A method for predicting the spread of social network information

Abstract

The present invention provides a method for predicting the spread of social network information. This method comprehensively considers three key factors in the social network, namely social influence, the temporal dynamic characteristics of information spread, and the similarity of non-connected user behaviors. By fusing three algorithms, namely exposure time sorting, gravitational sorting, and cascade similarity sorting, a model capable of predicting the spread of information in the social network is proposed. Among them, the exposure time sorting algorithm can not only ensure the temporal dynamic characteristics of information dissemination under one topic, but also take into account the large differences in the temporal dynamic characteristics of information dissemination under different topics. The gravitational sorting algorithm is derived from the social network topology, which can calculate the influence of all users participating in the topic on their neighbors and sort all users not participating in the topic according to the activation probability. The cascade similarity sorting is derived from the cascade data of each user, characterizes users with a vector, and measures the similarity between them using the cosine similarity of the vectors.

Description

Technical Field

[0001] The present invention belongs to the technical field of machine learning, and in particular relates to a method for predicting information diffusion in social networks. Background Art

[0002] In the natural world, networks of all shapes and sizes exist. The study of network structure has spanned centuries, with network science evolving from regular and random networks to complex networks. Complex networks are a generalized representation of complex systems. In the real world, everything from transportation systems to ecosystems can be abstracted into complex networks. Social networks are a form of complex network, representing an abstraction of user social relationships. With the development of internet technology, social networks have become ubiquitous. Social networks have become a hot topic of research, with the evolution of information dissemination within them being a key focus.

[0003] Information diffusion on social networks refers specifically to the process of information dissemination within social networks. Due to the inherent openness and unrestricted nature of online social networks, they have become a crucial medium for information dissemination worldwide, with information diffusion reaching unprecedented levels of activity. Consequently, research on information diffusion on social networks has garnered increasing attention. When one person receives a piece of information, their neighbors may be influenced to receive the same information. This phenomenon is ubiquitous on social networks, most notably through forwarding on Weibo. When a user posts a Weibo post, their fans or followers are likely to forward it. Information diffusion is often triggered by user behavior, so information diffusion can also be considered the spread of user behavior.

[0004] Information dissemination prediction refers to predicting the future trend of information dissemination based on the known existing information dissemination situation, following certain methods and rules, so as to estimate the final process and results of information dissemination. On this basis, it is convenient for us to intervene in the dissemination of information and seek benefits and avoid harm.

[0005] Based on different purposes, current research on social network information dissemination prediction is mainly divided into three categories: dissemination range prediction, user influence prediction, and user behavior prediction.

[0006] Regarding the research questions of the present invention, the existing technologies related to user behavior prediction are mainly introduced. User behavior prediction mainly focuses on the behavior of individual users. In social networks, this usually manifests as specific behaviors such as forwarding microblogs and joining groups. The research in this area mainly lies in determining whether users perform such operations. Zaman et al. were the first to start using collaborative filtering models to predict users' forwarding behavior based on their past behaviors. Chen et al. solved the data sparsity problem existing in collaborative filtering models by extracting keywords from microblog information, further improving the prediction effect. Hong et al., based on the different focuses of users, used the bag-of-words model to calculate text similarity. They believed that the greater the similarity, the greater the probability of users' forwarding, and used this to predict users' forwarding behavior. Benefiting from the rapid development of natural language processing, topic models based on latent Dirichlet distribution and their related variants have also been widely used in user behavior prediction. Liu et al. proposed a generative graph model that uses multi-layer link information and the text information content associated with each node in the network to mine the influence of topic content on forwarding behavior. Based on the obtained influence, they further studied how the discovered topic-level influence helps predict user behavior. Yang et al. statistically analyzed the characteristics of users' forwarding behavior and proposed introducing the factor graph model into the supervised learning method to predict users' future forwarding behavior through their historical forwarding behavior. Xu et al. transformed the forwarding behavior prediction problem into a binary classification problem, divided the factors affecting users' forwarding behavior into social relationship-based features, user-based features, and text-based features, trained through traditional machine learning methods (decision trees, SVM (support vector machines), logistic regression, etc.), and compared the effectiveness of each feature using the feature exclusion method, and found that social relationship-based features are more important in information dissemination. In addition, Luo et al. performed a top-K ranking on potential forwarders and carried out forwarding prediction based on the Pointwise ranking algorithm, and also obtained the conclusion that social relationship-based features are more important than other features.

