Reputation evaluation model in social networks based on information behavior

Abstract

Social network has become an important channel for people to obtain information.Trusted user information behavior is the key to build cyberspace security. A dynamic reputation evaluation method based on supervision feedback of user information behavior is helpful to promote social network self-discipline and achieve good community autonomy. The comprehensive reputation evaluation of each node integrates identity and behavior reputation. And the reputation is dynamically updated by setting the new node evaluation period and phased update mechanism. Identity reputation is calculated by information disclosure and network characteristics; Behavior reputation is calculated by information release and forwarding, and rewards or punishments will be given to self-correction of information behavior or blocking of bad information. The simulation results show that compared with the traditional trust evaluation mechanism, setting rewards and punishments guidance can improve the accuracy of reputation evaluation. At the same time, reputation incentive can also inhibit the interaction of bad information while promoting the consciousness of reporting.

Keywords

Social network reputation information behavior reputation rewards and punishments network security

1. Introduction

The information interaction between people makes social network become a kind of “content-based” relationship network nowadays. Social networks has become an important channel for Internet users to obtain information, and the concept of social media has emerged. Each node in social media is a producer, disseminator and receiver of content. The communication way of content also forms a disorderly and infinite situation of point-to-point, point-to-face and face-to-face, which makes social media governance more difficult than other networks [17]. It is undoubtedly a huge challenge that Information security is only supervised by the platform. However, each node in social media can be the publisher, supervisor and blocker of content. From the perspective of network node autonomy, we can control the source of security problems and build a good social network ecology. Similar to social network relationship, in the open social network environment, the reputation of nodes in the community and the trust between nodes are important network interpersonal relationships [19]. With the implementation of the “network security law” in June 2017 in China, the requirements for network real name authentication and network integrity construction are put forward, which also makes the reputation of social network will be more and more valued by users. If the information disclosure and information behavior are integrated to evaluate its reputation, it can produce stronger binding force on the behavior of nodes and achieve better behavior self-discipline. At the same time, the reward and punishment feedback of social behavior will be integrated into the reputation model design, so that the node reputation can be further dynamically adjusted. In our work, the dynamic reputation model based on user’s information behavior not only guarantees its self-discipline, but also helps to guide node’s positive information behavior. This research work can promote the mutual supervision of low-quality, bad and false information in social media, and better realize the autonomous mechanism of group prevention and control.

2. Related work

Information dissemination in social networks is basically a low threshold and barrier free mode. At the same time, the authenticity and effectiveness of information content in social networks are difficult to guarantee. These characteristics make the security problems of social media more prominent. Relevant national legal departments and network platforms have issued a lot of relevant regulations and norms [5]. However, in the complex social network environment, user dishonesty and malicious behavior are inevitable. Designing corresponding algorithms by the means of information technology is also an important aspect of social network governance. Reputation and trust mechanism are the common solutions to ensure the security of social network [8]. Information behavior can be all human behaviors related to information source, information acquisition, utilization and dissemination. The network information behavior is also a kind of social behavior. The users play different roles in the network society, and information exchange are still the interactions between people [13]. Publishing false information, bad information, uncivilized comments, malicious slander and vicious language attacks are easy to spread quickly by network groups. The convenience of content generation in social networks requires users to self-judge and correct information behavior [10]. Orzan and others used sociology, psychology and behaviorism to study users’ information behavior, revealing users’ internal personality traits, motivation, cognition, etc., and external social attributes such as trust and culture have a significant impact on information behavior [18]. Different application scenarios also have different types of information behaviors. for example, from the perspective of knowledge management, the blog applications includes knowledge answer, tag user comments, etc. If from the perspective of the business perspective, it includes product review, rating, recommendation, etc. If from the perspective of the communication science, it includes information contribution, search and acquisition, post sharing, and exchange comments [14]. In the identification of rumors spread on social networks, the user information behavior is generally decomposed into several stages of information generation, transmission, judgment and influence [15,16]. The influence of network nodes is related to the strength of social relations, which can be indicated by the number of users’ followers, the number of fans, the frequency of interaction, the number of forwarding and the number of comments [22]. In addition, Kwon and others pointed out that in the research on the information contribution behavior of social network, the user’s sense of responsibility and consciousness can be reflected by the information behaviors such as user self-information disclosure and social search verification [11]. Zhang examined a reward and punishment mechanism based on the theory of planned behavior and deterrence. And his research revealed that adding punishment mechanism on social platform can also promote users to make active judgment and reduce the possibility of spreading Internet rumors [7].

Trust management and reputation mechanism has been widely studied and applied in computer network, communication, e-commerce, recommendation system, cloud computing and other fields [4]. Trust measurement in social networks can be done by direct trust, indirect trust, authentication trust, recommendation trust and so on [24]. The trust evaluation method is more inclined to the local trust relationship between nodes, and it reflects a mutual trust relationship. Reputation can be understood as the global trust measurement of nodes in the whole virtual community, and it is similar to social trust, which is a cumulative social asset [20]. There have been many studies on trust and reputation evaluation of social networks. We refer to some typical studies and summarize the relevant research results. According to the topology of message diffusion in user social network, Yuzhen gives the user’s final credibility through the trust of recommended nodes in the integrated network [26]. Huang and Yu evaluated the trust of microblog community users from two aspects of authority and authenticity. And their trust evaluation elements include gender, authentication, number of fans, number of followers, etc.. [6,25]. Srinivasan proposed to evaluate the interaction trust of mobile social network users by fuzzifying the virtual attributes of subjective perception, and attenuation according to the interaction time [23]. In terms of improving the user participation of mobile group governance perception network users, some studies integrated the reputation model to encourage user participation, which not only improves the network task processing efficiency, but also reduces the processing cost [9].

