Mobile QR Code QR CODE

  1. (Basic Teaching Department, Zhengzhou Railway Vocational and Technical College, Zhengzhou 450000, China caili78922@163.com )



Information dissemination, Interlayer coupling, Maximize, Node sorting, Sequential network

1. Introduction

People’s daily lives are significantly impacted by their participation in online social networks [1]. Through online social networks, people can acquire and disseminate information that can be communicated and exchanged online [2]. The important nodes of a social network have a greater impact on the function and structure of the network, and the mining of important nodes is the focus of current social network analysis. Mining important nodes in social networks have specific applications in different fields [3]. Most studies have been based on static networks to mine important nodes. Among them, research on node ranking mainly includes topological structure-based recognition methods and dynamics-based recognition methods [4]. Although these research results are fruitful, there are also problems. For example, the social network is simplified into a static graph, ignoring the time characteristics of the social network [5]. Based on these problems, this paper proposes the modeling method of the super adjacency matrix temporal network and the influence maximization algorithm of the temporal network based on interlayer coupling intensity attenuation. At the same time, the centrality of the eigenvector is introduced to assess its importance. In addition, attenuation factors are also introduced to characterize the interlayer coupling strength. The first part of this paper is an overview of the research on mining and the ranking of essential nodes in online social networks. The second part outlines the construction of the research model. The third part shows the design of the influence-maximization algorithm and visualization. The fourth part reports the conclusion of the study.

2. Related Work

Online social networks have become increasingly essential because of the proliferation of mobile devices and advances in Internet technology. Hence, there has been increasing research on the mining and ranking of important nodes in online social networks. Tian and Chan [6] proposed a spatio-temporal attention wave network multi-step prediction model to capture the cumbersome temporal and spatial correlations and predict road traffic conditions. The model was embedded by developing self-learning nodes requiring zero prior knowledge of graphs. The model possessed a good prediction effect on multiple data sets. Yu [7] proposed a systemic assessment based on traffic volume in the network to assess the key points of industrial system nodes. The method involved the establishment of an evaluation index and the cascading effect model. This technique was useful for determining which nodes in industrial systems are the most crucial and designing an optimization scheme based on reducing consumption. Yang et al. [8] proposed one way to extract features for fault diagnosis based on the spatiotemporal graph to obtain the high-level characteristics from vibration data. This method extracts the feature vector from the space-time graph and transforms the space-time graph into a single vector to construct a hypergraph. This method has a good effect on extracting high-level features. An and Shen [9] proposed a model for seeing three dimensions in vehicle-mounted networks to examine the coupling and spread of information in visual 3D perception. The framework uses information ellipsoids and represents the coupling with visual infographics. This method provided guidance for the development of 3D visual perception technology. Kumar et al. [10] suggested a natural heuristic method based on an improved forest fire model. This method could mimic information dissemination via internet social networks. Under the verification of multiple data sets, this method performed well in predicting information diffusion. Singh et al. [11] constructed an influence-maximization method to address the issue of online social network influence maximizing. The algorithm avoided the problem of early convergence through a probabilistic environment. The proposed algorithm had many advantages.

Wang et al. [12] proposed a trust-based solution to simulate the propagation of competitive influence. This scheme involves a diffusion model of competition influence based on trust and an efficient algorithm of competition influence maximization. This method is more effective than other algorithms. To predict depression, experts, such as Uddin et al. [13], proposed a model based on deep temporal and spatial features and multiple layers of bidirectional long short-term memory. The model obtained the time information by integrating the time-median pool. The predictive effectiveness of the proposed model was strong. Casarin et al. [14] proposed a new multilayer network analysis method to analyze the relationship between multiple drivers in the oil market. This method involves a Bayesian graph vector autoregressive model. This method can reveal the relationship between the driving factors in the oil market. Fang et al. [15] proposed a spatiotemporal slow-graph convolutional network model to identify the skeleton operation. They designed the skeletal data for the adjacency matrix of the graph. The model can be used to design a small spatiotemporal graph by the junction of adjacent frames. This model could capture the long-range and short-range spatio-temporal joint relationships and complete action recognition at low cost. Ghasemi and Ahmadi [16] used the metadata in online social networks to predict the time pattern based on the average propagation ability of each user in the network, manage the cache in online social networks, and optimize the underlying cache management mechanism. This method could better manage the cache in online social networks. Thabit et al. [17] proposed a model based on trust management and data protection to protect data in online social networks. Trust management in this model can identify the accuracy of trust factors. This model has good performance and computational efficiency and effectively solves privacy risks such as data leakage.