[0007] Currently, research on social network prediction is in full swing, but most models ignore the fact that information dissemination is essentially a time-dynamic process. In existing research on time-dynamic information dissemination, Satio et al. transformed user behavior prediction into a maximum likelihood estimation problem by studying the dissemination mechanisms of the independent cascade model and the linear threshold model, and used the expectation maximization algorithm (EM algorithm) to solve it. Although this method takes into account the time factor, due to its time complexity, it cannot be used to process massive data. Goyal et al. proposed two time-varying models based on continuous time (CT) and discrete time (DT) to calculate social influence and used them to predict dynamic information dissemination under a unified threshold model. However, under this model, the prediction of information dissemination largely depends on the activation threshold of users, which is difficult to set in practice. In addition, due to the approximate simulation mechanisms of the CT and DT models for calculating social influence themselves having side effects on the prediction performance, this method has great defects. Based on game theory, Li et al. proposed the GT model and conducted user behavior prediction under a unified threshold model. In the GT model, they regarded each node as an agent and judged the future state of the node by comparing the benefits brought by different behaviors. Considering time dynamics, the GT model proposed a time-varying model based on users' historical information to calculate social influence. Therefore, at different time points, the benefits of different behaviors of the same user also have time-dynamic characteristics. However, the GT model still has the cold start problem caused by data sparsity. Table 1 shows the comparison of existing user behavior prediction models considering time-dynamic characteristics in terms of time complexity, accuracy, information acquisition, and application ability.

[0008] Behavior prediction model Time complexity Accuracy rate Information acquisition Application ability Likelihood estimation model Highest High Difficult Small-scale network Continuous-time model High Lower Easy Small-scale network Discrete-time model Lower Low Easy Medium-scale network Game theory model Low High Difficult Large-scale network

[0009] Table 1 Summary of the Invention

[0010] To solve the above problems, the present invention provides a method for predicting information diffusion based on time-dynamic characteristics in a social network, and the present invention adopts the following technical solutions:

[0011] The present invention provides a method for predicting the diffusion of social network information, which is used to predict the information diffusion based on the time dynamic characteristics in a social network. It is characterized by including the following steps: Step S1, fusing an exposure time sorting algorithm, a gravity sorting algorithm, and a cascade similarity sorting algorithm to construct an information diffusion prediction model; Step S2, using the to-be-predicted topic, the set of to-be-predicted users related to the to-be-predicted topic, and historical propagation information as the input of the information diffusion prediction model; Step S3, obtaining an exposure time sorting array of the to-be-predicted topic based on the exposure time sorting algorithm; Step S4, deriving the participating users of the to-be-predicted topic based on the social network topology structure, calculating the influence of the participating users and the activation rate of the non-participating users using the gravity sorting algorithm to obtain a gravity sorting array; Step S5, calculating the similarity between users for each user in the social network topology structure using the cascade similarity sorting algorithm, and obtaining a cascade similarity ranking; Step S6, predicting whether the set of to-be-predicted users will participate in the to-be-predicted topic according to the exposure time sorting array, the gravity sorting array, and the cascade similarity ranking.

[0012] In a method for predicting the diffusion of social network information provided by the present invention, it may also have the following technical feature: The process of obtaining the exposure time sorting array is as follows: statistically analyzing the total exposure time distribution of all users participating in the to-be-predicted topic, calculating the total exposure time of all users not participating in the to-be-predicted topic, calculating the activation probability of all users not participating in the to-be-predicted topic based on the total exposure time distribution and the total exposure time, and sorting the activation probabilities in descending order to obtain the exposure time sorting array.

[0013] In a method for predicting the diffusion of social network information provided by the present invention, it may also have the following technical feature: The process of obtaining the gravity sorting array is as follows: Let u and v be two users having a path in the directed graph of the social network, and let path uv be the shortest path length from u to v, and U be the set of users participating in the topic. Then, the resultant gravity for user v is defined as:

[0014]

[0015] In the formula, d u is the out-degree of user u, d v is the in-degree of user v. Sort all non-participating users according to the resultant gravity to obtain the gravity sorting array.

