Content perception susceptibility estimation method based on graph neural network
By applying the adaptive convolutional network (SACN) model of graph neural networks in social networks, combining user historical behavior and social network structure data, the problem of difficulty in accurately estimating user susceptibility in the prior art is solved, and higher evaluation accuracy and lower computing costs are achieved.
Patent Information
- Application Number
- CN202411843119.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-13
- Publication Date
- 2025-04-15
AI Technical Summary
The prior art is difficult to accurately estimate the user's susceptibility to specific content in social networks, cannot fully consider user interests and content characteristics, and the subject-based method has problems of content representation ambiguity and high calculation cost.
Adaptive convolutional network (SACN) model based on graph neural network is adopted to obtain user historical forwarding behavior and social network structure data, build user behavior feature matrix and adaptive adjacency matrix, perform convolution operations to generate fusion feature matrix, and then evaluate user susceptibility to text content.
It improves the accuracy and robustness of susceptibility assessment, avoids content representation ambiguity, and reduces computational complexity, significantly improving prediction accuracy in high-density and complex mode network data.
Smart Images
Figure CN119829939A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a content-aware susceptibility estimation method based on a graph neural network. Background Art
[0002] With the rapid development of social media, users' behavior patterns on social networks have become increasingly complex, and predicting the message forwarding paths between users is a challenge. The user's behavior history includes forwarding various messages and accepting forwardings from other users. The industry calls the user's sensitivity to seed users (the source of message delivery) or the probability that the user is influenced by the seed user susceptibility. How to accurately estimate the user's susceptibility to specific content has become a core issue in social network analysis. Existing research mainly focuses on using deep learning technology to mine network structure features, but fails to effectively integrate user behavior history and network structure information. In addition, although the topic-based prediction method can estimate the scope of influence from the perspective of different topics, the content is fuzzily clustered into a limited number of topics, which has representation ambiguity and limits the accuracy of the model.
[0003] A lot of research has been done on the problem of evaluating the susceptibility of social networks, including diffusion model-based methods and topic-based methods. Diffusion model-based methods mainly use diffusion models to simulate the diffusion process in the network to predict node susceptibility. Classical diffusion models include independent cascade models and linear threshold models. Monte Carlo simulation is used to perform a large number of repeated diffusion simulations in the network, and the results are summed and averaged to obtain the susceptibility of the node. Topic-based methods predict the susceptibility of users by identifying topics of interest to them. For example, topic classification models such as LDA are used to fuzzy cluster text content, and then estimate the user's attention to different topics. This method can mine potential topic preferences from user-generated content and help understand the distribution of user interests. However, this fuzzy clustering method simplifies content into discrete topics, which has ambiguity in content representation.
[0004] The existing susceptibility assessment algorithms still have the following shortcomings:
[0005] 1. Although the traditional diffusion-based model can capture the basic characteristics of social networks, it cannot fully consider the user's interest in specific content when facing content-aware susceptibility assessment, which affects the accuracy of the prediction.
[0006] 2. The topic-based fuzzy clustering method represents the message content spread in social networks as a collection of discrete topics, which introduces the problem of content representation ambiguity. At the same time, the computational cost increases with the increase in the number of clustering categories.
[0007] 3. Although graph neural network methods perform well in processing complex network structures, they rely too much on topological structure information, ignore the diversity of user interests, and cannot accurately reflect users' sensitivity to specific content based on their interests.
[0008] Therefore, it is necessary to provide a new method for predicting social network susceptibility, which uses a self-designed graph neural network model to estimate the susceptibility of users to seed users under single text conditions. Summary of the invention
[0009] According to the technical problems raised above, a method for estimating content-aware susceptibility based on graph neural network is provided. The present invention mainly uses the adaptive convolutional network SACN model to estimate the user's susceptibility to content, which can be applied to social network content recommendation, advertising marketing and public opinion analysis and control.
[0010] The technical means adopted by the present invention are as follows:
[0011] A content-aware susceptibility estimation method based on graph neural network, comprising:
[0012] S1. Obtain user historical forwarding behavior and social network structure data, construct user historical behavior dataset and adjacency matrix, and establish adaptive convolutional network SACN model;
[0013] S2, using the Node2Vec algorithm to obtain the network structure characteristics of the user nodes in the adjacency matrix in S1, and construct an adaptive adjacency matrix;
[0014] S3, constructing a user behavior feature matrix based on the user historical behavior data set in S1, and performing pruning, deletion and dimensionality reduction operations on the user behavior feature matrix;
[0015] S4, taking the adaptive adjacency matrix in S2 and the user behavior feature matrix in S3 as input to perform convolution operation to generate a fused feature matrix;
[0016] S5, based on the fusion feature matrix in S4, evaluate the user's susceptibility to the text content and output a susceptibility score;
[0017] S6. Feedback the susceptibility score in S5 to the adaptive convolutional network SACN model, optimize the fusion feature matrix through the back propagation algorithm, and iteratively optimize the adaptive convolutional network SACN model to estimate the user's susceptibility to the content.
