A method and device for generating an information diffusion graph based on user similarity

Abstract

The present invention discloses a method and device for generating an information diffusion graph based on user similarity, which relates to the technical field of network information diffusion. A candidate user set for generating an information diffusion graph is generated as needed, and a standardized feature matrix is ​​determined; the importance ranking of the candidate users is determined based on the standardized feature matrix and the diffusion type label of each candidate user; a user similarity matrix is ​​calculated; and the number of candidate users is determined as the diffusion step length. T , and obtain the initial information propagation graph; from the number of steps T Starting from step 1, reversely iterate to generate the edges connecting the candidate users in the initial information diffusion graph. The initial information diffusion graph after iteration is determined as the information diffusion graph of the candidate user set. This method can generate an accurate information diffusion graph.

Description

Technical Field

[0001] The present invention relates to the technical field of network information dissemination, and in particular to a method and device for generating an information dissemination graph based on user similarity. Background Art

[0002] In today's digital age, the rise of social media and online platforms has enabled information dissemination to become more widespread and rapid. To better understand and analyze the spread of this information, studying how information propagates among users is crucial. Currently, information diffusion on social networks is often modeled as a random process on a propagation network. The topology of a propagation network is a complex network composed of nodes and edges. Nodes typically represent individuals or entities, and edges represent the propagation relationships between them, i.e., the actual path of information dissemination. Furthermore, many downstream tasks, such as localizing information sources and detecting disinformation, require a large amount of propagation network topologies as training sets to data-drivenly explore underlying propagation patterns. However, in real-world scenarios, monitoring and acquiring realistic propagation network structures consumes significant resources, resulting in a relatively small amount and scale of data. This can lead to drawbacks such as limiting model generalization, increasing the risk of overfitting, and reducing model robustness. Furthermore, due to privacy and security concerns, obtaining publicly available real-world propagation network data is becoming increasingly difficult. To address this difficulty, propagation graph generation methods aim to simulate and emulate the information propagation process, thereby generating realistic propagation network topologies.

[0003] Currently, methods for generating propagation maps fall into two main categories: simulation methods based on information propagation models and graph generation methods based on deep learning. However, regardless of the method, current methods for generating propagation maps fail to fully consider user interaction, resulting in insufficient fidelity in the generated maps. Summary of the Invention

[0004] Based on this, it is necessary to provide a method and device for generating an information diffusion graph based on user similarity to address the above technical problems, so as to solve the defects of the existing technology that it is unable to detect and identify false texts in a timely and effective manner, and does not fully consider the problem of user interaction, thereby improving the accuracy of the generated information diffusion graph.

[0005] The present invention adopts the following technical solutions:

[0006] The present invention provides a method for generating an information diffusion graph based on user similarity, comprising:

[0007] Generate a candidate user set for the information diffusion graph as needed and determine a standardized feature matrix; the candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features in multiple dimensions; each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension;

[0008] Determine the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user;

[0009] Calculate the user similarity matrix based on the standardized feature matrix;

[0010] The number of candidate users is determined as the diffusion step length T , and obtain the initial information propagation graph; the initial information propagation graph includes candidate user nodes of all candidate users;

[0011] From the number of steps T -1 until the number of steps is 1, reverse iteration is performed to generate the edges connecting the candidate users in the initial information propagation graph, and the initial information propagation graph after iteration is determined as the information propagation graph of the candidate user set.

[0012] Optionally, a candidate user set of the information diffusion graph is generated as needed, and a standardized feature matrix is ​​determined, including:

[0013] For the features of any dimension, normalize the features of all candidate users in that dimension to obtain the normalized features of each user in that dimension;

[0014] A standardized feature matrix is ​​constructed based on the normalized features of all users in each dimension; each element in the standardized feature matrix is ​​the normalized feature of the corresponding candidate user in the corresponding dimension.

[0015] Optionally, determining the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user includes:

[0016] Determine the importance score of each dimension’s features based on the standardized feature matrix and the communication type label of each candidate user;

[0017] The importance ranking of candidate users is determined based on the importance score of the features of each dimension and the standardized feature matrix; the candidate users with higher importance scores are ranked higher.

