A User Node Selection Method Based on Multi-Task Evolutionary Optimization for Enhancing the Information Dissemination Range of Social Networks

Abstract

This invention discloses a user node selection method based on multi-task evolutionary optimization of network information propagation, comprising the following steps: Step A, using multiple low-cost heuristics to quickly mine different "important node sets" from the original graph, and using these sets to induce multiple subgraphs, forming an initial search space for multiple sub-tasks; Step B, the main task and sub-tasks are carried out in parallel and collaboratively, with the main task optimizing the seed set on the original graph; Step C, the main task drives unidirectional knowledge transfer; Step D, determining the relationship between the sub-tasks and the main task. This invention significantly reduces the effective search space, and in high-frequency iteration scenarios, it can reduce the overall running time and resource consumption to an engineeringable range. It avoids the optimality loss caused by search perspective bias by relying on the main task, exhibits better solution quality and result stability under different types of networks, improves overall robustness and convergence reliability, and is applicable to multiple diffusion models and multiple network types, thus expanding its applicability.

Description

Technical Field

[0001] This invention specifically relates to a user node selection method based on multi-task evolution optimization of network information propagation. Background Technology

[0002] Influence Maximization (IM) is a crucial problem in social network analysis and combinatorial optimization. Its goal is to select a set of seed nodes with a budget of k from the network, given a diffusion model, to maximize the reach (desired number of activated nodes) of information, opinions, or behaviors. This problem has significant applications in scenarios such as precision marketing, public opinion propagation and suppression, social recommendation, public safety incident information dissemination, and emergency notification delivery.

[0003] Under classic diffusion models (such as Independent Cascade (IC), Weighted Cascade (WC), and Linear Thresholding (LT), the Imaging Problem (IM) has been proven to be NP-hard. While traditional greedy approximation algorithms can provide theoretical approximation guarantees, they typically require extensive Monte Carlo simulations to evaluate the influence of candidate seed sets, leading to a sharp increase in computational cost with network size. This makes it difficult to meet the real-time and resource-consumption requirements of large-scale online services. Sampling-based RIS-type algorithms alleviate the complexity problem to some extent, but still face challenges in ultra-large-scale networks, including high memory consumption, complex engineering implementation, and sensitivity to graph structure changes.

[0004] In recent years, heuristic methods (Degree, PageRank, K-core, etc.) can quickly provide a set of "relatively important nodes" at extremely low cost. However, they neglect the overlapping propagation coverage between nodes, the cascading effect of diffusion paths, and cooperative relationships, and the quality of the solutions often fails to meet the requirements of high-demand applications. On the other hand, metaheuristic optimization methods such as evolutionary algorithms (EAs) have strong global search capabilities and adaptability. However, as the network size increases, the dimensionality of the search space expands dramatically. Crossover and mutation operators can easily cause the population to get stuck in local optima or the search efficiency to decrease in high-dimensional space, resulting in problems such as slow convergence and insufficient stability.

[0005] Therefore, there is an urgent need for an influence maximization solution framework that can significantly reduce the effective search space while maintaining high solution quality, and can leverage multi-source prior knowledge to enhance global exploration capabilities, making it suitable for engineering deployment scenarios of large-scale social networks. Summary of the Invention

[0006] The purpose of this invention is to provide a user node selection method based on multi-task evolution optimization of network information propagation.

[0007] The technical solution to achieve the purpose of this invention is a user node selection method based on multi-task evolutionary optimization of network information propagation, comprising the following steps:

[0008] Step A: Use multiple low-cost heuristics to quickly mine different "important node sets" from the original graph, and use the above sets to induce multiple subgraphs, forming the initial search space for multiple sub-tasks;

[0009] Step B: The main task and sub-tasks are carried out in parallel and collaboratively. The main task optimizes the seed set on the original graph, while the sub-tasks optimize on their respective subgraphs. The sub-tasks generate high-quality individuals and serve as candidate "guide samples" for the main task.

