Information propagation source positioning method, device, equipment and medium

Abstract

The invention relates to the technical field of network security, and discloses an information propagation source positioning method and device, equipment and a medium. The method comprises the following steps: acquiring a social network relationship of information dissemination; wherein the social network relationship is composed of a plurality of user nodes and a plurality of connecting edges representing that any two user nodes have propagation interaction behaviors in the social network, and each user node has a corresponding portrait feature; establishing propagation cascading according to the social network relationship; obtaining dynamic propagation characteristics of each user node in the propagation cascade; analyzing whether the user comment of each user node is a propagation source or not and whether the user comment is a cause of the propagation source or not through a large language model; and the propagation source of information propagation is positioned by combining the portrait feature of each user node, the dynamic propagation feature, whether the user comments are the propagation source or not and the reason of whether the user comments are the propagation source or not, so that the accuracy is higher.

Description

Technical Field

[0001] The present invention relates to the field of network security technology, and particularly relates to a method, device, equipment and medium for locating information dissemination sources. Background Art

[0002] With the rapid development of the Internet and social media, profound changes have taken place in the production and lifestyle of human beings, and social media has gradually become the main platform for users to obtain and share information. The complex social network relationships established by social media make information dissemination fast, highly interactive and widely covered. While this convenient information dissemination method promotes economic and social progress, it is also accompanied by certain potential risks. For example, the spread of harmful information or network viruses may lead to the leakage of user privacy and property losses, pose a threat to the network security environment, and even endanger social stability and national security. Therefore, researching how to effectively locate the dissemination source is of great significance for suppressing the spread of harmful information, protecting public interests and social order.

[0003] In the field of dissemination source location, the widely used methods are graph algorithms and graph neural network algorithms, and usually the propagation snapshot cascade model is adopted without including text information. For example, the label propagation technology based on source saliency is used to locate the dissemination source; the model based on Graph Convolutional Networks (GCN) is used to locate multiple dissemination sources; the dynamic characteristics of propagation are considered before performing source inference. Although these methods have initially proved their effectiveness, with the rapid development of the Internet, more and more text comments are generated by social media users during the dissemination process. These text comments are not only the direct carriers for users to express their opinions and emotions, but also may become an important way for the spread and diffusion of rumors. Facing the increasing quantity and importance of text comments, the limitation that traditional methods cannot make full use of text information becomes more and more obvious. Therefore, how to effectively tap the potential of text comments has become a key issue.

[0004] With the continuous progress of artificial intelligence large model technology, large language models (LLMs) have demonstrated powerful language understanding and generation capabilities in the field of text analysis, bringing new possibilities for solving complex tasks. For example, by analyzing the language style and content features of rumors, more context-appropriate rumor texts can be generated. However, when directly applying LLMs to the dissemination source location task, due to their insufficient understanding of the complex dissemination patterns and network structures in the dissemination cascade, the accuracy of the location results is limited. Summary of the Invention

[0005] The object of the present invention is to provide a method, device, equipment and medium for information dissemination source localization, which can solve the technical problem that when applying the LLM to the dissemination source localization task, the accuracy of the localization result is limited due to the lack of understanding of the complex dissemination patterns and network structures in the cascaded data.

[0006] To solve the above technical problem, an embodiment of the present invention provides a method for information dissemination source localization, including the following steps: Obtain the social network relationship of information dissemination; wherein, the social network relationship is composed of multiple user nodes and multiple edges representing the presence of dissemination interaction behaviors between any two user nodes in the social network, and each user node has corresponding portrait features; Establish a dissemination cascade based on multiple target user nodes participating in information dissemination in the social network relationship, edges representing the dissemination interaction behaviors between the multiple target user nodes, neighbor nodes of the target user nodes, and adjacent edges of each target user node; For each target user node in the dissemination cascade, obtain the proportion of neighbor nodes participating in information dissemination among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information dissemination to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, and the proportion of neighbor nodes not participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, as the dynamic dissemination feature of the target user node; Analyze whether the user comment of each target user node is a dissemination source and the reason for being a dissemination source through a large language model; Locate the dissemination source of information dissemination by combining the portrait features, dynamic dissemination features, whether the user comment is a dissemination source, and the reason for the user comment being a dissemination source of each target user node.