[0018] Optionally, based on the standardized feature matrix and the communication type label of each candidate user, an importance score of the feature of each dimension is determined, including:

[0019] For the features of any dimension, the chi-square statistic of the features of each dimension is calculated based on the standardized feature matrix and the communication type label of each candidate user;

[0020] The chi-square statistic of the features of each dimension is determined as the importance score of the features of each dimension.

[0021] Optionally, the chi-square statistic is calculated as:

[0022] ;

[0023] in, Indicates the j The chi-square statistic of the dimension feature, Represents the first i The first candidate user j The normalized features of the dimension, 、 Represents the candidate users who participate in non-rumor spreading and rumor spreading respectively j The average value of the dimensionally normalized features.

[0024] Optionally, a user similarity matrix is ​​calculated based on the standardized feature matrix, including:

[0025] Based on the standardized feature matrix, the candidate user set is updated and the updated candidate user set is input into the graph variational autoencoder to obtain the user potential representation of each candidate user;

[0026] Based on the potential representation of each candidate user, calculate the similarity between any two candidate users among all candidate users;

[0027] Based on the similarity between any two candidate users among all candidate users, a user similarity matrix is ​​constructed.

[0028] Optionally, from the number of steps T Starting from step -1 until the number of steps reaches 1, reverse iteration is performed to generate the edges connecting the candidate users in the initial information propagation graph, specifically including:

[0029] In each iteration, the target candidate user node corresponding to the current step in the importance ranking is obtained, the user similarity matrix of the target candidate user node and the initial information propagation graph are input into the diffusion model, and the edges between the target candidate user node and other candidate user nodes in the initial information propagation graph are predicted;

[0030] The predicted edges are integrated into the initial information propagation graph. Based on the new initial information propagation graph, the edges of the candidate users in the initial information propagation graph are continuously generated iteratively until the iteration is completed.

[0031] Optionally, the process of constructing the diffusion model includes:

[0032] Obtain multiple historical communication graphs; the historical communication graphs include multiple historical users, communication paths between historical users, and user attribute information of historical users;

[0033] According to the historical propagation graph, calculate the importance ranking of historical users and the historical user similarity matrix;

[0034] The number of historical users is determined as the historical diffusion step length;

[0035] Starting from step 1 to the historical diffusion step, the absorption operation of historical user nodes is iteratively performed according to the historical propagation graph, the importance ranking of historical users and the historical user similarity matrix, and the diffusion model is trained in this process.

[0036] Optionally, based on the historical propagation graph, the importance ranking of historical users, and the historical user similarity matrix, the historical user node absorption operation is iteratively performed, and the diffusion model is trained in this process, including:

[0037] For the historical diffusion step number t Step 1: Select the historical user node with the least importance in the importance ranking ;

[0038] Recorded in the current historical propagation map Middle and historical user nodes Connected node labels ;

[0039] Historical user nodes Convert it to a masked state and connect it with all historical user nodes in the current historical propagation graph with masked edges to generate an intermediate graph with masked edges :

[0040] ;

[0041] in, Absorb Represents the operation of converting the historical user node to the absorbing state;

[0042] The middle picture Similarity matrix with historical users Input the denoising module of the diffusion model to predict historical user nodes and Which nodes in the have ever had edges connected to them?

[0043] ;

[0044] in, is the trainable parameter of the denoising module, is the predicted label;

[0045] By node label and predicted labels Calculate the mean squared error loss and optimize it by gradient descent .

[0046] The present invention provides a device for generating an information diffusion graph based on user similarity, comprising:

[0047] An acquisition module is used to generate a candidate user set for the information diffusion graph as needed and determine a standardized feature matrix; the candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features in multiple dimensions; each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension;

[0048] A ranking module is used to determine the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user;

[0049] A calculation module, used to calculate a user similarity matrix based on the standardized feature matrix;

[0050] The first determination module is used to determine the number of candidate users as the diffusion step length T , and obtain the initial information propagation graph; the initial information propagation graph includes candidate user nodes of all candidate users;

[0051] Iteration module, used to start from the number of steps T -1 until the number of steps reaches 1, reverse iteration is performed to generate the edges connecting the candidate users in the initial information propagation graph;

[0052] The second determining module is used to determine the initial information propagation graph completed by iteration as the information propagation graph of the candidate user set.

[0053] The present invention provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned method for generating an information propagation graph based on user similarity.