[0010] Step C: The main task drives unidirectional knowledge transfer. After each iteration of the main task, some individuals are transferred from each subtask to the main task, indirectly improving the optimal value of the main task.

[0011] Step D: Determine the relationship between the subtask and the main task. If the subtask and the main task have low similarity and the transfer has no effect, introduce a subgraph expansion.

[0012] When the migration probability is below a threshold, the subgraph is automatically expanded to improve structural similarity.

[0013] As the subgraph gradually expands to the same size as the main graph, the subtasks should flow unidirectionally to the main task to ensure that the main task is not diluted.

[0014] A further preferred option is: in step A, a directed or undirected graph is given. ,

[0015] in It is the set of nodes in the network, where n represents the total number of nodes; It is the set of edges in the network. Represents the number of sides;

[0016] Each edge Associated with a diffusion probability , represents a node Successfully activated node The probability of;

[0017] Select seed set In the diffusion model, the influence range after the propagation ends is denoted as σ(S), and the objective is:

[0018]

[0019] Each edge in the model With activation probability ,exist Seed node activation; in At this time, the newly activated node in the previous time step attempts to activate its neighbors independently; propagation terminates when no new activated node is found.

[0020] A further preferred approach is: in step A, for the original network, at least two heuristic importance algorithms are used to generate a set of candidate nodes and construct a subgraph;

[0021] Each subtask has its own independently configured evolutionary algorithm parameters.

[0022] A further preferred embodiment is that in step B, the multi-task parallel collaboration includes individual encoding: an individual represents a seed set, using a node sequence / set encoding of length 𝑘.

[0023]

[0024] Duplicate nodes are not allowed in the encoding.

[0025] Population initialization: First, maintain a candidate node list TabooList. Second, construct individuals one by one: randomly select nodes to add to individuals. Then, define Similarity(u, v) based on topological similarity or neighborhood overlap, and add nodes that are too similar to the selected nodes to TabooList to avoid subsequent selection. Finally, if NodeList is empty or cannot meet the length 𝑘, fill it with the remaining nodes to obtain an initial population with higher diversity.

[0026] Crossover operator: from parent individual Choose one as the benchmark Randomly select intersection point i ∈ [0,k], swap the segments after the intersection point. If the swap introduces duplicate nodes, perform conflict resolution on the duplicate items and generate feasible descendants to be added to the offspring population.

[0027] Mutation operator: Replaces a position in an individual. Mutation candidates come from the "feasible region pool", which consists of nodes that do not appear in the individual. This ensures that there are no duplicates after mutation. Heuristic weights can be used to guide mutation further.

[0028] Local search operator: for each node in the chromosome The first-order neighbor set is taken as a candidate replacement. If the fitness of the individual is improved after replacement, it is accepted and the search is terminated early. In order to control resource consumption, an early stop counter LOC_THRESHOLD is introduced: if the number of consecutive attempts reaches the threshold and there is still no improvement, the search is returned empty / stopped.

[0029] A further optimization is as follows: In step C, the task similarity is approximated by comparing the current best individual in the main task with the current best individual in the subtask.

[0030] The set of nodes for the optimal individual in the main task: The set of nodes for the optimal individual in subtask i: B Similarity can be defined as:

[0031]

[0032] It can also be replaced by metrics such as coverage gain correlation and influence increment similarity;

[0033] This is used to measure whether "the transfer of an individual to a subtask actually helps improve the main task," thereby suppressing negative transfer.

[0034] In a single migration, if an individual migrating into the subtask performs better than an individual of the corresponding level in the current population of the main task under the evaluation of the main task, it is counted as a positive contribution PosTrans. The total number of migrations is SumTrans, and the reward score is calculated using an exponentially weighted moving average (EWMA).

[0035]

[0036] Where w∈[0,1] is the smoothing coefficient;

[0037] The migration probability is obtained by combining task similarity and reward score. The typical form is:

[0038]

[0039] in For configurable coefficients,

[0040] Migration execution method:

[0041] For each subtask i, press The number of individuals sampled for migration, MigrateSize, is fixed or adaptive. Several high-quality individuals are selected from the subtask population to migrate into the main task. The main task evaluates the fitness of the migrated individuals and merges them with the current population.