Many studies have been conducted on mining and ranking important nodes in online social networks. On the other hand, these studies also have problems, such as simplifying social networks into static graphs and ignoring the temporal characteristics of social networks. Therefore, this paper proposes a network modeling method with interlayer coupling-intensity attenuation and time-network-influence maximization. The centrality of eigenvectors is introduced to evaluate the importance of nodes. In addition, attenuation factors are also introduced to characterize the interlayer coupling strength.

3. Node Ranking and Influence Maximization Algorithm Design based on Interlayer Coupling Strength Attenuation

A super adjacency matrix-time-network modeling method was constructed to sort the nodes of the time network. The model includes the attenuation factor to describe the interlayer coupling strength more accurately. An algorithm combining heuristic and greedy strategy was proposed to determine the impact of the nodes of a temporal network.

3.1 Construction of Superadjacency Matrix Temporal Network Modeling Method based on Interlayer Coupling Intensity Attenuation

Finding critical nodes in a network has long been a popular area of study. A technique for modeling temporal networks that uses interlayer coupling intensity attenuation, the super-based adjacency matrix (ASAM), was proposed to represent the changing process of the actual network more appropriately, taking the multilayer graph temporal network model as the foundation. In the specific application of temporal networks, individual users are usually regarded as nodes in the network and generally defined as $G=\left(V,F\right)\,.$ Where $V=\left\{v_{1},v_{2},\ldots ,v_{N}\right\}$ represents the node set that contains all nodes. $\left[0,m\right]$ is the time period; The triplet $f_{t}=\left(i,j,t\right)$ is the set of edges of the temporal network on $\left[0,m\right]$; $i$ and $j$ denote the nodes in the network; $t$ is the time to establish the connections for nodes. $F=\left\{f_{1},f_{2},\ldots ,f_{t}\right\}$ is the network edge set of all triples. The temporal network may be split into $T$ time windows using time intervals $w$, and the temporal network becomes a discrete slice network $G_{1},G_{2},\ldots ,G_{T}$. Sequential network modeling is the basis of sequential network node ranking and influence maximization algorithms. Temporal network modeling is divided mainly into three types: time aggregation graph, snapshot-based temporal network model, and temporal network model based on obvious path flow [18]. The time aggregate graph model can show the dynamic evolution of the actual network more realistically, but it is not easy to store and calculate the matrix. Snapshot-based temporal network models are divided into a multilayer graph temporal network model and a time window model [19]. Fig. 1 illustrates the time window model.

Fig. 1. Time window graph model with a time window size of 2.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig1.png

Fig. 2. Multilayer graph temporal network model with a time window size of 2.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig2.png

Fig. 3. Time series network model based on obvious runoff.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig3.png

The time interval of the time window model at this time was 2, which affects the time window size (Fig. 1). A, B, C, and D are the nodes of a temporal network, and the interactions among nodes are represented by the lines connecting them. Snapshots at different times are displayed as a time aggregation graph in chronological order. The model can show the evolution process of the event. Fig. 2 presents the multilayer diagram sequential network model.

Fig. 2 denotes the relationship between layers. The model also includes four network nodes, A, B, C, and D, and lines represent the interaction among nodes. The multilayer graph temporal network model generally comprises intra-layer and inter-layer relationships. The interactions among nodes are represented mainly by intra-layer relationships, and there are corresponding snapshots. In adjacent snapshots, the connection relationship between the corresponding nodes is expressed mainly through the inter-layer relationship. Fig. 3 shows the sequential network model based on obvious channel runoff.

The time-series network model based on obvious path runoff contains more nodes and more connections among nodes (Fig. 3). The horizontal axis below the model represents the time scale, which results from further refinement of the model in the time dimension. In addition, the interaction between nodes occurs only between two adjacent time layers, and the nodes only connect themselves in the preceding and following time layers. Overall, the evaluation methods of node importance can be split into three categories: evaluation based on network effectiveness changes, evaluation based on the propagation model, and evaluation based on the correlation [20]. The evaluation calculation based on network performance changes is expressed in Eq. (1):