The research on reputation in our work which is mainly based on the research of social network information behavior and social network trust and reputation. Although abundant research results have been produced in two independent research directions, from the perspective of social media governance, there is a lack of research on reputation evaluation model integrating information behavior and behavior reward and punishment feedback. The reputation of social network node provides a user profile for the behavior of users in the virtual society, which is conducive to promoting users to maintain their own good image and stimulate their sense of responsibility and consciousness of network information behavior [2]. Therefore, reputation is not only a cumulative performance of user’s social behavior, but also a means to supervise and guide user’s behavior.

The research hopes to use the user information disclosure and information interaction behavior as the data source of node reputation evaluation, and solve some problems. The first is how to extract the characteristic elements of node global trust to build reputation evaluation model? The second is how to use information behavior to dynamically adjust node reputation? The third is how to supervise and manage the user’s information behavior by reputation? The main contribution of the work is as follows: (1) Node’s reputation is fused based on user information disclosure, network characteristics and historical information behavior. (2) Dynamically reputation updating is realized by simulation node information release, forward, delete, report and other dynamic information behavior. (3) According to the information release quality, self-blocking, verification and judgment, the reward and punishment factors are added; and the reputation will be modified dynamically by feedback, which can encourage and guide social network users to form good information behavior.

3. Construction of reputation evaluation model

3.1. Working principle of reputation model

In the evaluation of node reputation, the information behavior data can be obtained from three aspects: identity reputation, behavior reputation, and reputation reward and punishment. Identity reputation of user is to achieve user identification and ensure that social network nodes are true and effective; behavior reputation is a comprehensive evaluation of historical information interaction behavior; reputation reward and punishment is using behavior information feedback to guide node to adjust information behavior, and it is a positive expectation of node behavior. Therefore, the reputation ( $RP$ ) of social network nodes is defined as follow: it is a comprehensive credibility evaluation based on user identity, information behavior and behavior feedback in social network, and is also an evaluation of expected information behavior of nodes with effective identity in social network.

In social network, nodes’ mutual attention and information dissemination can form a network relationship. The reputation model takes the information disclosure and the position status in social networks as node identity reputation; the cumulative of information interaction between users as node behavior reputation; and different feedback attitudes generated by information interaction as reputation reward and punishment adjustment mechanism to compute node reputation. The calculation process of various reputation is shown in Fig. 1.

Fig. 1.

Reputation calculation framework.

Reputation in different dimensions includes relevant influencing factors, which can be obtained from behavior data for calculation. The basis and calculation rules of each influence factor are further elaborated below.

3.2. Identity reputation

The identity reputation $IR$ of a node is mainly determined by the disclosure of user information, the location status in the social network and the authentication of the user identity by the platform.

3.2.1. User disclosure

User disclosure $UDI$ : the personal information that users need to fill in social platform generally includes name, gender, birthday, location, industry, education level, etc.. Users’ disclosure of personal information mainly aims to maintain and develop interpersonal relationships, and also affects social network trust [3]. This paper evaluates the disclosure index from three dimensions, including the honest, rationality and authority of the disclosure information. Honest means that the more complete the data, the more users can pay attention to their virtual identity and be more responsible for their information behavior. Authority refers to the indicators with the ability to judge and identify information, which will affect the value recognition of other social nodes, such as education, occupation and professional title. However, at the same time of calculating integrity and authority, it is necessary to verify the authenticity of the information filled in by users. Since the public information of social platform will not set authentication for all information filled in, a rationality probability value can be set to determine whether the information is reasonable according to common sense rules, such as naming rules, time range, information conflict, and etc..

In the process of identity reputation calculation, ${UI}_{i}$ represents the integrity of information disclosure of node i, $fy {(i)}_{j}$ indicates whether the item j’s content of node i is filled or not, 0 represents not filled in, 1 represents filled in, and $w_{j}$ represents the necessary weight of the content.

${UA}_{i}$ refers to the authority of information disclosure of node i, and the authority attribute list set is $a = {A_{1}, A_{2} \dots A_{n}}$ . The authority weight of each information is quantified as ${af}_{j}$ according to the attribute value. For example, referring to the user information of social network, $w_{j}$ represents the authority weight of the content. As shown in Table 1, common authority influencing factors are set for discrete annotation.