[0042] A further preferred approach is to use a heuristic importance algorithm comprising the following steps:

[0043] Step 1: Data Acquisition and Modeling;

[0044] Acquire social network data, construct graph G=(V,E), prepare the parameters required for the diffusion model, determine the seed budget k, determine the evaluator type, and set the parameters for the main task and subtasks;

[0045] Step 2: Multi-subgraph extraction and task initialization;

[0046] Generate candidate node sets Vi respectively, construct sub-task subgraph Gi=G[Vi] for each Vi, establish sub-task set Ti, establish main task Tmain in the original graph, and use the "low overlap initialization strategy" to initialize the main task and each sub-task population respectively to ensure diversity and feasibility;

[0047] Step 3: The main task and sub-tasks evolve in parallel;

[0048] The main task executes one generation of evolution: selection—crossover—mutation—local search—fitness assessment—elite preservation. Subtasks execute several generations of evolution in parallel, with the same process as the main task but with lighter parameters.

[0049] Step 4: Calculation of transfer probability and one-way knowledge transfer;

[0050] Read the current best individual BestChrommain for the main task and the best individual BestChromi for each subtask, calculate the similarity Sim(main,i), and update the reward score based on the positive contribution of the transferred individuals to the main task. Calculate the migration probability Based on this, the number of migrations and the set of migration individuals are determined, and the subtask individuals are migrated unidirectionally into the main task, and the fusion and update of the main task is completed.

[0051] Step 5: Subgraph adaptive expansion and update;

[0052] like <μ, calculate the target subgraph size and the number of new nodes Δn, perform expansion based on the K-hop neighborhood set of the optimal seed set of the subtask, and generate a new subgraph. Update the subtask structure and population;

[0053] Step 6: Termination and Output;

[0054] If the main task reaches the maximum number of iterations or meets the convergence condition, it terminates and outputs the set of seed nodes S corresponding to the optimal individual in the main task.

[0055] Compared with existing technologies, this invention has the following positive effects: By decomposing the original large-scale network task into a main task and multiple subtasks generated by an efficient initiation method, this invention unidirectionally migrates effective individuals generated by subtasks to the main task during parallel evolution. It dynamically measures the contribution of subtasks through a migration probability mechanism and suppresses negative migration. When the contribution of a subtask is insufficient, it triggers subgraph expansion, allowing subtasks to gradually converge towards the structure of the main task, continuously providing effective exploration directions. Ultimately, it obtains a high-quality seed set under controllable computational overhead. This significantly reduces the effective search space, significantly lowers the proportion of invalid evaluations, and increases the output per unit of computing power. In high-frequency iteration scenarios, it can reduce the overall running time and resource consumption to an engineeringable range. Furthermore, it avoids the optimality loss caused by search perspective bias by relying on the main task, thus exhibiting better solution quality and result stability under different types of networks. It reduces the dependence on a single heuristic or single parameter configuration and improves overall robustness and convergence reliability. It is applicable to multiple diffusion models and multiple network types, thus expanding its applicability. Attached Figure Description

[0056] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0057] Figure 1 This is a structural block diagram of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Example

[0060] See Figure 1 A user node selection method based on multi-task evolutionary optimization of network information propagation includes the following steps:

[0061] Step A: Use multiple low-cost heuristics to quickly mine different "important node sets" from the original graph, and use these sets to induce multiple subgraphs, forming the initial search space for multiple sub-tasks; In Step A, given a directed or undirected graph... ,

[0062] in It is the set of nodes in the network, where n represents the total number of nodes; It is the set of edges in the network. Represents the number of sides;

[0063] Each edge Associated with a diffusion probability , represents a node Successfully activated node The probability of;

[0064] Select seed set In the diffusion model, the influence range after the propagation ends is denoted as σ(S), and the objective is:

[0065]

[0066] Each edge in the model With activation probability ,exist Seed node activation; in At this time, the newly activated node in the previous time step attempts to activate its neighbors independently; propagation terminates when no new activated node is found.

