can achieve twice the results with half the effort by applying it to ideological education.

Virtual reality scene simulation

Colleges and universities can use virtual reality technology to simulate the scene of civic education. Through the virtual reality scene, the audience can watch the process of relevant events in an immersive way and improve their understanding of the content of ideological education.

Visualization and Immersion Learning

Based on virtual reality technology, colleges and universities can provide audiences with an immersive learning experience, stimulating their interest in learning through visual and auditory stimulation. The audience can interact with the virtual characters to enhance the learning experience.

Practical and Simulation Training

Based on virtual reality technology, colleges and universities can solve the dilemma that it is difficult to combine the traditional communication methods of Civic and Political Education with the actual situation, and provide a new way of practice and simulation training for the audience, thus greatly reducing the cost of practical activities. For example, audiences can participate in teamwork, decision-making, ethical issue exploration, and other activities in the virtual reality environment to improve their practical abilities and problem-solving skills.

Interdisciplinary cooperation and resource sharing

Based on virtual reality technology, colleges and universities can realize interdisciplinary cooperation and resource sharing, combine Civic Education with other disciplines and fields, so as to enrich the content and form of Civic Education and promote the innovation and development of Civic Education. Colleges and universities apply virtual reality technology to civic and political education, which can overcome the many limitations of traditional dissemination channels for civic and political education. The technology helps to improve the attractiveness and educational effect of Civic and Political Education, provide audiences with a better and more compatible practice environment, and promote the overall development of the audience.

Integration of online educational resources for cross-regional dissemination

Integrating online education resources and realizing cross-regional dissemination is one of the innovative application strategies of “wisdom +” Civic and political education dissemination. By integrating and utilizing online education resources, colleges and universities can realize cross-regional dissemination and coverage of Civic and Political Education. The following are the specific steps:

Resource Integration and Sharing

Colleges and universities can integrate high-quality Civic and Political Education resources from various regions, including teaching courses, teaching materials, case studies, lectures, etc.; through the establishment of a unified online education platform or resource library, these resources can be shared and made available to audiences and teachers nationwide.

Online Teaching and Learning

Colleges and universities should actively utilize online education platforms to conduct learning activities. This enables audiences to learn Civics content, participate in online discussions and interactions, and complete assignments and exams, etc. via the Internet. Teachers can conduct teaching and assessment through the online education platform, guide the audience to learn online, and improve Civics education.

Live broadcasting and recording promotion

Colleges and universities can use live broadcasting technology to transmit real-time images of lectures, seminars and other activities of Civic and Political Education to audiences in remote areas. At the same time, colleges and universities can also record and broadcast these activities, promote and share them through online education platforms, for audiences to watch and learn at any time.

Social Recommendations

With the prevalence of social networks, social recommendation that incorporates audience social relationships has become a research priority for personalized services. Social recommendation utilizes additional audience-audience social networks to mitigate sparsity of audience-item interaction data and improve recommendation performance. The underlying reason is that audiences are influenced by self-centered social networks, and audiences with social connections tend to have similar preferences, based on social influence theory. Early approaches used this relationship directly as a regularizer to constrain the final audience representation, or utilized it as an input to augment the original audience embedding. From a graph learning perspective, early social recommendation tasks can be viewed as modeling the first-order neighbors of each audience. However, in practice, an audience may be influenced by other factors, such as their friends’ friends. In the past, neglecting the diffusion of higher-order social influences of an audience may result in poor recommendations. Due to its ability to model how audiences are affected by the recursive social diffusion process, graph neural networks (GNNs) have become a popular choice for modeling audience social relationships in social recommendation tasks.

In social recommendation, audiences are involved in two types of relationships: one is audience-item interaction and the other is audience social relationship. In GNN-based social recommendation, the input representations of the model are usually audience-item interaction graphs, audience-audience social graphs, and joint graphs that combine the two.