(1)
$ V\left(G\right)=\frac{E\left(G\right)-E\left(G\backslash \sigma \right)}{E\left(G\right)} $

where $E\left(G\right)$ is the network timing performance before important nodes are deleted; $E\left(G\right)$ is the deletion ratio of the nodes; $E\left(G\right)$ is the network timing performance after the cumulative deletion of $E\left(G\right)$ proportional nodes. The network timing performance is usually evaluated using the network timing efficiency and the maximum connectivity component of network timing. The network timing efficiency was calculated using Eq. (2):

(2)
$ E_{t}\left(G\right)=\frac{1}{N\left(N-1\right)}\sum _{u\neq v}\frac{1}{d_{t}\left(u,v\right)} $

where $d_{t}\left(u,v\right)$ denotes the temporal distance between nodes opposite sides $\left(u,v\right)$ of time $t$; $N$ denotes the total quantity of nodes. Eq. (3) shows the calculation of the maximum connected component of network timing.

(3)
$ LCC^{T}=\frac{1}{m}\sum _{\partial =1}^{m}\frac{\max _{1\leq \partial \leq k}\left(\left| \left.C_{\partial }^{\psi }\right| \right.\right)}{N} $

where $\left| \left.C_{\partial }^{\psi }\right| \right.\left(1\leq \partial \leq k\right)$ denotes the quantity of nodes in $\partial $ connected graph in the $\psi $ snapshot, and$k$ is the number of connected components in each snapshot. The evaluation based on the correlation mainly involves the Kendall correlation coefficient and accuracy. The Kendall correlation coefficient was calculated using Eq. (4):

(4)
$ \tau =\frac{n_{c}-n_{d}}{0.5\times \xi \times \left(\xi -1\right)} $

where$\xi $ is the logarithm of the sequence; $n_{c}$ is a sequence consistent quantity; $n_{d}$ is the number of sequence inconsistencies. The value interval of $\tau $ is $\left[-1,1\right]$, and the greater the link, the higher the value. The accuracy rate was calculated using Eq. (5):

(5)
$ precision=\frac{\left| \left.P_{\upsilon }\cap R_{\upsilon }\right| \right.}{\left| \left.R_{\upsilon }\right| \right.} $

where $P_{\upsilon }$ and $R_{\upsilon }$ represent the sorting results of the first$\upsilon $ point in real sorting and predicted sorting respectively. The multilayer coupling-network-analysis technique was used in the development process of the Supra-adjacency Matrix (SAM). The SAM matrix is expressed in Eq. (6).

(6)
$ SAM=\left[\begin{array}{llll} A^{\left(1\right)} & \omega I & 0 & \ldots \\ \omega I & A^{\left(2\right)} & \omega I & \ldots \\ 0 & \omega I & A^{\left(3\right)} & \ldots \\ \ldots & \ldots & \ldots & \ldots \end{array}\right] $

where $A^{\left(1\right)},A^{\left(2\right)},\ldots ,A^{\left(T\right)}$ denotes intra-layer connection; $\omega I$ is interlayer connectivity; $\omega $ is an adjustable parameter; $I$ is the identity matrix. The enhanced similarity index (ESI) is a new suggested local similarity index for measuring inter-layer coupling, which was proposed after integrating fixed parameters and local similarity index. The ESI was calculated using Eq. (7).

(7)
$ ESI_{i}^{\left(t,t+1\right)}=0.5+\frac{\sum _{j}a_{ij}^{t}+\sum _{j}a_{ij}^{t+1}}{2N}+\frac{\left| CN\left._{i}^{\left(t,t+1\right)}\right| \right.}{N} $

When $\sum _{j}a_{ij}^{t}=1$, nodes $i$ and $j$ in time layer $G_{T}$ have connected edges. $CN_{i}^{\left(t,t+1\right)}$ denotes the number of common neighbors of node $i$ on both $G_{T}$ networks. Eq. (8) was used to calculate $CN_{i}^{\left(t,t+1\right)}$.

(8)
$ CN_{i}^{\left(t,t+1\right)}=\left| \left.\Gamma _{i}^{t}\cap \Gamma _{i}^{t+1}\right| \right. $

where the First-order neighbor set was denoted by $\Gamma _{i}^{t}$. An ASAM model was proposed to measure the interlayer coupling strength of the adjacent-layer and cross-layer network nodes. In addition, the attenuation coefficient was incorporated into the ASAM model to describe the interlayer coupling intensity more accurately. The attenuation factor was calculated using Eq. (9).