[0054] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, the above-mentioned method for generating an information propagation graph based on user similarity is implemented.

[0055] At least one of the above technical solutions adopted by the present invention can achieve the following beneficial effects:

[0056] First, the importance ranking of candidate users is determined based on the standardized feature matrix and the communication type label, and the correlation between user attributes and communication type is taken into account, so as to screen out candidate users who play a key role in information dissemination. Candidate users with great communication influence are ranked at the top, which facilitates giving priority to the connection relationship of key candidate user nodes when constructing the information dissemination graph and improving the accuracy of the information dissemination graph; on this basis, the reverse iterative process selects the target candidate user node from the importance ranking, combines its user similarity matrix and the initial dissemination graph to predict the edge, which can accurately characterize the information dissemination path, because the target candidate user node is the core of dissemination, and similar users are more likely to generate dissemination edges; each iteration integrates the predicted edge into the dissemination graph and continues to iterate, gradually improving the structure of the dissemination graph and continuously optimizing the dissemination path prediction, so that the final generated information dissemination graph is closer to the actual dissemination situation. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0058] Figure 1 A schematic flow chart of a method for generating an information diffusion graph based on user similarity provided by the present invention;

[0059] Figure 2 A schematic diagram of a discrete forward process provided by the present invention;

[0060] Figure 3 A schematic diagram of a computer device for implementing a method for generating an information diffusion graph based on user similarity provided by the present invention. DETAILED DESCRIPTION

[0061] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present invention and corresponding drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0062] The technical solutions provided by various embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0063] Figure 1 The following is a flow chart of a method for generating an information diffusion graph based on user similarity in the present invention, which specifically includes the following steps:

[0064] S101, generate a candidate user set of the information diffusion graph as needed and determine a standardized feature matrix; the candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features of multiple dimensions; each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension.

[0065] The candidate user set can be used Indicates that, Represents the candidate user set and candidate feature set Represents user attribute information, including description, blue badge certification status, location, registration date, number of posts, number of fans, and number of followings.

[0066] Optionally, a candidate user set of the information diffusion graph is generated as needed, and a standardized feature matrix is ​​determined, including: for the features of any dimension, the features of all candidate users in that dimension are normalized to obtain the normalized features of each user in that dimension; based on the normalized features of all users in each dimension, a standardized feature matrix is ​​constructed; each element in the standardized feature matrix is ​​the normalized feature of the corresponding candidate user in the corresponding dimension.

[0067] The Min-Max normalization method is used to normalize the features of each dimension in the user attribute information and output a standardized feature matrix to ensure that each feature is in the range of [0, 1].

[0068] S102: Determine the importance ranking of the candidate users based on the standardized feature matrix and the communication type label of each candidate user.

[0069] Optionally, the importance ranking of candidate users is determined based on the standardized feature matrix and the communication type label of each candidate user, including: determining the importance score of the features of each dimension based on the standardized feature matrix and the communication type label of each candidate user; determining the importance ranking of candidate users based on the importance score of the features of each dimension and the standardized feature matrix; the feature value with a larger importance score corresponds to a higher ranking of the candidate user.

[0070] In one embodiment, the importance score of the features of each dimension is determined based on the standardized feature matrix and the communication type label of each candidate user, including: for the features of any dimension, according to the standardized feature matrix and the communication type label of each candidate user, calculating the chi-square statistic of the features of each dimension; and determining the chi-square statistic of the features of each dimension as the importance score of the features of each dimension.

[0071] The calculation formula of the chi-square statistic is:

[0072] (1);

[0073] in, Indicates the j The chi-square statistic of the dimension feature, Represents the first i The first candidate user j The normalized features of the dimension, 、 Represents the candidate users who participate in non-rumor spreading and rumor spreading respectively j The average value of the dimensionally normalized features.

[0074] The importance score of the feature of each dimension is calculated according to formula (1).

[0075] For each candidate user, the feature values ​​of the corresponding dimensions of each candidate user are compared in turn according to the ranked feature importance (obtained from the standardized feature matrix). Specifically, for example, taking the candidate user as an example with 7 dimensional features, the feature importance scores of each dimension are ranked from large to small as follows: 、 、 、 、 、 and , first compare The corresponding eigenvalue will have the maximum The candidate users with the smaller value are ranked first, and so on. The candidate users with the same value are ranked in the back order; if there are multiple candidate users If the values ​​are the same, compare and so on, until Through this layer-by-layer comparison, each candidate user is finally i Generate a unique one-hot encoding representation based on feature ranking .