In addition, to enhance the representation of audience preferences by using audience social relations, there are usually two strategies to combine the information from these two networks. One is to learn the corresponding audience representations from these two networks separately and then integrate them into the final audience preference vector, and the other is to combine the two networks into a unified network structure and apply GNN to disseminate the information. The advantage of the first strategy is twofold: on the one hand, the depth of the diffusion process of the two networks can be distinguished due to the separate treatment; on the other hand, any state-of-the-art approach for audience-item interaction graphs can be directly applied, especially for homogeneous graphs such as social networks, for which the GNN techniques are ideally suited to simulate the influence process, as they were originally proposed for homogeneous graphs. For the integration of audience representations learned from two relationships, there are two main mechanisms, namely linear and nonlinear combinations. The advantage of the second strategy is that the diffusion of higher-order social influence in the audience social graph and the diffusion of interest in the audience-item interaction graph can be simulated in a unified model, where both types of information reflect audience preferences at the same time, and a commonly used integration strategy is the hierarchical aggregation model.

Graph Convolutional Networks

Graph Convolutional Network (GCN) is a neural network model for processing graph-structured data. Compared to traditional neural network models, GCN specifically models nodes and edges in graph data, which is suitable for various graph data applications scenarios such as molecular structures, social networks, recommendation systems, etc. The basic idea of GCN is to compute node representations by utilizing the neighboring information of nodes to capture the topology of the graph and associations between nodes. Its core operations include feature propagation and aggregation. Firstly, the features of each node will be passed and aggregated by the features of the neighboring nodes, and then combined with its own features to obtain the updated node representation. This process can be recursively performed by multilayer GCN to obtain more global information about the graph structure.

Specifically, it is assumed that there is a graph denoted as G={V,E}, where V is the set of nodes and E is the set of edges [16]. For each node, it has node features denoted as xi and the adjacency matrix is utilized to represent the connectivity between the nodes.The computation of the GCN layer layer is shown in Equation (1). H(l+1)=σ(D^12A^D^12H(i)W(l))

where H(i) is the node representation matrix of layer (1), H(0) = X, X are the node feature matrices. D˜12A˜D˜12 is the normalized adjacency matrix, which aims to normalize the adjacency matrix A to balance the contribution of nodes with different degrees to the information aggregation. W(l) is the learnable weight matrix for layer (1), A^=A+I is the adjacency matrix (A) plus the normalized matrix for self-connectivity, which serves to enable each node to retain its own information even when aggregating information from its neighbors. D^ is the degree matrix of A^, and σ(·) is the activation function, usually ReLU or other nonlinear function, which aims to enhance the expressive power of the model. In each layer, the above formula propagates and aggregates the features of the nodes through the neighbor matrix and combines them with their own features and weights to obtain a new node representation.

Graph Attention Networks

Graph Attention Network (GAT) is a deep learning model for graph data learning, which utilizes the attention mechanism to efficiently capture the relationships and feature representations between nodes in a graph.The key idea of GAT is to introduce an attention mechanism that allows each node to assign different attention weights to the features of its neighboring nodes when aggregating them, so that the aggregation weights of the features can be dynamically adjusted according to the relationships between each pair of nodes, thus more accurately capturing important relationships among nodes and resulting in better robustness and performance of the model. Thus, the important relationships between nodes are captured more accurately, resulting in better model robustness and better performance. The calculation process is shown in Eqs. (2) to (2). eij=LeakyReLU(a[WhiWhj]) αij=exp(eij)kNiexp(eik) hi=σ(jNiαijWhj)

where a is the adaptive attention mechanism parameter vector, W is the learnable weight matrix, hi, hj are the feature representations of node i, j, respectively, and ∥ denotes the splicing operation. Ni is the set of neighboring nodes of node i, and hi is the feature representation of the updated node i. σ() is the activation function, and the commonly used ones include ReLU, sigmoid and tanh.

Meanwhile, in order to enable the self-attention to represent the nodes stably, the Multi-Head Graph Attention Network (Multi-HeadGAT) mechanism is introduced to improve the model’s representation ability, and each attention head can focus on different relationships and feature combinations. Its calculation process is shown in Equation (5). eijk=LeakyReLU(ak[WkhiWkhj]),αijk=exp(eijk)k=1KlNexp(eilk)

where eijk denotes the attention weight between node i and node j computed by the knd attention head, ak is the adaptive attention mechanism parameter vector of the kth attention head. Wk is the learnable weight matrix of the kth attention head. hik denotes the feature representation of node i updated by the kth attention head. Finally, the features of multiple heads are spliced or averaged to obtain a richer node representation. The computation process is shown in Eqs. (6) and (7): hik=k=1Kσ(jNiαijkWkhj) hik=σ(1Kk=1KjNiαijkWkhj)