[0007] Optionally, the step of locating the dissemination source of information dissemination by combining the portrait features, dynamic dissemination features, whether the user comment is a dissemination source, and the reason for the user comment being a dissemination source of each target user node includes: Perform normalization processing on the portrait features and dynamic dissemination features of each target user node and then splice them; Use a graph convolutional network based on a self-loop attention mechanism to process the spliced portrait features and dynamic dissemination features to generate a dissemination feature representation of each target user node; Locate the dissemination source of information dissemination by combining the dissemination feature representation of each target user node, whether the user comment is a dissemination source, and the reason for the user comment being a dissemination source; Among them, the graph convolutional network based on the self-loop attention mechanism processes the spliced portrait features and dynamic propagation features through the following formula: ; ; ; In the formula, is the propagation feature representation, represents the activation function of the graph convolutional network, is through propagation cascade , is the identity matrix, is corresponding degree matrix, is the preset first weight matrix, is the spliced portrait features and dynamic propagation features, is a diagonal matrix, and , is a real matrix with dimension , is the total number of user nodes in the propagation cascade; is the diagonal matrix constructor, is the th user node self-loop attention coefficient, represents neighbor nodes of, is the softmax normalization result of the self-loop attention coefficient of the th user node, is a single-layer backpropagation BP neural network applied to the head of each attention mechanism, is a non-linear activation function, is the exponential function, is the preset second weight matrix, is the th propagation feature representation.

[0008] Optionally, the source of information propagation is located by combining the propagation feature representation of each target user node, whether the user comment is a propagation source, and the reason why the user comment is a propagation source, including: Through the trained BERT model, the text content of the large language model analyzing whether the user comment is a propagation source and the reason for being a propagation source is converted into a reason feature representation; Combining the propagation feature representation of each target user node and the reason feature representation, the source of information propagation is located; Among them, the enhanced reason feature representation is obtained through the following formula: ; In the formula, is the enhanced cause feature representation, is the same batch of positive sample pairs belonging to the same propagation cascade as , represents the cause feature representation of negative samples from non - propagation cascades, represents the random sampling operation, is the cosine similarity evaluation function, represents the contrast learning process.

[0009] Optionally, the combining of the propagation feature representation and the cause feature representation of each target user node to locate the propagation source of information dissemination includes: Using the differentiable masking technique to filter out the invalid cause feature representations in the enhanced cause feature representation to obtain the effective cause feature representation; Combining the propagation feature representation of each target user node and the effective cause feature representation to locate the propagation source of information dissemination; Among them, the differentiable masking technique filters out the invalid cause feature representations in the enhanced cause feature representation through the following formula: ; In the formula, is the effective cause feature representation, represents element - wise multiplication, is the preset attenuation coefficient, is the differential approximation result of discrete feature selection, represents the enhanced cause feature representation, represents the f - th feature of user v, represents the corresponding feature whether it is masked, represents the masking feature , represents the retained feature ; The differential approximation result of discrete feature selection is obtained based on the following feature masking method of Gumbel Softmax: ; ; In the formula, is the differentiable approximation method for discrete feature selection, represents linearly transforming the cause feature representation of each user comment into a generation logarithm through a binary classification decision network, is used to control the continuity degree of the differential approximation result, and respectively represent real - number matrices of corresponding dimensions, To ensure that the output is a one-hot vector during the forward propagation process and to retain differentiability during the backward propagation process.

[0010] Optionally, combining the propagation feature representation and the effective cause feature representation of each target user node to locate the propagation source of information dissemination, including: Adopting a cross-modal attention mechanism to fuse the propagation feature representation and the effective cause feature representation; Locating the propagation source of information dissemination through the fused propagation feature representation and effective cause feature representation corresponding to each target user node; Among them, the cross-modal attention mechanism fuses the propagation feature representation and the effective cause feature representation through the following formula: In the formula, is the propagation feature representation, is the cause feature representation filtered by the differentiable masking technique, represents the propagation feature representation after being processed by the cross-modal attention mechanism, represents the cause feature representation after being processed by the cross-modal attention mechanism, represents the concatenated propagation feature representation and cause feature representation, represents that the softmax function normalizes the multi-layer perceptron (MLP) classification scores into probability values, represents the probability of being predicted as a non-propagation source, and respectively represent the query matrices applied to and , and respectively represent the key matrices applied to and , and respectively represent the value matrices applied to and , represents the feature dimension.