[0076] For example, if there are a total of 5 sorted candidate users, then the one-hot encoding representations of these 5 candidate users are 10000, 01000, 00100, 00010, and 00001 respectively.

[0077] S103: Calculate the user similarity matrix based on the standardized feature matrix.

[0078] Optionally, a user similarity matrix is ​​calculated based on the standardized feature matrix, including: updating the candidate user set based on the standardized feature matrix, and inputting the updated candidate user set into a graph variational autoencoder to obtain a user potential representation of each candidate user; calculating the similarity between any two candidate users among all candidate users based on the user potential representation of each candidate user; and constructing a user similarity matrix based on the similarity between any two candidate users among all candidate users.

[0079] User potential representation , where the graph variational autoencoder is , is the standardized feature matrix, k is the hidden layer dimension; the user potential representations of all candidate users constitute the latent space.

[0080] The user similarity matrix is ​​obtained by calculating the latent space through the attention mechanism .

[0081] in, The calculation formula of the elements in can be expressed as: , is the attention mechanism, It is a splicing operation. It is i candidate users and j The similarity between candidate users; For the i The user potential representation of candidate users, Indicates the j The user potential representation of each candidate user.

[0082] S104, determining the number of candidate users as the diffusion step length T , and obtain the initial information propagation graph; the initial information propagation graph includes the candidate user nodes of all candidate users.

[0083] According to the number of candidate users Determine the diffusion step size , and initialize to generate the initial information propagation graph , as the initial probability distribution for back diffusion.

[0084] S105, from the number of steps T -1 until the number of steps is 1, reverse iteration is performed to generate the edges connecting the candidate users in the initial information propagation graph, and the initial information propagation graph after iteration is determined as the information propagation graph of the candidate user set.

[0085] Among them, the number of steps is T-1 until the number of steps is 1, and the edges of the candidate users in the initial information propagation graph are generated by reverse iteration, specifically including: in each iteration process, obtaining the target candidate user node corresponding to the current step number in the importance ranking, inputting the user similarity matrix of the target candidate user node and the initial information propagation graph into the diffusion model, and predicting the edges between the target candidate user node and other candidate user nodes in the initial information propagation graph; integrating the predicted edges into the initial information propagation graph, and based on the new initial information propagation graph, continuing to iteratively generate the edges of the candidate users in the initial information propagation graph until the iteration is completed.

[0086] The candidate user ranked first in the importance ranking corresponds to step 1, and so on. The candidate user ranked last in the importance ranking corresponds to step 2. T step.

[0087] Specifically, for the t Step 1: Select the importance order Corresponding candidate user node ; The candidate user node Similarity matrix and Initial information propagation diagram of the first step Input the trained diffusion model to predict candidate user nodes The initial information diffusion diagram of the previous step Which nodes in has an edge? ， in, are the parameters of the denoising module in the trained diffusion model, is the predicted edge; and candidate user nodes Integrate into graph Get the first Step diagram After each step is completed, the final social communication graph is output. .

[0088] In one embodiment, the process of constructing the diffusion model includes the following steps:

[0089] S201, obtaining multiple historical communication graphs; the historical communication graphs include multiple historical users, communication paths between historical users, and user attribute information of historical users.

[0090] Historical Communication Map ,in, Represents a user collection; Indicates that from historical users To historical users The propagation path of information from historical users Flow to historical users ; Feature set Represents user-specific user attribute information, including description, blue badge verification status, location, registration date, number of posts, number of fans, and number of followings.

[0091] S202: Calculate the importance ranking of historical users and the historical user similarity matrix based on the historical communication graph.

[0092] The basic method for calculating the importance ranking of historical users and the normalized feature matrix in this embodiment is the same as that for calculating the importance ranking of candidate users and the historical user similarity matrix in the above embodiment, and will not be repeated in this embodiment.

[0093] Update the historical propagation graph based on the standardized feature matrix of historical users , the updated history propagation graph is input into the graph variational autoencoder (GVAE) to encode the user potential representation Z: .