Research on Civics Content Recommendation Algorithm Based on Graph Neural Networks
Attention Memory Module

Audience-ideological and political content attention memory module

The memory matrix in the Audience-Ideological and Political Content Attention Memory Module (UP-AMM) is represented by ML×d, L represents the number of memory slices of the memory matrix, and it can be seen that the memory matrix is composed of L memory slices, and the lth memory slice is represented by Mld, modeling the audience’s preference for a certain aspect of ideological and political content, and d represents the dimension of each memory slice, which is equal to the dimension of the audience embedding vector or the ideological and political content embedding vector. The input of the UP-AMM module is the embedding vector tuple (ui,pj,or) of the audience, the ideological and political content and the audience’s evaluation of the ideological and political content, wherein ui represents the embedding vector of the audience i, pj represents the embedding vector of the ideological and political content j, or represents the audience i’s evaluation of the ideological and political content j and the embedding vector of r. The output of the UP-AMM module is an information vector vij, which represents the information vector of the audience’s i‘s preference for ideological and political content j.

Different audiences will have different evaluations of the same Civic content, assuming that audience i evaluates Civic content j as r. In order to improve the modeling of the audience’s preference for Civic content, this module splices the Civic content embedding vector pj with the evaluation embedding vector o, and obtains the Civic content-evaluation joint embedding vector qj through the Multi-layer Perceptron, as shown in Eq. (8) [17]. qjMLP([pjor])

where the qj vector dimension is the same as pj and or. The MLP adopts a two-layer structure in its implementation.

The different memory slices in the memory matrix can be regarded as storing the audience’s preference information on different aspects of the Civic content. Therefore, the Civic Content-Evaluation Joint Embedding Vector qj and the Memory Slice Mi can be combined through Eq. (9) to obtain the preference information of the Civic Content j stored in the Memory Slice l. mji=qjMi

where ⊙ represents the product of vector elements, mjt has the same dimension as qj, and there are L vectors of mji related to ideological and political content j.

Preference information is closely related to the audience. The preference information stored in the L memory slices of Civic Content j is obtained from equation (9), but the importance of preference information varies from person to person, so the attention mechanism is used here to distinguish the preferences that are more important to this audience and assign a higher weight to them, and ultimately integrate all the preferences: αij*=(uiqj)Kl αijt=exp(αijt*)l[1,L]exp(αijt*) vij=l[1,L]αijtmijt

where • denotes the inner product of vectors, and Kld denotes the vector of key values associated with Ml, the number of which is also the same as the number of Ml. By using the above equation, the final information vector vij is obtained from the output of UP-AMM module.

The UP-AMM module targets each audience-civic content interaction to mine the audience’s intrinsic multifaceted preferences for the civic content, and learns an optimal intrinsic preference interpretation vector for each audience-civic content interaction. Ideally, this intrinsic preference interpretation vector captures the hidden semantics between each interaction.

Audience-Audience Attentional Memory Module

The Audience-Audience Attention Memory Module (UU-AMM) is used to capture the multifaceted influences of neighboring audiences on the target audience, and its structure is similar to the UP-AMM. This module uses memory matrix M to store the multifaceted influences of neighboring audiences on the target audience, and the inputs of the module are the embedding vectors of the target audience i and the embedding vectors of the neighboring audiences j that are socially related to the audience i, denoted by ui and uj, respectively. The output of the module is the information vector vij, representing the multifaceted influence of audience i by audience j.

The UU-AMM module is similar in principle and structure to the UP-AMM module, with the difference that the input of the UU-AMM module is the audience embedding vector instead of the joint civic content-appraisal embedding vector in the UP-AMM module.The computational procedure of the UUAMM module is shown in Eqs. (13)-(16). mjt=ujMl αijt*=uiujKl αijt=exp(αijt*)l[1,L]exp(αijt*) vij=l[1,L]αijtmj

The UU-AMM module is controlled by the audience-audience social relationships, making the learned multifaceted influence vectors unique to each audience-audience pair. Thus, this can be interpreted as an exclusive optimal translation vector learned for each audience-audience interaction relationship.