(9)
$ \varphi =e^{-\mu \cdot \Delta t} $

where $\mu $ is the adjustable parameter; $\Delta t$ is the space of time between successive epochs.

(10)
../../Resources/ieie/IEIESPC.2024.13.6.632/eq10.png

where the time values $t_{1}$ and $t_{2}$ are associated with the time-layer network, and $1\leq t_{2}-t_{1}\leq 4$. The research used the centrality of eigenvectors for the calculation to express the importance of nodes, as expressed in Eq. (11).

(11)
$ \theta _{\alpha \beta }=\eta _{N\left(\beta -1\right)+\alpha } $

where $\eta $ is the feature vector; $\theta _{\alpha \beta }$ is the element in the column $\alpha $ of row $\beta $ of the matrix, displaying the eigenvector centrality of node $i$ on the temporal layer $\beta $.

3.2 Design of Influence Maximization Algorithms and Visualization Tools for Temporal Networks

The information dissemination model can predict the future dissemination trend of information and can visualize the process of information dissemination. The Independent Cascade (IC) model and the Linear Threshold (LT) model are influence models, both of which may replicate the process of information dissemination. Fig. 4 shows the propagation mechanism of the IC model.

where $V_{1}$ to $V_{7}$ are nodes, and the value between nodes is the propagation probability $p_{u,v}\in \left[0.1\right]$. Green indicates that the node is currently active, and red indicates that the node is currently inactive. First, decide the parent node of the active state. The second step is to determine whether to activate other nodes according to the size of the propagation probability. The LT model mainly completes the propagation process through the influence weight and influence threshold. There are four kinds of infectious disease models, and the network nodes in these four models have four states. Different models have different node state transitions. Fig. 5 shows the node state transitions of each of the four models.

Fig. 4. Propagation process of the IC model.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig4.png

Fig. 5. Node state transitions for each of the four models.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig5.png

S, I, R, and E represent the susceptible, infected, recovered, and exposed states of the node, respectively (Fig. 5). $\delta $ is the probability of infection; $\gamma $is the probability of cure; $\varepsilon $ is the probability of infection. The susceptible infected (SI) nodes in the SI model have a $\delta $ probability of becoming infected. The SI nodes in the susceptible (SIS) model are likely to $\delta $ transform into infected nodes; A node in an infected state has a $\gamma $ probability of transforming into a susceptible state. The susceptible infected recovered (SIR) nodes have a $\delta $ probability of transforming into susceptible. A node in a susceptible state has a $\gamma $ probability of transforming into a recovered state. The susceptible exposure infected recovered (SEIR) nodes in the SEIR model have a $\delta $ probability of being exposed. A node in the exposed state has an $\varepsilon $ probability of transforming into an infected state. A node in an infected state has a $\gamma $ probability of transitioning to a recovered state. A discrete combinatorial optimization problem in mathematics is the impact maximization issue. The initial seed set selected in the temporal network is expressed as Eq. (12).

(12)
$ \phi ^{*}=\underset{\left| \left.\phi \right| \right.=\mathrm{\hslash },\phi \subseteq V}{argmax}\left\{\sigma \left(\phi \right)\right\} $

where $\phi $ is a set of nodes found in the temporal network; $\phi $ is also known as seed collections; $\sigma \left(\phi \right)$ denotes the range of influence of the seed collection. $\left| \left.\phi \right| \right.$ is the quantity of nodes in $\phi $, and $\mathrm{\hslash }$ denotes the value of $\left| \left.\phi \right| \right.$. The marginal gain of the seed set $\phi $ is expressed as Eq. (13).

(13)
$ \sigma _{v}\left(\phi \right)=\sigma \left(\phi \cup \left\{v\right\}\right)-\sigma \left(\phi \right) $

where $\sigma _{v}\left(\phi \right)$ is the size of the marginal benefit of the node. The main metrics of influence maximization algorithms are the time complexity and the influence range of the seed set. The propagation probability was calculated using Eq. (14).

(14)
$ p_{u,v}=0.01+\frac{d_{u}+d_{v}}{N}+\frac{CN\left(u,v\right)}{N} $

where $d_{v}$ denotes the degree of the node; $CN\left(u,v\right)$ indicates the number of common neighbors among nodes. Eq. (15) expresses the likelihood that node $u$ will affect node $v$.

