An influence maximization method based on temporal network community detection
Through the time series network community detection and community division methods, the problem of insufficient accuracy of maximum influence seed node selection in the time series network is solved, and more efficient seed node selection is achieved, network timing information and structural characteristics are captured, and influence is improved. Maximize the accuracy and efficiency of the algorithm.
Patent Information
- Application Number
- CN202110924140.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-08-12
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-08-12
AI Technical Summary
The existing method of maximizing influence takes into account insufficient timing information in the timing network, insufficient accuracy in selecting seed nodes, and low algorithm efficiency.
Using a method based on time-series network community detection, through node embedding and community division, the micro and macro dynamics of the time-series network are learned, node similarity is calculated, and high-priority nodes in important communities are selected as seed nodes.
It improves the accuracy and efficiency of selecting seed nodes in the timing network, captures network structure and timing information, and enhances the accuracy of the influence maximization algorithm.
Smart Images

Figure SMS_17 
Figure SMS_51 
Figure SMS_53
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of Internet technology, in particular to the field of time series network analysis technology, and specifically relates to an influence maximization method based on time series network community detection. Background Art
[0002] In recent years, with the rapid development of the Internet and social media, the scale of online social networks has become larger and larger, and their dissemination scale and efficiency have far exceeded those of traditional social networks. A small number of nodes in the network may have an increasing impact on the entire network. Therefore, how to select a set of seed nodes to maximize the scope of information dissemination has attracted widespread attention. This problem is influence maximization. The problem of influence maximization was first proposed in the field of advertising marketing to solve the problem of viral marketing. They focused on how to select a group of influential people as initial users so that the company's products can be promoted to the greatest extent among the user group. Now, influence maximization is also widely used in scenarios such as recommendation systems, information diffusion, time detection, expert discovery, and link prediction.
[0003] Influence maximization specifically refers to finding a fixed number of initial active nodes in a given network, and through a specific propagation model, maximizing the number of final active nodes. Therefore, when solving the problem of influence maximization, we need to focus on both network structure and propagation model. Most of the current research on influence maximization is based on static networks. However, the actual network is not static, and its topology changes over time. Static networks ignore these changes and the time order in which the changes occur, which will result in the loss of certain network features. In order to take into account changes in the network over time, our influence maximization algorithm is based on a network whose structure changes over time, namely a time series network.
[0004] In summary, most solutions to influence maximization are also based on static networks, and there are three main types: greedy algorithms, heuristic methods, and community partitioning methods. In 2003, Kempe et al. first applied the greedy algorithm to solve the influence maximization problem. The algorithm starts with an empty set of seed nodes and iteratively adds nodes with the largest marginal benefit compared to the current set of seed nodes. This method has a high time complexity and cannot handle large-scale networks. Researchers have proposed a series of heuristic methods. Chen et al. proposed a degree reduction algorithm, which is based on penalizing the neighbor nodes of seed nodes. In addition, Morone et al. proposed a method for selecting seed nodes by considering the degree of distant nodes. However, the heuristic method is highly dependent on the network structure, and its accuracy varies greatly for networks with different structures. In most networks, the accuracy of heuristic methods in selecting seed nodes is much lower than that of greedy algorithms. In order to ensure accuracy as much as possible, Wang et al. proposed a CGA algorithm based on community partitioning to solve the influence maximization problem. Summary of the invention
[0005] The purpose of the present invention is to provide an influence maximization method based on time series network community detection, so as to solve the shortcomings of the existing influence maximization seed node selection method, such as insufficient consideration of time series information, insufficient accuracy in selecting seed nodes, and low algorithm efficiency.
[0006] The method of the present invention specifically comprises:
[0007] Step (1) Node embedding, including micro-dynamic embedding and macro-dynamic embedding:
[0008] For a given timing network and ε are the node set and edge set respectively, is a time series, each time series edge (u i ,u j ,t)∈ε represents node u i and node u j At time t, an edge is generated, and ε is the set of edges; the goal of temporal network embedding is to learn a mapping function where d is the embedding dimension, and represents the field of real numbers;
[0009] The microscopic dynamics are embedded by minimizing the following objective function To record the microscopic dynamics in the temporal network:
[0010] Where p represents node u i A historical neighbor before time t, and They are node u i and u j The historical neighbor set before time t;
[0011] The macrodynamics embedding is achieved by minimizing the following objective function To record the macroscopic dynamics in the time series network:
[0012] Δe(t)=e(t)-e(t-1); e(t) is the cumulative number of edges up to time t, and Δe′(t) is the number of new edges generated at time t;
[0013] Unified objective function The weight of macro-dynamic embedding η∈[0,1];
[0014] Step (2) Community detection: A clustering method that does not specify the number of communities is used to divide the nodes in the network into different communities. First, the network is embedded, and based on the embedding results, the similarity between each node pair is calculated. Then, each node is regarded as a community. If the similarity between two nodes is the largest, the two nodes are divided into the same community. The newly created community is regarded as a node, and the previous step is continued until the end condition is reached.