[0067] A further preferred approach is: In step A, for the original network, at least two heuristic importance algorithms are used to generate a candidate node set and construct a subgraph; typical examples include, but are not limited to:

[0068] Degree / HighestDegree: Sort by node degree and select the top few nodes; PageRank: Sort by random walk steady-state distribution and select the top few nodes; Core / K-shell: Decompose by kernel to obtain high-shell nodes; It can also be extended to betweenness, closeness, H-index, leaderRank, truss, etc.

[0069] Each subtask can be configured with its own evolutionary algorithm parameters, which can differ from those of the main task to improve efficiency. For example, the subtask population size can be set to a certain proportion of that of the main task (e.g., 1 / 2); the fitness evaluation of the subtask can prioritize the use of a fast surrogate model, while the main task can use a more accurate evaluator; and the subtask can perform multiple iterations within the same generation of the main task.

[0070] Each subtask has its own independently configured evolutionary algorithm parameters.

[0071] Step B involves the main task and subtasks working in parallel. The main task optimizes the seed set on the original graph, while the subtasks optimize on their respective subgraphs. The subtasks generate high-quality individuals, which serve as candidate "guide samples" for the main task. In Step B, the multi-task parallel collaboration includes individual encoding: an individual represents a seed set, encoded using a node sequence / set of length λ.

[0072]

[0073] Duplicate nodes are not allowed in the encoding.

[0074] Population initialization: First, maintain a candidate node list TabooList. Second, construct individuals one by one: randomly select nodes to add to individuals. Then, define Similarity(u, v) based on topological similarity or neighborhood overlap, and add nodes that are too similar to the selected nodes to TabooList to avoid subsequent selection. Finally, if NodeList is empty or cannot meet the length 𝑘, fill it with the remaining nodes to obtain an initial population with higher diversity.

[0075] Crossover operator: from parent individual Choose one as the benchmark Randomly select intersection point i ∈ [0,k], swap the segments after the intersection point. If the swap introduces duplicate nodes, perform conflict resolution on the duplicate items and generate feasible descendants to be added to the offspring population.

[0076] Mutation operator: Replaces a position in an individual. Mutation candidates come from the "feasible region pool", which consists of nodes that do not appear in the individual. This ensures that there are no duplicates after mutation. Heuristic weights can be used to guide mutation further.

[0077] Local search operator: for each node in the chromosome The first-order neighbor set is taken as a candidate replacement. If the fitness of the individual is improved after replacement, it is accepted and the search is terminated early. In order to control resource consumption, an early stop counter LOC_THRESHOLD is introduced: if the number of consecutive attempts reaches the threshold and there is still no improvement, the search is returned empty / stopped.

[0078] Step C: Main task-driven unidirectional knowledge transfer. After each iteration of the main task, some individuals are transferred from each subtask to the main task. The benefits of the transfer are divided into: direct improvement: the transferred individuals are better than the current population of the main task in one transfer; indirect improvement: the transferred individuals produce better offspring after participating in crossover / mutation, thereby indirectly improving the optimal value of the main task.

[0079] In step C, the task similarity is approximated by comparing the current best individual in the main task with the current best individual in the subtask: and the node set of the best individual in the main task is: The set of nodes for the optimal individual in subtask i: B Similarity can be defined as:

[0080]

[0081] It can also be replaced by metrics such as coverage gain correlation and influence increment similarity;

[0082] This is used to measure whether "the transfer of an individual to a subtask actually helps improve the main task," thereby suppressing negative transfer.

[0083] In a single migration, if an individual migrating into the subtask performs better than an individual of the corresponding level in the current population of the main task under the evaluation of the main task, it is counted as a positive contribution PosTrans. The total number of migrations is SumTrans, and the reward score is calculated using an exponentially weighted moving average (EWMA).