Table 1
Example of user authority attribute indicators

Attribute Attribute interpretation Discrete rules

Profession Profession category Professional occupation: 1, student: 0.5, non-professional occupation: 0;

Organization Organization size Large-sized organization: 1, small and medium-sized organizations: 0.5, and unorganized: 0;

Education Education and knowledge level Above graduate: 1, college and secondary school: 0.5, below medium 0;

Rank Professional and technical rank Senior: 1, intermediate: 0.5, junior: 0;

Attribute	Attribute interpretation	Discrete rules
Profession	Profession category	Professional occupation: 1, student: 0.5, non-professional occupation: 0;
Organization	Organization size	Large-sized organization: 1, small and medium-sized organizations: 0.5, and unorganized: 0;
Education	Education and knowledge level	Above graduate: 1, college and secondary school: 0.5, below medium 0;
Rank	Professional and technical rank	Senior: 1, intermediate: 0.5, junior: 0;

$P_{j}$ is used to represent the rationality of the content. Through experience and common sense, regular expressions or common rules are constructed to judge the content. The default rationality is 1. If the unreasonable evaluation rule is offended, the score will be deducted. For example, take the name as an example in Table 2.

Table 2

Examples of attribute rationality judgment rules

Attribute	Unreasonable judgment rules
Name	If nationality is China and include non Chinese characters: $P_{j} = 0$
	If nationality is Chinese and the length is less than 2: $P_{j} = 0$
	If nationality is China and include non Chinese characters: $P_{j} = 0$
	If ethnic Groups is Han and length > 4: $P_{j} = 0.5$
	If ethnic Groups is Han and the first character cannot match the surname database sample: $P_{j} = 0.5$

The formula to calculate ${UI}_{i}$ is as expression (1), if the property is not set with rationality verification, $P_{j}$ is directly set to 1; otherwise, the value is set according to the rules. $\begin{matrix} (1) & {UI}_{i} = \sum_{j = 1}^{n} (fy {(i)}_{j} * w_{j} * P_{j}) \end{matrix}$

The formula to calculate ${UA}_{i}$ is like expression (2). Authority attributes only calculate the options in authority attribute list a, and adjust them according to the rationality of attributes. $\begin{matrix} (2) & {UA}_{i} = \sum_{j = 1}^{n} ({af}_{j} * w_{j} * P_{j}) j \in a \end{matrix}$

3.2.2. Clustering value

Clustering value $CLD$ : There is a clustering phenomenon in social behavior,and nodes who pay attention to high reputation also imply they are more likely to obtain high-quality information. And it has a positive effect on their information behavior. If the amount of node i that paying attention to other nodes is n, then node i’s social attention can be calculated by the average reputation value of all nodes, and the number of participating nodes is taken as the reinforcement factor s to reduce the error caused by obtaining high average value with a small number of concerns. The empirical value N is set as the average attention scale, and $n / N$ is taken as the strengthening coefficient. When n is greater than N, s is directly recorded as 1. ${UF}_{i}$ is calculated as expression (3), where $s = n / N$ , when $n ⩾ N$ , $s = 1$ . $\begin{matrix} (3) & {CLD}_{i} = \frac{1}{n} \sum_{j = 1}^{n} {RP}_{i \to j} * s \end{matrix}$

3.2.3. Influence value

Influence value $IFD$ : if a node has more fans, the node will have an obvious effect on information diffusion. And it also has a higher the degree of influence. The influence value is also a kind of voting recognition of other nodes in social network. The quantization level is divided according to the number of fans. The level standard is a list of $MF = {(0, m_{1}) : 0, [m_{1}, m_{2}) : 0.1 \dots [m_{10}, + \infty) : 1}$ . The calculation method of ${IFD}_{i}$ is as expression (4).Where $f m a p$ is the mapping function of influence degree. That is, the influence degree is the value of the quantization mapping list $MF$ corresponding to the acquisition node i. $\begin{matrix} (4) & {IFD}_{i} = fmap (i, MF) \end{matrix}$

3.2.4. Certification level

Certification level ${CTL}_{i}$ : Social platform provides node authentication mechanism. It mainly in the form of personal authentication and official authentication. Personal authentication includes interest authentication, “personal media” authentication and identity authentication. Official authentication includes government, enterprise, media, institution, public welfare, etc.. The influence of different authentication on reputation is set according to different subject. If the certification quantification level is $MA = {not certified: 0, individual certification: 0.5, official certification: 1}$ , it can also be adjusted according to the platform certification difficulty and certification requirements. The calculation method of ${CTL}_{i}$ is as expression (5), where $fmap$ is the authentication indicator mapping function to obtain the value of authentication quantization mapping table M corresponding to node i. $\begin{matrix} (5) & {CTL}_{i} = fmap (i, MA) \end{matrix}$

3.3. Behavior reputation

The social reputation $SR$ of a node can be determined by the interaction behavior of the node’s location characteristics and historical information in the social network.

3.3.1. Content interaction value

Content interaction value $SC$ : a node forwards a post is a social network interaction at each time. The forwarding, liking, commenting and trampling behaviors given by other nodes are all the feedback given after receiving information. Four score levels are set as (forwarding 3, commenting 2, liking 1 and stepping on $- 1$ ). The reputation value of interaction node affects the validity of rating. And the interaction given by high reputation node is more valuable for node reputation promotion. ${Rp}_{j}$ is the reputation value of the node giving the j-th feedback, and ${fs}_{j}$ is the quantitative score of the given j-th feedback. Therefore, the content interaction degree of node i is weighted to accumulate the interaction reputation of all the nodes that give comments. ${SC}_{i}$ is calculated as formula (6). $\begin{matrix} (6) & {SC}_{i} = \sum_{j = 1}^{n} {RP}_{j} * ({fs}_{j} / 3) \end{matrix}$

