Social network key user mining system and method based on propagation feature deep reinforcement learning

Through a deep reinforcement learning method based on propagation characteristics, using the graph attention mechanism and dual-deep value network of propagation tendency perception, the user's propagation representation and value function are calculated, and the problem of small impact range in the existing technology is solved, and the problem of in-depth reinforcement learning method is realized. Effectively explore key users and expand the scope of information dissemination in social networks.

CN118916553APending Publication Date: 2024-11-08School of Political Science, National Defense University of the Chinese People's Liberation Army
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410918801.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-10
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing technology has a small range of information impact in social networks, fails to effectively consider the user's communication characteristics, and cannot effectively explore key users in social networks.

Method used

Using a deep reinforcement learning method based on propagation characteristics, the user's propagation representation is learned through the graph attention mechanism of propagation tendency perception, and the value function that measures the cumulative impact of the expected is calculated. The propagation representation and value function are trained using the dual-deep value network (DDQN) Select the key users with the greatest value.

Benefits of technology

Under real circumstances, it can effectively expand the scope of information dissemination and improve the dissemination performance of information on social networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916553A_ABST
    Figure CN118916553A_ABST
Patent Text Reader

Abstract

The invention discloses a social network key user mining system and method based on propagation feature deep reinforcement learning, and the system comprises a propagation representation extraction unit, a value function calculation unit, a training unit, and a key user selection unit. The training module is used for training a propagation feature deep reinforcement learning model based on a graph attention mechanism, the model extracts propagation characterization of a user from a training sample network and calculates a value function based on the propagation characterization, and the value function of the user is calculated according to real-time input through the trained propagation feature deep reinforcement learning model in an online stage. And based on a value function, the key user with the largest information spreading range is rapidly mined and pushed. According to the method, the propagation characteristics of the user are learned by utilizing a graph attention mechanism, the value function for measuring cumulative influence expectation is calculated, DDQN is adopted to train value function parameters, the key user is selected based on the value function, and good influence propagation performance can be obtained under the real condition.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a technology in the field of neural network applications, specifically a network key user mining system and method based on deep reinforcement learning of propagation features. Background Art

[0002] Mining key users on social networks refers to how to select a certain number of users from the massive number of users on social networks, and change the probability of them being activated by a certain information that is being spread by giving them incentives, so that the total number of users activated by the information on the social network is maximized. Deep reinforcement learning based on propagation characteristics can more quickly and reasonably mine key users on social networks. Summary of the invention

[0003] In view of the shortcomings of the prior art, such as small information influence range, no consideration of user propagation characteristics under the real propagation model, and inability to be used for mining key users in social networks, the present invention proposes a social network key user mining system and method based on deep reinforcement learning of propagation characteristics. The propagation representation of users is learned by using a graph attention mechanism that perceives propagation tendency, and a value function that measures the expected cumulative influence is calculated. A dual deep value network (DDQN) is used to train the neural network parameters in the propagation representation and the value function. The user value is evaluated based on the trained value function, and key users are selected according to the value size, which can achieve good influence propagation performance in real situations.

[0004] The present invention is achieved through the following technical solutions:

[0005] The invention relates to a network key user mining system based on deep reinforcement learning of propagation features, comprising: a propagation representation extraction unit, a value function calculation unit, a training unit and a key user selection unit, wherein: the propagation representation extraction unit calculates the propagation tendency perception attention weights between users according to the user initial representation, the user influence strength and the user propagation tendency threshold information through the propagation tendency perception attention mechanism, and calculates the user propagation representation by fusing the user's neighbor propagation representation according to the propagation tendency perception attention weights, so as to realize the update extraction of all user propagation representations; the value function calculation unit calculates the value function of each user according to the propagation representation information extracted from each user, so as to realize the estimation of the value of each user; the training unit performs iterative selection of key users, loss function calculation and back propagation optimization processing according to the propagation representation extraction and value function calculation results of the user, so as to realize the training and optimization of the neural network parameters involved in the propagation representation extraction unit and the value function calculation unit; the key user selection unit calculates the user's propagation representation according to the real-time input through the trained propagation representation extraction unit, calculates the user's value function through the trained value function unit, and realizes the mining of the key users who promote the largest information dissemination range based on the value function ranking of all users.

