Influence blocking maximization algorithm based on overlapping interest community detection

Abstract

The invention provides an influence blocking maximization algorithm based on overlapping interest community detection. Comprising a competition propagation model considering multiple factors and an algorithm based on the model and overlapping interest community detection, and is used for multi-information competition propagation process modeling and positive information seed node selection in a social network. Competition propagation process modeling and influence blocking maximization are hotspots of social network analysis, a traditional model is expanded based on IC and LT, and it is difficult to accurately describe a real propagation process. According to a traditional community-based algorithm, communities are divided based on a topological structure, and user interests and preferences are ignored. According to the method, interests and preferences of the users, and information interaction time delay and credibility among the users are comprehensively considered, so that a competition propagation process can be better modeled; a traditional community concept is expanded to an interested community, and the community detection accuracy is improved; and the selection of positive information seed nodes is guided through an overlapping interest community structure, so that an expected negative information influence blocking effect is achieved in a short time.

Description

Technical Field

[0001] The invention belongs to the field of social network analysis, and specifically proposes a competitive propagation model considering multiple factors and an influence blocking maximization algorithm based on the model and overlapping interest community detection. Background Art

[0002] In recent years, social platforms such as Weibo, Facebook, and Twitter have developed rapidly, and the number of registered users and active users has continued to increase. Users have established social connections through social behaviors such as following, thus building a large-scale social network. The emergence of social networks has facilitated information interaction and sharing, and information dissemination has shown the characteristics of fast speed and wide range. Given that user decisions are influenced by neighbors, it is possible to target a small number of users in social networks and utilize the complex social connections between users to achieve large-scale information dissemination. The problem of information maximization (IM) has emerged. While the rapid development of social networks has brought convenience, some negative information will also spread rapidly and widely. Given the urgency of blocking the influence of negative information, it is necessary to minimize its impact within a certain time constraint. In this context, a variant of the IM problem: Influence Blocking Maximization under Time constraint (T-IBM) has attracted attention and been widely studied. The problem is described as follows: a set of seed nodes of negative information is known and has been propagated in the social network for a period of time. Seed nodes are selected to publish positive information to the social network. The two compete for propagation in the social network according to the specified propagation model. Within the given time constraint, positive information occupies a dominant position in the propagation. On the one hand, the inhibitory effect of positive information on negative information is considered, and on the other hand, the influence range of positive information itself is considered.

[0003] The T-IBM problem involves the competitive propagation of positive and negative information in social networks. The traditional single information propagation model (independent cascade model, linear threshold model) is no longer applicable. It is necessary to describe the multi-information propagation model in which multiple pieces of information compete and propagate in social networks at the same time. The existing multi-information propagation models are mainly extensions of the classic single information propagation model: the multi-information propagation model based on the independent cascade model and the multi-information propagation model based on the linear threshold model. In some social network environments: users may be interested in specific topics, and their interest in different topics varies. The information fed back to the social network has a certain topic tendency; when a neighbor forwards a piece of information to a user, the user does not receive it immediately, but there is a certain time delay; trust is established between users in information interaction, and a trust calculation mechanism is needed to evaluate the degree of trust between users. Moreover, social networks usually reflect a community structure: users within the community are closely connected, and users outside the community are sparsely connected.

[0004] For the solution of T-IBM problem, there are some algorithms that block the spread of negative information by publishing positive information. Similar to the solution of IM problem, they can be divided into three categories: approximate algorithms that can provide performance guarantee but the efficiency needs to be improved, community-based algorithms that effectively use community structure to improve scalability, and heuristic algorithms that are efficient but cannot provide performance guarantee. Existing community-based algorithms usually only consider the topological structure of social network, ignoring the influence of user interest preferences on the formation of community structure. The community divided by interest preference can be defined as interest community: users within the interest community are closely connected and have high interest preference similarity, while users outside the interest community are sparsely connected and have low interest preference similarity; at the same time, users may belong to multiple interest communities, that is, interest communities may overlap.