The joint audience-evaluation embedding vectors are obtained without using the attentional memory module: vijMLP([uior])

On the other hand, it is not reasonable to analyze the contents of the Civics and Politics in terms of its multifaceted preferences, so the Attentional Memory Module is employed only in the process of information dissemination in the audience node, which completes the first stage of the information dissemination process of the graphical neural network - the work of information construction.

Attention aggregation module

In order to solve the problems caused by equal weight aggregation, this subsection designs the Attention Aggregation Module (AAM) to dynamically distinguish the importance of neighboring nodes, differentially aggregate the information propagation vectors of neighboring nodes, and complete the second stage of the information propagation process of graph neural networks - information aggregation. Specifically, combined with the attention mechanism, the implementation in this module is: βij*=MLP([vivij]) βij=exp(βij*)jN(i)exp(βij*) xi=1|N(i)|jN(i)βijvij

where vij denotes the output vector of the Attention Memory Module, vi denotes the target node embedding vector, βij* denotes the similarity between the output vector vij of the Attention Memory Module and the node embedding vector vi, xi denotes the output vector of the node i in the AAM Module, and N(i) denotes the set of neighboring nodes of the node i in the Audience-Audience Social Graph or the Audience-Thinking Content Dichotomous Graph.

Model Prediction and Learning

The model prediction module is the last module in the GAMSN algorithm, which takes over the audience node hidden vectors and the Civics content node hidden vectors output from the attention aggregation module, and obtains the audience’s predicted score for the Civics content through the computation of the model prediction module, which indicates the audience’s preference of Civics content, and at the same time, the difference between the predicted score and the actual score reflects the algorithmic model’s recommended Accuracy. Specifically, this module first splices the audience embedding vectors and the Civics content embedding vectors obtained from the attention aggregation module, and then obtains the audience’s predicted ratings of the Civics content by means of a multilayer perceptron: r^ijMLP([xixj])

where r^ij represents the predicted rating of audience i on the ideological content j.

In order to make the results predicted by the model closer to the real value, model training, i.e., the parameter learning process, is needed. Through the error feedback between the predicted value and the real value, back propagation is carried out to adjust the parameters of the model and gradually make the predicted value closer to the real value in order to achieve the purpose of accurate prediction. The mainstream practice is to first set a suitable objective function, i.e., loss function, and then optimize it using an optimization algorithm. The loss function used in this algorithm is shown in equation (22): L=12|O|i,jO(r^ijrij)2

Where, O denotes the data sample set, r^ij denotes the predicted rating of audience i on the Civic Content j, and rij denotes the true rating of audience i on the Civic Content j. The square of the difference between the model’s predicted ratings and the true ratings of audience i for Civic content j in the sample is calculated, and then the same operation is performed for each sample tuple in the data sample set, and finally summed to obtain the value of the loss function.

The model parameters were updated in the negative gradient direction using the RMSProp optimization algorithm.

GAMSN algorithm flow

The GAMSN algorithm utilizes the information about the interaction between the audience and the ideological content and the information about the social relationship between the audience and the audience for the prediction of the audience’s rating of the ideological content.

In summary, the graph-attention-memory social network algorithm based on graph neural networks proposed in this chapter has the following main contributions:

Modeling the audience’s preference for a certain aspect intrinsic to the Civic politics content, and setting the weight of this aspect’s preference and integrating all the preferences through the attention mechanism, so as to learn an optimal multifaceted preference vector representation for each audience-Civic politics content interaction relationship, completing the construction of information vectors in the process of information dissemination of graph neural networks, and ultimately improving the recommendation ability of the algorithm.

Modeling the influence factors of the audience on the target audience in the social network, and distinguishing the intensity of the influence factors of different aspects on the target audience through the attention mechanism, so as to learn the multifaceted influence factors of the audience on the audience in the social relationship, and constructing the information dissemination vector. The information dissemination vector is propagated by the information dissemination of graph neural networks to achieve the purpose of using social relations to improve the recommendation effect.

Based on the graph neural network, we aggregate the multifaceted preference information of the neighbor nodes in the audience-Civic content dichotomous graph and the multifaceted influencing factors information of the neighbor nodes in the audience-audience social graph, and use the attention mechanism to differentiate the importance of the different neighbor information for differentiated processing. The recommendation performance of the algorithm can be improved by aggregating information more effectively by graph neural network.