[0018] Furthermore, the step S1 specifically includes:
[0019] Obtain a dataset containing user historical forwarding behaviors and social network structures, and represent the user historical behavior dataset as The social network structure data set is represented as an adjacency matrix A; the node set and directed edge set of the social network G are represented as G=(V, E);
[0020] The user historical behavior dataset The diffusion instance contained in each data is represented as D, then D is a weakly connected subgraph of G; the adjacency matrix A, where the element A in the adjacency matrix ij =1 indicates that there is a one-way forwarding behavior between user i and user j, otherwise A ij =0.
[0021] Furthermore, the step S2 specifically includes:
[0022] Use the Node2Vec algorithm to construct the initial structure embedding matrix M∈R N×d , obtain the network structure characteristics of the user nodes in the adjacency matrix, N = |V| represents the total number of nodes in the social network G, d represents the embedding dimension; the transposed matrix of the initial structure embedding matrix M is M T ;
[0023] Constructing an adjacency matrix with self-loops
[0024]
[0025] Where L is the identity matrix, λ represents the weight assigned to the self-loop in the adjacency matrix;
[0026] Construct a combination of M and The composite matrix
[0027]
[0028] Among them, ⊙ represents element-wise multiplication, and K represents the degree of involvement of structural features in the task;
[0029] Use the degree matrix to scale the adjacency matrix to prevent nodes with high degrees from generating vectors that are too large, and construct an adaptive adjacency matrix
[0030]
[0031] in, Represents the adaptive adjacency matrix.
[0032] Furthermore, the step S3 specifically includes:
[0033] Use the RoBERTa algorithm to obtain the embedding vectors of all forwarded content texts in social networks, with an embedding dimension of d and a specified number of rows of q;
[0034] Traversal Get the user's historical forwarded message text and use the corresponding embedding vector as the initial feature matrix X of user i i ∈R n×d The row vector of , n represents the number of messages that user i has forwarded;
[0035] For each X i , when n<q, calculate the average vector size of each row embedding, select the row vector corresponding to the maximum value as the representative row vector, and copy the row vector to fill the matrix X i , until n = q; when n>q, select the minimum value corresponding to the row vector of the matrix X i Perform trimming and deletion operations until n=q;
[0036] The sorted initial feature matrix X i The row vectors are expanded and concatenated, and the dimension is reduced through the fully connected layer to generate the user behavior feature matrix I.
[0037] Furthermore, the step S4 specifically includes:
[0038] Monte Carlo simulation is used to propagate multiple times on each directed edge of the social network, and the content of each propagation is randomly generated; the average of the Monte Carlo simulation results is used as the susceptibility label of each directed edge; the label is input into the convolution layer, and the convolution operation is performed to generate the fusion feature matrix H.
[0039] Furthermore, the step S5 specifically includes:
[0040] According to the fusion feature matrix H, under the condition of content c, the similarity of the feature matrix between users is calculated to measure the similarity between users; and the similarity between the head node feature and content c is calculated;
[0041] The weighted average is calculated for the similarity of the feature matrix between users and the similarity between the head node feature and the content c, and then normalized:
[0042]
[0043] Among them, P(u→v|c) represents the susceptibility of user u to seed node v, sigmoid represents the activation function, w represents the adjustable weight, and n u Denotes the interest feature matrix I of user u u The number of rows, n v Represents the interest feature matrix I of user v v number of rows.
[0044] Furthermore, the step S6 specifically includes:
[0045] The loss function MSE is calculated based on the difference between the susceptibility score and the actual label; the fusion feature matrix is updated using the gradient descent method; the susceptibility score output by S5 is fed back to the adaptive convolutional network SACN model, the model parameters are optimized through the back propagation algorithm, and the process from S4 to S5 is repeated until the adaptive convolutional network SACN model converges or reaches the predetermined maximum number of iterations.