[0016] In a method for predicting the diffusion of social network information provided by the present invention, the following technical features may also be included. The process of obtaining the cascade similarity is as follows: Extract the user IDs sorted by timestamp in each cascade and use them as a sentence. Using one-hot encoding and the skip-gram algorithm, embed the ID of each user into a 128-dimensional vector for optimization training. Set the window size to 10. The cascade similarity between labels h between u and v is measured by the cosine similarity between u and v:

[0017]

[0018] Assume that the set of transferred users followed by v is S u , then the transfer probability of user v is proportional to and sort the users to be predicted in descending order according to their similarity to obtain the cascade similarity ranking.

[0019] Function and effect of the invention

[0020] According to the method for predicting the diffusion of social network information provided by the present invention, a new model capable of predicting the diffusion of information in a social network is proposed by integrating the exposure time sorting algorithm, the gravitational sorting algorithm, and the cascade similarity sorting algorithm. Among them, the exposure time sorting algorithm can not only ensure the time dynamic characteristics of information dissemination under one topic, but also consider the problem that the time dynamic characteristics of information dissemination under different topics vary greatly; the gravitational sorting algorithm is derived from the social network topology structure, and can calculate the comprehensive influence of all users participating in the topic on their neighbors and sort all users not participating in the topic according to the activation probability; the cascade similarity sorting is derived from the cascade data of each user, uses a vector to describe the user, and measures the similarity between them using the cosine similarity of the vector.

[0021] Therefore, the method for predicting the diffusion of social network information of the present invention comprehensively considers three key factors: social influence in the social network, time dynamic characteristics of information diffusion, and similarity of non-connected user behavior. Combining the characteristics of the real social network environment and complex network research, a prediction model that can accurately predict user behavior based on time dynamic characteristics in the social network is given. Using this prediction model for user behavior prediction not only improves the prediction effect, but also has a great improvement in time complexity and has an extremely wide application prospect in the real social network environment, providing a certain reference significance for the research in the field of release time prediction. Brief description of the drawings

[0022] Figure 1 is a flowchart of the method for predicting the diffusion of social network information in an embodiment of the present invention;

[0023] Figure 2It is a schematic diagram of the framework of the information diffusion prediction model in the embodiment of the present invention;

[0024] Figure 3 It is a flowchart of the exposure time sorting algorithm in the embodiment of the present invention;

[0025] Figure 4 It is a flowchart of the gravitational sorting algorithm in the embodiment of the present invention;

[0026] Figure 5 It is a comparison chart of the average accuracy rates obtained by testing the CT, DT, GT, and TR models on the first experimental dataset in the embodiment of the present invention;

[0027] Figure 6 It is a comparison chart of the average recall rates obtained by testing the CT, DT, GT, and TR models on the first experimental dataset in the embodiment of the present invention;

[0028] Figure 7 It is a comparison chart of the average F1-Measures obtained by testing the CT, DT, GT, and TR models on the first experimental dataset in the embodiment of the present invention;

[0029] Figure 8 It is a comparison chart of the average accuracy rates obtained by testing the CT, DT, GT, and TR models on the second experimental dataset in the embodiment of the present invention;

[0030] Figure 9 It is a comparison chart of the average recall rates obtained by testing the CT, DT, GT, and TR models on the second experimental dataset in the embodiment of the present invention;

[0031] Figure 10 It is a comparison chart of the average F1-Measures obtained by testing the CT, DT, GT, and TR models on the second experimental dataset in the embodiment of the present invention;

[0032] Figure 11 It is a comparison chart of the average accuracy rates obtained by testing the CT, DT, GT, and TR models on the third experimental dataset in the embodiment of the present invention;

[0033] Figure 12 It is a comparison chart of the average recall rates obtained by testing the CT, DT, GT, and TR models on the third experimental dataset in the embodiment of the present invention;

[0034] Figure 13 It is a comparison chart of the average F1-Measures obtained by testing the CT, DT, GT, and TR models on the third experimental dataset in the embodiment of the present invention. Detailed implementation manners

[0035] In order to make the technical means, creative features, achieved objectives and effects of the present invention easy to understand, the following specifically describes the social network information diffusion prediction method of the present invention in conjunction with embodiments and drawings.