[0015] Step (3) Generate candidate seed nodes:
[0016] For a given timing network and a seed node set size k, a set of communities is obtained through community detection, that is, the clustering result C = {c1,c2,...,c R}, c r That is, the rth community, r = 1, 2, ..., R, R is the total number of communities, n r For the community r The number of nodes in the important community Select the top τ% nodes in the sum of node degree and similarity in each important community as candidate seed nodes, τ = 5-20, and the node degree is the number of neighbor nodes;
[0017] Step (4) Seed node selection:
[0018] (4-1) Seed node allocation: For a given sequential network and the size k of the seed node set, assuming C s ={c′1,c′2,...,c′ R′}, and each community c′ r′ There is n′ r′ nodes, then c′ r′ The number of seed nodes allocated in
[0019]
[0020] (4-2) Selecting seed nodes: Using the position score and centrality as evaluation indicators, the priority comparison rule is used to compare the priorities of the nodes in the candidate node set to determine the final seed node;
[0021] Position score It means any;
[0022] node Centrality SC is the community set in which seed nodes have been selected in community set C;
[0023] For two candidate nodes in a given community, they are compared according to the priority comparison rule. By collecting the quota (c r ″) high priority nodes to obtain the seed node set.
[0024] This patented method extends the problem of maximizing influence in static networks to time-series networks, and proposes an influence maximization algorithm based on community detection in time-series networks. First, the time-series network is embedded, and the nodes in the network are represented by low-dimensional vectors to capture the topological structure in the network. Then, the similarity between nodes is calculated through the node embedding vector, and the nodes are divided into different communities based on the similarity. Finally, seed nodes are selected from different communities to form the final seed node set.
[0025] The method of the present invention takes into account the time series information in the network and adopts the community division method to select seed nodes, which is the key point of the technical solution. When selecting the seed nodes with the greatest influence, the method of the present invention solves the shortcomings of the existing method for selecting seed nodes with the greatest influence, which is insufficient consideration of time series information and insufficient accuracy in selecting seed nodes, and enriches the method for selecting the seed node set with the greatest influence based on the time series network. DETAILED DESCRIPTION
[0026] The technical solution of the present invention is further described below by specific embodiments:
[0027] The present invention proposes a method for maximizing influence based on community detection in a time series network. Taking into account the time series information in the network and adopting a community division method to select seed nodes is the key point of the technical solution. When selecting the seed node with the greatest influence, the method of the present invention solves the shortcomings of the existing method for selecting seed nodes with the greatest influence, that is, insufficient consideration of time series information and insufficient accuracy in selecting seed nodes, and enriches the method for selecting seed nodes with the greatest influence based on a time series network.
[0028] The method is specifically:
[0029] Step (1) Node embedding: For a given temporal network and ε are the node set and edge set respectively, is a time series, each time series edge (u i ,u j ,t)∈ε represents node u i and u j At time t, an edge is generated, and E is the set of edges; the goal of temporal network embedding is to learn a mapping function where d is the embedding dimension, and represents the real number domain; the goal of the function f is to simulate the evolutionary pattern of the network in both the micro and macro dimensions of the temporal network, including micro dynamics embedding and macro dynamics embedding.
[0030] A. Microdynamics Embedding: Microdynamics describes a given temporal network The network structure formation process, edge formation time series (u i ,u j ,t) represents node u i and u j At time t, an edge is established as a temporal event, and the order number of the edge formation is m = 1, 2, ..., |ε|. Based on this, the temporal attention point process is used to record the microscopic dynamics of the temporal network;
[0031] First, for a time series edge o=(u i ,u j ,t), that is, a time series event, is embedded through the network to obtain the embedding matrix
[0032] For node u i The vector obtained by embedding is used to parameterize the intensity of the event. Based on the basic strength of the node itself and the historical influence of its neighbors, the temporal edge o = (u i ,u j ,t) Appearance intensity:
[0033] Among them, g(·) is a function that measures the similarity between two nodes. p represents node u i A historical neighbor before time t, t p Indicates the time when the historical neighbor p established the edge; and The node u i The historical neighbor node u p and node u j The historical neighbor node uq Embed the resulting vector; and They are node u i and u j The historical neighbor set before time t; α and β are two attention coefficients defined by the hierarchical temporal attention mechanism, and the time decay function κ(tt p )=exp(-δ i (tt p )),δ i >0 is the decay rate that is node-dependent and learnable.
