Two-stage topic-aware influence maximization method enhanced with deep learning

Abstract

The application discloses a two-stage topic-aware influence maximization method using deep learning enhancement in the field of social network information propagation technology, and comprises the following steps: S1, establishing a propagation model and problem modeling; S2, defining a node importance measurement index SRC; S3, step 3: candidate node set construction; S4, a training process of a node importance estimator NIE; S5, a seed optimization process based on an improved hybrid frog leap algorithm; and S6, optimal seed set output. The application forms a synergistic mechanism in three key links of propagation mechanism modeling, candidate node screening and influence estimation, significantly improves the calculation efficiency under the premise of guaranteeing the influence propagation effect. Through experiments on multiple typical network data sets, the method of the application shows performance advantages superior to existing methods in influence diffusion capacity and running efficiency, and provides a feasible solution for efficient seed selection problems in large-scale social networks.

Description

Technical Field

[0001] This invention relates to a two-stage topic-aware influence maximization method using deep learning enhancement in the field of social network information dissemination technology. Background Technology

[0002] In recent years, with the rapid development of online social networks, information dissemination has significantly increased in both scale and speed. How to achieve efficient information diffusion in complex networks has become an important research question. Influence Maximization (IM), as a fundamental research direction, aims to achieve widespread information dissemination by selecting a small number of key nodes. Furthermore, with the expansion of application scenarios, researchers have gradually discovered that the real-world information dissemination process not only depends on the network topology but is also significantly influenced by the information content and user interests. For example, users' activity levels and dissemination capabilities differ across different topics, making it difficult for traditional assumptions to accurately characterize real-world dissemination behavior.

[0003] To address this, Barbieri et al. proposed the Topic-aware Influence Maximization (TIM) problem. By introducing topic distribution and interest modeling, the propagation probability can reflect the matching relationship between users and content, thereby improving the model's expressive power. Building on this, related research has diversified the propagation mechanism from the perspectives of user similarity, content relevance, and structural feature fusion. At the algorithmic level, existing research has proposed various methods to balance efficiency and effectiveness, including heuristic-based fast selection methods, deep learning-based node importance evaluation methods, and two-stage optimization frameworks that reduce the search space through candidate sets. These methods alleviate the computational pressure in large-scale networks to some extent, but still have significant shortcomings. First, some methods only consider a single factor in propagation modeling, lacking a comprehensive characterization of multi-dimensional information and failing to reflect complex propagation dynamics. Second, in the candidate node selection process, the overlap of influence between nodes is often ignored, leading to a decrease in the propagation efficiency of the final seed set. Furthermore, insufficient coordination between global exploration and local development during optimization can easily cause the algorithm to get trapped in local optima. Therefore, how to construct a more accurate propagation model in topic-aware scenarios and design a seed selection method that balances computational efficiency and propagation effectiveness remains a key issue that urgently needs to be addressed. Summary of the Invention

[0004] The purpose of this invention is to provide a two-stage topic awareness influence maximization method using deep learning enhancement, which can build a more accurate propagation model in topic awareness scenarios and design a seed selection method that balances computational efficiency and propagation effect.

[0005] To achieve the above objectives, this invention provides a two-stage topic-aware influence maximization method enhanced by deep learning, comprising the following steps:

[0006] S1. Establish a propagation model and problem modeling;

[0007] Given a social network , It is a social network A set of nodes in the , where each node is represented as , Represents a node exist Interest preference vectors under a specific topic , It represents the connection relationship between users; each node has the possibility of being activated during the information propagation process, and spreads the information to its neighboring nodes through edges; in order to characterize the probabilistic characteristics of the information propagation process, the diffusion process is modeled based on the independent cascading propagation mechanism;

[0008] S2, Definition of node importance measurement indicators;

[0009] A new metric for measuring node importance was designed. This indicator is used to characterize the comprehensive influence of a node in network propagation. It takes into account the propagation characteristics of influence and alleviates the problem of influence overlap, enabling it to quickly and accurately identify important nodes.

[0010] S3. Construction of candidate node set;

[0011] After obtaining the importance metrics of all nodes, the node set... Sort by SRC value in descending order and select the top-ranked values. The nodes constitute the candidate node set ,in This process compresses the search space, allowing subsequent optimizations to be performed only in the high-potential node subspace, thereby significantly reducing computational complexity.