[0036] <Embodiment>

[0037] Figure 1 It is a flowchart of the social network information diffusion prediction method in an embodiment of the present invention.

[0038] As Figure 1 shown, the social network information diffusion prediction method includes the following steps:

[0039] Step S1, construct an information diffusion prediction model.

[0040] Step S2, use the topic to be predicted, the set of users to be predicted related to the topic to be predicted, and the historical propagation information as the input of the information diffusion prediction model.

[0041] Figure 2 It is a schematic framework diagram of the information diffusion prediction model in an embodiment of the present invention.

[0042] As Figure 2 shown, the information diffusion prediction model in this embodiment is obtained by fusing the exposure time sorting algorithm, the gravitational sorting algorithm, and the cascade similarity sorting algorithm.

[0043] Step S3, based on the exposure time sorting algorithm, obtain the exposure time sorting array of the topic to be predicted to obtain the influence status of the time dynamic factor.

[0044] In a social network, assume that a user posts a message about topic A. Then all the fans who follow this user can see this message, thus having an information channel to access topic A. For users who further choose to participate in topic A, there will be a time difference between when they access topic A and when they participate in topic A. This time difference will not be too short because users need reaction time, and it will not be too long because the information on the online social network has timeliness. That is to say, this time difference may be regular, and this time difference is called the exposure time (the time when the user accesses the topic).

[0045] Figure 3 It is a flowchart of the exposure time sorting algorithm in an embodiment of the present invention.

[0046] As Figure 3 shown, in this embodiment, let u be a user who has not participated in the topic, and V be the set of users who have participated in a certain topic and are followed by u. For a specific topic A, the total exposure time of u is defined as:

[0047]

[0048] where \(t\) is the current time, and \(t\) v is the time of user \(v\) participating in the topic.

[0049] To more accurately characterize the exposure time function, it is necessary to count the total exposure time distribution \(P(t)\) for all users participating in the topic.

[0050] For all users who did not participate in the topic, calculate their total exposure time, and then use \(P(t=\text{sum} u ) to calculate the activation probability. Then, sort these activation probabilities in descending order, and the obtained sorting is called the exposure time sorting.

[0051] This embodiment also measures the impact of time dynamic factors on user behavior by analyzing the distribution of the infection times of neighbor nodes around different nodes in the social network, that is, the impact status of time dynamic factors. Specifically:

[0052] In the historical propagation information, the sum of the infection times when each propagating user participates in the propagation \(\text{sum} v =\text{sum} v +t - t u . According to the sum of infection time data, the sum of infection time distribution \(P\) is statistically obtained:

[0053] When calculating the specific impact, calculate the sum of the infection times for the user to be predicted, and then obtain the activation probability of the sum of the infection times \(P(\varepsilon=\text{sum} v ).

[0054] The input of the exposure time sorting algorithm in this embodiment is: propagation cascade \(C\), current time \(t\), information \(A\), user set \(U\) d ; the output is: exposure time sorting \(X\), candidate set \(Ca\). The code is as follows:

[0055]

[0056]

[0057] Step S4, based on the social network topology structure, derive the participating users of the topic to be predicted, and use the gravity sorting algorithm to calculate the influence of the participating users and the activation rate of the non-participating users to obtain the gravity sorting array.

[0058] In a social network, there should be an influence among users. In particular, in a social network with only one-way follow relationships, this influence is specifically reflected in the influence of the followed user on their followers. In the universal gravitation formula, the masses of two objects are used for calculation. When migrated to a social network, the two objects correspond to the followed user and the follower, and the masses of the two objects should be correspondingly replaced by the out-degree of the followed user and the in-degree of the follower. The out-degree of the followed user and the in-degree of the follower are analogous to the roles of the masses of the objects in gravitation. The larger the out-degree of the followed user, the greater the influence of this node on other nodes. The larger the in-degree of the follower, the more inclined this node is to accept new information. The magnitude of the influence of the followed user on the follower should be proportional to these two factors. When there is no direct follow relationship between two nodes but an indirect follow relationship, it is also considered that the influence can be transmitted along the path corresponding to the indirect follow relationship, but it is much smaller than the intensity of direct following. This is why in the gravitational model calculation formula, the gravity is inversely proportional to the square of the path length.