3.3.2. Activity value

The activity value $SA$ : it is used to describe the frequency of nodes using social networks. It can be measured by ${fre}_{i}$ and ${dur}_{i}$ . In a time window $TP$ , ${fre}_{i}$ is the frequency of node i logging in the social network and ${dur}_{i}$ is the cumulative online time. Assuming the average value of login frequency is $FR$ , the node login frequency ( ${fre}_{i} / FR$ ) can be used as the enhancement coefficient. If ${fre}_{i}$ is greater than $FR$ , it is also directly recorded as 1. Then calculates the daily average online time proportion of nodes in $TP$ time window, and d represents the date. ${SA}_{i}$ calculation method is shown in expression (7). $\begin{matrix} (7) & {SA}_{i} = {fre}_{i} / FR * 1 / TP * \sum_{d = 0}^{TP} {({dur}_{i})}_{d} \end{matrix}$

3.3.3. Contribution value

Contribution value $SO$ : the original publishing information is an independent view of the node, which will contribute information value to the whole community. It can also reflect the node has subjective initiative, and can promote its reputation accumulation. If node i publishes original information as $pa$ , the cumulative interaction score is $fta$ , the total information released is $pb$ , and the total interaction score is $ftb$ . The relative popularity of original information $sd$ is defined as the ratio of the average number of original interactions ( $fta / pa$ ) to the total average number of interactions ( $ftb / pb$ ). The higher the popularity of original information accounts for the overall popularity, the higher the value contribution. Also consider the average size of original content N, we take the number of original content $n / N$ as the enhancement coefficient. When n is greater than N, it is also directly recorded as 1. The ${SO}_{i}$ calculation method is shown in expression (8). $\begin{matrix} (8) & {SO}_{i} = \frac{fta}{pa} / \frac{ftb}{pb} * s \end{matrix}$

3.4. Reputation reward and punishment

High reputation nodes in social networks will consciously maintain network stability and control negative information transmission in case of crisis. The platform can judge the reputation of nodes according to whether they have active information behavior to maintain social network security [21]. Consciousness can also be reflected in the careful forwarding of information, sharing before authenticity verification of information actively, and that is to say the user has a good information discrimination ability [1]. During information dissemination, “judgment” also emphasizes that nodes need to effectively screen the information released or forwarded. If the reputation rewards and punishments are used to guide the node’s active information strategy, it can more effectively promote node autonomy. In the reputation reward and punishment mechanism, three dimensions are set up: content quality, information self-correction and bad information transmission blocking. The specific reward and punishment rules are shown in Table 3, and the reward and punishment coefficient can also be adjusted according to the platform experience.

Table 3
Reputation reward and punishment rules

Dimension Information behavior Explanation Reward and punishment rule Coefficient r

Content verification Release original high quality content The published original content is set as high-quality content according to the number of hits and forwarding One article was approved for high-quality content $+ 1 u$

Release bad information Publishing bad information will be checked by the platform One article was issued for each verification $- 3 u$

Forward bad information Publishing bad information will be checked by the platform One article was issued for each verification $- 2 u$

Self-correction Delete bad information Delete and clarify before being verified by the platform One article was issued for each verification $- 1 u$

Transmission blocking Effective reporting Consciously check the information and report the bad information on the platform One item for each invalid verification $+ 2 u$

Invalid report No verification of information, no valid evidence, illegal reporting One item for each invalid verification $- 1 u$

Dimension	Information behavior	Explanation	Reward and punishment rule	Coefficient r
Content verification	Release original high quality content	The published original content is set as high-quality content according to the number of hits and forwarding	One article was approved for high-quality content	$+ 1 u$
Release bad information	Publishing bad information will be checked by the platform	One article was issued for each verification	$- 3 u$
Forward bad information	Publishing bad information will be checked by the platform	One article was issued for each verification	$- 2 u$
Self-correction	Delete bad information	Delete and clarify before being verified by the platform	One article was issued for each verification	$- 1 u$
Transmission blocking	Effective reporting	Consciously check the information and report the bad information on the platform	One item for each invalid verification	$+ 2 u$
Invalid report	No verification of information, no valid evidence, illegal reporting	One item for each invalid verification	$- 1 u$

The value of penalty coefficient r under various conditions is described in Table 3. The system has more punishment scenarios than rewards, which is to make the reputation of behavior decline faster in case of bad behavior, and make nodes more cautious about their information behavior. Reputation updating mechanism uses time series to punish and supervise reputation behavior and dynamically adjust reputation value. In the process of adjustment, if the nodes have higher reputation, they also have greater influence to inhibit bad information dissemination. The degree of reputation reward and punishment is directly proportional to the influence of nodes. Therefore, after add the reward and punishment coefficient, the calculation method of ${RP}_{i}$ is shown in expression (9), ${RP}_{i}^{'}$ is the last reputation value, and u is the penalty unit that can be set. $\begin{matrix} (9) & {RP}_{i} = {RP}_{i}^{'} * (1 + r * u) \end{matrix}$

If the number of user complaints received by the social platform exceeds the threshold, it will block the information. At the same time, after the information is manually reviewed and determined by the social platform, the bad information will be cleared. Then the reputation rewards and punishments will be given to the publishers, communicators and informants respectively. The normal propagation speed of information is 1. When the number of information complaint nodes reaches the specified threshold, the information speed is 0. If the social platform receives complaints, it will reduce the transmission speed of information.