[0011] An embodiment of the present invention also provides an information dissemination source location device, including: A social network acquisition module, configured to acquire the social network relationship of information dissemination; wherein, the social network relationship is composed of multiple user nodes and multiple edges representing that any two user nodes have a propagation interaction behavior in the social network, and each user node has a corresponding portrait feature; A propagation cascade establishment module, which is used to establish a propagation cascade according to multiple target user nodes participating in information propagation in the social network relationship, edges representing the propagation interaction behaviors between the multiple target user nodes, neighbor nodes of the target user nodes, and adjacent edges of each target user node; A propagation feature acquisition module, which is used to, for each target user node in the propagation cascade, obtain the proportion of neighbor nodes participating in information propagation among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information propagation to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information propagation to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, and the proportion of neighbor nodes not participating in information propagation to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, as the dynamic propagation features of the target user node; A user comment analysis module, which is used to analyze whether the user comment of each target user node is a propagation source and the reason for being a propagation source through a large language model; A propagation source localization module, which is used to localize the propagation source of information propagation by combining the portrait features, dynamic propagation features, whether the user comment is a propagation source, and the reason for the user comment being a propagation source of each target user node.

[0012] An embodiment of the present invention also provides a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, instructions executable by the at least one processor are stored in the memory, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute the above information propagation source localization method.

[0013] An embodiment of the present invention also provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the above information propagation source localization method is implemented.

[0014] The information propagation source localization method provided by the present invention has at least the following beneficial effects: The present invention locates the source of information dissemination by combining the portrait features, dynamic propagation features of user nodes, whether the user comments are the source of dissemination, and the reasons why the user comments are the source of dissemination, that is, multi-modal fusion features. Among them, the portrait features of user nodes can provide prior knowledge for understanding the propagation cascade structure. By analyzing whether the user comments of each user node are the source of dissemination and the reasons therefor, the large language model can perform semantic traceability analysis on the user comments, obtain the dissemination intention, sentiment tendency, and semantic relevance in the comment content, capture the potential diffusion motivation in the propagation cascade, and fuse it with the dynamic propagation features reflected by the propagation cascade, which can make up for the deficiency of the LLM in understanding the network topology structure and improve the accuracy of information dissemination source location. At the same time, user neighbor nodes are considered in the information dissemination cascade, realizing the expansion of the propagation cascade, which can more comprehensively capture the information dissemination path and further improve the accuracy of information dissemination source location. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] One or more embodiments are exemplarily illustrated by the pictures in the corresponding drawings. These exemplary illustrations do not constitute a limitation on the embodiments.

[0016] Figure 1 is a specific flowchart of a method for locating the source of information dissemination provided according to an embodiment of the present invention; Figure 2 is a schematic diagram of the principle of a method for locating the source of information dissemination provided according to an embodiment of the present invention; Figure 3 is a chart of a real social network dataset provided according to an embodiment of the present invention; Figure 4 is a comparison chart of F1-scores between the present invention and various source location methods on a real social network dataset provided according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. However, those of ordinary skill in the art can understand that many technical details are proposed in the embodiments of the present invention to help readers better understand the present invention. However, even without these technical details and various changes and modifications based on the following embodiments, the technical solutions required to be protected by the present invention can still be implemented. The following division of each embodiment is for convenience of description and should not constitute any limitation on the specific implementation manner of the present invention. Each embodiment can be combined and cross-referenced with each other on the premise of no contradiction.

[0018] One embodiment of the present invention relates to a method for locating an information dissemination source. The implementation details of the information dissemination source location method in this embodiment will be specifically described below. The following content is only the implementation details provided for convenient understanding and is not necessary for implementing this solution.

[0019] For the specific process and principle of the information dissemination source location method in this embodiment, reference can be made to Figure 1 and Figure 2 , including: Step 101: Obtain the social network relationships of information dissemination. Among them, the social network relationships are composed of multiple user nodes and multiple edges representing the existence of dissemination interaction behaviors between any two user nodes in the social network, and each user node has corresponding portrait features.

[0020] Specifically, let the social network relationship of information dissemination be ; in the formula, is the set of user nodes of the social network relationship , which is composed of multiple user nodes; is the set of edges of the social network relationship , and the edge represents the existence of dissemination interaction behaviors (such as comments or forwards) between any two user nodes in the social network; is the set of user portrait features of the social network relationship , which is composed of the portrait features of each user node, and the portrait feature of each user node contains the following several dimensions of information, that is, the verification status of the user (two-dimensional variable) , the number of posts or tweets of the user , the registration date , the number of fans , the number of follows , and the ratio of the number of fans to the number of follows .