[0006] The present invention relates to a method for mining key users of a social network based on the above system. In the model building and training stage, samples are collected from the social network and a training sample network is constructed to build a propagation representation extraction unit and a value function calculation unit. The propagation representation extraction unit is used to calculate the attention weight of propagation tendency perception and the propagation representation aggregation processing according to the user's initial representation, the user's influence strength and the user's propagation tendency threshold information to obtain the final user propagation representation result. The value function calculation unit is used to calculate the user's value function result according to the propagation representation information of all users. The training unit is used to train the neural network parameters in the propagation representation extraction unit and the value function calculation unit through key user iterative selection, loss function calculation, and back propagation optimization processing. In the model inference stage, a key user selection unit is constructed. The key user selection unit is used to calculate the user's propagation representation according to the real-time input through the trained propagation representation extraction unit. The user's value function is calculated through the trained value function unit, and the key user who promotes the largest information dissemination range is quickly mined based on the value function ranking of all users.

[0007] The training sample network is constructed in the following way: extracting the relationship data between users from the data of the online social media platform, constructing a global social network G(V,E) based on the extracted relationship data between users, randomly sampling l initial users from the social network G(V,E), and constructing l sub-networks from these initial users based on the breadth-first search strategy, and the size of each sub-network is limited to 50, where: V is the user set, and E is the relationship set between users.

[0008] The user propagation characterization result is obtained by:

[0009] Step 1) Randomly initialize the initial representation of each user

[0010] Step 2) For the target user v, the dynamic interaction influence weight of the neighbor u on v is calculated based on the attention mechanism, that is, in: is the attention vector, is to change the representation dimension from h k Convert to h k+1 The weight matrix of , [·,·] is the vector concatenation;

[0011] Step 3) Normalize the dynamic interaction influence weights of all neighbors on v using the softmax function with LeakyReLU activation Where: N in (v) is the input neighbor set of v;

[0012] Step 4) Calculate the aggregation of the propagation representations of all neighbors of v based on the propagation tendency-aware attention weights: in: is the attention weight of the propagation tendency perception, and is the weight parameter, w uv is the influence of u on v, θ v is the propagation tendency threshold of v, w uv and θ v It is formulated by users based on the actual situation of social networks. is the dynamic interaction influence weight;

[0013] Step 5) Calculate the propagation representation of v at the k+1 layer: in: and is the parameter weight, a v is the current activation state of v. When v is activated, a v =1, otherwise a v =0;

[0014] Step 6) Iterate the above steps 2)-5) K times to obtain the user's propagation representation

[0015] The value function is specifically: in: and is the value function parameter weight vector, the vector dimension is h K , is the set of all neural network parameters involved in the value function calculation process.

[0016] The training specifically includes:

[0017] Step i) setting the behavior neural network with parameter Θ and the target neural network with parameter Θ' and initializing the network parameters of the two to be consistent;

[0018] Step ii) The initial key user set is set to an empty set, Initialize the key user count variable t=0;

[0019] Step iii) uniformly sample a random number z in the range [0,1], when z < ε, from the set V\(U t-1 ∪S) randomly selects a user v t , otherwise calculate the value function of all users And select the user with the largest value function, that is Where: ε is the preset threshold, S is the source seed user set;

[0020] Step iv) Select the user vt Add key user set U t =U t-1 ∪{v t};

[0021] Step v) When t≥n, the current training sample Add to the playback buffer M, where: γ is the decay rate, r(U t-n+i ,v t-n+i ) is the reward function, indicating the selection of user v t-n+i Add key user set U t-n+i The amount of change in the post-propagation impact range;

[0022] Step vi) After looping the above steps iii) to v) b times, randomly sample training sample batches B~M from the playback buffer M, and optimize the loss function based on the back propagation algorithm: Where: b is the budget size of the key user set;

[0023] Step vii) loops the above steps ii) to vi) N times, and updates the target neural network parameters Θ' with the behavioral neural network parameters Θ every m times, wherein: N and m are preset training parameters.

[0024] The selection of the user with the largest value function refers to: using the trained propagation representation extraction unit and the value function calculation unit to calculate the value of each user on the social network Sort all users by value, and select the top b users with the largest value to form the key user set for dissemination. Technical Effects

[0025] The present invention extracts the user's propagation representation based on the graph attention mechanism of propagation tendency perception, so that the representation based on topological structure learning can capture the cascade propagation effect in information propagation, which can be used to calculate the user's value function and select key users according to the value function ranking. Compared with the existing key user selection technology, by using the above technical means to select a certain number of key users in the propagation path of the target information and adjust their propagation tendency threshold, the target information can be spread more widely on the social network. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic diagram of the system of the present invention;