[0059] Figure 4 It is the flowchart of the gravitational sorting algorithm in the embodiments of the present invention.

[0060] In this embodiment, as Figure 4 shown, let u and v be two users with a path in the directed graph of the social network. Let path uv be the shortest path length from u to v, and U be the set of users participating in the topic. Then the resultant gravitational force for user v is defined as:

[0061]

[0062] In the formula, d u is the out-degree of user u, and d v is the in-degree of user v.

[0063] Sort all the users who have not participated in the topic according to the resultant gravitational force, and the gravitational sorting array is obtained.

[0064] Generally speaking, users with lower sorting values in the sorting array have a higher probability of being infected. Whether it is the exposure time sorting or the gravitational sorting, neither of them can independently make a comprehensive assessment of the probability of a user's published information. Combining the exposure time sorting and the gravitational sorting for prediction is a very promising solution.

[0065] The code of the gravitational sorting algorithm in this embodiment is as follows:

[0066]

[0067]

[0068] Step S5: Use the cascaded similarity ranking algorithm to calculate the similarity between each pair of users in the social network topology structure, and obtain the cascaded similarity ranking.

[0069] In this embodiment, the cascaded similarity ranking is derived from the cascaded data of each user. A vector is used to characterize the user, and the cosine similarity of the vectors is used to measure the similarity between them. Specifically:

[0070] First, extract the user IDs sorted by timestamp in each cascade and use them as a sentence.

[0071] Then, use one-hot encoding and the skip-gram algorithm to embed the ID of each user into a 128-dimensional vector, similar to the word embedding widely used in natural language processing references. In the optimization training of the skip-gram algorithm, the window size is set to 10, following the common practice in existing research.

[0072] Second, the cascaded similarity of label h between u and v is measured by the cosine similarity between u and v, that is: Assume that the set of transferred users followed by v is S u , then the transfer probability of user v is proportional to is proportional to.

[0073] Finally, sort the candidate set, that is, the user to be predicted, in descending order according to their similarity to obtain the cascaded similarity ranking (CSR).

[0074] The input of the cascaded similarity ranking algorithm in this embodiment is: the propagation behavior record Ω, the candidate set Ca, the current time t, and the information A; the output is: the cascaded ranking Z. The code is as follows:

[0075]

[0076] Step S6: Predict whether the set of users to be predicted participates in the topic to be predicted according to the influence status of time dynamic factors, the gravitational ranking array, and the cascaded similarity ranking.

[0077] In this embodiment, three algorithms are fused to obtain a fused ranking algorithm. The input of this algorithm is: the propagation behavior record Ω, the candidate set Ca, and the information set Λ; the output is: the prediction results TP, FN, FP, TN. The code is as follows:

[0078]

[0079]

[0080]

[0081] In the prediction of release time, the topological structure property (gravitational value), exposure time, and non-connected user behavior similarity are three key factors.

[0082] As shown in the above code, the information diffusion prediction model of this embodiment combines the influences of these three factors on the prediction result, and defines the fusion ranking as:

[0083] R = α * GR + β * ETR + (1 - α - β) * CSR

[0084] In the formula, α and β are hyperparameters of the information diffusion prediction model, GR is the gravitational ranking value, ETR is the exposure time ranking, and CSR is the cascade similarity ranking.

[0085] Regarding the problem of user behavior prediction based on time dynamics in social networks, this embodiment conducts an experimental comparison between the information diffusion prediction model (TR) and the existing game theory model (GT). Specifically:

[0086] To verify the performance of this method on datasets of different scales, three datasets are selected. Two of them are from the Stanford Network Analysis Project (SNAP) group at Stanford University. The SNAP datasets are mainly published on the website http: / / snap.stanford.edu / data / . As a network analysis platform, SNAP provides a large amount of complex network-related data for complex network researchers to use freely. The Twitter network data is only a part of it. The Twitter dataset provided by SNAP provides the relationship network of a total of 456,626 Twitter users from July 1, 2012 to July 7, 2012 and 14,855,875 social relationship edges between them (i.e., the first experimental dataset). The other one is the relationship network of 465,017 Twitter users and 834,797 social relationship edges collected by Haewoon Kwak et al. from October 2009 to November 2009 (i.e., the second experimental dataset). The third dataset is the relationship network of 595,460 users and 14,273,311 social relationship edges collected by Weng et al. from March 24, 2012 to April 25, 2012 (i.e., the third experimental dataset).