[0046] Compared with the prior art, the present invention has the following advantages:
[0047] The content-aware susceptibility estimation method based on graph neural network provided by the present invention can take into account social network structure information and user interest preference for message content forwarding to estimate content susceptibility. By integrating the user's historical behavior characteristics and social network structure, the learning ability of the structural adaptive convolutional network (SACN) model is improved, and the accuracy and robustness of susceptibility assessment are improved. At the same time, the SACN model considers the content from a holistic perspective, avoids the ambiguity caused by discrete topics and reduces computational complexity. Experimental verification shows that the SACN model provided by the present invention can effectively improve the accuracy of susceptibility assessment. Compared with traditional methods, the method of the present invention shows higher accuracy when processing high-density and complex-pattern network data; compared with the method based on topic clustering, the method of the present invention reduces computational cost while improving the content-aware susceptibility estimation effect.
[0048] Based on the above reasons, the present invention can be widely promoted in fields such as artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0050] Figure 1 This is a flow chart of the content-aware susceptibility estimation method based on graph neural network of the present invention.
[0051] Figure 2 Schematic diagram of the model architecture in an embodiment of the present invention.
[0052] Figure 3 The SACN model in the embodiment of the present invention is compared with the prior art in terms of the content-perceived susceptibility estimation effect.
[0053] Figure 4 The figure is a comparison of the accuracy of the SACN model in the embodiment of the present invention for content-aware susceptibility estimation under different numbers of historical behaviors. DETAILED DESCRIPTION
[0054] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0055] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0056] like Figure 1 As shown, the present invention provides a content-aware susceptibility estimation method based on a graph neural network, comprising:
[0057] S1. Obtain user historical forwarding behavior and social network structure data, construct user historical behavior dataset and adjacency matrix, and establish adaptive convolutional network SACN model;
[0058] In specific implementation, as a preferred embodiment of the present invention, the step S1 specifically includes:
[0059] Obtain a dataset containing user historical forwarding behaviors and social network structures, and represent the user historical behavior dataset as The social network structure data set is represented as an adjacency matrix A; the node set and directed edge set of the social network G are represented as G=(V, E);
[0060] The user historical behavior dataset The diffusion instance contained in each data is represented as D, then D is a weakly connected subgraph of G; the adjacency matrix A, where the element A in the adjacency matrix ij =1 indicates that there is a one-way forwarding behavior between user i and user j, otherwise A ij =0.
[0061] In implementation, the meaning of each piece of data is that the seed node starts to publish the content c, and then the neighbor node forwards the content c, and the forwarding continues for multiple rounds until there is no neighbor to forward it again.
[0062] S2, using the Node2Vec algorithm to obtain the network structure characteristics of the user nodes in the adjacency matrix in S1, and construct an adaptive adjacency matrix;
[0063] In specific implementation, as a preferred embodiment of the present invention, step S2 specifically includes:
[0064] Use the Node2Vec algorithm to construct the initial structure embedding matrix M∈R N×d , obtain the network structure characteristics of the user nodes in the adjacency matrix, NN = |V| represents the total number of nodes in the social network G, d represents the embedding dimension; in general, d is 16, and the transposed matrix of the initial structure embedding matrix M is M T ;
[0065] Constructing an adjacency matrix with self-loops
[0066]
[0067] Where L is the identity matrix, λ represents the weight assigned to the self-loop in the adjacency matrix, which is set to 0.01 in the implementation;
[0068] Construct a combination of M and The composite matrix
[0069]
[0070] Among them, ⊙ represents element multiplication, and K represents the degree of involvement of structural features in the task; in implementation, the larger the K, the smaller the proportion of structural features contained in the adaptive adjacency matrix;
[0071] Use the degree matrix to scale the adjacency matrix to prevent nodes with high degrees from generating vectors that are too large, and construct an adaptive adjacency matrix
[0072]
[0073] in, Represents the adaptive adjacency matrix.
[0074] S3, constructing a user behavior feature matrix based on the user historical behavior data set in S1, and performing pruning, deletion and dimensionality reduction operations on the user behavior feature matrix;
[0075] In specific implementation, as a preferred embodiment of the present invention, step S3 specifically includes:
[0076] Use the RoBERTa algorithm to obtain the embedding vectors of all forwarded content texts in social networks, with an embedding dimension of d and a specified number of rows of q;
[0077] Traversal Get the user's historical forwarded message text and use the corresponding embedding vector as the initial feature matrix X of user i i ∈R n×d where n is the number of messages that user i has forwarded; the number of forwardings for different users is not the same.