[0021] Among them, The data comes from the public datasets Twitter15, Twitter16, and Weibo. Due to the lack of user portraits and comment information, in the implementation process of the present invention, a crawler tool is used to complete the user portraits and comment content, and a script is used to obtain the analysis results of the comment content through the API services of GPT-4o and GPT-4.

[0022] Step 102: Establish a propagation cascade based on multiple target user nodes participating in information dissemination in the social network relationship, the edges representing the dissemination interaction behaviors between the multiple target user nodes, the neighbor nodes of the target user nodes, and the adjacent edges of each target user node.

[0023] Specifically, observe the social network relationship at a certain moment to obtain a preliminary propagation cascade , and record the set of propagation sources , according to the propagation cascade of neighbor nodes, and expand it into a propagation subgraph , this propagation subgraph is the final propagation cascade of this embodiment.

[0024] Among them, the expansion and construction process of the propagation cascade is as follows; ; ; In the formula, is the set of all neighbor nodes of a certain user node, is the expansion of the neighbor nodes of the node snapshot (i.e., the user node) , is the expansion of the adjacent edges of all nodes in the edge set , indicates whether the user participates in the propagation cascade , indicates the participation of the th user, indicates the non - participation of the th user.

[0025] Step 103, for each target user node in the propagation cascade, obtain the proportion of neighbor nodes participating in information propagation among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information propagation among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information propagation among all neighbor nodes of the target user node with the largest total number of neighbor nodes, and the proportion of neighbor nodes not participating in information propagation among all neighbor nodes of the target user node with the largest total number of neighbor nodes, as the dynamic propagation characteristics of the target user node.

[0026] Specifically, construct the dynamic propagation feature representation , and obtain a * 7 matrix, where is the total number of users after expanding the neighbor nodes of the node snapshot .

[0027] The construction method of the dynamic propagation feature representation is as follows: ; In the formula, It is the propagator graph, denotes a certain neighbor user of denotes all neighbor users of is the operation of finding the number of elements in a set. is the ratio of the number of neighbors of the th user who participate in the propagation cascade to the total number of its neighbors, while is the ratio of the number of neighbors of the th user who do not participate in the propagation cascade to the total number of its neighbors. denotes the maximum value of the number of neighbors of all users. is the ratio of the number of neighbors of the th user who participate in the propagation cascade to the maximum value of the number of neighbors, while is the ratio of the number of neighbors of the th user who do not participate in the propagation cascade to the maximum value of the number of neighbors, denotes that the th user participates in the propagation cascade , denotes that the th user does not participate in the propagation cascade denotes the degree centrality of the

[0028] th user node. In one example, a graph convolutional network based on a self-loop attention mechanism is used to generate a propagation feature representation Figure 2 (see (a)), obtaining a

[0029] matrix of *14. The dynamic propagation feature representation of the th user is concatenated with the feature representation of the user portrait after normalization to obtain the propagation feature representation .

[0030] In this embodiment, a graph convolutional network based on a self-loop attention mechanism is used to generate a propagation feature representation : ; In the formula, is the sparse adjacency matrix constructed through the propagator graph , is the learnable weight matrix, , is the identity matrix, is the corresponding degree matrix, representing the activation function set by the GCN network. is a learnable diagonal matrix, is the total number of users after expanding the neighbor nodes of the node snapshot .

[0031] Among them, the learnable diagonal matrix is: ; In the formula, is the learnable weight matrix in the attention module, is the self-loop attention coefficient of the th user, is the softmax normalization result of the self-loop attention coefficient of the th user, is a single-layer BP neural network applied to the head of each attention mechanism, is a non-linear activation function, is the propagation feature representation of the th user, and the diagonal matrix is composed of the diagonalized and normalized self-loop attention coefficients.

[0032] Step 104, analyze whether the user comment of each target user node is a propagation source and the reason for being a propagation source through a large language model.

[0033] Specifically, use a large language model to analyze the reason why the user comment is a propagation source, and convert the analyzed text content into a feature representation using a pre-trained model , obtaining a *768 matrix.

[0034] Use the prompt engineering of the large language model to analyze whether the user comment is a propagation source and obtain the analyzed reason, and convert the analyzed text content into a feature representation using the pre-trained model BERT : In the formula, represents the reason analysis text of whether it is a propagation source obtained by analyzing the comments of each user through a large language model.