[0084]

[0085] Where w∈[0,1] is the smoothing coefficient;

[0086] The migration probability is obtained by combining task similarity and reward score. The typical form is:

[0087]

[0088] in For configurable coefficients,

[0089] Migration execution method:

[0090] For each subtask i, press The number of individuals sampled for migration, MigrateSize, is fixed or adaptive. Several high-quality individuals are selected from the subtask population to migrate into the main task. The main task evaluates the fitness of the migrated individuals and merges them with the current population.

[0091] Subgraph update strategies (negative migration suppression and adaptive expansion)

[0092] Triggering conditions

[0093] When the transition probability of subtask i remains consistently low, it is considered that the current subgraph of this subtask differs too much from the main task or its contribution is insufficient. A threshold μ is set: if... This triggers the subgraph expansion update; μ can be a fixed threshold or dynamically adjusted with iteration (e.g., gradually decreasing / increasing as the main task converges).

[0094] Subgraph size target and update frequency

[0095] Set the maximum number of updates for the subtask to Q, and the current number of updates is q. The target size of the subgraph can be achieved using a strategy that linearly approximates the size of the original graph, for example:

[0096]

[0097] Where N is the number of nodes in the original graph, N i To update the number of nodes in the subgraph before the update, This represents the target number of nodes after this update. This design allows the subgraph to expand gradually, potentially reaching a size equal to the original subgraph in extreme cases. Figure 1 This ensures that subtasks do not become permanently invalid.

[0098] The number of new nodes required in this round is:

[0099]

[0100] The heuristic importance algorithm includes the following steps:

[0101] Step 1: Data Acquisition and Modeling;

[0102] Acquire social network data, construct graph G=(V,E), prepare the parameters required for the diffusion model, determine the seed budget k, determine the evaluator type, and set the parameters for the main task and subtasks;

[0103] Step 2: Multi-subgraph extraction and task initialization;

[0104] Generate candidate node sets Vi respectively, construct sub-task subgraph Gi=G[Vi] for each Vi, establish sub-task set Ti, establish main task Tmain in the original graph, and use the "low overlap initialization strategy" to initialize the main task and each sub-task population respectively to ensure diversity and feasibility;

[0105] Step 3: The main task and sub-tasks evolve in parallel;

[0106] The main task executes one generation of evolution: selection—crossover—mutation—local search—fitness assessment—elite preservation. Subtasks execute several generations of evolution in parallel, with the same process as the main task but with lighter parameters.

[0107] Step 4: Calculation of transfer probability and one-way knowledge transfer;

[0108] Read the current best individual BestChrommain for the main task and the best individual BestChromi for each subtask, calculate the similarity Sim(main,i), and update the reward score based on the positive contribution of the transferred individuals to the main task. Calculate the migration probability Based on this, the number of migrations and the set of migration individuals are determined, and the subtask individuals are migrated unidirectionally into the main task, and the fusion and update of the main task is completed.

[0109] Step 5: Subgraph adaptive expansion and update;

[0110] like <μ, calculate the target subgraph size and the number of new nodes Δn, perform expansion based on the K-hop neighborhood set of the optimal seed set of the subtask, and generate a new subgraph. Update the subtask structure and population;

[0111] Step 6: Termination and Output;

[0112] If the main task reaches the maximum number of iterations or meets the convergence condition, it terminates and outputs the set of seed nodes S corresponding to the optimal individual in the main task.

[0113] Step D: Determine the relationship between the subtask and the main task. If the subtask and the main task have low similarity and the transfer has no effect, introduce a subgraph expansion.

[0114] When the migration probability is below a threshold, the subgraph is automatically expanded to improve structural similarity.

[0115] As the subgraph gradually expands to the same size as the main graph, the subtasks should flow unidirectionally to the main task to ensure that the main task is not diluted.