[0078] For each X i , when n<q, calculate the average vector size of each row embedding, that is, multiply the corresponding row vector by η and divide it by d, where η is a column vector of all 1s. Select the row vector corresponding to the maximum value as the representative row vector, which shows relatively balanced performance in multiple dimensions and has the highest overall evaluation. Copy the row vector to fill the matrix X i , until n = q; when n>q, select the minimum value corresponding to the row vector of the matrix X i Perform trimming and deletion operations until n=q;
[0079] The sorted initial feature matrix X i The row vectors are expanded and concatenated, and the dimension is reduced through the fully connected layer to generate the user behavior feature matrix I.
[0080] S4, taking the adaptive adjacency matrix in S2 and the user behavior feature matrix in S3 as input to perform convolution operation to generate a fused feature matrix;
[0081] In specific implementation, as a preferred embodiment of the present invention, step S4 specifically includes:
[0082] Monte Carlo simulation is used to perform an average of 10,000 propagations on each directed edge of the social network, and the content of each propagation is randomly generated; the average of the Monte Carlo simulation results is used as the susceptibility label of each directed edge; the label is input into the convolution layer, and the convolution operation is performed to generate the fusion feature matrix H. In the implementation, the number of convolution layers is 2, and the regularization coefficient is 5×10 -4 .
[0083] S5, based on the fusion feature matrix in S4, evaluate the user's susceptibility to the text content and output a susceptibility score;
[0084] In specific implementation, as a preferred embodiment of the present invention, step S5 specifically includes:
[0085] According to the fusion feature matrix H, under the condition of content c, the similarity of the feature matrix between users is calculated to measure the similarity between users; and the similarity between the head node feature and content c is calculated;
[0086] The weighted average is calculated for the similarity of the feature matrix between users and the similarity between the head node feature and the content c, and then normalized:
[0087]
[0088] Among them, P(u→v|c) represents the susceptibility of user u to seed node v, sigmoid represents the activation function, w represents the adjustable weight, and n u Denotes the interest feature matrix I of user u u The number of rows, n v Represents the interest feature matrix I of user v v number of rows.
[0089] In implementation, the source of the content is predicted based on the similarity matrix, and the accuracy of the model in predicting susceptibility is evaluated. The final susceptibility score is output to evaluate the user's susceptibility to specific content.
[0090] S6. Feedback the susceptibility score in S5 to the adaptive convolutional network SACN model, optimize the fusion feature matrix through the back propagation algorithm, and iteratively optimize the adaptive convolutional network SACN model to estimate the user's susceptibility to the content.
[0091] In specific implementation, as a preferred embodiment of the present invention, step S6 specifically includes:
[0092] The loss function MSE is calculated based on the difference between the susceptibility score and the actual label; the fusion feature matrix is updated using the gradient descent method; the susceptibility score output by S5 is fed back to the adaptive convolutional network SACN model, the model parameters are optimized through the back propagation algorithm, and the process from S4 to S5 is repeated until the adaptive convolutional network SACN model converges or reaches the predetermined maximum number of iterations.
[0093] Example
[0094] like Figure 1 As shown, the present invention provides a content-aware susceptibility estimation method based on graph neural network. When implemented, the model architecture of the method of the present invention is as follows Figure 2 shown.
[0095] In this example, the interaction pattern between users on social networks is explicitly modeled by introducing user historical behavior features, and the content of messages forwarded by users is extracted to construct interest features, which helps to learn the influence between users and makes the susceptibility assessment task more accurate. The experimental results on four real data sets (Yelp, YouTube, Twitter, Aminer) are as follows: Figure 3The results show that the SACN model has significant improvements in hit@1 and hit@3 indicators, which is an improvement over existing methods.
[0096] The present invention introduces user historical behavior, i.e., propagation subgraph set. This embodiment verifies the robustness of SACN by reducing the training set data, retaining 20%, 40%, 60% and 80% of the data while keeping other parameters unchanged. The experimental results are shown in Figure 4 This shows that although the amount of data available for learning is reduced, the model can still maintain a high accuracy.
[0097] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.
[0098] In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0099] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.
[0100] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0101] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.