[0035] In an example, enhance the comment feature representation through a contrast learning mechanism (see Figure 2 (b)), obtaining a *768 matrix.

[0036] Adopt a contrastive learning mechanism between rumors and non-rumors, and enhance the comment feature representation by minimizing the distance of positive similarity pairs and maximizing the distance of negative similarity pairs. : In the formula, is the enhanced feature representation, is the same batch of positive sample pairs belonging to the rumor cascade as , represents the negative sample feature representation from the non-rumor cascade, represents the random sampling operation, is the cosine similarity evaluation function, represents the contrastive learning process. Through contrastive learning, the distinguishability between rumor and non-rumor comments can be enhanced, and then the feature representation of comments can be enhanced.

[0037] In one example, the feature representation after filtering invalid comment features through differentiable masking technology (see Figure 2 (c)), a *768 matrix is obtained; The differentiable masking technology filters invalid comment features to obtain the feature representation The process is as follows: In the formula, represents element-wise multiplication, is the attenuation coefficient, is the differential approximation result of discrete feature selection, is the feature representation enhanced through contrastive learning, represents the feature needs to be masked. Conversely, if represents the feature is retained, is the feature representation after filtering invalid comment features through differentiable masking technology.

[0038] Among them, for the differential approximation process of discrete feature selection, a feature masking method based on Gumbel Softmax is used to achieve differentiable masking of invalid features. The formula is as follows: ; Among them, is the method of differentiable approximation for discrete feature selection, is the differential approximation result, It is a binary classification decision network that linearly transforms the feature representation of each comment into a logit to determine whether each feature is masked. It is to control the continuity degree of the differential approximation result, It is used to ensure that the output is a one-hot vector during the forward propagation process and retain differentiability during the backward propagation process. It is the feature representation enhanced by contrastive learning. It is the node snapshot The total number of users after expanding the neighbor nodes of

[0039] Step 105: Combine the portrait features, dynamic propagation features of each target user node, and the reason whether the corresponding user comment is a propagation source to locate the information propagation source.

[0040] In the specific implementation, a cross-modal attention mechanism is used to fuse the feature representations of propagation and comments (see Figure 2 (d)), to obtain a *782 matrix.

[0041] Use a cross-modal attention mechanism to fuse the propagation feature representation and the comment feature representation : ; ; ; In the formula, is the query matrix applied to , is the key matrix applied to , is the value matrix applied to , represents the dimension of the feature. is obtained by concatenating two optimized feature representations. represents that the softmax function normalizes the MLP classification score into a probability value, with a dimension of *2, represents the probability of being predicted as a non-propagation source, represents the probability of being predicted as a propagation source.

[0042] The group of nodes with the highest output probability is the predicted propagation source (see Figure 2 (e)): ; In the formula, is the node snapshot of the propagation cascade of

[0043] The information dissemination source location method of the present invention has the following beneficial effects: First, the present invention uses an LLM to assist in guiding the source location process, analyzes whether the comments are the reasons for the dissemination source using the LLM, and uses a contrastive learning mechanism between the analysis results of rumor and non-rumor comments to enhance the feature representation of the text. At the same time, a differentiable masking technique is adopted to filter out invalid features, thereby improving the robustness and accuracy of the model, and solving the limitations of traditional methods that cannot fully utilize text information and the limited accuracy when directly applying the LLM to source location. In addition, the large model-assisted source location model used in the present invention is feature-driven rather than topology-driven, so in the scenario where the social network dissemination structure changes, it still has good generalization, ensuring that the trained location model can be applied to different social networks or dissemination models;

[0044] Second, based on the historical relationship network in the real-world scenario, the present invention constructs a new dissemination cascade dataset, including user portraits, comments, and the comment analysis results of the LLM, and proves the practicality of these dataset features using ablation studies. In addition, the present invention uses a self-loop attention mechanism to refine the aggregation strategy of each user, improving the quality of the dynamic dissemination and user portrait feature representation. Experiments show that the location accuracy of the present invention is better than other latest methods;

[0045] Third, the expected benefits and commercial value after the transformation of the technical solution of the present invention are: using the present invention to perform the source location task under network space security can play a key role in applications such as tracing the origin of social media rumors and tracking the source of network viruses; specifically, mining and tracking the originators of rumors or viruses, etc., is crucial for cutting off the dissemination path, timely stopping losses, killing harmful information, and then purifying the network environment and maintaining social stability.