[0087] To verify the superiority of this method, it is compared with the following three existing user behavior propagation prediction methods: CT (extracted from A. Goyal, F. Bonchi, and L. V. Lakshmanan, "Learning influence probabilities in social networks," in Proceedings of the Third ACM International Conference on Web Search and Data Mining, 2010, pp. 241–250), DT (extracted from A. Goyal, F. Bonchi, and L. V. Lakshmanan, "Learning influence probabilities in social networks," in Proceedings of the Third ACM International Conference on Web Search and Data Mining, 2010, pp. 241–250), and GT (extracted from D. Li, S. Zhang, X. Sun, H. Zhou, S. Li, and X. Li, "Modeling information diffusion over social networks for temporal dynamic prediction," IEEE Transactions on Knowledge and Data Engineering, vol. 29, no. 9, pp. 1985–1997, Sep. 2017) models.

[0088] Among them, the CT and DT models consider the user's behavior as the sum of the influences of all its parent nodes and introduce a time delay factor to predict the specific time. The GT model regards the nodes in the social network as intelligent and rational individuals and calculates the corresponding benefits for different decisions that these individuals can make. By introducing the time-related benefits based on the propagation data, the GT model can be used to predict the information propagation process with temporal dynamic characteristics.

[0089] The specific experimental settings are as follows. In this method, a non - negative K - dimensional vector is used to represent the exposure time probability distribution. The larger the value of the hyperparameter K, the higher the prediction accuracy, but it also consumes more running time. Conversely, the smaller the value of the hyperparameter K, due to the reduction in vector length, information about the exposure time probability distribution is lost, and the prediction accuracy also decreases, but the running time consumed also decreases. In this experiment, considering the characteristics of the social network data collected, it is not appropriate to set K as a constant because the exposure time probability distributions vary greatly under different topics. Therefore, K varies with the maximum exposure time under different topics.

[0090] In addition, the concept of time steps is also required to predict the user participation time. Since the model makes predictions for each time step, increasing the number of time steps will make the predicted occurrence time more accurate. Because as the number of time steps increases, the time interval becomes smaller, and the predictions are more refined in terms of time, but it also requires more time. If the number of time steps is set too large, the running time will increase meaninglessly. For example, when the number of time steps is so large that the time interval is one second. If the number of time steps is set too small, the prediction results will not be practically meaningful. For example, when the number of time steps is too small and the time interval is one month, when predicting that a user will participate in information dissemination, this behavior can occur immediately after the prediction is made or one month later. Considering the rapid change of hot topics, this is meaningless.

[0091] Therefore, to comprehensively consider the running time and data characteristics, in the experiment of this embodiment, the number of time steps is set to 8, and the time interval length is determined accordingly with the number of time steps.

[0092] The experimental evaluation metrics used are Precision, Recall, and F1 - Measure, and their specific calculation formulas are as follows:

[0093]

[0094]

[0095]

[0096] In the formula, TP, FP, TN, and FN represent True Positive, False Positive, True Negative, and False Negative respectively.

[0097] The experimental results are respectively shown in Figures 5 to 13 It can be seen from the experimental results that the method proposed in the present invention is superior to other models in terms of Precision and F1 - Measure.

[0098] The GT model has a better recall rate because it has a lower activation threshold. As a result, a large number of potential users are predicted as positive. This setting greatly improves the recall rate performance of the GT model, but also significantly reduces its accuracy.

[0099] The above experiments demonstrate the superiority of the information diffusion prediction model (TR) of the present invention from two perspectives: time complexity and prediction effect. In a real social network environment, the scale of the user relationship network often reaches hundreds of thousands, millions, or even tens of millions. The improvement in time complexity makes the TR model have a wider application prospect in real social networks, which is of great significance.