[0094] The user similarity matrix is ​​obtained by calculating the latent space through the attention mechanism : ,in, is the attention mechanism, It is a splicing operation. is the similarity between historical users i and j.

[0095] The attention mechanism is implemented as a learnable vector with a dimension twice that of z; and After splicing, The dimension of is the same as a, then through the dot multiplication operation, that is, a and After multiplying the numbers in the corresponding positions and then summing them up, we finally get and The similarity between .

[0096] S203 , performing forward iterative training on the initial diffusion model according to the historical diffusion graph, the importance ranking of historical users, and the historical user similarity matrix to obtain a diffusion model.

[0097] Optionally, the initial diffusion model is forward iteratively trained according to the historical propagation graph, the importance ranking of historical users and the historical user similarity matrix to obtain the diffusion model, including: determining the number of historical users as the historical diffusion step; starting from the step number 1 to the step number being the historical diffusion step, iteratively performing the absorption operation of the historical user nodes according to the historical propagation graph, the importance ranking of historical users and the historical user similarity matrix, and training the diffusion model in this process.

[0098] Figure 2 Schematic diagram of a discrete forward process; optionally, according to the historical propagation graph, the importance ranking of historical users, and the historical user similarity matrix, iteratively absorbs historical user nodes, and trains the diffusion model in this process, including the following steps:

[0099] S301, for the historical diffusion step number t Step 1: Select the historical user node with the least importance in the importance ranking .

[0100] S302, record in the current historical propagation map Middle and historical user nodes Connected node labels .

[0101] S303, historical user nodes Convert it to a masked state and connect it with all historical user nodes in the current historical propagation graph with masked edges to generate an intermediate graph with masked edges :

[0102] (2);

[0103] in, Absorb Represents the operation of converting the historical user node to the absorbing state.

[0104] Historical user nodes Convert to masked state, that is, historical user nodes Switch to absorption state.

[0105] S304, the middle image Similarity matrix with historical users Input the denoising module of the diffusion model to predict historical user nodes and Which nodes in the have ever had edges connected to them?

[0106] (3);

[0107] in, is the trainable parameter of the denoising module, is the predicted label.

[0108] S305, through node labels and predicted labels Calculate the mean squared error (MSE) loss and optimize it by gradient descent .

[0109] (4);

[0110] in, is the loss function of the denoising module.

[0111] In one embodiment, the present application further provides a method for generating an information diffusion graph based on user similarity, comprising the following steps:

[0112] S401, input several historical propagation graphs , the importance of the normalized historical user attributes is ranked I through the chi-square test, and the historical user similarity matrix P is calculated.

[0113] S402, in the discrete forward process, according to the historical propagation graph G , node importance ranking I, user similarity matrix P trains the parameters of the denoising module in the diffusion model.

[0114] S403: Generate candidate user sets for the propagation graph as needed Calculate the importance ranking of candidate users and candidate user similarity matrix .

[0115] S404, in the reverse process, the candidate user similarity matrix As a starting point, the diffusion model trained based on the candidate user set is used to iteratively predict the connection probability between users, and the final generated social communication graph is output. .

[0116] The social graph generation method based on user similarity provided by the present invention constructs historical user attributes according to several historical communication graphs for importance ranking and historical user similarity matrix. Subsequently, in the discrete forward process, the noise is gradually absorbed according to the order of node influence, and converted into a node-level edge probability classification task to train the denoising module in the diffusion model. Then, the candidate user node features of the communication graph are generated as needed to calculate the candidate importance ranking and similarity matrix. Finally, in the reverse process, the user similarity matrix is ​​used as the starting point to generate a social communication graph through iterative denoising. On the one hand, this method overcomes the discrete graph structure modeling deviation caused by the continuous noise assumption of the traditional diffusion model. On the other hand, by correcting the generation process by user similarity prior, a social communication graph that is more in line with the actual situation is generated, which can better serve social network downstream tasks such as communication source positioning that require a large number of communication graphs for model training.

[0117] The beneficial effects of the present invention are embodied in the following aspects:

[0118] (1) This paper introduces a discrete forward diffusion process. Through the node-level edge probability classification task and the absorption state mechanism, it solves the problem that the traditional continuous diffusion model cannot effectively model the discrete structure of the social graph, and significantly improves the topological rationality of the generated graph and the ability to restore the true propagation pattern.