[0046] The following describes the effect of the information dissemination source location method of the present invention in combination with the data and charts in the experimental process: Figure 3 Shows the scale of the real social network dataset used in the present invention, Figure 4 Is the comparison of F1-score between the present invention and various source location methods on the real social network dataset. Figure 4In both cases, GPT-4 outperforms other deep learning-based methods and exhibits the best source localization performance among all current state-of-the-art methods, highlighting the great potential of LLMs in the field of rumor source localization. Then, compared with the best baseline model, GPT-4, the method proposed in the present invention, namely the large model-assisted rumor source localization method (CRSLL), shows an average improvement of 62.3% on real-world datasets. The key reasons for the significant improvement are as follows: The self-loop attention mechanism optimizes the aggregation strategy for each user, thereby enhancing the representation quality of propagation dynamics and user profiles; the contrastive learning with non-rumor comments improves the representation quality of rumor comments, and the learnable feature masking module further removes redundant features from high-dimensional comment embeddings. Through the above comparative experiments, it can be concluded that the method of the present invention demonstrates significant advantages compared with existing advanced methods.

[0047] The method of the present invention first constructs dynamic propagation and user profile feature representations based on propagation cascades, uses a graph convolutional network with a self-loop attention mechanism to generate propagation feature representations, and optimizes the aggregation strategy for each user. Further, the present invention combines the advantages of large language models in text semantic analysis and understanding, analyzes user comments, and determines the reasons for being a source user. By adopting a contrastive learning mechanism on the analysis results of rumor and non-rumor comments, the feature representation is enhanced; the redundant features of high-dimensional comment embeddings are further removed through differentiable masking technology. Finally, the present invention introduces a cross-modal attention mechanism, enabling the effective integration of information from propagation feature representations and comment feature representations, thereby generating richer and more comprehensive feature representations. After concatenation, it can be further transformed into a binary classification result through a fully connected layer and normalized into a probability value. The node with the largest final probability value is the rumor source. The present invention is feature-driven rather than topology-driven, enabling the model to be applied to new social networks without any prior knowledge, making the method of the present invention have practical guiding significance.

[0048] The step divisions of the above various methods are only for clear description. During implementation, they can be combined into one step or some steps can be split into multiple steps. As long as the same logical relationship is included, they are all within the protection scope of the present invention; adding insignificant modifications or introducing insignificant designs to the algorithm or process, but without changing the core design of its algorithm and process, are all within the protection scope of this invention.

[0049] Another embodiment of the present invention relates to an information propagation source localization device. The implementation details of the information propagation source localization device in this embodiment will be specifically described below. The following content is only the implementation details provided for convenient understanding and is not necessary for implementing this solution. The information propagation source localization device in this embodiment includes: A social network acquisition module for acquiring the social network relationships of information dissemination; wherein, the social network relationships are composed of multiple user nodes and multiple edges representing the presence of dissemination interaction behaviors between any two user nodes in the social network, and each user node has corresponding portrait features; A propagation cascade establishment module for establishing a propagation cascade based on multiple target user nodes participating in information dissemination in the social network relationships, edges representing the propagation interaction behaviors between the multiple target user nodes, neighbor nodes of the target user nodes, and adjacent edges of each target user node; A propagation feature acquisition module for, for each target user node in the propagation cascade, obtaining the proportion of neighbor nodes participating in information dissemination among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information dissemination to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, and the proportion of neighbor nodes not participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, as the dynamic propagation features of the target user node; A user comment analysis module for analyzing whether the user comment of each target user node is a dissemination source and the reason for being a dissemination source through a large language model; A dissemination source location module for locating the dissemination source of information dissemination by combining the portrait features, dynamic propagation features, whether the user comment of each target user node is a dissemination source, and the reason for the user comment being a dissemination source;

[0050] It is not difficult to find that this embodiment is the device embodiment corresponding to the above method embodiment, and this embodiment can be implemented in cooperation with the above method embodiment. The relevant technical details and technical effects mentioned in the above embodiment are still valid in this embodiment. To avoid repetition, they are not elaborated here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiment.

[0051] It is worth mentioning that each module involved in this embodiment is a logical module. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. In addition, to highlight the innovative part of the present invention, units not closely related to solving the technical problems proposed by the present invention are not introduced in this embodiment, but this does not mean that there are no other units in this embodiment.