[0027] Figure 2 It is a flow chart of the present invention;

[0028] Figure 3 This is a performance comparison chart of the impact range of the embodiment;

[0029] In the figure: (a) is a graph showing the influence scope on Twitter network changing with the budget number of key users; (b) is a graph showing the influence scope on caGr network changing with the budget number of key users; (c) is a graph showing the influence scope on Wiki network changing with the budget number of key users; (d) is a graph showing the influence scope on Epinions network changing with the budget number of key users; (e) is a graph showing the influence scope on Buzznet network changing with the budget number of key users; (f) is a graph showing the influence scope on Youtube network changing with the budget number of key users;

[0030] Figure 4 It is a comparison chart of the running time of the embodiment;

[0031] In the figure: (a) is the running time on the Twitter network and the graph showing the variation of the key user budget number; (b) is the running time on the caGr network and the graph showing the variation of the key user budget number; (c) is the running time on the Wiki network and the graph showing the variation of the key user budget number; (d) is the running time on the Epinions network and the graph showing the variation of the key user budget number; (e) is the running time on the Buzznet network and the graph showing the variation of the key user budget number; (f) is the running time on the Youtube network and the graph showing the variation of the key user budget number. DETAILED DESCRIPTION

[0032] like Figure 2 As shown, this embodiment relates to a method for mining key users of social networks based on deep reinforcement learning of propagation features, including the following steps:

[0033] Step 1: Extract the relationship data between users from the data of online social media platforms. Based on the extracted relationship data between users, use users as nodes and the relationships between users as edges to construct a global social network graph G(V,E), where V is the user set and E is the relationship set between users. Randomly sample l initial users on the network graph, where l is a parameter specified by the user. Based on these initial users, construct l subgraphs based on the breadth-first search strategy, and the size of each subgraph is limited to 50.

[0034] Step 2: Use the subgraph obtained in step 1 to extract the propagation representation based on the attention mechanism of the propagation tendency perception graph, specifically including:

[0035] 2.1) Randomly initialize the initial representation of each user For a target user v, the dynamic interaction influence weight of neighbor u on v is calculated based on the attention mechanism, that is, in: is the attention vector, is to change the representation dimension from h k Convert to h k+1The weight matrix of , [·,·] is the vector concatenation.

[0036] 2.2) Use the softmax function with LeakyReLU activation to normalize the dynamic interaction influence weights of all neighbors on v: Where: N in (v) is the input neighbor set of v.

[0037] 2.3) Calculate the aggregation of propagation features of all neighbors of v based on the attention weights perceived by the propagation tendency: in: is the attention weight of the propagation tendency perception, and is the weight parameter, w uv is the influence of u on v, θ v is the propagation tendency threshold of v, w uv and θ v It is formulated by users based on the actual situation of social networks. is the dynamic interaction influence weight.

[0038] 2.4) Calculate the propagation characteristics of v in the k+1 layer: in and is the parameter weight, a v is the current activation state of v. When v is activated, a v =1, otherwise a v =0.

[0039] 2.5) Loop steps 2.3) - 2.4) K times to get the user's propagation representation

[0040] Step 3: Use the user propagation representation extraction unit obtained in step 2 to calculate the value function, specifically: in: and is the value function parameter weight vector, the vector dimension is h K , is the set of all neural network parameters involved in the value function calculation process.

[0041] Step 4: Train the neural network, including:

[0042] 4.1) Initialize the behavior neural network with parameter Θ and the target neural network with parameter Θ': the parameters Θ and Θ' are consistent, and the initial key user set is set to an empty set. Initialize the key user count variable t=0.

[0043] 4.2) Uniformly sample a random number z in the range [0,1]. When z < ε, select a random number from the set V\(U t-1 ∪S) randomly selects a user v t , otherwise calculate the value function of all users And select the user with the largest value function, that is The selected user v t Add key user set U t =U t-1 ∪{v t}, where: ε is the preset threshold, S is the source seed user set.

[0044] 4.3) When t≥n, the current training sample Added to the playback buffer M, where γ is the decay rate, r(U t-n+i ,v t-n+i ) is the reward function, which means choosing user v t-n+i Add key user set U t-n+i The change in the scope of post-propagation influence can be estimated by performing Monte Carlo simulation on the subgraph according to the propagation model.

[0045] 4.4) After looping steps 4.2) and 4.3) b times, randomly sample training sample batches B~M from the playback buffer M, and optimize the loss function based on the back propagation algorithm: Where b is the budget size of the critical user set.