[0005] In summary, the existing multi-information competitive propagation models do not comprehensively consider user interest preferences, information interaction delays between users, and trust between users, and community-based algorithms ignore the impact of user interest preferences on the community (interest community) structure in social networks. Therefore, it is of great practical significance to study the general competitive propagation model close to real social networks and the algorithm based on interest communities. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes an influence blocking maximization algorithm based on overlapping interest community detection, including a competitive propagation model considering multiple factors and an influence blocking maximization algorithm based on the model and overlapping interest community detection. The universality of the competitive propagation model and the accuracy of community detection are improved, and the selection of positive information seed nodes is guided to effectively suppress the spread of negative information under given time constraints.

[0007] In a first aspect, a competitive communication model considering multiple factors is provided, including: social network modeling, mathematically describing users in a social network, social connections between users, and heterogeneous characteristics presented by the social network to obtain a social network model; calculating the trust between users based on the heterogeneous characteristics of the social network to improve the social network model; based on the social network model, defining all possible states of users in the social network and a transition mechanism between states; and describing a multi-information competitive communication process based on the social network model and the user state transition mechanism.

[0008] In an optional implementation, the heterogeneity characteristics presented by the social network include: user heterogeneity and social connection heterogeneity, user heterogeneity includes user interest preferences, which are modeled using topic preference vectors; social connection heterogeneity includes information interaction delay between users, information interaction times, and trust between users, wherein the information interaction delay between users is measured by the encounter probability m uv Modeling, m uv It represents the probability of user u and user v meeting at any time step. After user u forwards information to v, the corresponding meeting probability is checked at each time step. When the meeting probability is met, user u and user v meet, which means that user v has received the information forwarded by user u.

[0009] In an optional implementation, the calculation of the trust between users based on the heterogeneity characteristics of the social network is shown in the following formula:

[0010] t uv =sx_t uv +s uv +reli uv

[0011] Among them, t uv represents the trust of user v in neighbor u, sx_t uv The trust generated by familiarity depends on the number of information exchanges between users. The more frequent the information exchanges between users, the more familiar the users are with each other, and the higher the trust generated by familiarity. uv The trust generated by similarity depends on the similarity of interest preferences between users. The higher the similarity of interest preferences, the more common language there is between users, and the easier it is to accept the information forwarded by each other, and the higher the trust generated by the corresponding similarity. The information exchange between user v and neighbor u is not frequent, but neighbor u itself has a high degree of trustworthiness (manifested in that it frequently exchanges information with some neighbors, or it maintains information exchange with a large number of neighbors). At this time, user v also tends to trust neighbor u, so reliability uv It represents the trustworthiness of neighbor u for user v, and its calculation depends on the information interaction between neighbor u and other neighbors.

[0012] In an optional implementation, the definition of all possible states of users in a social network and the transition mechanism between states includes: defining the possible states of users based on the acceptance and forwarding of information by users in the social network: ① not accepting and not forwarding, ② accepting a certain message but not forwarding, ③ accepting a certain message and forwarding the message, and the states ② and ③ of different messages need to be defined differently. Here, for a unified description, the acceptance of information and forwarding of information are separated, that is, acceptance ≠ forwarding, because in real life, users also have certain interest preferences in forwarding information; the state transition mechanism should include: when the user is in state ①, receiving a certain message, what is the probability that the user will accept the information and forward it? If multiple messages are received, which message will the user choose to accept and whether to forward it? When the user is in state ②, receiving the accepted message again, what is the probability that the user will choose to forward the message? If other information is received other than the received information, whether to accept the new information and whether to forward it? The probabilities need to be calculated by comprehensively considering the user's interest preferences and the trust between users.

[0013] In an optional implementation, the description of the multi-information competitive propagation process based on the social network model and the user state transition mechanism includes: the node state in the seed node set of each information is ③, and the other node states are ①; the multi-information competitive propagation process is carried out in discrete steps, and at each time step, the node newly converted to state ③ in the previous time step forwards the corresponding information to all neighbor nodes that are not in state ③, and arrives at the neighbor node after a corresponding delay. After the neighbor node receives the information, it makes a state transition decision based on the information forwarded by other nodes; when the preset time constraint is reached or the termination condition is met (there is no node newly converted to state ③ and there is no node to be decided), the information propagation process ends.