[0116] Compared with existing technologies, this invention has the following positive effects: By decomposing the original large-scale network task into a main task and multiple subtasks generated by an efficient initiation method, this invention unidirectionally migrates effective individuals generated by subtasks to the main task during parallel evolution. It dynamically measures the contribution of subtasks through a migration probability mechanism and suppresses negative migration. When the contribution of a subtask is insufficient, it triggers subgraph expansion, allowing subtasks to gradually converge towards the structure of the main task, continuously providing effective exploration directions. Ultimately, it obtains a high-quality seed set under controllable computational overhead. This significantly reduces the effective search space, significantly lowers the proportion of invalid evaluations, and increases the output per unit of computing power. In high-frequency iteration scenarios, it can reduce the overall running time and resource consumption to an engineeringable range. Furthermore, it avoids the optimality loss caused by search perspective bias by relying on the main task, thus exhibiting better solution quality and result stability under different types of networks. It reduces the dependence on a single heuristic or single parameter configuration and improves overall robustness and convergence reliability. It is applicable to multiple diffusion models and multiple network types, thus expanding its applicability.

[0117] The standard parts used in this embodiment can be purchased directly from the market, and the non-standard structural parts described in the instruction manual can also be processed without any doubt based on existing technical common sense. At the same time, the connection methods of each component adopt mature conventional methods in the existing technology, and the machinery, parts and equipment all adopt conventional models in the existing technology, so they will not be described in detail here.

[0118] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, these obvious variations or modifications derived from the essential spirit of the present invention still fall within the scope of protection of the present invention.

Claims

1. A user node selection method based on multi-task evolutionary optimization of network information propagation, characterized in that: Includes the following steps: Step A: Use multiple low-cost heuristics to quickly mine different "important node sets" from the original graph, and use the above sets to induce multiple subgraphs, forming the initial search space for multiple sub-tasks; Step B: The main task and sub-tasks are carried out in parallel and collaboratively. The main task optimizes the seed set on the original graph, while the sub-tasks optimize on their respective subgraphs. The sub-tasks generate high-quality individuals and serve as candidate "guide samples" for the main task. Step C: The main task drives unidirectional knowledge transfer. After each iteration of the main task, some individuals are transferred from each subtask to the main task, indirectly improving the optimal value of the main task. Step D: Determine the relationship between the subtask and the main task. If the subtask and the main task have low similarity and the transfer has no effect, introduce a subgraph expansion. When the migration probability is below a threshold, the subgraph is automatically expanded to improve structural similarity. As the subgraph gradually expands to the same size as the main graph, the subtasks should flow unidirectionally to the main task to ensure that the main task is not diluted.

2. The user node selection method based on multi-task evolutionary optimization network information propagation according to claim 1, characterized in that: In step A, a directed or undirected graph is given. , in It is the set of nodes in the network, where n represents the total number of nodes; It is the set of edges in the network. Represents the number of sides; Each edge Associated with a diffusion probability , represents a node Successfully activated node The probability of; Select seed set In the diffusion model, the influence range after the propagation ends is denoted as σ(S), and the objective is: Each edge in the model With activation probability ,exist Seed node activation; in At this time, the newly activated node in the previous time step attempts to activate its neighbors independently; propagation terminates when no new activated node is found.

3. The user node selection method based on multi-task evolutionary optimization of network information propagation according to claim 2, characterized in that: In step A, for the original network, at least two heuristic importance algorithms are used to generate a set of candidate nodes and construct a subgraph; Each subtask has its own independently configured evolutionary algorithm parameters.