[0034] The exponential function is used to transform the occurrence intensity to obtain the occurrence intensity which is a positive real number.
[0035] Determine the local influence from neighboring nodes:
[0036] Node u i The historical neighbor node u p With node u j The similarity between u p For the time series event o=(u i ,u j ,t) depends on node u i , and changes with the change of neighborhood structure, the local attention coefficient
[0037] Its intermediate parameters in, For the concatenation operation, note that the vector a∈R 2d , W is the local weight matrix, and the sigmoid function σ(x)=exp(x) / (1+exp(x)).
[0038] Determine the global influence from all neighbors:
[0039] Node u i With node u j The global attention coefficient of the formed edge Among them, node u i The intermediate parameters of attention Node u j The intermediate parameters of attention Node u i A collection of all historical neighbors' information Node u j A collection of all historical neighbors' information Average attenuation of past events s(·) is a single-layer neural network that aggregates information through neighbors Get aggregate embeddings with the average time decay of past events As input;
[0040] Then, at time t, node u i and u j The probability of establishing an edge between
[0041]
[0042] By minimizing the following objective function To record the microscopic dynamics in the temporal network:
[0043]
[0044] B. Macrodynamic embedding:
[0045] For a given timing network Macrodynamics refers to the evolution of network size. The number of edges in the network is used to represent the network size. The time series of the number of edges is e(t) is the cumulative number of edges up to time t;
[0046] The number of new edges generated at time t Where n(t) is the cumulative number of nodes up to time t. According to the power law distribution of network density in network evolution, the average number of reachable neighbors can be obtained: is the linear sparse coefficient, γ is the power law sparse index; any node u i The probability of connecting to other nodes at time t θ is the time decay exponent.
[0047] r(t) depends not only on the timing information in the network, but also on the network structure. First, as the network density increases, the number of edges in the network will also increase, and the rate of edge addition will decrease, that is, the connection rate will decay with the increase of the network time step. Second, the establishment of edges promotes the evolution of the network structure, and the connection rate is related to the network structure. Therefore, in order to capture these timing and structural characteristics in the network embedding, the time decay function t is used. θ and node embeddings to parameterize the connection probability.
[0048] By minimizing the following objective function To record the macroscopic dynamics in the time series network: Δe(t)=e(t)-e(t-1).
[0049] Since macro and micro dynamics mutually drive the evolution of temporal networks and thus affect network embedding learning alternately, the following unified model is used to capture the network topology formation process and the evolution pattern of network size: Unified objective function The weight η∈[0,1] of the macro-dynamics embedding.
[0050] Step (2) Community detection: A clustering method that does not specify the number of communities is used to divide the nodes in the network into different communities. First, the network is embedded, and based on the embedding results, the similarity between each pair of nodes is calculated. Then each node is regarded as a community, and if the similarity between two nodes is the largest, the two nodes are divided into the same community. The newly created community is regarded as a node, and the previous step is continued until the end condition is reached. In order to determine what the end condition is, the concept of modular gain is introduced. Specifically:
[0051] (2-1) Calculate the similarity between nodes: For each There is a d-dimensional vector To represent; then the two adjacent nodes u i and u j Similarity
[0052] (2-2) Calculate modular gain: For a given sequential network Assume that the clustering result C = {c1, c2, ..., c R}, c r That is, for the rth community, r = 1, 2, ..., R, R is the total number of communities, then the modular gain Among them, c r The sum of the similarities of the midpoints c r The sum of similarities between the intermediate node and other nodes in the network The sum of similarities between any two nodes in the network
[0053] For Network Given two different clustering results C and C′, the modularity gain ΔQ from C to C′ is C→C′ =Q(C′)-Q(C), assuming that the clustering results of the last iteration and this iteration are C and C′ respectively. If the modularity gain from C to C′ is negative, clustering stops.
[0054] Step (3) Generate candidate nodes: Since the social network in reality is very large, before selecting a set of seed nodes with the greatest influence, how to narrow down the candidate seed node set is also an important issue. Compared with simply considering large communities with a large number of nodes as important communities, important communities are distinguished based on the following definition.