Results and Discussion
Algorithm performance experiment

The datasets used in the experiments are those that are publicly available on the Internet and widely used in collaborative filtering research, covering POI recommendation, item recommendation, website recommendation, and many other aspects. Table 1 displays the characteristics of every dataset, which includes basic descriptions like data size and sparsity degree.

Overview of data sets

Data set size Sparse degree describe
Gowalla 29864*4099 0.00085 Poi recommends data sets
Amazon-book 47516*91227 0.00054 Amazon book scoring data set
TSD 25913*57794 0.00034 Amazon electronic product scoring data set
NATCD 1895*4487 0.0065 NATCD music data set
Delicious 1866*69224 0.0016 User-url data set
Yelp2018 31666*38045 0.0016 The yelp contest data set in 2018

In the dataset, Gowalla and Yelp2018 data do not contain rating information. A 0 or 1 in the data represents whether or not there is an audience interaction with the item. The data in the two Amazon datasets are the audience’s ratings of the items, with a distribution of scores ranging from 1 to 5. Except for the multitask learning model that learns the ratings, the rest of the models convert the ratings into 0s and 1s during the data preprocessing process.The data values in the NATCD, on the other hand, represent the length of time that the audience has been playing a particular piece of music, and to a certain extent, the degree of the audience’s preference can be represented by the duration of time, which is a way to show the feedback of the The Delicious dataset is the dataset of the audience and the corresponding browsing preference URLs. In both NATCD and Delicious, there are both audience-item interaction matrices outside and social relationship matrices between audiences. One step of data censoring is also required in the data preprocessing process. Consistent with experiments in other work, a minimum number of interactions 10 is kept in the different datasets, and if the number of interactions for an audience (item) is less than 10, then the node is deleted to avoid interference from extreme nodes.

Experiments on Propagating Recommendation Prediction Tasks

Two GNN modeling frameworks based on multi-task learning are given in Chapter 2. The two architectures respectively add the multi-task learning module on top of the existing graph neural network model as a way to sub-task to learn interactivity and scoring information. In this paper, LightGCN is used for the underlying model, and MMoE module and MLP module are added on top of LightGCN model.

In this experiment, the GNN is fixed to two layers. After repeated hyper-parameter debugging to take the optimal results, the hyper-parameter α (the weight of the prediction scoring score item) is set to 0.1.The experimental results on the representative NATCD New Era Civics Curriculum Database (NATCD) and the Civics Content Learning Thematic Dataset (TSD) datasets are given here.In the Civics Content Learning Thematic Dataset, the results on the Civics Content Learning Thematic Dataset are consistent with that on the e-Civics Content Dataset. In order to save space, only the former experimental results are given.The experimental results on NATCD and TSD are shown in Table 2 and Table 3 respectively.In the process of data preprocessing, the NATCD data are mapped into 1-5 by the method of bucket division of the original length of listening to music, and the higher the score value represents the greater interest.

NATCD data set experiment results

Model recall@20 recall@40 recall@60 ndcg@20 ndcg@40 ndcg@60
LightGCN 0.283 0.398 0.472 0.221 0.264 0.284
LightGCN+mlp 0.293 0.4 0.476 0.229 0.264 0.287
LightGCN+MMoE 0.204 0.3 0.372 0.148 0.179 0.202

TSD data set experiment results

Model recall@20 recall@40 recall@60 ndcg@20 ndcg@40 ndcg@60
LightGCN 0.042 0.065 0.082 0.033 0.043 0.044
LightGCN+mlp 0.043 0.068 0.083 0.036 0.039 0.047
LightGCN+MMoE 0.022 0.039 0.05 0.016 0.022 0.025

Through the experimental results, it can be found that when the mlp multitasking architecture is used, the results of the experiments on the NATCD dataset are slightly improved compared to LightGCN, but the indexes will be decreased if the MMoE architecture is used. The reason is that the MMoE model has a large number of parameters to be trained, and too many parameters affect the learning of the embedding vectors of the graph neural network. On the other hand, the mlp architecture only plays an auxiliary role in the learning of scores, and directly applying the output of the GNN to the learning of interactive information will not produce the phenomenon of gradient disappearance, and too many parameters will not affect the learning of interactive information. In addition, the multitasking learning model of MLP architecture does not work on the Civic Content Learning topic dataset to enhance the effect. The reason is that the value on the NATCD dataset represents the time that the audience listens to a certain musician, and the length of time spans a wide range, but the length of time can clearly reflect the size of the audience’s interest. On the other hand, the value on the Amazon series dataset represents the audience’s rating of a specified item, with a value distribution of 1-5, which is more difficult to learn, and the sparseness of the dataset makes the amount of interaction information much larger than the amount of information contained in the ratings, and at the same time, utilizing the mlp module to learn the only rating information brings too much noise to improve the predictive effect of the model. For this type of dataset, there is no need to build a new module to learn the rating information.