3.5. Integrated reputation mapping rules

(1) Reputation dynamic update.

The identity reputation is relatively stable, and it doesn’t update frequency once initialized. When setting a long update period $tp$ , such as 1 week or 1 month. In this way, it can reduce the computing cost of the system, and prevent some nodes from gaining instant reputation by simply attracting fans or following to others. Behavior reputation and its feedback punishment is an effective means of monitoring social behavior of node, which is more important. It can timely punish reputation update according to behavior operation, and dynamically update the comprehensive reputation with a short time window, such as one day.

(2) Initialize and update behavior reputation of new nodes

Due to no information interaction, the initial reputation of the new node is too low. The basic identity information can be registered in the node, and the evaluation period of the new node is set as $TC$ . During the evaluation period, the node uses the median value of behavior reputation ${sp}_{med}$ to initialize the behavior reputation. If the reputation decline speed $sd$ exceeds the threshold $tr$ , the system will give a warning and let the node enter the closed period, and set its reputation 0. If the reputation does not increase and $Δ RP < 0$ , the system will deprive the behavior of the initial value of reputation. If the reputation has increased, the reputation will be dynamically updated as normal. Setting evaluation period and closing period can further regulate the behavior of nodes and prevent malicious operations through registered nodes.

(3) Reputation integration

When calculating the comprehensive reputation, each indicator of identity reputation and behavior reputation can be fused according to a certain weight. Each indicator of identity reputation is a normalized value ${UA, UI, CLD, IFD, CTL}$ . The identity reputation $IR = UA + UI + CLD + IFD + CTL$ , and the value range of addition and sum is between $[0, 5]$ . Behavior reputation $SR = SC + SA + SO$ , it is only based on the information behavior within the time window $TP$ . If the time of behavior occurred outside the time window, it cannot be used as the basis for reputation calculation. The weights of identity reputation and behavior reputation are $w_{i r}$ and $w_{s r}$ respectively. Since the reputation can be accumulated with node behavior in a time window $TP$ , the value range is $[0, + \infty]$ after adding behavior reputation, reward and punishment mechanism. The maximum accumulated reputation of nodes in the community is $RPM$ as reference, and the comprehensive reputation is obtained by normalizing the reputation. The node’s lifetime in the community is recorded as $TS$ , and the expression of comprehensive reputation calculation method is as formula (10) in different survival periods. $\begin{matrix} (10) & {RP}_{i} = \{\begin{matrix} \frac{(w_{i r} * IR + w_{s r} * SR) * (1 + r * u)}{RPM} & {TS}_{i} > TC and SR in TP \\ \frac{(w_{i r} * IR + w_{s r} * {SP}_{med}) * (1 + r * u)}{RPM} & {TS}_{i} ⩽ TC and Δ {RP}_{i} > 0 \\ \frac{w_{i r} * IR}{RPM} & {TS}_{i} ⩽ TC and {Sd}_{i} < tr and Δ {RP}_{i} ⩽ 0 \\ 0 & {TS}_{i} ⩽ TC and {Sd}_{i} ⩾ tr \end{matrix} \end{matrix}$

4. Experiment

4.1. Experimental index

4.1.1. Reputation calculation error

In the traditional social trust model, public information and interactive behavior of nodes have been used as the basis of direct and indirect information calculation. In the experiment, we take the reputation calculation error $RCE$ as the indictor [12], and the calculation is as expression (11). This paper calculates the reputation calculation error in three scenarios, including the reward and punishment mechanism of information behavior, the reputation update and newly added node reputation management mechanism. $\begin{matrix} (11) & RCE = \frac{1}{n} \sum_{i = 1}^{n} | {RP}_{i} - R_{i} | \end{matrix}$

${RP}_{i}$ is the reputation value of node i calculated by the model; $R_{i}$ is the reputation value of the node should be set in the experiment, and $| {RP}_{i} - R_{i} |$ is the error of calculation deviation. The average error $RCE$ of N nodes in this category is taken as the reputation accuracy calculation indictor. If the model has a higher accuracy, it has a lower $RCE$ value, then a more credible value can reflect the real reputation level of community members.

4.1.2. Proportion of contaminated nodes

If the social platform does not identify and manage the content, bad information will continue to affect new nodes in the social network. We define the proportion of contaminated nodes $PCE$ as the number of nodes in the social network that have been affected before the bad information is verified. The reputation reward and punishment mechanism will guide the nodes to give feedback in the process of bad information dissemination. According to life experience, there is a certain positive correlation between the value incentive coefficient and the probability of actively checking bad information. Under the condition of different probability of reporting willingness, the social platform will suppress, review and eliminate bad information, and update the reputation of corresponding nodes according to the reward and punishment coefficient. The calculation formula of $PCE$ is computed by formula (12). p is the probability condition of different willingness to complain, and ${FN}_{p}$ represents different complaint probability scenarios. $pm {(i)}_{j}$ is the number of nodes polluted by the node i in he j-th propagation. We accumulate the number of nodes while information is not been terminated. If the $PCE$ is smaller, the reputation incentive mechanism is effective to suppress the bad information. $\begin{matrix} (12) & PCE = {FN}_{P} (\sum_{j = 1}^{n} pm {(i)}_{j}) \end{matrix}$

4.2. Experiment process

In order to verify the accuracy of the model trust calculation, two types of nodes are simulated in the experiment. The main work is to verify the influence of information behavior and behavior punishment on node reputation and behavior constraint. Therefore, we initialize the identity reputation, and set different types of information behavior rules. We simulated the behavior of the program to calculate the difference of indicators. The main experimental parameters are described as follows.

