《计算传播学导论》读书笔记：第六章 网络传播模型与机器学习框架

There are mainly two methods to study the dissemination of information in online social networks, one is to use mathematical modeling for analysis and simulation, and the other is to use actual data for data mining and machine learning method for prediction. There are two main reasons for using the first model method to study information communication: first, the process of information dissemination is very complex, and it is very difficult to study without using the model for abstraction and simplification; second, in the early research of information dissemination, the collection, acquisition and storage of empirical data are difficult, and the cost is very high, and the use of model method greatly reduces the demand for data.

The research paradigm of information communication based on machine learning can improve the accuracy of the prediction of network information dissemination and deepen our understanding of each link of information dissemination and its influencing factors.

The dissemination of information on the Internet is a complex behavior of human beings on social networking sites and social media, which is affected and limited by many factors. Based on the topological structure of the social network, some information dissemination models are used to study the information dissemination on the network, which is a common method in the early research of information dissemination.

The diffusion of a new technology in the social system requires four necessary conditions: innovation, communication channels, time and social system. Diffusion is a kind of social exchange behavior, so diffusion must occur in the social system with structural and functional characteristics, and the process can be planned or spontaneous.

According to the adoption time of innovative entity adopters, adopters are divided into the following five categories: innovators, early adopters, early adopters, late adopters and backwards.

In a typical infectious disease model, the status of individuals in a population can be divided into the following categories:

In the traditional classical model of disease transmission, a basic assumption is that all individuals in the population are completely mixed, that is, an individual has an equal chance of contact with any individual in the population per unit time. However, in the context of social networks, the individual can only contact his friends (neighbor nodes), so the infected individual can only directly transmit the virus to the nodes directly connected to it.

The steady state value in the above equation is also called epidemic state in epidemiology.

There is a strong similarity between information transmission and disease transmission. For a long time, it is generally believed that the spread of information between individuals is similar to pathogens, each time exposed to friends who know the information, it will potentially cause other individuals to become infected (that is, knowing information), and the individual’s response to repeated exposure to information is more complex than imagined.

At the same time, there is also a significant difference between the spread of information and the spread of disease, which is rooted in people’s cognitive limitations in processing information. In the transmission of pathogens, individuals with more friends are more likely to be infected with the disease, while such people are less likely to be infected in the spread of information. This is because the amount of information increases with the number of friends a user follows, and highly connected users are less likely to notice a particular message, and they tend to need stronger social signals to take action than low-value users. In the early information communication research, there are many scholars who use the disease transmission model. With the in-depth understanding of the mechanism of information transmission and the refinement of the study of transmission model, there are fewer and fewer studies on the use of disease transmission model to study information transmission.

The threshold model, also known as the threshold model, assumes that everyone crosses a threshold when deciding to participate in an activity. The existence of this threshold is determined by the cost and benefit of the action. when the cost of the action is equal to the benefit, it reaches the threshold of the behavior.

The threshold model can be used to analyze the role of social influence (or interpersonal role) in information dissemination. Social influence refers to the mutual influence between friends, and the “contagion” of this behavior is based on social learning or imitation. Social influence plays an important role in the information dissemination of social media.

For information dissemination, for each message forwarder, when he forwards the message, we can also calculate the proportion of his friends who have forwarded the message, which is the diffusion threshold (or diffusion threshold) of the node.

The diffusion threshold can be defined as the number of times the user has been exposed to the message when the message forwarder is forwarded, that is, the number of friends of the user who have forwarded the message.

In this model, when an individual becomes active, every inactive individual in his neighbor will be activated with a certain probability. A simple example is the independent cascade model, in which each initial active node produces its own independent diffusion cascade, which is independent of each other and does not interfere with each other. In the model, the probability of an individual being activated by a nearest active neighbor is independent of the probability of the neighbor set that tried to activate it in the past.

An information source sends information to friends, who in turn send information to their friends, which in turn goes on, and the number of friends sent meets a specific distribution.

One of the basic properties of the branching process is that if a generation of individuals no longer propagate the message, the message will disappear. Therefore, the branch process model actually has only two possibilities of information diffusion: one is that no one continues to spread information, so the information disappears after a limited number of steps; the other is that people continue to spread information. This process continues indefinitely on the social network.

From the basic reproduction number, information diffusion is at least related to two factors: one is the result of the network, and the other is people’s willingness to spread information.

The diffusion of information is related to the heterogeneous patterns of human behavior and is driven by the decisions of the participants, so their transmission dynamics will be different from the traditional infectious disease models.

To analyze the relationship between features and information dissemination, the traditional research method is to carry out regression analysis, first calculate a single feature statistic in the network, and then use the feature statistics to fit the information dissemination index in the network. This approach does not take into account the interrelationship and influence of intra-class and inter-class feature statistics, and the results between different networks are not comparable. Based on the information dissemination prediction framework of machine learning, firstly, the multi-category and multi-feature statistics of the empirical network, namely feature engineering, are calculated, and then the influence and interrelation of each feature statistics, namely feature selection, are calculated. and the static, weighted and dynamic statistics are analyzed systematically and comprehensively, and the best feature selection method is selected. finally, the information dissemination prediction is realized by supervised two-classification or regression machine learning algorithm.

The research on the prediction of information communication under the background of big data involves four steps: data acquisition and preprocessing, feature engineering, feature selection and evaluation criteria. For data acquisition, there are two main ways to analyze the information diffusion of a topic. One way is to describe the diffusion of all the information including a topic; the other is to analyze only the complete diffusion of a piece of information and get the information of each forwarder in chronological order. and the forwarding of messages between the forwarders, so that the diffusion tree of a single message can be constructed.

In many studies, it mainly focuses on the second way to analyze the dissemination of information, that is, to describe the characteristics of a single piece of information diffusion. In addition to describing many aspects of a single message, we can also analyze the various factors that affect the dissemination of information.

With the popularity of smart phones, spatial information in online social networks has become more and more accessible, and local information is becoming more and more important for targeted marketing and personalized services. News spreads among cities every day through social networks, and the spatial location of users’ friends will greatly affect the area to which the information will be diffused. as a result, some users can spread to a particular geographical area more easily and more effectively than other users. Some metrics in traditional social network research can be extended to spatial social networks, such as spatial degree, spatial compactness centrality, spatial efficiency and local clustering coefficient. These measures can quantitatively describe the impact of social connections on the diffusion of information in a region.