(15)
$ p_{uv}=0.01+\frac{E_{u}}{E_{u}+E_{v}},E_{u}+E_{v}\neq 0 $

where the eigenvector centrality of network node $u$ is shown by $E_{u}$. The eigenvector centrality of network node $E_{v}$ is shown by $v$. Because the IC model is proposed based on a static network and cannot be applied directly, it was improved in this research. Fig. 6 presents the propagation process of the improved IC model in the G3 slicing website.

The circles in Fig. 6 represent the nodes. The green, yellow, blue, and red circles are the seed nodes, active nodes, nodes to be activated, and inactive nodes, respectively. The propagation probability is mostly dependent on the activity of the node. The temporal-network combined heuristic greedy (TCHG) algorithm was adopted. The algorithm combines the benefits of greedy and heuristic algorithms. The selection of seed nodes can be split into the heuristic and greedy stages. In the previous stage, nodes with greater influence values are chosen for the candidate seed set. In the later stage, the algorithm selects the nodes with more sway in the candidate seed set and takes them as the final seed set. The timing network visualization tool is implemented mainly using Python language and a Python-based graphical user interface. The visualization tool has six main functional modules: data reading and writing module, network parameter calculation module, network topology mapping module, ASAM method module, TCHG algorithm module, and seed node information-propagation process module.

Fig. 6. Propagation process of the IC model in slicing network G3.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig6.png

4. Analysis of Node Ranking and Influence Maximization Algorithm Results based on Interlayer Coupling Strength Attenuation

The comparison algorithm and experimental data set were selected, and the parameters of the attenuation factor were set, starting from the two perspectives of the maximum connected component of the time series and the change in network performance. The effectiveness of the TCHG algorithm was tested by selecting the comparison algorithm and experimental data set and starting from two dimensions of influence range of seed set and algorithm running time.

4.1 Analysis of Superadjacency Matrix Temporal Network Model Results based on Interlayer Coupling Strength Attenuation

The SAM model and the serial-network model based on the node similarity supra adjacency matrix (SSAM) were selected for comparison to evaluate the performance of the ASAM model. The Enrons and Highschool3 datasets were selected to ensure accurate comparison findings. The attenuation adjustable parameter of the attenuation factor was set to 1. The validity verification of the ASAM model mainly starts from the time maximum connection component of the network performance change after node deletion. Fig. 7 compares the three model timing maxima and the components on the Enrons dataset.

In Fig. 7, when the centrality value of the deleted feature vector was the first 25 nodes of the network, according to the SAM model, the maximum and component may go as high as 0.7473 and as low as 0.7345. The maximum and minimum values of the SSAM model with its components were 0.7473 and 0.7311, respectively. The maximum and minimum values of the maximum and component of the ASAM model were 0.744 and 0.730, respectively. In Fig. 7(b), when the centrality value of the deleted feature vector was the first 35 nodes of the network, the maximum and minimum values of the maximum and component of the SAM model were 0.7455 and 0.7274, respectively. The highest and lowest values of the maximum and component of the SSAM model were 0.7455 and 0.728. respectively. The maximum and minimum values of the maximum and component of the ASAM model were 0.7412 and 0.7251, respectively. In both cases, the maximum and components of the ASAM model were smaller than those of the other two models, indicating that when removed from the network, the selected critical nodes of the model have a disproportionately large effect. Moreover, the ASAM model offers some clear benefits when locating pivotal nodes. Fig. 8 compares the time maxima and components of the three models on the Highschool3 dataset.

In Fig. 8, when the centrality value of the deleted feature vector was the first 20 nodes of the network, the highest and lowest values of the maximum and component of the SAM model were 0.9205 and 0.9199, respectively. The highest and lowest values of the maximum and component of the SSAM model were 0.9205 and 0.9199, respectively. The highest and lowest values of the maximum and component of the ASAM model were 0.9175 and 0.9163, respectively. In Fig. 8(b), when the centrality value of the deleted feature vector was the first 40 nodes of the network, the maximum and minimum values of the maximum and component of the SAM model were 0.915 and 0.9109, respectively. The maximum and lowest values for the maximum and minimum components of the SSAM model were 0.915 and 0.9109, respectively. The highest and lowest values of the maximum and component of the ASAM model were 0.9052 and 0.9034, respectively. The study compared the network performance changes of the three models to assess the effects of ASAM models.