[0055] First, select important communities: For a given time series network and a seed node set size k, a set of communities is obtained through community detection, that is, the clustering result C = {c1,c2,…,c R},n r For the community r The number of nodes in the important community
[0056] Based on the selected important communities, the central nodes of the community seem to be candidate seed nodes, but the hub nodes connecting multiple communities should also be considered, because the hub nodes can easily transmit influence to different communities. Therefore, in addition to taking the nodes with large degree in the community as candidate nodes, the nodes with large sum of node similarities are also defined as candidate nodes. Finally, the top 10% of the sum of node degree and similarity in each important community are selected as candidate seed nodes, and the node degree is the number of neighbor nodes.
[0057] Step (4) Seed node selection:
[0058] (4-1) Seed node allocation: Since the size of a community is an important factor in determining an important community, a large community is likely to spread its influence to more nodes than a small community. The total number of nodes in an important community is used to determine the number of seed nodes allocated to each important community. For a given time series network and the size k of the seed node set, assuming C s ={c′1,c′2,…,c′ R′}, and each community c′ r′ There is n′ r′ nodes, then c′ r′ The number of seed nodes allocated in
[0059]
[0060] (4-2) Selecting seed nodes: After determining C s After counting the number of seed nodes in each important community, we start selecting seed nodes in important communities. When selecting seed nodes, we should consider not only the central nodes but also the hub nodes. Therefore, we use the position score and hub degree as evaluation indicators. Finally, we use a priority comparison function to compare the priorities of the nodes in the candidate node set and determine the final seed nodes.
[0061] Position score: The node The position score is defined as u i Number of connected communities, location scores It means any;
[0062] For nodes Centrality SC is the community set in which seed nodes have been selected in community set C;
[0063] For two candidate nodes in a given community, the priority comparison is based on the following rules:
[0064] a. When comparing two non-hub nodes, the node with a larger degree has a higher priority; if the degrees of two non-hub nodes are the same, the node with a larger sum of similarities has a higher priority;
[0065] b. When comparing two hub nodes, the node with a higher position score has a higher priority; if the position scores of two hub nodes are the same, the node with a higher degree has a higher priority; if the degrees are the same, the node with a higher sum of similarities has a higher priority. In addition, the hub degree of the hub node must be greater than the hub threshold, because a low hub threshold means that it covers too many communities;
[0066] c. When comparing hub nodes and non-hub nodes, the node with a larger degree has a higher priority; if the degrees of two nodes are the same, the node with a larger sum of similarities has a higher priority. If the degrees are also the same, the hub node has a higher priority.
[0067] By collecting quota(c′ r′ ) high priority nodes to obtain the seed node set.
[0068] The contents described in this embodiment are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiment. The protection scope of the present invention also extends to the equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for maximizing influence based on temporal network community detection. It is characterized in that The method is specifically: Step (1) Node embedding, including micro-dynamic embedding and macro-dynamic embedding: For a given timing network and ε are the node set and edge set respectively, is a time series, each time series edge (u i ,u j ,t)∈ε represents node u i and node u j At time t, an edge is generated, and ε is the set of edges; the goal of temporal network embedding is to learn a mapping function where d is the embedding dimension, and represents the field of real numbers; The microscopic dynamics are embedded by minimizing the following objective function To record the microscopic dynamics in the temporal network: Where p represents node u i A historical neighbor before time t, and They are node u i and u j The historical neighbor set before time t; The macrodynamics embedding is achieved by minimizing the following objective function To record the macroscopic dynamics in the time series network: Δe(t)=e(t)-e(t-1); e(t) is the cumulative number of edges up to time t, and Δe′(t) is the number of new edges generated at time t; Unified objective function The weight η∈[0,1] of macro-dynamic embedding; Step (2) Community detection: A clustering method that does not specify the number of communities is used to divide the nodes in the network into different communities. First, the network is embedded, and based on the embedding results, the similarity between each node pair is calculated. Then, each node is regarded as a community. If the similarity between two nodes is the largest, the two nodes are divided into the same community. The newly created community is regarded as a node, and the previous step is continued until the end condition is reached. Step (3) Generate candidate seed nodes: For a given timing network and a seed node set size k, a set of communities is obtained through community detection, that is, the clustering result C = {c 1 ,c 2 ,...,c R }, c r That is, the rth community, r = 1, 2, ..., R, R is the total number of communities, n r For the community r The number of nodes in the important community Select the top τ% nodes in the sum of node degree and similarity in each important community as candidate seed nodes, τ = 5-20, and the node degree is the number of neighbor nodes; Step (4) Seed node selection: (4-1) Seed node allocation: For a given sequential network and the size k of the seed node set, assuming C s ={c 1 ′,c 2 ′,...,c′ R′ }, and each community c′ r′ There is n′ r′ nodes, then c′ r′ The number of seed nodes allocated in (4-2) Selecting seed nodes: Using the position score and centrality as evaluation indicators, the priority comparison rule is used to compare the priorities of the nodes in the candidate node set to determine the final seed node; Position score It means any; node Centrality SC is the community set in which seed nodes have been selected in community set C; For two candidate nodes in a given community, they are compared according to the priority comparison rule. By collecting the quota (c′ r′ ) high priority nodes to obtain the seed node set.