[0014] In a second aspect, an influence blocking maximization algorithm based on the competitive propagation model considering multiple factors and overlapping interest community detection is provided, including: based on the social network model, identifying a complete subgraph composed of users with high centrality and interest preference similarity as an initial community; based on the social network model and the initial community, detecting the overlapping interest community structure in the social network based on label propagation; based on the social network model and the overlapping interest community structure, extracting overlapping nodes and non-overlapping nodes and constructing a set of candidate seed nodes according to influence values; and selecting a final set of seed nodes based on the competitive propagation model considering multiple factors and the set of candidate seed nodes.

[0015] In an optional embodiment, the identification of a complete subgraph composed of users with high centrality and interest preference similarity as the initial community includes: arranging the nodes in the social network in descending order of centrality, and visiting them in sequence to construct the initial community; if the centrality of the currently visited node is greater than the average centrality, joining the current initial community; evaluating the importance of neighbor nodes to the currently visited node based on the interest preference similarity and centrality with the currently visited node, selecting the most important neighbor nodes to join the initial community, and retaining the common neighbor nodes of the currently visited node and the newly added neighbor nodes; continuously searching for the nodes with the highest importance to the current initial community from the common neighbor nodes to add to the initial community and deleting non-common neighbor nodes until the set of common neighbor nodes is empty.

[0016] In an optional embodiment, the method for detecting overlapping interest community structures in social networks based on label propagation includes: setting a unique label for the initial community, the nodes in the initial community share the same label, and the attribution coefficient is 1; in each iteration, the label set of each node in the current iteration is updated based on the label set of the neighboring node in the previous iteration, until the minimum number of nodes belonging to each label no longer changes. The method for updating the label set of the node in the current iteration based on the label set of the neighboring node in the previous iteration includes: summarizing all labels appearing in the label set of the neighboring node and adding the corresponding attribution coefficients, and weighting the attribution coefficient according to the relationship strength of the node to the community corresponding to the label (the number of social connections between the node and the community) and the similarity of interest preferences.

[0017] In an optional implementation, the extracting overlapping nodes and non-overlapping nodes and constructing a candidate seed node set according to the influence value includes: extracting overlapping nodes and non-overlapping nodes based on the overlapping interest community structure, and calculating their influence values ​​on the information propagation process respectively, wherein the overlapping nodes are simultaneously associated with multiple communities (C v : The community set connected to the overlapping nodes) is located at the hub of information dissemination between multiple communities, and its influence value depends on the relationship strength r with each community v→c And the average similarity of interest preferences with each node in the community s B→c ;

[0018]

[0019] The non-overlapping nodes are located inside the community, and the impact of their two-hop neighbor nodes on the information propagation process is considered, including the trust value and the information interaction delay.

[0020] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0021] 1) The present invention proposes a competitive communication model that considers multiple factors, comprehensively analyzes the impact of user interest preferences, information interaction delay between users and trust between users on information dissemination, and in particular, proposes a new method for calculating trust between users based on the principle of trust generation between people in social psychology; by reasonably designing a node state transition mechanism, the process of competitive dissemination of multiple pieces of information in a social network can be effectively modeled.

[0022] 2) The present invention performs overlapping interest community detection based on label propagation, extends traditional overlapping community detection to overlapping interest community detection, and effectively utilizes user interest preferences to improve detection efficiency and accuracy.