In Fig. 9, when the centrality value of the deleted feature vector was the first 25 nodes of the network, the minimum value of network performance change in the SAM and SSAM models was 0, and the highest possible values were 0.017 and 0.214, respectively. The maximum and minimum values of network performance change in the ASAM model were 0.229 and 0.0042, respectively. In Fig. 9(b), when the centrality value of the deleted feature vector was the first 35 nodes of the network, the minimum value of the network performance change in the SAM model and SSAM model was 0.0021, and the maximum values were 0.0267 and 0.026, respectively. The maximum and minimum values of network performance change in the ASAM model were 0.0298 and 0.008, respectively. The network performance of the ASAM model changed the most in both cases, suggesting that the ASAM model has a better effect.

Fig. 7. Comparison of the three model timing maximums along with their components on the Enrons dataset.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig7.png

Fig. 8. Comparison of three model timing maxima and components on the Highschool3 dataset.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig8.png

Fig. 9. Comparison of network performance changes for the three models on the Enrons dataset.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig9.png

4.2 Analysis of the Results of the Influence Maximization Algorithm of Temporal Network

Four data sets and four comparison algorithms were selected. The data sets were the manufacturing dataset, highschool3 dataset, emil-EU-core dataset, and collegeMsg dataset. The comparison algorithms were the DegreeDiscount, NewDiscount, and Neighbourhood Size (NS) algorithms. The comparison angle has the subset influence range and the seed set selection time. Fig. 10 presents the comparative analysis of the respective spheres of the effect in the Highschool3 and manufacturing data sets of the four algorithms.

In Fig. 10, under the Highschool3 dataset, the maximum and minimum influence ranges of the seed set of the DegreeDiscount were 118 and 30, respectively. The maximum and minimum influence ranges of the seed set of the NewDiscount algorithm were 148 and 50, respectively. The maximum and minimum influence ranges of the seed set of the NS algorithm were 188 and 58, respectively. The maximum and minimum influence ranges of the seed set of the TCHG algorithm were 237 and 108, respectively. In Fig. 10(b), under the manufacturing data set, the DegreeDiscount algorithm had a maximum and minimum 85 and 30-degree sphere of effect for its seed set, respectively. The maximum and minimum influence range of the seed sets of the NewDiscount algorithm were 99 and 47, respectively. The maximum and minimum influence range of the seed sets of the NS algorithm was 115 and 87, respectively. The maximum and minimum influence range of the seed sets of the TCHG algorithm seed set were 123 and 90, respectively. The seed set of the TCHG algorithm had the largest influence range.

In Fig. 11, under the Emil-Eu-core dataset, the maximum and minimum influence range of the DegreeDiscount algorithm seed set was 250 and 70, respectively. The maximum and minimum influence range of the seed sets of the NewDiscount algorithm were 360 and 140, respectively. The maximum and minimum influence ranges of the NS algorithm seed sets were 540 and 400, respectively. The maximum and minimum influence range of the seed sets of the TCHG algorithm seed set were 680 and 530, respectively. In Fig. 11(b), the maximum and minimum reach of the DegreeDiscount algorithm seed set were 320 and 128, respectively, under the CollegeMsg dataset. The maximum and minimum influence range of the seed set of the NewDiscount algorithm were 380 and 200, respectively. The maximum and minimum effective influence range of the NS algorithm seed sets were 770 and 590, respectively. The maximum and minimum influence range of the TCHG algorithm seed set were 960 and 730, respectively. Hence, the seed set of the TCHG algorithm has the largest influence range.

In Table 1, when 30 seed nodes were selected, the maximum and minimum running lengths of DegreeDiscount were 0.029 s and 0.014 s, respectively. The maximum and minimum running lengths of NewDiscount were 0.358 s and 0.043 s, respectively. The maximum and minimum running lengths of NS were 0.352 s and 0.080 s, respectively. The maximum and minimum running lengths of the TCHG were 25.656 s and 5.203 s, respectively. When 50 seed nodes were selected, the maximum and minimum DegreeDiscount run times were 0.018 s and 0.016 s. The maximum and minimum running lengths of NewDiscount were 0.367 s and 0.053 s, respectively. The maximum and minimum running lengths of NS were 0.179 s and 0.083 s, respectively. The maximum and minimum running lengths of TCHG were 58.302 s and 25.656 s, respectively. The TCHG performed better when addressing the issue of maximizing influence in temporal networks.