[0012] S4. Training process of the Node Importance Estimator (NIE);

[0013] To further improve the efficiency of seed selection, a node importance estimator is introduced to approximate the marginal gain of nodes or combinations of nodes, thereby avoiding a large amount of Monte Carlo simulation computation.

[0014] S5. Seed optimization process based on improved hybrid frog jumping algorithm;

[0015] In the candidate node set The improved hybrid frog-leaping algorithm is used for global search to obtain the optimal seed combination. Specifically, it includes two collaborative processes: hierarchical local development and global information exploration, thereby improving the combinatorial optimality of the solution while ensuring search efficiency.

[0016] S6, Output the optimal seed set;

[0017] When the algorithm reaches the maximum number of iterations or the fitness converges, it outputs the seed set corresponding to the current best individual. As the final solution for maximizing influence, this set can achieve the maximum expected influence diffusion under the propagation model.

[0018] As a further improvement to the present invention, the specific details of modeling the diffusion process based on the independent cascading propagation mechanism in S1 are as follows.

[0019] Set nodes At any moment If activated, it will be in the next moment. With probability Try to activate its neighboring nodes The propagation process follows the one-trial principle, meaning that each edge triggers a propagation attempt only when the source node is first activated.

[0020] ;

[0021] Indicates in the topic Down neighboring nodes Based on topic vectors Interest vector and its own interest vector The activation probability obtained; where , There are two variable parameters; It is the original activation probability of the edge. It involves similarity calculation; based on this, a seed set is defined. Its scale is ;make Indicates a given query, For the query vector, the influence range function is denoted as... , indicating in the query Under the condition of the lower node set Scope of influence; Let represent the maximum weighted expected range of influence. Then, the problem of maximizing influence based on topic perception can be formally defined as follows:

[0022] (1).

[0023] 3. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 2, characterized in that: the specific content of S2 is as follows,

[0024] S2.1 Sampling of the social network G: For each edge (u,v) in the social network G, with probability... Randomly delete the edge, and the remaining edges form the edge set E' of the sub-social network, thus obtaining the sub-social network G'=(V',E'); repeat the sampling SN times to obtain the set of sub-social networks. ;

[0025] S2.2, Traverse each sub-social network in the sub-social network set SG. For each node v, a depth-first traversal algorithm is used to determine its position in the sub-social network. All nodes reachable from v are listed in the dictionary, and this node information is stored in a dictionary. middle;

[0026] S2.3 After traversal, calculate the reachability centrality value (SRC) of each node v's sub-social network; initially, the SRC of each node is defined as the sum of the number of reachable nodes in all sub-social networks, i.e.:

[0027] (2).

[0028] 4. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 3, characterized in that: in S2.3, to alleviate the influence overlap problem, a dynamic discount mechanism is introduced based on SRC(v), the specific steps of which are as follows:

[0029] 1. Initialize the seed set S to be empty, and execute |V| rounds of iteration;

[0030] 2. Based on the current SRC value of each node, select the node v with the largest SRC value, i.e.:

[0031] (3);

[0032] 3. Add node v to the seed set S, and record the reachability centrality value of the final sub-social network of this node:

[0033] (4);

[0034] 4. For a newly added seed node v, update its SRC value for all its one-hop neighbors u ∈ N(v) according to the following formula:

[0035] (5);

[0036] 5. Among them, Indicated in sub-social networks The size of the intersection of the sets of reachable nodes of node u and node v;

[0037] 6. Proceed to the next iteration. After the iteration ends, the final reachability centrality value of the sub-social network of each node v is... (v); Through the above process, the reachability centrality values ​​of the sub-social networks of all nodes can be obtained.

[0038] 5. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 4, characterized in that: the specific content of S4 is as follows,

[0039] S4.1 First, construct the training dataset. For the candidate node set... The nodes in the simulation are used to calculate their true influence gain through multiple independent cascaded propagation simulations, resulting in training sample pairs:

[0040] (6);

[0041] in Represents a node The feature vector contains structural features including its SRC value and degree information. This represents the true marginal influence of a node, which is estimated through multiple propagation simulations.