[0046] 4.5) Repeat steps 4.2) and 4.4) N times, and update the target neural network parameters Θ' with the behavioral neural network parameters Θ every m times, where N and m are parameters specified by the user.

[0047] Step 5: Use the neural network trained in step 4 to calculate the value of each user on the social network, sort all users by value, and select the top b users with the largest value to form the key user set for dissemination.

[0048] After specific practical experiments, the CPU Intel (R) Xeon (R) CPU E5-2630 v4 @ 2.20GHz, GPU NVIDIA TITAN RTX, RAM 128.00GB, operating system centos7, simulation code environment python environment, experiments were conducted on real-world network datasets to evaluate the effect of the method, including: Twitter, caGr, Buzznet, Wiki, Epinions and Youtube, among which: Twitter, caGr, and Buzznet are from Rosi et al. in 2015 at the International Conference on The paper "The network data repository with interactive graph analytics and visualization" published in the Annual AAAI Conference on Artificial Intelligence, Wiki, Epinions and Youtube from the website https: / / snap.stanford.edu / snap / index.html .

[0049] This embodiment is represented by KPRL in the simulation experiment. The simulation comparison of this embodiment with existing key user discovery methods includes: the IMM method proposed by Tang et al. in "Influence maximization in near-linear time: AMartingale approach" published in the international conference "ACM SIGMOD International Conference on Management of Data" in 2015, the OPIM-C method proposed by Tang et al. in the article "Online processing algorithms for influence maximization" published in the international conference "ACM SIGMOD International Conference on Management of Data" in 2018, the S2V-DQN method proposed by Khalil et al. in the article "Learning combinatorial optimization algorithms over graphs" published in the international conference "Conference on Neural Information Processing Systems" in 2017, and the S2V-DQN method proposed by Manchanda et al. in the international conference "Conference on Neural Information Processing Systems" in 2020. The GCOMB method was proposed in the paper "GCOMB: Learning budget-constrained combinatorial algorithms over billion-sized graphs" published in the "Conference on Neural Information Processing Systems", and Li et al. proposed the PIANO method in the paper "PIANO: Influence maximization meets deep reinforcement learning" published in the international journal "IEEE Transactions on Computational Social Systems" in 2022.

[0050] In the specific implementation, the propagation threshold of each user is randomly selected from the range of (0,1). The budget b of the number of key users on all training data is set to 5. For each test data, the budget b is set to b∈{10,20,30,40,50}. The number of iterative learning of user propagation features K is set to 3. On each network, 50 users are randomly selected as the source seed users of the information initiation.

[0051] like Figure 3 As shown, the performance comparison of the propagation influence range of this embodiment and multiple existing methods on six data sets, namely Twitter, caGr, Buzznet, Wiki, Epinions and Youtube. As can be seen from the figure, the performance of this embodiment is the best on all data sets, among which: the performance of IMM and OPIM-C is relatively close, and IMM is slightly better than OPIM-C in general. The performance of S2V-DQN, GCOMB and PIANO is unstable on different data sets. This may be because the network feature extraction modules of these methods can extract features of different qualities on different networks. The propagation feature learning module specially constructed in this embodiment can adapt well to different network structures, so its performance is stably better than other methods on different networks.

[0052] like Figure 4As shown, it is a comparison of the running time of this embodiment and multiple existing methods on six data sets of Twitter, caGr, Buzznet, Wiki, Epinions and Youtube. As can be seen from the figure, among all the tested methods, OPIM-C takes the least time, followed by PIANO. The running speed of this embodiment is slightly slower than PIANO, which may be because this method has more parameters and needs to calculate the dynamic interaction influence weights between users. But the running time of this embodiment is acceptable, even on the million-size data set Youtube, its running time is less than 10 seconds. The S2V-DQN method runs slowly, and its running time increases approximately linearly with the size of the key user budget. GCOMB runs slower than S2V-DQN on the Twitter and caGr data sets. This may be because GCOMB is proposed for combinatorial optimization problems on very large networks, but for smaller networks, the additional computational overhead of GCOMB makes it run longer.

[0053] In summary, the present invention uses the graph attention mechanism to learn the user's propagation characteristics and calculates the value function that measures the expected cumulative influence. It uses DDQN to train the value function parameters and selects key users based on the value function, which can achieve good influence propagation performance in real situations.