[0052] Another embodiment of the present invention relates to a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute the information dissemination source location method in the above embodiments.

[0053] Wherein, the memory and the processor are connected by a bus. The bus may include any number of interconnected buses and bridges, and the bus connects various circuits of one or more processors and the memory together. The bus may also connect various other circuits such as peripheral devices, voltage regulators, and power management circuits, etc., which are well known in the art, and thus will not be further described herein. The bus interface provides an interface between the bus and the transceiver. The transceiver may be an element or multiple elements, such as multiple receivers and transmitters, and provides a unit for communicating with various other devices on the transmission medium. The data processed by the processor is transmitted over the wireless medium through the antenna. Further, the antenna also receives data and transmits the data to the processor.

[0054] The processor is responsible for managing the bus and general processing, and can also provide various functions, including timing, peripheral interface, voltage regulation, power management, and other control functions. The memory can be used to store data used by the processor when executing operations.

[0055] Another embodiment of the present invention relates to a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the above method embodiments are implemented.

[0056] That is, those skilled in the art can understand that all or part of the steps in implementing the above method embodiments can be completed by instructing relevant hardware through a program. The program is stored in a storage medium and includes several instructions to enable a device (which may be a single-chip microcomputer, a chip, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0057] Those of ordinary skill in the art can understand that the above embodiments are specific embodiments for implementing the present invention, and in practical applications, various changes can be made in form and details without departing from the spirit and scope of the present invention.

Claims

1. A method for locating an information dissemination source, characterized in that, The method includes: Obtaining the social network relationships of information dissemination; wherein, the social network relationships are composed of multiple user nodes and multiple edges representing the presence of dissemination interaction behaviors between any two user nodes in the social network, and each user node has corresponding portrait features; Establishing a propagation cascade based on multiple target user nodes participating in information dissemination in the social network relationships, the edges representing the dissemination interaction behaviors between the multiple target user nodes, the neighbor nodes of the target user nodes, and the adjacent edges of each target user node; For each target user node in the propagation cascade, obtaining the proportion of neighbor nodes participating in information dissemination among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information dissemination to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of the target user nodes, and the proportion of neighbor nodes not participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of the target user nodes, as the dynamic dissemination features of the target user node; Analyzing whether the user comments of each target user node are a dissemination source and the reasons for being a dissemination source through a large language model; Locating the dissemination source of information dissemination by combining the portrait features, dynamic dissemination features, whether the user comments are a dissemination source, and the reasons for the user comments being a dissemination source of each target user node.

2. The information dissemination source location method according to claim 1, characterized in that The step of locating the dissemination source of information dissemination by combining the portrait features, dynamic dissemination features, whether the user comments are a dissemination source, and the reasons for the user comments being a dissemination source of each target user node includes: Performing normalization processing on the portrait features and dynamic dissemination features of each target user node and then splicing them; Using a graph convolutional network based on a self-loop attention mechanism to process the spliced portrait features and dynamic dissemination features to generate a dissemination feature representation for each target user node; Locating the dissemination source of information dissemination by combining the dissemination feature representation of each target user node, whether the user comments are a dissemination source, and the reasons for the user comments being a dissemination source; Among them, the graph convolutional network based on the self-loop attention mechanism processes the spliced portrait features and dynamic dissemination features through the following formula: ； ； ； In the formula, in the formula, is the propagation feature representation, represents the activation function of the graph convolutional network, is through propagation cascading , is the identity matrix, is the corresponding degree matrix, is the preset first weight matrix, is the concatenated portrait feature and dynamic propagation feature, is a diagonal matrix, and , is a real matrix of dimension the total number of user nodes in the propagation cascade; is the diagonal matrix constructor, is the th self-loop attention coefficient of the user node represents the neighbor nodes of is the softmax normalization result of the self-loop attention coefficient of the th user node, is a single-layer backpropagation BP neural network applied to the head of each attention mechanism, is a non-linear activation function, is the exponential function, is the preset second weight matrix, is the propagation feature representation of 3. The information dissemination source location method according to claim 2, characterized in that, The step of locating the dissemination source of information dissemination by combining the dissemination feature representation of each target user node, whether the user comments are a dissemination source, and the reasons for the user comments being a dissemination source includes: Converting the text content of the large language model analyzing whether the user comments are a dissemination source and the reasons for being a dissemination source into a reason feature representation through a trained BERT model; Locating the dissemination source of information dissemination by combining the dissemination feature representation of each target user node and the reason feature representation; Among them, the enhanced reason feature representation is obtained through the following formula: ； In the formula, is the enhanced cause feature representation, is the same batch of positive sample pairs belonging to the same propagation cascade as , represents the cause feature representation of negative samples from non - propagation cascades, represents the random sampling operation, is the cosine similarity evaluation function, represents the contrast learning process.