Fig. 10. Comparison of the influence areas of the four algorithms in seed sets under Highschool3 and Manufacturing datasets.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig10.png

Fig. 11. Comparison of the influence of four algorithms on the seed set under the emil-EU-core and CollegeMsg datasets.

../../Resources/ieie/IEIESPC.2024.13.6.632/fig11.png

Table 1. Comparison of Running Time of Four Algorithms in Four Data Sets.

Datasets

HighSchoool3

Manufacturing

Emil-Eu-core

ColegeMsg

Seed set size

30

50

30

50

30

50

30

50

DegreeDiscount

0.014

0.016

0.016

0.018

0.029

0.016

0.016

0.017

NewDiscount

0.199

0.232

0.043

0.053

0.198

0.358

0.358

0.367

NS

0.080

0.083

0.099

0.168

0.352

0.134

0.134

0.179

TCHG

12.187

58.302

5.203

26.906

23.341

25.656

25.656

26.703

5. Conclusion

This study proposed an ASAM model and TCHG algorithm to explore important nodes in social networks. This paper introduced the eigenvector centrality and attenuation factor into the ASAM model, and the likelihood of transmission among the nodes was defined based on eigenvector centrality. On the Enrons dataset, the maximum connectivity components of the ASAM model under the no-case scenario were 0.744 and 0.7412, respectively. On the Highschool3 dataset, the maximum connectivity components of the ASAM model under no conditions were 0.9175 and 0.9052, respectively. The maximum connectivity component of the ASAM model was inferior to the comparison model, suggesting that the model possesses obvious strengths in identifying important nodes. The maximum values of the network performance change in the ASAM model were 0.229 and 0.02998, respectively, and the overall network performance change in the ASAM model was superior to the comparison model, showing the stronger impact of the model. The maximum seed set influence ranges of the TCHG algorithm in the Manufacturing and Highschool3 data sets were 237 and 123, respectively. The maximum seed set impact ranges under the emil-EU-core and CollegeMsg datasets were 680 and 960, respectively. The seed set influence range of the TCHG algorithm was larger than that of the contrast algorithm, which also shows that the algorithm has a better effect. Under the four data sets, the maximum running times of the TCHG algorithm were 25.656 s and 58.302 s, respectively, which were higher than those of other algorithms. The TCHG algorithm was more effective in maximizing the influence in temporal networks.