(1) Attributes of node: in the experiment, the total number of nodes is set to 200, the proportion of malicious nodes gradually changes from 10% to 50%, and other nodes are honest nodes. Each node is randomly connected with 20% nodes in the community.

(2) Behavior rules of Node: malicious nodes mainly release and forwarding of bad information, and invalid reporting; honest nodes release high-quality information, delete and clarify in time after publishing bad information. Honest nodes also give real feedback to other nodes, and check and report when they see bad information. Due to the occasional cognitive bias, there are some abnormal behaviors in reality. In addition, the feedback evaluation between nodes also occurs with a certain probability. From practical experience, human behavior has clustering characteristics, which makes the probability of feedback interaction between similar nodes is higher than that between different nodes.

(3) Malicious node behavior rules: The node publishes 50 pieces of information and forwards 50 pieces of information, and the proportion of bad information is 90%. Malicious nodes will give false feedback to connected malicious nodes with a probability of 50% and give false feedback to connected honest nodes with a probability of 20%. They will complain with a certain probability, but 90% of the complaints are invalid.

(4) Honest node behavior rules: The node publishes 50 pieces of information and forwards 50 pieces of information, and the proportion of normal information is 90%, the proportion of high-quality information is 5%, the proportion of bad information is 5%.

(5) Information interaction behavior rules:All connected nodes carry out information forwarding and reporting according to the node category and report with a certain probability. Assuming that the number of nodes reporting bad information exceeds the system threshold, the information will be deemed as bad information. The social platform makes a verification label for this information, punishes the posting or forwarding person for reputation, and rewards the informant for reputation.

Fig. 2.

Malicious node’s RCE.

4.3. Experimental results

4.3.1. Reputation calculation error

Fig. 2 to Fig. 4 show the reputation calculation error (RCE) of different types of nodes under different proportion (PMN) under three models. The model includes feedback reputation (FBR), behavior reputation without reward and punishment (NWRP) and reputation reward and punishment(WRP). In Fig. 2, when evaluating the reputation of malicious nodes, FBR feedback reputation comes from the feedback evaluation of nodes. Malicious nodes get more feedback from similar nodes, so they can obtain high false reputation at the beginning. With the increase of the proportion of malicious nodes, the reputation error will increase accordingly. However, the growth of reputation is limited by the real evaluation of honest nodes at a certain rate, and the growth rate is not large. When the proportion of malicious nodes is low, the reputation is mainly the weighted integration of network connected nodes if there is no reputation reward and punishment (NWRP). Because there are few malicious nodes in the community, and there are few opportunities for mutual interaction to accumulate reputation, the RCE is relatively low. However, with the increasing proportion of malicious nodes, its reputation value is easy to be interacted by more malicious nodes. If the number of credit nodes is balanced, the system will even evaluate it as a good faith node. However, with the model with reputation reward and punishment mechanism, once the honest node browses a node publishing or forwarding bad information in the community, it will make effective report with a probability of nearly 90%. Because the simulated social network connection setting is relatively balanced, the information will be recognized by the honest node in the traditional circulation. Assuming that the platform can determine the validity of the complaint according to the complaint behavior, the reputation of malicious nodes will be punished by a certain coefficient, and the reputation error of malicious nodes will have random error around 10%.

Fig. 3.

Honest node’s RCE.

In Fig. 3, when evaluating the reputation of honest nodes, they get more feedback from similar nodes, so the change trend of reputation error is similar to that of malicious nodes, the value does not vary greatly with the proportion of malicious nodes. A small amount of reputation errors in FBR are mainly caused by the low probability false feedback of malicious nodes and the accidental reverse behavior of honest nodes. The NWRP model will not punish the occasional misinformation behavior of the honest node. Therefore, if the probability of malicious node is low, the interaction between malicious node and honest node has little impact on its reputation, and the RCE is the lowest. However, with the increasing number of malicious nodes, the probability of interaction with malicious nodes will also increase, which will improve the RCE of honest nodes. Malicious nodes will publish and forward bad information in the community, an increasing of information interaction will occur between malicious nodes and honest nodes. And the interaction between honest nodes is relatively limited, so the overall growth rate is not high. The WRP model will encourage the honest node to report bad information effectively. When the number of malicious nodes in social network increases, it will breed more bad information and give more opportunities to honest nodes to get reporting reward. Therefore, with the increase of malicious nodes, the RCE of credit nodes decreases. Even if there is a small amount of RCE, it should be caused by the reverse behavior of setting corresponding probability in the experiment.

Fig. 4.

RCE of all nodes.

Fig. 5.

PCE under different complaint probabilities.