[0042] S3.2 Based on this, construct a regression model. This is used to fit the mapping relationship between node features and their influence; the model training objective is to minimize the error between the predicted value and the true value.

[0043] (7);

[0044] After training, the NIE model can quickly predict the influence contribution of nodes, thus replacing the high-cost propagation simulation calculation in subsequent optimization processes.

[0045] 6. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 5, characterized in that: the specific content of S5 is as follows,

[0046] S5.1, First, in the candidate node set Initialize the population Each individual Represents a size of The seed set; the fitness value of each individual is calculated based on the node importance estimator (NIE). The population was then divided into groups based on fitness, ranked from highest to lowest. Meme group ;

[0047] S5.2 In the process of hierarchical local development, for the worst solution within each meme group A two-stage progressive update strategy was designed; first, a basic improvement process was executed, which involved moving towards the optimal solution of the current meme set. Learn to replace and update the worst solution;

[0048] S5.3 After completing the local development within each meme group, the global information exploration phase begins; this is achieved by introducing the global optimal solution. Guide the search direction;

[0049] S5.4 After several rounds of local development and global exploration, all meme groups are reshuffled, that is, the population is reordered and re-divided into meme groups to break the original search structure, realize the full exchange of information between different meme groups, and thus avoid the algorithm from getting stuck in local optima.

[0050] 7. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 6, characterized in that: the specific content of S5.2 is as follows,

[0051] S5.2.1 First, execute the basic improvement process by moving towards the optimal solution of the current meme set. Learning involves replacing and updating the worst-case solution; specifically, from... Randomly select nodes from, and simultaneously from Randomly remove the corresponding node and construct a new solution:

[0052] (8);

[0053] like Then use Replace the original worst solution; otherwise, proceed to the next stage.

[0054] S5.2.2 Based on this, a local depth optimization process is introduced to enhance the fine-grained search capability of the solution; this process centers on the current solution and searches the candidate node set. Multiple neighborhood search operations are performed for any candidate node. Calculate which node replaces any existing node. Marginal revenue afterward:

[0055] (9);

[0056] The replacement operation with the maximum positive gain is selected to perform the update; this process is repeated until there is no replacement that can improve fitness, thus achieving a local optimum.

[0057] 8. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 7, characterized in that: the specific content of S5.3 is as follows,

[0058] The worst solution in each meme set Construct an update strategy based on the global optimal solution, that is, from Select a node to replace, thus generating a new solution:

[0059] (10);

[0060] in , This update mechanism enables local search results to converge towards the globally optimal region.

[0061] Compared with existing technologies, the beneficial effects of this invention lie in its ability to effectively solve the problem of accurate identification of key users in large-scale social networks by constructing a topic-aware independent cascade model and fusing a seed node selection algorithm based on deep reinforcement learning. This is achieved by constructing a multi-dimensional topic-aware independent cascade model (MD-TIC) that integrates user-content matching and user-user interest similarity. Based on this model, a deep learning-enhanced two-stage algorithm (DLETS) is designed. In the first stage of the algorithm, the propagation characteristics of information in the network are fully considered, and the influence overlap problem is alleviated, selecting a high-quality set of candidate nodes. Furthermore, to improve processing speed, a node importance estimator (NIE) based on SGFormer is designed, significantly improving the algorithm's running speed. In the second stage, an improved shuffled frog leaping algorithm (ISFLA) is proposed by introducing a hierarchical local development operator and a global information exploration operator, ensuring the algorithm's effectiveness. Attached Figure Description

[0062] Figure 1 This is a flowchart of the DLETS algorithm of the present invention.

[0063] Figure 2 This invention compares the influence spread of different candidate node selection algorithms across multiple datasets.

[0064] Figure 3The running time for different algorithms of this invention to select k=30 nodes in multiple datasets.

[0065] Figure 4 This is a comparison of the influence spread of the DLETS algorithm of this invention with all baseline algorithms. Detailed Implementation

[0066] The present invention will be further described below with reference to the accompanying drawings:

[0067] like Figure 1-4 The two-stage topic-aware influence maximization method using deep learning enhancement, as shown, includes the following steps:

[0068] S1. Establish a propagation model and problem modeling;

[0069] Given a social network , It is a social network A set of nodes in the , where each node is represented as , Represents a node exist Interest preference vectors under a specific topic , It represents the connection relationship between users; each node has the possibility of being activated during the information propagation process, and spreads the information to its neighboring nodes through edges; in order to characterize the probabilistic characteristics of the information propagation process, the diffusion process is modeled based on the independent cascading propagation mechanism;

[0070] Set nodes At any moment If activated, it will be in the next moment. With probability Try to activate its neighboring nodes The propagation process follows the one-trial principle, meaning that each edge triggers a propagation attempt only when the source node is first activated.