Visualization analysis

In this section, the embedding vectors of this paper’s algorithm are shown through visualization. After GNN propagation and BPR loss training, the embedding vectors of audience and items are obtained. The closer the embedding vectors of the two are to the audience’s preference for the item, the smaller the smoothing loss between the embedding vectors will be. Therefore, for the central node audience and the K items whose computed inner products are the largest, the embedding vectors of these nodes are visualized in reduced dimensions, and the distribution of the embedding vectors of audience and items is observed. The K is set to 20 in the experiment, the embedding vectors of the Gowalla dataset are used for the examples in the figure, and the dimensionality reduction method is the TSNE method, which is widely used in the visualization field. The embedding vector is downscaled to two dimensions using this method. The visualization results are shown in Fig. 1. As a comparison, the NGCF model embedding results are visualized as well. Where Fig. 1(a) shows the embedding vector visualization results obtained from this paper’s model and Fig. 1(b) shows the results obtained from the NGCF model. Where the triangular morphological items represent the audiences, and the rest of the colors represent the K items that are closest to different audiences.

Figure 1.

Compare the visual results of the Gowalla data set

The data points in the figure are close together because of the relationship between the embedding vectors. By observing the visualization results, it is found that the embedding vectors of items preferred by a single audience are distributed in close proximity to that audience. Compared with the NGCF model, the embedding vectors of the audience-preferred items obtained from the training of the model in this paper are more proximate to the embedding vectors of that audience, and the smoothing loss is smaller. The visualization analysis results can also explain that this paper’s model can achieve better recommendation results.

Integration into social network information analysis

The model in this paper is able to achieve excellent prediction results on the collaborative filtering task, and a step of extended experimentation is conducted in this section. The model of this paper incorporates social information, and the aggregation process is shown in Fig. 2. Consider both social network information and similarity network information in the propagation process of second-order audience embedding vectors, and observe whether it can achieve better experimental results.

Figure 2.

Social information and similarity information dissemination

Experiments incorporating social network information are conducted on the NATCD and Delicious datasets to examine the structure of this paper’s model incorporating social network information, with 1,000 rounds of iterations and a learning rate fixed at 0.001. In order to examine the characteristics of the networks, second-order propagation of GNNs is conducted, i.e., first-order relationships are accomplished on the original audience-item interaction graphs, and second-order relationships are accomplished by utilizing multiple networks on the The experimental results on the NATCD and Delicious datasets correspond to Tables 4 and 5, respectively.

The prediction of the NATCD data set is compared

Model recall@20 recall@40 recall@60 ndcg@20 ndcg@40 ndcg@60
GNN 0.288 0.407 0.477 0.226 0.266 0.287
Social 0.286 0.394 0.471 0.226 0.263 0.286
This model 0.295 0.409 0.482 0.23 0.269 0.293

The prediction of the Delicious data set is compared

Model recall@20 recall@40 recall@60 ndcg@20 ndcg@40 ndcg@60
GNN 0.084 0.121 0.143 0.082 0.098 0.107
Social 0.086 0.122 0.142 0.084 0.098 0.112
This model 0.085 0.127 0.146 0.084 0.105 0.114

In the comparison method, social refers to the second-order propagation by utilizing social relationships instead of similarity relations. Through experiments, it can be found that after fusing the similarity propagation relationship and social relationship, the prediction results of the model in this paper are improved over the baseline model. To analyze the reason, the similarity propagation matrix is constructed by using the similarity degree of the audience’s preferred items, while the social network information is taken from the real social information, and there are some differences in the information contained in the two. After capturing more valuable information, the model’s prediction results are enhanced.