The overall $RCE$ in Fig. 4 shows that increasing reputation rewards and punishments will not improve the accuracy of reputation evaluation only when the proportion of malicious nodes in social networks is small. On the contrary, the mechanism of rewards and punishments makes the nodes suffer punishment after accidental errors, and the more severe reputation evaluation makes the relative deviation. However, with the increasing proportion of malicious nodes, the $RCE$ of both honest nodes and malicious nodes are decreasing,and the overall RCE is also declining. In other models, when the number of malicious nodes increases, the $RCE$ of different nodes has increased to a certain extent, especially for malicious nodes. Therefore, once the proportion of malicious nodes increases, the $RCE$ of both types of nodes of WRP is significantly lower than that of the model relying solely on feedback FBR and NWRP.

4.3.2. Contaminated nodes under different probabilities

Due to the lack of the sense of responsibility to actively maintain community governance, the behavior was set as the proportion of random reports following the interactive rule. In the experiment, it is assumed that the reward mechanism of reputation will stimulate the honest nodes to report bad information. There are many factors involved in the relationship between the degree of reward and the probability of stimulating reporting in practice,such as node activity, cognitive ability, social responsibility, etc.. In this experiment, we pay more attention to the probability of willingness, and other complex factors are not considered. The experiment tested the spread times of bad information under five different probability reporting conditions ${20 %, 40 %, 60 %, 80 %, 100 %}$ . As shown in Fig. 5, we compares the PCE index values of honest nodes under five different reporting probabilities. In the figure, 10%-AC and 50%-AC represents the situation of setting 10% or 50% malicious nodes with medium threshold; 10%-HAC and 50%-HAC represents the situation of setting 10% or 50% malicious nodes with high threshold.

The experiment compares the number of community nodes that can be polluted before being blocked under two complaint thresholds when a piece of bad information has two different malicious node ratios of 10% to 50%. In the experiment, two different malicious node ratios of 10% and 50% and two complaint thresholds are set respectively. The threshold takes the average connection size of nodes as the threshold reference value, and sets AC and HAC thresholds. For example, in the experiment, the connection rate of nodes is set to 20%, the average connection is 40 nodes, and the threshold of complaint nodes is 40 and 20 times. According to the results in Fig. 5, under five different complaint probability conditions, the number of infected nodes decreases with the increase of complaint Intention. When the proportion of malicious nodes is 10%, there are fewer infected nodes in both threshold scenarios than when has 50% percent of malicious nodes. In the case of 50% malicious nodes, a bad message will soon be received by all nodes in the platform when the complaint intention is not high. Even if the node has the highest willingness to complain, more than half of the nodes will receive it when the threshold is a little higher. With the increase of malicious nodes, the number of bad information sources will also increase. The outbreak of multi-point transmission will further aggravate the number of infected nodes.

As the proportion of malicious nodes in the community changes from 10% to 50%, the adverse information published and forwarded in the community should be identified and communication impact needs to be assessed. According to the experimental results, when the proportion of malicious nodes is less, the proportion of bad information in the community is also lower. the total number of bad information dissemination is relatively small before being identified by the system. Similarly, with the increase of the proportion of malicious nodes, the bad information will be more, and the spread area and frequency will be higher. According to the proportion of five complaints willingness, the higher the probability of complaint intention, the lower the number of bad information spread in the community, and the faster it can be screened out by the system. Reputation reward and punishment can encourage nodes to improve their willingness to complain. Therefore, if the platform properly adjusts the threshold of the number of complaints, the number of nodes in the community infected by bad information can be greatly reduced, and bad information can also be screened out by the platform faster.

5. Conclusion

Reputation in social networks is a measure of the credibility of user behavior, and it is an important means to govern social network and promote user self-discipline. The purpose of this paper is to encourage the effective disclosure of user identity information and construct a dynamic reputation update mechanism to guide users’ information behavior. Experiments show that when there are some malicious nodes in the community, the introduction of reputation mechanism can more accurately evaluate whether the nodes are trusted and identify malicious nodes. If the platform can give more permissions to high reputation nodes, reputation rewards and punishments will guide nodes to timely self-correct and participate in group supervision, which can effectively inhibit the bad information behavior of malicious nodes. Since only two types of nodes are set up in the experiment, more agents and their behavior rules will be studied in the next step, and the evolution process of reputation behavior of different types of newly added nodes will be analyzed to further improve the reputation model.

Footnotes

Acknowledgements

This paper is supported by the Ministry of Education Humanities and Social Sciences Youth Fund Project, “Research on social media communication governance based on trust and content risk perception” (20yjc860032).

Conflict of interest

None to report.

References

Bar-Isaac and

Deb, What is a good reputation? Career concerns with heterogeneous audiences, International Journal of Industrial Organization34(1) (2014), 44–50. doi:10.1016/j.ijindorg.2014.02.012.

Chen and

Wan, Social trust and corporate social responsibility: Evidence from China, Corporate Social Responsibility and Environmental Management27(2) (2020), 485–500. doi:10.1002/csr.1814.

Chen,

Xia and

Wang, A novel trust-based community detection algorithm used in social networks, Chaos Solitons & Fractals108 (2018), 57–65. doi:10.1016/j.chaos.2018.01.025.

Gambhir and

Gambhir, Trust analysis techniques in online social networks a review, International Journal of Computer Science and Engineering7(3) (2019), 254–258. doi:10.26438/ijcse/v7i3.254258.