[0071] ;

[0072] Indicates in the topic Down neighboring nodes Based on topic vectors Interest vector and its own interest vector The activation probability obtained; where , There are two variable parameters; It is the original activation probability of the edge. It involves similarity calculation; based on this, a seed set is defined. Its scale is ;make Indicates a given query, For the query vector, the influence range function is denoted as... , indicating in the query Under the condition of the lower node set Scope of influence; Let represent the maximum weighted expected range of influence. Then, the problem of maximizing influence based on topic perception can be formally defined as follows:

[0073] (1).

[0074] S2, Definition of Node Importance Metric SRC;

[0075] A novel node importance metric, SRC, is designed to characterize the comprehensive influence of nodes in network propagation. This metric is constructed based on the local structural information of nodes and models the potential propagation capability of nodes by fusing the degree information of nodes and their neighborhood structural features.

[0076] S2.1 Sampling of the social network G: For each edge (u,v) in the social network G, with probability... Randomly delete the edge, and the remaining edges form the edge set E' of the sub-social network, thus obtaining the sub-social network G'=(V',E'); repeat the sampling SN times to obtain the set of sub-social networks. ;

[0077] S2.2, Traverse each sub-social network in the sub-social network set SG. For each node v, a depth-first traversal algorithm is used to determine its position in the sub-social network. All nodes reachable from v are listed in the dictionary, and this node information is stored in a dictionary. middle;

[0078] S2.3 After traversal, calculate the reachability centrality value SRC(v) of each node v in its sub-social network; initially, the SRC(v) of each node is defined as the sum of the number of reachable nodes in all sub-social networks, i.e.:

[0079] (2).

[0080] To mitigate the overlap of influences, S2.3 introduces a dynamic discount mechanism based on SRC(v), with the following specific steps:

[0081] 1. Initialize the seed set S to be empty, and execute |V| rounds of iteration;

[0082] 2. Based on the current SRC value of each node, select the node v with the largest SRC value, i.e.:

[0083] (3);

[0084] 3. Add node v to the seed set S, and record the reachability centrality value of the final sub-social network of this node:

[0085] (4);

[0086] 4. For a newly added seed node v, update its SRC value for all its one-hop neighbors u ∈ N(v) according to the following formula:

[0087] (5);

[0088] 5. Among them, Indicated in sub-social networks The size of the intersection of the sets of reachable nodes of node u and node v;

[0089] 6. Proceed to the next iteration. After the iteration ends, the final reachability centrality value of the sub-social network of each node v is... Through the above process, the reachability centrality values ​​of the sub-social networks of all nodes can be obtained.

[0090] S3. Construction of candidate node set;

[0091] After obtaining the SRC values ​​of all nodes, the node set is... according to Sort in descending order and select the top-ranked items. The nodes constitute the candidate node set ,in This process compresses the search space, allowing subsequent optimizations to be performed only in the high-potential node subspace, thereby significantly reducing computational complexity.

[0092] S4. Training process of the Node Importance Estimator (NIE);

[0093] To further improve the efficiency of seed selection, a node importance estimator is introduced to approximate the importance index of nodes, thereby avoiding a large amount of computation.

[0094] S4.1 First, construct the training dataset. For the candidate node set... The nodes in the simulation are used to calculate their true influence gain through multiple independent cascaded propagation simulations, resulting in training sample pairs:

[0095] (6);

[0096] in Represents a node The feature vector contains structural features including its SRC value and degree information. This represents the true marginal influence of a node, which is estimated through multiple propagation simulations.

[0097] S3.2 Based on this, construct a regression model. This is used to fit the mapping relationship between node features and their influence; the model training objective is to minimize the error between the predicted value and the true value.

[0098] (7);

[0099] After training, the NIE model can quickly predict the influence contribution of nodes, thus replacing the high-cost propagation simulation calculation in subsequent optimization processes.