[0023] 3) The present invention proposes an influence blocking maximization algorithm based on a competitive propagation model and overlapping interest community detection. On the basis of detecting the overlapping interest community structure in 2), the role and influence of overlapping nodes and non-overlapping nodes are analyzed, different influence calculation methods are proposed, and positive information seed nodes are screened out. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. Among them:

[0025] Figure 1 It is a node state transition diagram in the competitive propagation model considering multiple factors in the present invention;

[0026] Figure 2 This is a flow chart of the influence blocking maximization algorithm based on overlapping interest communities in the present invention:

[0027] Figure 3 This is a flow chart of the initial community identification algorithm in the present invention;

[0028] Figure 4 It is a flow chart of the overlapping interest community detection algorithm based on label propagation in the present invention;

[0029] Figure 5 The performance improvement of the proposed method compared with the commonly used heuristic algorithms on four real social network datasets is shown in FIG. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. The embodiments are only used to illustrate and explain the present invention, and do not limit the present invention. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0031] Figure 1 The node state transition diagram in the competitive propagation model considering multiple factors of the present invention is shown. For example, assuming that two pieces of information A and B are competing for propagation in a social network at the same time (which can be extended to more than two pieces of information), the model defines: a node in a social network may be in the following five states:

[0032] inactive: The node has not received, accepted, or forwarded any information;

[0033] accept-A (accept-B): The user receives information A (B) and accepts the information but does not forward it;

[0034] active-A(active-B): User forwards information A(B).

[0035] Exemplary, combined Figure 1 Taking information A (the same applies to information B) as an example, the state transition process of the node is introduced:

[0036] inactive->accept-A->active-A: The initial node is in the inactive state. At this time, it is affected by information A and needs to determine whether the node accepts and further forwards information A. Consider the combined trust of all neighboring nodes that forward information A to the node and the topic fit of the node to information A; if it is affected by both information A and information B, they need to be considered separately;

[0037] accept-A->active-A: The initial node is in the accept-A state, that is, when the node receives information A, the topic preference for information A alone is not enough to make it forward information A. At this time, it is affected by information A again and needs to determine whether the node forwards information A. Considering the number of neighbor nodes that have forwarded information A to the node up to this moment, the impact is cumulative;

[0038] accept-A->accept-B->active-B: The initial node is in the accept-A state. At this time, it is affected by information B and needs to determine whether the node will accept information B and further forward information B. First, compare the topic fit between the node and information A and B. If the topic fit with information A is higher, if you want the node to accept information B, you need to make up for the difference in topic fit between information A and B by the combined trust of all neighboring nodes that forward information B to the node; otherwise, it is a gain.

[0039] Figure 2 The influence blocking maximization algorithm process based on the competitive propagation model considering multiple factors and overlapping interest community detection is demonstrated. The algorithm is divided into four stages: the first stage: initial community identification; the second stage: overlapping interest community detection based on label propagation; the third stage: constructing a set of candidate seed nodes; the fourth stage: selecting the final set of seed nodes.

[0040] In the first stage, based on the social network model, a complete subgraph composed of users with high centrality and similarity of interest preferences is identified as the initial community. Exemplarily, node degree is used as the centrality measure. First, all nodes in the social network are arranged in descending order of degree and visited sequentially; if the node degree is greater than the average degree of nodes in the social network, it is added to the current initial community and the node with the highest importance to the currently visited node is selected from the neighbor nodes to join the current initial community; the public neighbor node set of the current initial community is constructed, and if the public neighbor node set is not empty, the node with the highest importance to the current initial community is found in the public neighbor node set to join the current initial community, and the public neighbor node set is updated. Repeat the above process until the public neighbor node set is empty; determine whether the number of nodes in the current initial community is greater than 3 (if the number of nodes is too small and not representative), if so, add it to the initial community set.

[0041] In the second stage, based on the social network model and the initial community, the overlapping interest community structure in the social network is detected based on label propagation. First, a unique label is set for each initial community, and the nodes in the initial community share the same label, and the attribution coefficient is 1; the node order is disrupted, and the node random access sequence is obtained and sequentially accessed; the label set of the current access node in this iteration is updated according to the label set of the neighbor nodes of the current access node in the previous iteration, including: summarizing all labels that appear in the label set of the neighbor nodes of the current access node, and adding the corresponding attribution coefficients; weighting the attribution coefficient according to the relationship strength and interest preference similarity between the current access node and the community corresponding to each label to obtain the weighted attribution coefficient; recording the maximum weighted attribution coefficient, deleting the label with the weighted attribution coefficient / maximum weighted attribution coefficient < a given threshold; if the node has been visited, determine whether the termination condition is met: the minimum number of members of each interest community label remains unchanged in the previous and next iterations, and if it is met, the nodes holding the same label are added to the same interest community, and some nodes may hold multiple labels and belong to multiple interest communities, thereby forming an overlapping interest community structure.