Gladyshev,

Marrington and

Baggili, Digital forensics and cyber crime, in: ICDF2C 2013. Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering, Vol. 132. Springer, Cham, 2013. doi:10.1007/978-3-319-14289-0_9.

Huang and

Xiong , Evaluation of microblog users’ influence based on pagerank and users behavior analysis, Advances in Internet of Things.3(2) (2013), 34–40. doi:10.4236/ait.2013.32005.

Huiping,

Xinhao and

Zhiwei, Impact of sanction mechanism on behavioral intention to identify Internet rumor, Journal of Intelligence35(12) (2016), 47–51.

Jazaieri,

Logli Allison,

Camposet al., Content, structure, and dynamics of personal reputation: The role of trust and status potential within social networks, Group Processes & Intergroup Relations22(7) (2019), 964–983.

Kareva,

Morin and

Karev, Preventing the tragedy of the commons through punishment of over-consumers and encouragement of under-consumers, Bull. Math. Biol.75(4) (2013), 565–588. doi:10.1007/s11538-012-9804-3.

10.

M.S.

Khan,

A.W.A.

Wahab,

Herawan,

Mujtaba,

Danjuma and

M.A.

Al-Garadi, Virtual community detection through the association between prime nodes in online social networks and its application to ranking algorithms, IEEE Access4 (2017), 9614–9624. doi:10.1109/ACCESS.2016.2639563.

11.

Kwon and

A.J.

Onwuegbuzie, Modeling the factors affecting individuals’ use of community networks: A theoretical explanation of community-based information and communication technology use, Journal of the American Society for Information and Technology56(14) (2014), 1525–1543. doi:10.1002/asi.20247.

12.

Lax and

M.L.S.

Giuseppe, CellTrust: A reputation model for C2C commerce, Electron Commerce Research8 (2008), 193–216. doi:10.1007/s10660-008-9019-8.

13.

C.M.

Liao,

S.H.

You and

Y.H.

Cheng, Network information analysis reveals risk perception transmission in a behaviour-influenza dynamics system, Epidemiology & Infection143(1) (2015), 23–36. doi:10.1017/S0950268814000430.

14.

M.J.J.

Lin,

S.W.

Hung and

C.J.

Chen, Fostering the determinants of knowledge sharing in professional virtual communities, Computers in Human Behavior25(4) (2009), 929–939. doi:10.1016/j.chb.2009.03.008.

15.

Y.H.

Liu,

X.L.

Jin and

H.W.

Shen, Towards early identification of online rumors based on long short-term memory networks, Information Processing & Management56(4) (2019), 1457–1467. doi:10.1016/j.ipm.2018.11.003.

16.

Z.Y.

Liu,

Zhang,

T.U.

Cunchao and

M.S.

Sun, Statistical and semantic analysis of rumors in Chinese social media, Scientia Sinica (Informationis)45(12) (2015), 1536. doi:10.1360/N112015-00243.

17.

Myneni,

N.K.

Cobb and

Cohen, Finding meaning in social media: Content-based social network analysis of quitnet to identify new opportunities for health promotion, Studies in Health Technology and Informatics192(1) (2013), 807–811. doi:10.3233/978-1-61499-289-9-807.

18.

Orzan,

Macovei and

A.M.

Radu, Privacy and trust in social media: An empirical investigation of users’ willingness to share personal information online, Social Responsibility, Ethics and Sustainable Business127(8) (2013), 749–754. doi:10.1017/S0022215113001473.

19.

Popescul and

Georgescu, Social networks security in universities: Challenges and solutions, Annals of the Alexandru Ioan Cuza University – Economics62(s1) (2015), 53–63. doi:10.1515/aicue-2015-0036.

20.

E.D.

Raj and

L.D.D.

Babu, An enhanced trust prediction strategy for online social networks using probabilistic reputation features, Neurocomputing219(1) (2016), 412–421. doi:10.1016/j.neucom.2016.09.036.

21.

Ren,

Li,

Xiang,

Cui and

Sakurai, Evolution of cooperation in reputation system by group-based scheme, Journal of Supercomputing63(1) (2013), 171–190. doi:10.1007/s11227-010-0498-8.

22.

Siqing,

Mengni and

Xiaobo, Analysis of the key influencing factors of haze information dissemination behavior and motivation in wechat, Interlending & Document Supply45(1) (2017), 21–29. doi:10.1108/IDD-09-2016-0029.

23.

Srinivasan and

Naseera, Trust and location based service in mobile social networks – a survey, Multiagent and Grid Systems14(3) (2018), 263–282.

24.

Urena,

Chiclana,

R.A.

Carrascoet al., Leveraging users’ trust and reputation in social networks, Procedia Computer Science162(12) (2019), 955–962. doi:10.1016/j.procs.2019.12.073.

25.

Yu,

Jia,

Tao,

Xue and

Lee, An efficient trust evaluation scheme for node behavior detection in the Internet of things, Wireless Personal Communications An International Journal93(2) (2017), 571–587. doi:10.1007/s11277-016-3802-y.

26.

Yu,

Zhu and

Shen, Trust model and simulation of E-commerce based on social networks, Journal of System Simulation30(8) (2018), 304–312. doi:10.16182/j.issn1004731x.joss.201808036.