REFERENCES

1 
W. Yan, Q. Liu, R. Chen, and M. Zhang, ``Favoritism or equality: Difference analysis of users’ utilization of academic social networks for top research corporations,'' Online Information Review, Vol. 45, no. 1, pp. 240-260, 2020.DOI
2 
G. Anand, S. Srivastava, and A. Shandilya, ``Recurrent neural networks in predicting the popularity of online social networks content: A review,'' ECS Transactions, Vol. 107, no. 1, pp. 19991-20003, 2022.DOI
3 
C. Wang, H. Zhu, and B. Yang, ``Composite behavioral modeling for identity theft detection in online social networks,'' IEEE Transactions on Computational Social Systems, Vol. 9, no. 2, pp. 428-439, 2022.DOI
4 
Y. Liu, J. Wang, H. He, G. Huang, and W. Shi, ``Identifying important nodes affecting network security in complex networks,'' International Journal of Distributed Sensor Networks, Vol. 17, no. 2, pp. 1560-1571, 2021.DOI
5 
Halevy, C. Canton-Ferrer, H. Ma, U. Ozertem, P. Pantel, and M. Saeidi, ``Preserving integrity in online social networks,'' Communications of the ACM, Vol. 65, no. 2, pp. 92-98, 2022.DOI
6 
Tian, and W.-K.-V. Chan, ``Spatial-temporal attention wavenet: A deep learning framework for traffic prediction considering spatial-temporal dependencies,'' IET Intelligent Transport Systems, Vol. 15, no. 4, pp. 549-561, 2021.DOI
7 
Z. Yu, ``Evaluation of node importance and invulnerability simulation analysis in complex load- network,'' Neurocomputing, Vol. 416, no. 11, pp. 158-164, 2020.DOI
8 
C. Yang, Z. Kaibo, and J. Liu, ``SuperGraph: Spatial-temporal graph-based feature extraction for rotating machinery diagnosis,'' IEEE Transactions on Industrial Electronics, Vol. 69, no. 4, pp. 4167-4176, 2021.DOI
9 
Q. An, and Y. Shen, ``On the information coupling and propagation of visual 3D perception in vehicular networks with position uncertainty,'' IEEE Transactions on Vehicular Technology, Vol. 70, no. 12, pp. 13325-13339, 2021.DOI
10 
S. Kumar, M. Saini, M. Goel, and B.-S. Panda, ``Modeling information diffusion in online social networks using a modified forest-fire model,'' Journal of Intelligent Information Systems, Vol. 56, no. 2, pp. 355-377, 2020.DOI
11 
S.-S. Singh, K. Singh, A. Kumar, and B. Biswas, ``ACO-IM: Maximizing influence in social networks using ant colony optimization,'' Soft Computing: Fundamentals and Applications, Vol. 24, no. 13, pp. 10181-10203, 2020.DOI
12 
F. Wang, J.-H. She, O. Yasuhiro, W.-J. Jiang, G.-Y. Min, G.-J. Wang, and M. Wu, ``Maximizing positive influence in competitive social networks: A trust-based solution,'' Information Sciences, Vol. 546, no. 1, pp. 559-572, 2021.DOI
13 
M.-A. Uddin, J.-B. Joolee, and Y.-K. Lee, ``Depression level prediction using deep spatiotemporal features and multilayer Bi-LTSM,'' IEEE Transactions on Affective Computing, Vol. 13, no. 2, pp. 864-870, 2020.DOI
14 
R. Casarin, M. Iacopini, G. Molina, E.-T. Horst, and R. Rigobon, ``Multilayer network analysis of oil linkages,'' Econometrics Journal, Vol. 23, no. 2, pp. 269-296, 2020.DOI
15 
Z. Fang, X. Zhang, T. Cao, Y. Zheng, and M. Sun, ``Spatial-temporal slowfast graph convolutional network for skeleton-based action recognition,'' IET Computer Vision, Vol. 16, no. 3, pp. 205-217, 2021.DOI
16 
Ghasemi, and A. Ahmadi, ``Cache management in content delivery networks using the metadata of online social networks,'' Computer Communications, Vol. 189, no. 5, pp. 11-17, 2022.DOI
17 
S. Thabit, Y. Lianshan, Y. Tao, and A.-B. Abdullah, ``Trust management and data protection for online social networks,'' IET Communications, Vol. 16, no. 12, pp. 1355-1368, 2022.DOI
18 
F. Yi, M. Huo, J. Li, and J. Yu, ``Time‐series transcriptomics reveals a drought‐responsive temporal network and crosstalk between drought stress and the circadian clock in foxtail millet,'' The Plant Journal, Vol. 110, no. 4, pp. 1213-1228, 2022.DOI
19 
C. Tong, L. Zhang, H. Li, and Y. Ding, ``Temporal inception convolutional network based on multi-head attention for ultra-short-term load forecasting,'' IET Generation, Transmission & Distribution, Vol. 16, no. 8, pp. 1680-1696, 2022.DOI
20 
X. Wang, M. Cheng, J. Eaton, C.-J. Hsieh, and F. Wu, ``Fake node attacks on graph convolutional networks,'' Journal of Computational and Cognitive Engineering, Vol. 1, no. 4, pp. 165-173, 2022.DOI
Li Cai
../../Resources/ieie/IEIESPC.2024.13.6.632/au1.png

Li Cai was born on July 27, 1973, female. Native to Huangchuan County, Xinyang City, Henan Province. Han ethnicity. She obtained a Bachelor's degree in Chinese Language and Literature from Zhengzhou University in June 1995, and a Master's degree in Literature and Art from Zhengzhou University in June 2006. Her research directions is Chinese language and literature, communication and culture. She served as an internship lecturer at Zhengzhou Railway Education Center from August 1995 to August 1996; Lecturer at Zhengzhou Railway Vocational and Technical College from August 1996 to July 2019; From August 2019 to present, Director of the Chinese Language Teaching and Research Office of Zhengzhou Railway Vocational and Technical College. She has published more than 20 papers, including 6 Chinese core papers and 1 CSSCI paper. She has participated in the compilation of 7 textbooks and works, including 6 as deputy chief editors and 1 as chief editor. Presided over and participated in more than 30 projects, including 2 projects at provincial level or above, 7 projects at departmental level or above, and 1 project at national level. She won 1 second prize of Excellent Achievement Award for Humanities and Social Sciences Research of the Education Department.