4. The information dissemination source location method according to claim 3, characterized in that, The step of locating the dissemination source of information dissemination by combining the dissemination feature representation of each target user node and the reason feature representation includes: Using a differentiable masking technique to filter out the invalid reason feature representations in the enhanced reason feature representation to obtain valid reason feature representations; Combining the propagation feature representation and the effective cause feature representation of each target user node to locate the propagation source of information dissemination; Among them, the differentiable masking technology filters out the invalid cause feature representation in the enhanced cause feature representation through the following formula: ； In the formula, is the effective cause feature representation, represents element multiplication, is the preset attenuation coefficient, is the differential approximation result of discrete feature selection, represents the enhanced cause feature representation, represents the f-th feature of user v, represents the corresponding feature whether it is blocked, represents the blocked feature , represents the retained feature ; The differential approximation result of discrete feature selection is obtained based on the following feature masking method of Gumbel Softmax: ； ； In the formula, is a method for differentiable approximation of discrete feature selection, represents linearly transforming the reason feature representation of each user comment into a generation logarithm through a decision network for binary classification, is used to control the continuity degree of the differential approximation result, and respectively represent real number matrices of corresponding dimensions, is used to ensure that the output is a one-hot vector during the forward propagation process and retain differentiability during the backpropagation process.

5. The information dissemination source location method according to claim 4, characterized in that, The step of combining the propagation feature representation and the effective cause feature representation of each target user node to locate the propagation source of information dissemination includes: Adopting a cross-modal attention mechanism to fuse the propagation feature representation and the effective cause feature representation; Locating the propagation source of information dissemination through the fused propagation feature representation and effective cause feature representation corresponding to each target user node; Among them, the cross-modal attention mechanism fuses the propagation feature representation and the effective cause feature representation through the following formula: Wherein, is the propagation feature representation, is the cause feature representation filtered by the differentiable masking technique, represents the propagation feature representation after being processed by the cross-modal attention mechanism, represents the cause feature representation after being processed by the cross-modal attention mechanism, represents the concatenated propagation feature representation and cause feature representation, represents that the softmax function normalizes the classification scores of the multi-layer perceptron MLP into probability values, represents the probability of being predicted as a non-propagation source, and respectively represent the query matrices applied to and ; and respectively represent the key matrices applied to and ; and respectively represent the value matrices applied to and ; represents the feature dimension.

6. An information dissemination source location device, characterized in that The device includes: A social network acquisition module for acquiring the social network relationship of information dissemination; wherein, the social network relationship is composed of multiple user nodes and multiple edges representing the existence of propagation interaction behaviors between any two user nodes in the social network, and each user node has a corresponding portrait feature; A propagation cascade establishment module for establishing a propagation cascade according to multiple target user nodes participating in information dissemination in the social network relationship, edges representing propagation interaction behaviors between multiple target user nodes, neighbor nodes of the target user nodes, and adjacent edges of each target user node; A propagation feature acquisition module for, for each target user node in the propagation cascade, obtaining the proportion of neighbor nodes participating in information dissemination among all neighbor nodes of the target user node to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes not participating in information dissemination to the total number of neighbor nodes of the target user node, the proportion of neighbor nodes participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, and the proportion of neighbor nodes not participating in information dissemination to the largest total number of neighbor nodes among all neighbor nodes of all target user nodes, as the dynamic propagation feature of the target user node; A user comment analysis module for analyzing whether the user comment of each target user node is the propagation source and the reason for being the propagation source through a large language model; A propagation source location module for combining the portrait feature, dynamic propagation feature, whether the user comment is the propagation source, and the reason for the user comment being the propagation source of each target user node to locate the propagation source of information dissemination.

7. A computer device, characterized in that, Includes: At least one processor; And a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the information dissemination source location method according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the information dissemination source location method according to any one of claims 1 to 5.

Images