[0100] Functions and effects of the embodiment

[0101] According to the social network information diffusion prediction method provided in this embodiment, this method comprehensively considers three key factors: social influence in the social network, the time-dynamic characteristics of information diffusion, and the similarity of non-connected user behaviors. Combining the characteristics of the real social network environment and complex network research, by integrating the exposure time sorting algorithm, the gravitational sorting algorithm, and the cascade similarity sorting algorithm, it can cover the prediction of the dissemination range of social network information, the prediction of user influence, and the prediction of user behavior, and proposes a model that can accurately predict the information diffusion based on time-dynamic characteristics in the social network. At the same time, the experimental results on the dataset fully demonstrate that the information diffusion prediction model proposed in this embodiment exceeds the existing advanced algorithms in the field in terms of performance, computational complexity, and robustness, providing a certain reference significance for the research in the field of release time prediction.

[0102] In the embodiment, the exposure time sorting algorithm can not only ensure the time-dynamic characteristics of information dissemination under one topic, but also consider the problem that the time-dynamic characteristics of information dissemination under different topics vary greatly. At the same time, by analyzing the distribution of the infection time of neighbor nodes around different nodes in the social network, the impact of time-dynamic factors on user behavior is measured.

[0103] In the embodiment, the gravitational sorting algorithm is derived from the social network topology structure, which can calculate the comprehensive influence of all users participating in the topic on their neighbors and sort all users not participating in the topic according to the activation probability.

[0104] In the embodiment, the cascade similarity sorting is derived from the cascade data of each user, uses a vector to describe the user, and measures the similarity between them using the cosine similarity of the vectors, so as to obtain the similarity of non-connected user behaviors in the social network.

[0105] The above embodiments are only used to illustrate the specific implementation manners of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. A method for predicting the spread of social network information, characterized in that, Taking into account three key factors: social influence in the social network, the time-dynamic characteristics of information diffusion, and the similarity of non-connected user behaviors, and combining the characteristics of the real social network environment and complex network research, including the following steps: Step S1, fuse the exposure time sorting algorithm, the gravitational sorting algorithm, and the cascade similarity sorting algorithm to construct an information diffusion prediction model; Step S2, use the topic to be predicted, the set of users to be predicted related to the topic to be predicted, and the historical propagation information as the input of the information diffusion prediction model; Step S3, obtain the exposure time sorting array of the topic to be predicted based on the exposure time sorting algorithm; Step S4, derive the participating users of the topic to be predicted based on the social network topology structure, and use the gravitational sorting algorithm to calculate the influence of the participating users and the activation rate of the non-participating users to obtain the gravitational sorting array; Step S5, use the cascade similarity sorting algorithm to calculate the similarity between users in the social network topology structure for each user, and obtain the cascade similarity ranking; Step S6, predict whether the set of users to be predicted participates in the topic to be predicted according to the exposure time sorting array, the gravitational sorting array, and the cascade similarity ranking, wherein, the process of obtaining the exposure time sorting array is: Statistically analyze the total exposure time distribution of all users participating in the topic to be predicted, and calculate the total exposure time of all users not participating in the topic to be predicted, Calculate the activation probability of all users not participating in the topic to be predicted based on the total exposure time distribution and the total exposure time, Sort the activation probabilities in descending order to obtain the exposure time sorting array.

2. A method for predicting information diffusion in a social network according to claim 1, characterized in that: Among them, The process of obtaining the gravitational sorting array is: Let u and v be two users with a path in the social network directed graph, and let path uv be the shortest path length from u to v. Let U be the set of users participating in the topic. Then the resultant gravitational force on user v is defined as: where d u is the out-degree of user u, and d v is the in-degree of user v Sort all users not participating in the topic according to the resultant gravitational force to obtain the gravitational sorting array.

3. A method for predicting information diffusion in a social network according to claim 2, characterized in that: Among them, The process of obtaining the cascade similarity is: Extract the user IDs sorted by timestamp in each cascade and use them as a sentence, Use one-hot encoding and skip-gram algorithm to embed the ID of each user into a 128-dimensional vector for optimization training, and set the window size to 10, The cascade similarity between u and v for label h is measured by the cosine similarity between u and v: Suppose the set of transferred users followed by v is S u , then the transfer probability of user v is proportional to is proportional to Sort the users to be predicted in descending order according to their similarity to obtain the cascade similarity ranking.

Images