[0100] S5. Seed optimization process based on improved hybrid frog jumping algorithm;

[0101] In the candidate node set The improved hybrid frog-leaping algorithm is used for global search to obtain the optimal seed combination. Specifically, it includes two collaborative processes: hierarchical local development and global information exploration, thereby improving the combinatorial optimality of the solution while ensuring search efficiency.

[0102] S5.1, First, in the candidate node set Initialize the population Each individual Represents a size of The seed set; the fitness value of each individual is calculated based on the node importance estimator (NIE). The population was then divided into groups based on fitness, ranked from highest to lowest. Meme group ;

[0103] S5.2 In the process of hierarchical local development, for the worst solution within each meme group A two-stage progressive update strategy was designed; first, a basic improvement process was executed, which involved moving towards the optimal solution of the current meme set. Learn to replace and update the worst solution;

[0104] S5.2.1 First, execute the basic improvement process by moving towards the optimal solution of the current meme set. Learning involves replacing and updating the worst-case solution; specifically, from... Randomly select nodes from, and simultaneously from Randomly remove the corresponding node and construct a new solution:

[0105] (8);

[0106] like Then use Replace the original worst solution; otherwise, proceed to the next stage.

[0107] S5.2.2 Based on this, a local depth optimization process is introduced to enhance the fine-grained search capability of the solution; this process centers on the current solution and searches the candidate node set. Multiple neighborhood search operations are performed for any candidate node. Calculate which node replaces any existing node. Marginal revenue afterward:

[0108] (9);

[0109] The replacement operation with the maximum positive gain is selected to perform the update; this process is repeated until there is no replacement that can improve fitness, thus achieving a local optimum.

[0110] S5.3 After completing the local development within each meme group, the global information exploration phase begins; this is achieved by introducing the global optimal solution. Guide the search direction;

[0111] The worst solution in each meme set Construct an update strategy based on the global optimal solution, that is, from Select a node to replace, thus generating a new solution:

[0112] (10);

[0113] in , This update mechanism enables local search results to converge towards the globally optimal region.

[0114] S5.4 After several rounds of local development and global exploration, all meme groups are reshuffled, that is, the population is reordered and re-divided into meme groups to break the original search structure, realize the full exchange of information between different meme groups, and thus avoid the algorithm from getting stuck in local optima.

[0115] S6, Output the optimal seed set;

[0116] When the algorithm reaches the maximum number of iterations or the fitness converges, it outputs the seed set corresponding to the current best individual. As the final solution for maximizing influence, this set can achieve the maximum expected influence diffusion under the propagation model.

[0117] The core of this invention lies in proposing a deep learning-enhanced two-stage topic-aware influence maximization method, which achieves high-quality selection of seed nodes in large-scale social networks while balancing propagation modeling accuracy and computational efficiency. This method first extends the traditional independent cascade model by constructing a multidimensional topic-aware propagation model (MD-TIC), integrating user interests, user similarity, and content topic matching relationships to more realistically depict the dynamic process of information propagation in the network. Based on this, a matching mechanism between topic query vectors and node interest vectors is introduced, dynamically calculating the activation probability of edges, giving the propagation process significant topic sensitivity and personalized characteristics.

[0118] This invention designs a two-stage optimization framework (DLETS algorithm). In the first stage, by constructing a sub-social network reachability centrality (SRC) index and combining sub-social network sampling with a dynamic discount strategy, the true influence capability of nodes in the propagation process is effectively measured, and the problem of overlapping influence among nodes is explicitly mitigated. Based on this, a deep learning enhancement mechanism is introduced to construct a node importance estimator (NIE). Through a combination of offline training and online inference, efficient evaluation of node importance is achieved, thereby quickly screening out a set of high-potential candidate nodes and significantly reducing the search space.

[0119] In the second stage, this invention proposes an improved hybrid frog leaping optimization algorithm (ISFLA) to further select the optimal seed combination from the candidate node set. This algorithm enhances the utilization of high-quality solutions in local search by fusing hierarchical local exploitation operators and global exploration operators, while simultaneously improving global search capabilities through a random perturbation mechanism to avoid getting trapped in local optima. During the iteration process, the algorithm comprehensively considers the cooperative propagation effect of node combinations, achieving a dynamic balance between exploration and exploitation, thereby obtaining a high-quality seed node set.