[0054] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A network key user mining system based on deep reinforcement learning of propagation features, characterized in that: include: A propagation representation extraction unit, a value function calculation unit, a training unit and a key user selection unit, wherein: the propagation representation extraction unit calculates the propagation tendency-aware attention weights between users according to the user's initial representation, the user's influence strength and the user's propagation tendency threshold information through the propagation tendency-aware attention mechanism, and calculates the user's propagation representation by fusing the user's neighbor propagation representation according to the propagation tendency-aware attention weights, so as to realize the update extraction of all user's propagation representations; the value function calculation unit calculates the value function of each user according to the propagation representation information extracted from each user, so as to realize the estimation of each user's value; the training unit performs iterative selection of key users, loss function calculation, and back-propagation optimization processing according to the user's propagation representation extraction and value function calculation results, so as to realize the training and optimization of the neural network parameters involved in the propagation representation extraction unit and the value function calculation unit; the key user selection unit calculates the user's propagation representation according to the real-time input through the trained propagation representation extraction unit, calculates the user's value function through the trained value function unit, and realizes the mining of key users who promote the largest range of information dissemination based on the value function ranking of all users.

2. A method for mining key users of social networks based on the system of claim 1, characterized in that: In the model building and training stage, samples are collected from the social network and a training sample network is constructed to build a propagation representation extraction unit and a value function calculation unit. The propagation representation extraction unit is used to calculate the attention weight of propagation tendency perception and the propagation representation aggregation processing according to the user's initial representation, user influence strength and user propagation tendency threshold information to obtain the final user propagation representation result. The value function calculation unit is used to calculate the user's value function result according to the propagation representation information of all users. The training unit is used to train the neural network parameters in the propagation representation extraction unit and the value function calculation unit through key user iterative selection, loss function calculation, and back propagation optimization processing. In the model inference stage, a key user selection unit is constructed. The key user selection unit is used to calculate the user's propagation representation according to the real-time input through the trained propagation representation extraction unit, and the user's value function is calculated through the trained value function unit. Based on the value function ranking of all users, the key users who promote the largest range of information dissemination are quickly mined.

3. The method for mining key users in a social network according to claim 2, characterized in that: The training sample network is constructed in the following way: extracting the relationship data between users from the data of the online social media platform, constructing a global social network G(V,E) based on the extracted relationship data between users, randomly sampling l initial users from the social network G(V,E), and constructing l sub-networks from these initial users based on the breadth-first search strategy, where: V is the user set and E is the relationship set between users.

4. The method for mining key users in a social network according to claim 2, characterized in that: The user propagation characterization result is obtained by: Step 1) Randomly initialize the initial representation of each user Step 2) For the target user v, the dynamic interaction influence weight of the neighbor u on v is calculated based on the attention mechanism, that is, in: is the attention vector, is to change the representation dimension from h k Convert to h k+1 The weight matrix of , [·,·] is the vector concatenation; Step 3) Normalize the dynamic interaction influence weights of all neighbors on v using the softmax function with LeakyReLU activation Where: N in (v) is the input neighbor set of v; Step 4) Calculate the aggregation of the propagation representations of all neighbors of v based on the propagation tendency-aware attention weights: in: is the attention weight of the propagation tendency perception, and is the weight parameter, w uv is the influence of u on v, θ v is the propagation tendency threshold of v, w uv and θ v It is formulated by users based on the actual situation of social networks. is the dynamic interaction influence weight; Step 5) Calculate the propagation representation of v at the k+1 layer: in: and is the parameter weight, a v is the current activation state of v. When v is activated, a v =1, otherwise a v =0; Step 6) Iterate the above steps 2)-5) K times to obtain the user's propagation representation 5. The method for mining key users in a social network according to claim 2, characterized in that: The value function is specifically: in: and is the value function parameter weight vector, the vector dimension is h K , is the set of all neural network parameters involved in the value function calculation process.

6. The method for mining key users in a social network according to claim 2, characterized in that: The training specifically includes: Step i) setting the behavior neural network with parameter Θ and the target neural network with parameter Θ' and initializing the network parameters of the two to be consistent; Step ii) The initial key user set is set to an empty set, Initialize the key user count variable t=0; Step iii) uniformly sample a random number z in the range [0,1], when z < ε, from the set V\(U t-1 ∪S) randomly selects a user v t , otherwise calculate the value function of all users And select the user with the largest value function, that is Where: ε is the preset threshold, S is the source seed user set; Step iv) Select the user v t Add key user set U t =U t-1 ∪{v t }; Step v) When t≥n, the current training sample Add to the playback buffer M, where: γ is the decay rate, r(U t-n+i ,v t-n+i ) is the reward function, indicating the selection of user v t-n+i Add key user set U t-n+i The amount of change in the post-propagation impact range; Step vi) After looping the above steps iii) to v) b times, randomly sample training sample batches B~M from the playback buffer M, and optimize the loss function based on the back propagation algorithm: Where: b is the budget size of the key user set; Step vii) loops the above steps ii) to vi) N times, and updates the target neural network parameters Θ' with the behavioral neural network parameters Θ every m times, wherein: N and m are preset training parameters.