2. The method for maximizing influence based on time series network community detection as claimed in claim 1, It is characterized in that The microscopic dynamics embedding in step (1) is specifically: First, for a time series edge o=(u i ,u j ,t), and get the embedding matrix through network embedding For node u i Embedded vector, time series edge o = (u i ,u j ,t) appears intensity for: Among them, g(·) is a function that measures the similarity between two nodes. t p Indicates the time when the historical neighbor p established the edge; and The node u i The historical neighbor node u p and node u j The historical neighbor node u q The vector obtained by embedding; α and β are two attention coefficients defined by the hierarchical temporal attention mechanism, and the time decay function κ(tt p )=exp(-δ i (tt p )),δ i >0 is the decay rate that is related to the node and can be learned; the intensity of occurrence is transformed using an exponential function to obtain the occurrence intensity of positive real numbers Determine the local influence from neighboring nodes: node u i The historical neighbor node u p With node u j The similarity between u p For o=(u i ,u j ,t) depends on node u i , and changes with the change of neighborhood structure, the local attention coefficient Intermediate parameters For the concatenation operation, note that the vector W is the local weight matrix, sigmoid function σ(x) = exp(x) / (1+exp(x)); Determine the global influence of node u from all neighbors: i With node u j The global attention coefficient of the formed edge Among them, node u i The intermediate parameters of attention Node u j The intermediate parameters of attention Node u i A collection of all historical neighbors' information Node u j A collection of all historical neighbors' information Average attenuation of past events s(·) is a single-layer neural network that aggregates information through neighbors Get aggregate embeddings with the average time decay of past events As input; then, at time t, node u i and u j The probability of establishing an edge between By minimizing the following objective function To record the microscopic dynamics in the temporal network:
3. The method for maximizing influence based on time series network community detection as claimed in claim 2, It is characterized in that The specific embedding of macrodynamics in step (1) is: during the evolution of network size, the number of edges in the network is used to represent the network size, then the time series of the number of edges is The number of new edges generated at time t Among them, n(t) is the cumulative number of nodes up to time t, is the average number of reachable neighbors, is the linear sparse coefficient, γ is the power law sparse index; any node u i The probability of connecting to other nodes at time t θ is the time decay exponent; by minimizing the following objective function To record the macroscopic dynamics in the time series network: Δe(t)=e(t)-e(t-1).
4. The method for maximizing influence based on temporal network community detection as claimed in claim 3, It is characterized in that Step (2) is specifically: (2-1) Calculate the similarity between nodes: For each There is a d-dimensional vector To represent; then the two adjacent nodes u i and u j Similarity (2-2) Calculate modular gain: For a given sequential network Assume that the clustering result C = {c 1 ,c 2 ,...,c R }, then the modular gain Among them, c r The sum of the similarities of the midpoints c r The sum of similarities between the intermediate node and other nodes in the network The sum of similarities between any two nodes in the network For Network Given two different clustering results C and C′, the modularity gain ΔQ from C to C′ is C→C′ =Q(C′)-Q(C), assuming that the clustering results of the previous iteration and the clustering results of this iteration are C and C′ respectively; if the modular gain from C to C′ is negative, clustering stops.
5. The method for maximizing influence based on time series network community detection as claimed in claim 4, It is characterized in that In step (4), for two candidate nodes in a given community, the priority comparison rule is as follows: a. When comparing two non-hub nodes, the node with a larger degree has a higher priority; if the degrees of two non-hub nodes are the same, the node with a larger sum of similarities has a higher priority; b. When comparing two hub nodes, the node with a higher position score has a higher priority; if the position scores of two hub nodes are the same, the node with a larger degree has a higher priority; if the degrees are the same, the node with a larger sum of similarities has a higher priority; in addition, the hub degree of the hub node must be greater than the hub threshold, because a low hub threshold means that it covers too many communities; c. When comparing hub nodes and non-hub nodes, the node with a larger degree has a higher priority; if the degrees of two nodes are the same, the node with a larger sum of similarities has a higher priority; if the degrees are also the same, the hub node has a higher priority.