[0120] In summary, this invention, through an overall technical approach of "propagation modeling—candidate selection—combinatorial optimization," significantly improves computational efficiency while ensuring propagation effectiveness. Given a social network... Topic Search Seed collection scale Its optimization objective is to select the optimal seed set under topic constraints. This method aims to maximize the expected impact range of information under the MD-TIC propagation model. It effectively enhances the impact expansion capability in topic-aware scenarios and is applicable to large-scale complex network environments.

[0121] To verify the effectiveness of the proposed Deep Enhancement Two-Stage Topic Awareness Influence Maximization (DLETS) algorithm, the algorithm was validated on four simulated datasets and two real datasets (FilmTrust, Aren-Email, Hamster-Friend, Stelzl, AMiner, and Yelp). Figure 2 The results show the comparison of the influence spread of the candidate node selection method as the seed set size k changes. Figure 3 The study presents a comparison of the runtime of different node selection methods on six datasets. Figure 4 By comparing the influence expansion of DLETS with baseline algorithms (SAW_ASA, DPSOMIM, DeepIM, TA-Degree), the performance advantage of DLETS in the overall seed selection task is verified.

[0122] Figure 2 (a)-(f) respectively show the comparison results of the influence spread of candidate node selection methods as the seed set size k changes. From the overall trend, the influence spread of all methods increases with increasing k, but at the same k value, the proposed SRC and NIE algorithms consistently show better propagation performance. Using the Yelp dataset (… Figure 2 For example, when k=50, the influence expansion of both the SRC and NIE algorithms reaches 1800, while TA-Degree is only 1700. This difference stems from the deep modeling of the dynamic characteristics of influence propagation by the SRC and NIE algorithms: SRC effectively avoids the defect of topological centrality ignoring the overlap of activation probability and influence through sub-social network sampling and dynamic discounting mechanisms. It is worth noting that in the Hamster-Friend dataset ( Figure 2 In (c), a key turning point occurs when k=25—before this, the difference in influence expansion between SRC and NIE is small (approximately 5-10 activated nodes), but afterwards, NIE's advantage increases. This is because as the seed set size increases, the problem of overlapping influences between nodes in the network becomes increasingly complex. NIE, through the fitting ability of its deep learning model to the global propagation law, can more efficiently identify nodes with complementary influences, thereby avoiding the local bias that SRC may produce due to its reliance on sub-network sampling.

[0123] Figure 3The results show a comparison of the running times of different node selection methods on six datasets, with smaller values ​​on the vertical axis indicating higher algorithm efficiency. The results clearly show that the NIE algorithm outperforms MCDM, DMC, and RCC on all datasets, and is significantly more efficient than the SRC algorithm—NIE's running time is only about 1 / 100th of SRC's. The core reason for this efficiency advantage lies in NIE's "offline training—online inference" mode: in the offline stage, the SGFormer network fully learns the mapping rules of node importance, and in the online selection of candidate nodes, it only needs to perform forward propagation to quickly output the node importance score. In contrast, the MCDM, DMC, and RCC methods rely on complex iterative calculations of topological features, resulting in longer running times than NIE.

[0124] Figure 4 By comparing the influence scaling of DLETS with baseline algorithms, the performance advantage of DLETS in the overall seed selection task is verified. Figure 4 The comparison results show that DLETS's influence expansion is significantly higher than other algorithms across all seed set sizes k. Taking the stelzl dataset as an example, when k=50, DLETS's influence expansion reaches approximately 140, while SAW_ASA and TA-degree are 120 and 130 respectively. In the AMiner academic collaboration network, DLETS's advantage is even more pronounced. This difference stems from DLETS's two-stage optimization framework: the first stage uses NIE to screen high-quality candidate nodes, effectively reducing the search space. The second stage, the ISFLA algorithm, uses two optimization operators to ensure both in-depth mining of high-quality node combinations in the candidate set and avoid getting trapped in local optima.