7. The method for mining key users in a social network according to claim 6, characterized in that: The selection of the user with the largest value function refers to: using the trained propagation representation extraction unit and the value function calculation unit to calculate the value of each user on the social network Sort all users by value, and select the top b users with the largest value to form the key user set for dissemination.

8. The method for mining key users of a social network according to any one of claims 2 to 7, characterized in that: include: Step 1: Extract the relationship data between users from the data of online social media platforms, and construct a global social network graph G(V,E) based on the extracted relationship data between users, taking users as nodes and the relationships between users as edges, where V is the user set and E is the relationship set between users; randomly sample l initial users on the network graph, where l is a parameter specified by the user; construct l subgraphs from these initial users based on the breadth-first search strategy, and the size of each subgraph is limited to 50; Step 2: Use the subgraph obtained in step 1 to extract the propagation representation based on the attention mechanism of the propagation tendency perception graph, specifically including: 2.1) Randomly initialize the initial representation of each user For a target user v, the dynamic interaction influence weight of neighbor u on v is calculated based on the attention mechanism, that is, in: is the attention vector, is to change the representation dimension from h k Convert to h k+1 The weight matrix of , [·,·] is the vector concatenation; 2.2) Use the softmax function with LeakyReLU activation to normalize the dynamic interaction influence weights of all neighbors on v: Where: N in (v) is the input neighbor set of v; 2.3) Calculate the aggregation of propagation features of all neighbors of v based on the attention weights perceived by the propagation tendency: in: is the attention weight of the propagation tendency perception, and is the weight parameter, w uv is the influence of u on v, θ v is the propagation tendency threshold of v, w uv and θ v It is formulated by users based on the actual situation of social networks. is the dynamic interaction influence weight; 2.4) Calculate the propagation characteristics of v in the k+1 layer: in and is the parameter weight, a v is the current activation state of v. When v is activated, a v =1, otherwise a v =0; 2.5) Loop steps 2.3) - 2.4) K times to get the user's propagation representation Step 3: Use the user propagation representation extraction unit obtained in step 2 to calculate the value function, specifically: in: and is the value function parameter weight vector, the vector dimension is h K , is the set of all neural network parameters involved in the value function calculation process; Step 4: Train the neural network, including: 4.1) Initialize the behavior neural network with parameter Θ and the target neural network with parameter Θ': the parameters Θ and Θ' are consistent, and the initial key user set is set to an empty set. Initialize the key user count variable t=0; 4.2) Uniformly sample a random number z in the range [0,1]. When z < ε, select a random number from the set V\(U t-1 ∪S) randomly selects a user v t , otherwise calculate the value function of all users And select the user with the largest value function, that is The selected user v t Add key user set U t =U t-1 ∪{v t }, where: ε is the preset threshold, S is the source seed user set; 4.3) When t≥n, the current training sample Added to the playback buffer M, where γ is the decay rate, r(U t-n+i ,v t-n+i ) is the reward function, which means choosing user v t-n+i Add key user set U t-n+i The change in the post-propagation impact range. The reward function value can be estimated by performing a Monte Carlo simulation on the subgraph according to the propagation model. 4.4) After looping steps 4.2) and 4.3) b times, randomly sample training sample batches B~M from the playback buffer M, and optimize the loss function based on the back propagation algorithm: Where b is the budget size of the key user set; 4.5) looping steps 4.2) and 4.4) N times, and updating the target neural network parameters Θ' with the behavior neural network parameters Θ every m times, where N and m are parameters specified by the user; Step 5: Use the neural network trained in step 4 to calculate the value of each user on the social network, sort all users by value, and select the top b users with the largest value to form the key user set for dissemination.

Citation Information

Patent Citations

  • Method and device for discovering important users on the basis of spreading networks

    CN105956925A

  • Social network influence prediction method and device based on graph neural network

    CN113792937A

  • Social network equilibrium influence maximization system and method based on deep reinforcement learning

    CN117036077A