[0119] (2) This paper introduces the user similarity matrix as the starting point of back diffusion and combines it with social context priors (such as user attributes and propagation labels) to eliminate the distribution bias caused by traditional Gaussian noise initialization, making the generated graph more consistent with the data distribution of actual social networks.

[0120] (3) This paper uses a lightweight graph attention network as a denoising module to achieve efficient edge probability prediction and denoising iteration, reducing computational complexity while ensuring generation quality, and supporting the rapid generation of large-scale social graphs.

[0121] Table 1 shows the performance comparison results of the proposed method on three real datasets and other methods. Table 1 shows the generation performance comparison of the proposed method on Twitter15, Twitter16 and Weibo datasets. Experimental results show that the proposed method (PDDM) significantly outperforms existing methods in both degree distribution (Deg.) and graph edit distance (Dis.) MMD indicators, with an average reduction of 15.60% compared to the optimal baseline. The improved generation ability of PDDM can be attributed to three main factors: (1) The introduced discrete forward diffusion process enhances the learning ability to capture discrete structural features from the perspective of edge connection probability, which provides a suitable strategy for topology generation. (2) The use of a new reverse starting point mitigates the random Gaussian noise bias and uses user similarity as a probability matrix to ensure more accurate structure generation with social context. (3) The application of the back-diffusion process in the probability distribution space not only corrects the coarse-grained generation, but also introduces robustness to enhance the diversity of the generated graph.

[0122] Table 1

[0123]

[0124] Table 2 shows how the propagation data generated by our method enhances model performance in downstream tasks (such as localization of the source of the spread). Four sets of data were used in the experiments to ensure the rigor of the study. The original set trained the localization model on 90% of the Twitter propagation data and tested it on the remaining 10%. The augmented set generated 1,000 additional propagation graphs using our method (PDDM) and comparison methods (DAVA and GRAPHARM), adding these graphs to the training set. The control set simulated 1,000 snapshots using traditional SI, SIR, IC, and LT models, adding these simulated data to the training set. As shown in the table, training with simulated data from traditional models leads to reduced performance in downstream tasks related to real-world propagation scenarios, indicating the limited applicability of these models to practical tasks. In contrast, augmentation with real generated data leads to improved performance, with the best results observed when using propagation data generated by PDDM. This highlights the importance of realistic graph generation and underscores the effectiveness of PDDM in improving model performance on downstream tasks in social networks.

[0125] Table 2

[0126]

[0127] In summary, the information propagation graph generation method based on user similarity provided by the present invention redefines the forward process as a discrete process of node denoising and edge generation, and transforms the task of the denoising module into connection probability learning of node-level tasks. In addition, a new starting point for the reverse process is adopted, which can better utilize the social context by incorporating user similarity as a probability matrix. These developments reduce the reverse starting point bias and enhance the robustness of the model. Compared with existing methods, better graph generation effects are achieved, and the generated propagation graph can better improve the performance of the model in downstream tasks, which makes the social graph generation method based on user similarity of the present invention have practical guiding significance.

[0128] When applying the information diffusion graph generation method based on user similarity provided by the present invention, it is not necessary to Figure 1 The steps are executed in the order shown. The specific execution order of the steps can be determined according to needs, and the present invention does not limit this.

[0129] The above is a method for generating an information diffusion graph based on user similarity provided by one or more embodiments of the present invention. Based on the same idea, the present invention also provides a corresponding device for generating an information diffusion graph based on user similarity, which includes:

[0130] An acquisition module is used to generate a candidate user set for the information diffusion graph as needed and determine a standardized feature matrix; the candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features in multiple dimensions; each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension;

[0131] A ranking module is used to determine the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user;

[0132] A calculation module, used to calculate a user similarity matrix based on the standardized feature matrix;

[0133] The first determination module is used to determine the number of candidate users as the diffusion step length T , and obtain the initial information propagation graph; the initial information propagation graph includes candidate user nodes of all candidate users;

[0134] Iteration module, used to start from the number of steps T -1 until the number of steps reaches 1, reverse iteration is performed to generate the edges connecting the candidate users in the initial information propagation graph;

[0135] The second determining module is used to determine the initial information propagation graph completed by iteration as the information propagation graph of the candidate user set.