Effectiveness of Civic Communication Application

In this section, 56,693 ideological contents and 205,097 events are obtained by simulating the dissemination of ideological contents on the social network within 30 moments, and analyzing the data of the ideological content creation crowd, the distribution of the emotion of the ideological contents, the characteristics of the dissemination of the ideological contents and the distribution of the period of the dissemination of the ideological contents to obtain the overall dissemination trend. The retweet volume statistics for all Civic and political content on this network are shown in Figure 3.

Figure 3.

Ideological and political content Forward statistics

As can be seen in Figure 3, the top 120 or so Civic and political content with the highest forwarding volume occupy all Civic and political content forwarding events, from 120 to 1500 volume exists but the value is very low, very few, and there is almost no forwarding data after 1500. As the social network set up in this paper is a 5000-people network, it will generate a large amount of time cost to count the propagation law of all Civic Politics contents when performing propagation calculations, in order to save time and resources, this paper selects the representative top 100 Civic Politics contents with the highest retweeting volume to analyze the propagation law.

Civics Content Creation Crowd Analysis

Table 6 displays the statistics on the number of Civics content created by different audience groups for all Civics content and the top 100 Civics content with the highest forwarding amount.

Different user groups create content ratios

Type Distribution Core contributor Active participant General participant Diver
Full content Quantity 33725 19337 6686 396
Proportion 56.07% 32.15% 11.12% 0.66%
Top 100 content Quantity 58 36 6 0
Proportion 58.00% 36.00% 6.00% 0.00%

From Table 6, it can be seen that the core contributors occupy the majority of the Civics content on the whole network, amounting to more than 50%, and in the top 100 Civics content with the highest retweeting volume, the core contributors’ creation volume is 59. Secondly, the active participants account for more than 30% of the contribution volume. The contribution of core contributors and active participants has reached more than 80%, nearly 90%, of the total Civic-Political content on the web. This value is also more apparent in the top 100 retweets. In addition, the conditions of audience type show that the number of core contributors accounts for 10% of the total number of people on the whole network, and the number of active participants accounts for 20% of the total number of people on the whole network, i.e., 30% of the people create 90% of the contents, which indicates that a large amount of contents are created by a small number of audiences, which is in line with the real network rules nowadays.

Relative Emotional Analysis of Civics Content

Relative sentiment expresses the overall emotional tendency of the audience on the current social network, which includes four main situations: neutral, positive, negative, and conflict. When the proportion of audiences with neutral attitudes is higher than the number of audiences with positive attitudes and higher than the number of audiences with negative attitudes, the overall sentiment tendency of the network is neutral. When the number of audiences with positive attitudes in the network is higher than the number of audiences with negative attitudes and the number of audiences with neutral attitudes, the overall affective tendency of the network is positive. When the number of audiences with negative attitudes in the network is higher than the number of audiences with positive attitudes and the number of audiences with neutral attitudes, the overall affective tendency of the network is negative. Other than the above three cases, there is conflict. The statistics of different emotional tendencies in all Civics contents and the top 100 Civics contents in terms of retweets are shown in Table 7.

Affective type proportion

Type Distribution Neutrality Positive Negativity Conflict
Full content Quantity 1657 1043 955 234
Proportion 42.61% 26.82% 24.56% 6.02%
Top 100 content Quantity 40 32 26 2
Proportion 40.00% 32.00% 26.00% 2.00%

The highest proportion of the network’s overall sentiment type is neutral, amounting to 40%, followed by positive and negative, with a small difference between the two, each accounting for about 25% or so, and the least is conflict, accounting for less than 10%. Table 3-6 shows that most of the general audience is in a non-vocal state, but the overall proportion of positive and neutral in all the published Civics content is close to 70%, indicating that the overall tendency of the network is in a healthy and optimistic direction. Observing the emotional distribution of the top 100 Civics contents, the proportion with neutral and positive tendencies is even more than 70%, and the negative and conflict tendencies have a decreasing trend, indicating that the emotional tendency of the audience creating the top 100 Civics contents is more positive than the overall network tendency.

Whether from the analysis of the creation crowd or from the analysis of network emotion distribution, the overall network trend is developing well, which is in line with the current network cognition and development law and fits the reality, and the network status of the top 100 Civic and Political contents in terms of forwarding volume is more obvious, and the analysis of its dissemination characteristics can be used as a substitute for the prediction of the overall environment of the network.