[0042] In the third stage, based on the social network model and the overlapping interest community structure, overlapping nodes and non-overlapping nodes are extracted and a set of candidate seed nodes is constructed according to the influence value. The influence value of the overlapping node depends on the node and the connected community C v The strength of the relationship v→c and topic preference similarity s B→c .

[0043]

[0044] Exemplarily, the relationship strength is obtained by comparing the number of edges connecting the overlapping nodes to the community with the total number of edges of the overlapping nodes and the total number of edges of the nodes in the community. The topic preference similarity is the average of the topic preference similarities between the overlapping nodes and all nodes in the community. For non-overlapping nodes, the trust and encounter probability of their two-hop neighbor nodes are considered. The higher the trust, the more likely the neighbor node is to accept the information forwarded by it and expand the dissemination range. The higher the encounter probability, the shorter the delay of information interaction, so as to achieve a larger dissemination range in a shorter time.

[0045] In the fourth stage, the final seed node set is selected based on the competitive propagation model considering multiple factors and the candidate seed node set. Exemplarily, a greedy strategy is adopted to continuously select nodes with the largest marginal gain from the candidate seed node set to add to the final seed node set. Regarding the calculation of the influence range, a Monte Carlo simulation is adopted to input the seed node set of each information into the competitive propagation model considering multiple factors, and the average value is calculated by running 1000 times.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application is described in detail with reference to the above embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the above embodiments, or to replace some or all of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A competitive propagation model considering multiple factors, characterized in that: The following heterogeneous characteristics of social networks are comprehensively considered: user interest preferences, the delay of information interaction between users, and the number of information interactions to model social networks; The trust between users is calculated based on the heterogeneous characteristics of social networks; based on the improved social network model, the status of users in the social network and the transition mechanism between statuses are defined; finally, the multi-information competitive dissemination process is described.

2. The method of calculating the trust between users based on the heterogeneity characteristics of social networks according to claim 1, characterized in that: Combining the principle of trust between people in social psychology, we can calculate the trust t of user v towards neighbor u from three dimensions. uv Analysis: sx_t uv The trust generated by familiarity depends on the number of information exchanges between users. The more frequent the information exchanges between users, the more familiar they are with each other, and the higher the trust generated by familiarity. uv The trust generated by similarity depends on the similarity of interest preferences between users. The higher the similarity of interest preferences, the more common language there is between users, and the easier it is to accept the information forwarded by each other, and the higher the trust generated by the corresponding similarity. uv The degree of trustworthiness of neighbor u for user v is higher when neighbor u frequently exchanges information with some neighbors or exchanges information with a large number of neighbors. Even if the information exchange between user v and neighbor u is not frequent, user v tends to trust neighbor u. t uv =sx_t uv +s uv +reli uv 3. The method of defining the status of users in a social network and the transition mechanism between the statuses according to claim 1, characterized in that: Based on the user's acceptance and forwarding of information in social networks, the possible states of users are defined as follows: ① not accepting and not forwarding, ② accepting a piece of information but not forwarding it, ③ accepting a piece of information and forwarding it. Among them, the acceptance of information and forwarding of information are separated, that is, acceptance ≠ forwarding, because in real life, users also have certain interest preferences when forwarding information; based on the user state definition, a state transition mechanism is designed.