[0136] For the specific definition of the information diffusion graph generation device based on user similarity, please refer to the definition of the information diffusion graph generation method based on user similarity above, which will not be repeated here. The various modules in the above-mentioned information diffusion graph generation device based on user similarity can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0137] The present invention also provides a computer-readable storage medium, which stores a computer program, which can be used to execute the above Figure 1 The proposed method for generating information diffusion graph based on user similarity.

[0138] The present invention also provides Figure 3 The structural diagram of the computer equipment shown in FIG. Figure 3 As shown in the figure, at the hardware level, the computer device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory. Of course, it may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The proposed method for generating information diffusion graph based on user similarity.

[0139] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware using a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes in the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0140] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of the present invention.

Claims

1. A method for generating an information diffusion graph based on user similarity, characterized in that: include: Generate a candidate user set of the information diffusion graph as needed and determine the standardized feature matrix; The candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features of multiple dimensions; Each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension; Determine the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user; Calculate the user similarity matrix based on the standardized feature matrix; The number of candidate users is determined as the diffusion step length T, and an initial information propagation graph is obtained; the initial information propagation graph includes candidate user nodes of all candidate users; Starting from step T-1 to step 1, reversely iterate to generate the edges connecting the candidate users in the initial information propagation graph, and determine the initial information propagation graph after iteration as the information propagation graph of the candidate user set; Starting from step number T-1 until step number 1, reverse iteratively generate the edges of candidate users in the initial information diffusion graph. Specifically, in each iteration, obtain the target candidate user node corresponding to the current step in the importance ranking, input the user similarity matrix of the target candidate user node and the initial information diffusion graph into the diffusion model, and predict the edges between the target candidate user node and other candidate user nodes in the initial information diffusion graph; integrate the predicted edges into the initial information diffusion graph, and continue to iteratively generate the edges of candidate users in the initial information diffusion graph based on the new initial information diffusion graph until the iteration is completed; The diffusion model construction process includes: obtaining multiple historical diffusion graphs; the historical diffusion graphs include multiple historical users, the diffusion paths between historical users, and the user attribute information of historical users; calculating the importance ranking of historical users and the historical user similarity matrix based on the historical diffusion graphs; determining the number of historical users as the historical diffusion step; starting from step 1 to step 1 equal to the historical diffusion step, iteratively absorbing historical user nodes based on the historical diffusion graphs, the importance ranking of historical users, and the historical user similarity matrix, and training the diffusion model in this process; Based on the historical propagation graph, historical user importance ranking, and historical user similarity matrix, the historical user node absorption operation is iteratively performed, and the diffusion model is trained in this process, including: For the t-th step of historical diffusion, select the historical user node v with the least importance in the importance sorting t ; Recorded in the current historical propagation graph G = (V, E, F) and the historical user node v t Connected node label Y t ; The historical user node v t Convert it to a masked state and connect it with all historical user nodes in the current historical propagation graph with masked edges to generate an intermediate graph G with masked edges t : G t =Absorb(G,v t ); Among them, Absorb represents the operation of converting the historical user node to the absorbing state; The middle graph G t And the historical user similarity matrix P is input into the denoising module of the diffusion model to predict the historical user node v t With G t Which nodes in the have ever had edges connected to them? Among them, θ is the trainable parameter of the denoising module, is the predicted label; By node label Y t and predicted labels Calculate the mean squared error loss and optimize θ via gradient descent.

2. The method according to claim 1, characterized in that Generate a candidate user set for the information diffusion graph as needed and determine the standardized feature matrix, including: For the features of any dimension, normalize the features of all candidate users in that dimension to obtain the normalized features of each user in that dimension; A standardized feature matrix is ​​constructed based on the normalized features of all users in each dimension; each element in the standardized feature matrix is ​​the normalized feature of the corresponding candidate user in the corresponding dimension.

3. The method according to claim 1, characterized in that Based on the standardized feature matrix and the communication type label of each candidate user, the importance ranking of the candidate users is determined, including: Determine the importance score of each dimension’s features based on the standardized feature matrix and the communication type label of each candidate user; The importance ranking of candidate users is determined based on the importance score of the features of each dimension and the standardized feature matrix; the candidate users with higher importance scores are ranked higher.