Characterization of the dissemination of Civics content

In this paper, we choose spreading speed, spreading range, spreading distance and spreading durability to express the spreading characteristics of ideological and political content. The calculation rule for dissemination speed is: the moment when ideological content is first forwarded minus the moment when the content is created. The calculation rule of spreading range is: the number of audiences who forwarded the ideological content when it was forwarded for the first time. Distance of dissemination is calculated as follows: the distance of dissemination is increased by one for each audience that the ideological content is forwarded to, i.e., passes through an audience. The calculation rule for transmission durability is: the moment when the ideological content was last forwarded minus the moment when it was first forwarded. And through the dissemination characteristics of the top 100 forwarding volume of the Civics content by calculating the maximum value, minimum value, mean value and standard deviation as shown in Table 8. The dissemination distance and persistence of the top 100 forwarded civic and political contents are plotted as shown in Figure 4.

Forward quantity propagation line diagram

Fastest Slowest Mean Standard deviation
Propagation velocity 0 1 0.14 0.327
Propagation range 1 4 1.26 0.539
Propagation distance 7 99 55.4 30.726
Propagation persistence 0 27 13.42 9.313
Figure 4.

The spread of the spread of the Top 100

As can be seen from Table 8, the top 100 Civic and Political contents in terms of retweeting volume will all be retweeted in a relatively short period of time, and the range of the first time they are retweeted is small, and the dissemination speed and the range of the dissemination are kept at a stable level. Combining Table 8 and Figure 4, observing the dissemination speed, the top 100 Civic and political contents with retweeting volume have a fast audience response, basically they will be retweeted immediately after creation, and at the slowest, they will be retweeted after only one moment. Observing the dissemination range, the maximum value is 3, the minimum value is 1, and the average value is 1.25, with a relatively small standard deviation, which shows that the top 100 Civic and political contents are not aimed at a wide range of audiences, and the range of audiences who forwarded them for the first time is relatively small. Observing the spreading distance, the difference between the minimum value and the maximum value is large, and the standard deviation is large, which indicates that the fluctuation is large and subject to external influence factors. Observe the transmission persistence, which represents the influence of the discussion degree of the existence of the information, with high and unstable ups and downs, indicating that the changes are also larger for different Civic and Political contents.

Analysis of the dissemination cycle of political content

The development of each state of affairs is characterized by cyclicality, and each cycle corresponds to the corresponding communication law, which generally includes three states: latent period, explosive period, and receding period. To observe the dissemination cycle of political content, this paper analyzes the dissemination cycle of the top 10 political content in terms of retweets, and the results are shown in Figure 5.

Figure 5.

The propagation cycle of the forward quantity of the TOP 10

It can be seen that the dissemination cycle of the Civic and political content with high retransmission volume in the network is generally 3-4, all of which have experienced latency - outbreak - recession, and the outbreak period is generally in the middle of the whole cycle or in the front position, and rarely appears the phenomenon of staggered backwardness, and the peak volume of transmission single period were all above 1000. Civic and political content with high retweeting volume is prone to trigger the outbreak of public opinion in social networks, and mastering the pattern of the dissemination cycle of high retweeting volume of Civic and political content is a good means of controlling and predicting public opinion.

Conclusion

In this paper, in order to realize the innovation of ideological and political education dissemination mode, we propose the innovative application strategy of “Artificial Intelligence +” ideological and political education dissemination, and construct the ideological and political content recommendation algorithm based on graph neural network. The algorithm is tested and the communication effect of the innovative strategy is analyzed. It is found that the embedding vectors of the items preferred by the audience obtained after the training of the model in this paper are more similar to the embedding vectors of this audience, and the smoothing loss is smaller. The visualization analysis results can also explain that the model in this paper can achieve better recommendation results. Whether from the analysis of the creative population or from the analysis of network sentiment distribution, the overall network trend is well developed, in line with the current network cognition and development laws, fit the reality, and the top 100 forwarding volume of the ideological and political content of its network status is more obvious, analyzing its dissemination characteristics can be a substitute for the prediction of the network environment as a whole, so as to improve the ideological and political education dissemination efficiency and effect.

Sprache:
Englisch
Zeitrahmen der Veröffentlichung:
1 Hefte pro Jahr
Fachgebiete der Zeitschrift:
Biologie, Biologie, andere, Mathematik, Angewandte Mathematik, Mathematik, Allgemeines, Physik, Physik, andere