4. According to claim 3, the state transition mechanism is designed based on the user state definition, characterized in that: A node in state ① does not receive any information. When receiving a single message, it decides whether to accept the message based on the combined trust of the neighboring nodes that forwarded the message to the node and the topic fit between the node and the message. Further, it decides whether to forward the message based on the topic fit. If a node in state ① receives multiple messages, it makes judgments in the above manner. A node in state ② receives a message but does not forward it. If it receives the message again, it decides whether to forward the message based on the number of neighboring nodes that forwarded the message to the node, i.e., the impact accumulates. If a node in state ② receives other information besides the received information, it needs to determine whether the node abandons the received information and accepts the new information instead. It needs to calculate the combined trust of the neighboring nodes that forward the new information to the node and compare the topic fit between the node and the received information and the new information. If the topic fit with the received information is higher, the combined trust is needed to make up for the difference in the topic fit between the node and the two messages. On the contrary, trust works together to create a positive effect.

5. An influence blocking maximization algorithm based on overlapping interest community detection, characterized in that: A four-stage algorithm framework: Phase 1: Identify the complete subgraph composed of users with high centrality and interest preference similarity as the initial community; Phase 2: Detect overlapping interest community structures in social networks based on label propagation; The third stage: extract overlapping nodes and non-overlapping nodes based on the overlapping interest community structure and construct a set of candidate seed nodes according to the influence value; The fourth stage: select the final seed node set, and use a greedy strategy to continuously select the nodes with the largest marginal gain from the candidate seed node set to add to the final seed node set. Regarding the calculation of the influence range, Monte Carlo simulation is adopted, and the seed node set of each information is input into the competitive propagation model that considers multiple factors, and the average value is calculated by running it 1000 times.

6. According to claim 5, a complete subgraph composed of users with high centrality and interest preference similarity is identified as an initial community, characterized in that: Arrange the nodes in the social network in descending order of centrality, and visit them in sequence to build the initial community; if the centrality of the currently visited node is greater than the average centrality, join the current initial community; evaluate the importance of neighbor nodes to the currently visited node based on the similarity of interest preferences and centrality with the currently visited node, select the neighbor nodes with the highest importance to join the initial community, and retain the common neighbor nodes of all nodes in the current initial community; Continuously search for the nodes with the highest importance to the current initial community from the public neighbor nodes, add them to the initial community and delete non-public neighbor nodes until the public neighbor node set is empty.

7. The method for detecting overlapping interest communities in social networks based on label propagation according to claim 5, characterized in that: A unique label is set for the initial community, and nodes in the initial community share the same label, and the attribution coefficient is 1; In each iteration, the label set of each node in the current iteration of the node is updated based on the label set of the neighboring nodes in the previous iteration, including: summarizing all labels that appear in the label set of the neighboring nodes of the current access node, and adding the corresponding attribution coefficients, weighting the attribution coefficients according to the relationship strength and interest preference similarity between the current access node and the community corresponding to each label to obtain the weighted attribution coefficient, recording the maximum weighted attribution coefficient, and deleting labels whose weighted attribution coefficient / maximum weighted attribution coefficient < a given threshold; until the minimum number of nodes belonging to each label no longer changes; nodes holding the same label belong to the same interest community, and some nodes may hold multiple labels and belong to multiple interest communities, thus forming an overlapping interest community structure.

8. According to claim 5, overlapping nodes and non-overlapping nodes are extracted based on the overlapping interest community structure and a candidate seed node set is constructed according to the influence value, characterized in that: Based on the overlapping interest community structure, overlapping node sets and non-overlapping node sets are extracted. Overlapping nodes are simultaneously associated with multiple communities (C v : The community set connected to the overlapping node v) is located at the hub of information dissemination among multiple communities, and its influence value depends on the relationship strength r with each community v→c And the average similarity of interest preferences with each node in the community s B→c ; The non-overlapping nodes are located inside the community, and the impact of their two-hop neighbor nodes on the information propagation process is considered, including trust and encounter probability. The higher the trust, the more likely the neighbor nodes are to accept the information forwarded by them and expand the propagation range. The higher the encounter probability, the shorter the delay of information interaction, so as to achieve a larger propagation range in a shorter time.

Images