4. The user node selection method based on multi-task evolutionary optimization of network information propagation according to claim 3, characterized in that: In step B, multi-task parallel collaboration includes individual encoding: an individual represents a seed set, encoded using a node sequence / set of length 𝑘. Duplicate nodes are not allowed in the encoding. Population initialization: First, maintain a candidate node list TabooList. Second, construct individuals one by one: randomly select nodes to add to individuals. Then, define Similarity(u, v) based on topological similarity or neighborhood overlap, and add nodes that are too similar to the selected nodes to TabooList to avoid subsequent selection. Finally, if NodeList is empty or cannot meet the length 𝑘, fill it with the remaining nodes to obtain an initial population with higher diversity. Crossover operator: from parent individual Choose one as the benchmark Randomly select intersection point i ∈ [0, k], swap the segments after the intersection point. If the swap introduces duplicate nodes, perform conflict resolution on the duplicate items and generate feasible descendants to be added to the offspring population. Mutation operator: Replaces a position in an individual. Mutation candidates come from the "feasible region pool", which consists of nodes that do not appear in the individual. This ensures that there are no duplicates after mutation. Heuristic weights can be used to guide mutation further. Local search operator: for each node in the chromosome The first-order neighbor set is taken as a candidate replacement. If the fitness of the individual is improved after replacement, it is accepted and the search is terminated early. In order to control resource consumption, an early stop counter LOC_THRESHOLD is introduced: if the number of consecutive attempts reaches the threshold and there is still no improvement, the search is returned empty / stopped.

5. The user node selection method based on multi-task evolutionary optimization of network information propagation according to claim 4, characterized in that: In step C, the task similarity is approximated by comparing the current best individual in the main task with the current best individual in the subtask: The set of nodes for the optimal individual in the main task: The set of nodes for the optimal individual in subtask i: B Similarity can be defined as: Alternatively, it can be replaced with metrics such as coverage gain correlation and influence increment similarity; This is used to measure whether "the transfer of an individual to a subtask actually helps improve the main task," thereby suppressing negative transfer. In a single migration, if an individual migrating into the subtask performs better than an individual of the corresponding level in the current population of the main task under the evaluation of the main task, it is counted as a positive contribution PosTrans. The total number of migrations is SumTrans, and the reward score is calculated using an exponentially weighted moving average (EWMA). ， Where w∈[0,1] is the smoothing coefficient; The migration probability is obtained by combining task similarity and reward score. The typical form is: ， in For configurable coefficients, Migration execution method: For each subtask i, press The number of individuals sampled for migration, MigrateSize, is fixed or adaptive. Several high-quality individuals are selected from the subtask population to migrate into the main task. The main task evaluates the fitness of the migrated individuals and merges them with the current population.

6. The user node selection method based on multi-task evolutionary optimization of network information propagation according to claim 5, characterized in that: The heuristic importance algorithm includes the following steps: Step 1: Data Acquisition and Modeling; Acquire social network data, construct graph G=(V,E), prepare the parameters required for the diffusion model, determine the seed budget k, determine the evaluator type, and set the parameters for the main task and subtasks; Step 2: Multi-subgraph extraction and task initialization; Generate candidate node sets Vi respectively, construct sub-task subgraph Gi=G[Vi] for each Vi, establish sub-task set Ti, establish main task Tmain in the original graph, and use "low overlap initialization strategy" to initialize the main task and each sub-task population respectively to ensure diversity and feasibility; Step 3: The main task and sub-tasks evolve in parallel; The main task executes one generation of evolution: selection—crossover—mutation—local search—fitness evaluation—elite preservation. Subtasks execute several generations of evolution in parallel, with the same process as the main task but with lighter parameters. Step 4: Calculation of transfer probability and one-way knowledge transfer; Read the current best individual BestChrommain for the main task and the best individual BestChromi for each subtask, calculate the similarity Sim(main,i), and update the reward score based on the positive contribution of the transferred individuals to the main task. Calculate the migration probability Based on this, the number of migrations and the set of migration individuals are determined, and the subtask individuals are migrated unidirectionally into the main task, and the fusion and update of the main task is completed. Step 5: Subgraph adaptive expansion and update; like <μ, calculate the target subgraph size and the number of new nodes Δn, perform expansion based on the K-hop neighborhood set of the optimal seed set of the subtask, and generate a new subgraph. Update the subtask structure and population; Step 6: Termination and Output; If the main task reaches the maximum number of iterations or meets the convergence condition, it terminates and outputs the set of seed nodes S corresponding to the optimal individual in the main task.

Images