[0125] This invention addresses the problem of maximizing influence in large-scale social networks by constructing a multi-dimensional topic-aware independent cascade model and proposing a seed selection method that integrates deep learning and intelligent optimization search. First, by introducing the SRC (Side Controller Recognition) node importance index, this invention effectively characterizes the potential propagation capabilities of nodes from the network structure, significantly compressing the search space during the candidate node selection stage and improving the targeting and efficiency of subsequent optimization processes. Furthermore, this invention designs a node importance estimator (NIE), which learns the relationship between node structural features and their actual influence to achieve rapid prediction of node marginal gains. Compared to traditional methods that rely heavily on Monte Carlo simulations, this approach significantly reduces computational overhead while maintaining evaluation accuracy, making the algorithm applicable to larger-scale, complex network scenarios. In terms of optimization solution, this invention is based on an improved hybrid frog-leaping algorithm, introducing a search mechanism that combines hierarchical local exploitation with global information exploration, effectively avoiding the algorithm from getting trapped in local optima, thereby improving overall search performance and solution stability.

[0126] In summary, this invention establishes a synergistic mechanism across three key stages: propagation mechanism modeling, candidate node selection, and influence estimation, significantly improving computational efficiency while ensuring the effectiveness of influence propagation. Experimental verification on multiple typical network datasets demonstrates that the method of this invention outperforms existing methods in both influence diffusion capability and operational efficiency, providing a practical solution to the efficient seed selection problem in large-scale social networks.

[0127] This invention is not limited to the above embodiments. Based on the technical solutions disclosed herein, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A two-stage topic-aware influence maximization method enhanced by deep learning, characterized by: Includes the following steps, S1. Establish a propagation model and problem modeling; Given a social network , It is a social network A set of nodes in the , where each node is represented as , Represents a node exist Interest preference vectors under a specific topic , It represents the connection relationship between users; each node has the possibility of being activated during the information propagation process, and spreads the information to its neighboring nodes through edges; in order to characterize the probabilistic characteristics of the information propagation process, the diffusion process is modeled based on the independent cascading propagation mechanism; S2, Definition of node importance measurement indicators; A novel node importance metric, SRC, was designed to characterize the comprehensive influence of nodes in network propagation. This metric considers the propagation characteristics of influence while mitigating the problem of influence overlap, and can quickly and accurately identify important nodes. S3. Construction of candidate node set; After obtaining the importance metrics of all nodes, the node set... Sort by SRC value in descending order and select the top-ranked values. The nodes constitute the candidate node set ,in This process compresses the search space, allowing subsequent optimizations to be performed only in the high-potential node subspace, thereby significantly reducing computational complexity. S4. Training process of the Node Importance Estimator (NIE); To further improve the efficiency of seed selection, a node importance estimator is introduced to approximate the importance index of nodes, thereby avoiding a large amount of computation. S5. Seed optimization process based on improved hybrid frog jumping algorithm; In the candidate node set The improved hybrid frog-leaping algorithm is used for global search to obtain the optimal seed combination. Specifically, it includes two collaborative processes: hierarchical local development and global information exploration, thereby improving the combinatorial optimality of the solution while ensuring search efficiency. S6, Output the optimal seed set; When the algorithm reaches the maximum number of iterations or the fitness converges, it outputs the seed set corresponding to the current best individual. As the final solution for maximizing influence, this set can achieve the maximum expected influence diffusion under the propagation model.

2. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 1, characterized in that: The specific details of modeling the diffusion process based on the independent cascading propagation mechanism in S1 are as follows. Set nodes At any moment If activated, it will be in the next moment. With probability Try to activate its neighboring nodes The propagation process follows the one-trial principle, meaning that each edge triggers a propagation attempt only when the source node is first activated. ； Indicates in the topic Down neighboring nodes Based on topic vectors Interest vector and its own interest vector The activation probability obtained; where , There are two variable parameters; It is the original activation probability of the edge. It involves similarity calculation; based on this, a seed set is defined. Its scale is ;make Indicates a given query, For the query vector, the influence range function is denoted as... , indicating in the query Under the condition of the lower node set Scope of influence; Let represent the maximum weighted expected range of influence. Then, the problem of maximizing influence based on topic perception can be formally defined as follows: (1)。 3. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 2, characterized in that: The specific content of S2 is as follows. S2.1 Sampling of the social network G: For each edge (u,v) in the social network G, with probability... Randomly delete the edge, and the remaining edges form the edge set E' of the sub-social network, thus obtaining the sub-social network G'=(V',E'); repeat the sampling SN times to obtain the set of sub-social networks. ; S2.2, Traverse each sub-social network in the sub-social network set SG. For each node v, a depth-first traversal algorithm is used to determine its position in the sub-social network. All nodes reachable from v are listed in the dictionary, and this node information is stored in a dictionary. middle; S2.3 After traversal, calculate the reachability centrality value SRC(v) of each node v in its sub-social network; initially, the SRC(v) of each node is defined as the sum of the number of reachable nodes in all sub-social networks, i.e.: (2)。 4. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 3, characterized in that: To mitigate the overlap of influences, S2.3 introduces a dynamic discount mechanism based on SRC(v), with the following specific steps:

1. Initialize the seed set S to be empty, and execute |V| rounds of iteration; 2. Based on the current SRC value of each node, select the node v with the largest SRC value, i.e.: (3)； 3. Add node v to the seed set S, and record the reachability centrality value of the final sub-social network of this node: (4)； 4. For a newly added seed node v, update its SRC value for all its one-hop neighbors u ∈ N(v) according to the following formula: (5)； 5. Among them, Indicated in sub-social networks The size of the intersection of the sets of reachable nodes of node u and node v; 6. Proceed to the next iteration. After the iteration ends, the final reachability centrality value of the sub-social network of each node v is... Through the above process, the reachability centrality values ​​of the sub-social networks of all nodes can be obtained.

5. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 4, characterized in that: The specific content of S4 is as follows. S4.1 First, construct the training dataset. For the candidate node set... The nodes in the simulation are used to calculate their true influence gain through multiple independent cascaded propagation simulations, resulting in training sample pairs: (6)； in Represents a node The feature vector contains structural features including its SRC value and degree information. This represents the true marginal influence of a node, which is estimated through multiple propagation simulations. S3.2 Based on this, construct a regression model. This is used to fit the mapping relationship between node features and their influence; the model training objective is to minimize the error between the predicted value and the true value. (7)； After training, the NIE model can quickly predict the influence contribution of nodes, thus replacing the high-cost propagation simulation calculation in subsequent optimization processes.

6. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 5, characterized in that: The specific details of S5 are as follows. S5.1, First, in the candidate node set Initialize the population Each individual Represents a size of The seed set; The fitness value of each individual is calculated based on the node importance estimator (NIE). The population was then divided into groups based on fitness, ranked from highest to lowest. Meme group ; S5.2 In the process of hierarchical local development, for the worst solution within each meme group A two-stage progressive update strategy was designed; first, a basic improvement process was executed, which involved moving towards the optimal solution of the current meme set. Learn to replace and update the worst solution; S5.3 After completing the local development within each meme group, the global information exploration phase begins; this is achieved by introducing the global optimal solution. Guide the search direction; S5.4 After several rounds of local development and global exploration, all meme groups are reshuffled, that is, the population is reordered and re-divided into meme groups to break the original search structure, realize the full exchange of information between different meme groups, and thus avoid the algorithm from getting stuck in local optima.

7. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 6, characterized in that: The specific content of S5.2 is as follows: S5.2.1 First, execute the basic improvement process by moving towards the optimal solution of the current meme set. Learning involves replacing and updating the worst-case solution; specifically, from... Randomly select nodes from, and simultaneously from Randomly remove the corresponding node and construct a new solution: (8)； like Then use Replace the original worst solution; otherwise proceed to the next stage. S5.2.2 Based on this, a local depth optimization process is introduced to enhance the fine-grained search capability of the solution; this process centers on the current solution and searches the candidate node set. Multiple neighborhood search operations are performed for any candidate node. Calculate which node replaces any existing node. Marginal revenue afterward: (9)； The replacement operation with the maximum positive gain is selected to perform the update; this process is repeated until there is no replacement that can improve fitness, thus achieving a local optimum.

8. The two-stage topic-aware influence maximization method using deep learning enhancement as described in claim 7, characterized in that: The specific content of S5.3 is as follows: The worst solution in each meme set Construct an update strategy based on the global optimal solution, that is, from Select a node to replace, thus generating a new solution: (10)； in , This update mechanism enables local search results to converge towards the globally optimal region.

Images