[0102] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A content-aware susceptibility estimation method based on graph neural network, characterized in that: include: S1. Obtain user historical forwarding behavior and social network structure data, construct user historical behavior dataset and adjacency matrix, and establish adaptive convolutional network SACN model; S2, using the Node2Vec algorithm to obtain the network structure characteristics of the user nodes in the adjacency matrix in S1, and construct an adaptive adjacency matrix; S3, constructing a user behavior feature matrix based on the user historical behavior data set in S1, and performing pruning, deletion and dimensionality reduction operations on the user behavior feature matrix; S4, taking the adaptive adjacency matrix in S2 and the user behavior feature matrix in S3 as input to perform convolution operation to generate a fused feature matrix; S5, based on the fusion feature matrix in S4, evaluate the user's susceptibility to the text content and output a susceptibility score; S6. Feedback the susceptibility score in S5 to the adaptive convolutional network SACN model, optimize the fusion feature matrix through the back propagation algorithm, and iteratively optimize the adaptive convolutional network SACN model to estimate the user's susceptibility to the content.
2. The method for estimating content-aware susceptibility based on graph neural network according to claim 1, characterized in that: The step S1 specifically includes: Obtain a dataset containing user historical forwarding behaviors and social network structures, and represent the user historical behavior dataset as The social network structure dataset is represented as an adjacency matrix A; the node set and directed edge set of the social network G are represented as G = (V, E); The user historical behavior dataset The diffusion instance contained in each data is represented as D, then D is a weakly connected subgraph of G; the adjacency matrix A, where the element A in the adjacency matrix ij =1 indicates that there is a one-way forwarding behavior between user i and user j, otherwise A ij =0.
3. The content-aware susceptibility estimation method based on graph neural network according to claim 1 is characterized in that: The step S2 specifically includes: Use the Node2Vec algorithm to construct the initial structure embedding matrix M∈R N×d , obtain the network structure characteristics of the user nodes in the adjacency matrix, N = |V| represents the total number of nodes in the social network G, d represents the embedding dimension; the transposed matrix of the initial structure embedding matrix M is M T ; Constructing an adjacency matrix with self-loops Where L is the identity matrix, λ represents the weight assigned to the self-loop in the adjacency matrix; Construct a combination of M and The composite matrix Among them, ⊙ represents element-wise multiplication, and K represents the degree of involvement of structural features in the task; Use the degree matrix to scale the adjacency matrix to prevent nodes with high degrees from generating vectors that are too large, and construct an adaptive adjacency matrix in, Represents the adaptive adjacency matrix.
4. The content-aware susceptibility estimation method based on graph neural network according to claim 1 is characterized in that: The step S3 specifically includes: Use the RoBERTa algorithm to obtain the embedding vectors of all forwarded content texts in social networks, with an embedding dimension of d and a specified number of rows of q; Traversal Get the user's historical forwarded message text and use the corresponding embedding vector as the initial feature matrix X of user i i ∈R n×d The row vector of , n represents the number of messages that user i has forwarded; For each X i , when n<q, calculate the average vector size of each row embedding, select the row vector corresponding to the maximum value as the representative row vector, and copy the row vector to fill the matrix X i , until n=q; when n>q, select the row vector corresponding to the minimum value to perform a trimming and deletion operation on the matrix i until n=q; The sorted initial feature matrix i is expanded and concatenated as row vectors, and the dimension is reduced through the fully connected layer to generate the user behavior feature matrix I.
5. The content-aware susceptibility estimation method based on graph neural network according to claim 1 is characterized in that: The step S4 specifically includes: Monte Carlo simulation is used to propagate multiple times on each directed edge of the social network, and the content of each propagation is randomly generated; the average of the Monte Carlo simulation results is used as the susceptibility label of each directed edge; the label is input into the convolution layer, and the convolution operation is performed to generate the fusion feature matrix H.
6. The method for estimating content-aware susceptibility based on graph neural network according to claim 1, characterized in that: The step S5 specifically includes: According to the fusion feature matrix H, under the condition of content c, the similarity of the feature matrix between users is calculated to measure the similarity between users; and the similarity between the head node feature and content c is calculated; The weighted average is calculated for the similarity of the feature matrix between users and the similarity between the head node feature and the content c, and then normalized: Among them, P(u→v|c) represents the susceptibility of user u to seed node v, sigmoid represents the activation function, w represents the adjustable weight, and n u Denotes the interest feature matrix I of user u u The number of rows, n v Represents the interest feature matrix I of user v v number of rows.
7. The content-aware susceptibility estimation method based on graph neural network according to claim 1 is characterized in that: The step S6 specifically includes: The loss function MSE is calculated based on the difference between the susceptibility score and the actual label; the fusion feature matrix is updated using the gradient descent method; the susceptibility score output by S5 is fed back to the adaptive convolutional network SACN model, the model parameters are optimized through the back propagation algorithm, and the process from S4 to S5 is repeated until the adaptive convolutional network SACN model converges or reaches the predetermined maximum number of iterations.