4. The method according to claim 3, characterized in that Based on the standardized feature matrix and the communication type label of each candidate user, the importance score of the features in each dimension is determined, including: For the features of any dimension, the chi-square statistic of the features of each dimension is calculated based on the standardized feature matrix and the communication type label of each candidate user; The chi-square statistic of the features of each dimension is determined as the importance score of the features of each dimension.

5. The method according to claim 4, characterized in that The formula for calculating the chi-square statistic is: in, represents the chi-square statistic of the feature of the j-th dimension, f' ij represents the normalized feature of the jth dimension of the i-th candidate user in the normalized feature matrix, μ j0 、μ j1 They represent the average values ​​of the j-dimensional normalized features of candidate users who are not involved in rumor spreading and those who are involved in rumor spreading, respectively.

6. The method according to claim 1, wherein According to the standardized feature matrix, the user similarity matrix is ​​calculated, including: Based on the standardized feature matrix, the candidate user set is updated and the updated candidate user set is input into the graph variational autoencoder to obtain the user potential representation of each candidate user; Based on the potential representation of each candidate user, calculate the similarity between any two candidate users among all candidate users; Based on the similarity between any two candidate users among all candidate users, a user similarity matrix is ​​constructed.

7. A device for generating an information diffusion graph based on user similarity, characterized in that: include: An acquisition module is used to generate a candidate user set of the information diffusion graph as needed and determine a standardized feature matrix; The candidate user set includes multiple candidate users and user attribute information corresponding to each candidate user; the user attribute information includes features of multiple dimensions; Each element in the standardized feature matrix is ​​the feature of the corresponding candidate user in the corresponding dimension; A ranking module is used to determine the importance ranking of candidate users based on the standardized feature matrix and the communication type label of each candidate user; A calculation module, used to calculate a user similarity matrix based on the standardized feature matrix; A first determination module is configured to determine the number of candidate users as a diffusion step length T and obtain an initial information propagation graph; the initial information propagation graph includes candidate user nodes of all candidate users; Iteration module, which is used to iterate backward from step number T-1 to step number 1 to generate the edges connecting candidate users in the initial information propagation graph; Starting from step number T-1 until step number 1, reverse iteratively generate the edges of candidate users in the initial information diffusion graph. Specifically, in each iteration, obtain the target candidate user node corresponding to the current step in the importance ranking, input the user similarity matrix of the target candidate user node and the initial information diffusion graph into the diffusion model, and predict the edges between the target candidate user node and other candidate user nodes in the initial information diffusion graph; integrate the predicted edges into the initial information diffusion graph, and continue to iteratively generate the edges of candidate users in the initial information diffusion graph based on the new initial information diffusion graph until the iteration is completed; The diffusion model construction process includes: obtaining multiple historical diffusion graphs; the historical diffusion graphs include multiple historical users, the diffusion paths between historical users, and the user attribute information of historical users; calculating the importance ranking of historical users and the historical user similarity matrix based on the historical diffusion graphs; determining the number of historical users as the historical diffusion step; starting from step 1 to step 1 equal to the historical diffusion step, iteratively absorbing historical user nodes based on the historical diffusion graphs, the importance ranking of historical users, and the historical user similarity matrix, and training the diffusion model in this process; Based on the historical propagation graph, historical user importance ranking, and historical user similarity matrix, the historical user node absorption operation is iteratively performed, and the diffusion model is trained in this process, including: For the t-th step of historical diffusion, select the historical user node v with the least importance in the importance sorting t ; Recorded in the current historical propagation graph G = (V, E, F) and the historical user node v t Connected node label Y t ; The historical user node v t Convert it to a masked state and connect it with all historical user nodes in the current historical propagation graph with masked edges to generate an intermediate graph G with masked edges t : G t =Absorb(G,v t ); Among them, Absorb represents the operation of converting the historical user node to the absorbing state; The middle graph G t And the historical user similarity matrix P is input into the denoising module of the diffusion model to predict the historical user node v t With G t Which nodes in the have ever had edges connected to them? Among them, θ is the trainable parameter of the denoising module, is the predicted label; By node label Y t and predicted labels Calculate the mean squared error loss and optimize θ via gradient descent; The second determining module is used to determine the initial information propagation graph completed by iteration as the information propagation graph of the candidate user set.

Images