Cross-platform information tracing method and device

Abstract

The invention discloses a cross-platform information traceability method and device, and the method comprises the steps: carrying out the cross-platform user alignment of users related to information propagation during the traceability of information, and adding a virtual edge between the aligned users; the method comprises the following steps: distributing node tags for users in a plurality of platforms, and then calculating a tag value of tag information iteratively propagating among nodes by each node; when convergence is achieved, the iteration is stopped; and taking the local peak value with the highest label value as a potential information source node through the obtained convergence result. Aiming at the difficulty of information traceability of a plurality of social platforms, the invention provides a cross-network user alignment method to add'virtual edges' among different platforms, so that cross-platform information propagation behaviors can be considered among the platforms through the'virtual edges'. Meanwhile, the method is based on a source significance theory, an actual source node is automatically identified under the condition of not depending on a propagation model, and the limitation of a method based on the propagation model is eliminated.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of network information processing technology, and in particular to a cross-platform information tracing method and device. Background Art

[0002] In the current context of the booming development of social networks, the rapid growth of information dissemination and communication has brought positive impacts to society. However, this rapid dissemination is also accompanied by a series of problems, one of which is that it is becoming increasingly difficult to identify the source of information. New media and self-media platforms allow information to be disseminated multiple times, making it difficult for regulatory platforms to accurately trace the true source of information. In order to solve this problem, existing information tracing methods are mainly divided into two categories: analysis based on propagation models and analysis based on text content.

[0003] First, propagation model-based analysis relies on setting a propagation model to define the propagation pattern of information. This type of approach uses various models, such as the susceptible-infected (SI) model and the susceptible-infected-recovered (SIR) model, to reveal the propagation path of information in the network. By analyzing the network structure and propagation pattern, this approach can help regulators better understand how information spreads in social networks. However, these propagation models may have certain limitations when facing complex situations in the real world, because the spread of information in social networks is usually affected by multiple factors.

[0004] Secondly, text content-based analysis focuses on building multiple features of rumor content to distinguish true information from false information. These features include the language style of rumor text, characteristics of the propagation model, and user behavior. By training an automatic classifier, this method can analyze a given text to determine whether it is a rumor. This type of research has a positive effect on monitoring the spread of rumors and identifying accounts suspected of spreading rumors. However, it should be noted that the definition of rumors can be relatively subjective, so diversity and complexity need to be considered when building a classifier.

[0005] Although past research has made some progress in the field of information traceability, the existing technology still has some significant defects and shortcomings, which limit its applicability in dealing with complex information dissemination scenarios.

[0006] (1) First, most current information tracing technologies are mainly based on a single social platform (such as Weibo or Twitter). With the rapid popularization and continuous advancement of social networks, information dissemination has spanned multiple platforms, showing a cross-platform trend. However, existing technologies have not fully considered the complexity of information dissemination between multiple platforms, thus limiting the practical application of information tracing technology. Solving the problem of cross-platform information dissemination is crucial to building a more comprehensive and accurate information tracing system. Therefore, more attention needs to be paid to the research and technological innovation of cross-platform information dissemination.

[0007] (2) Secondly, propagation model-based methods usually assume that the underlying propagation model is fixed and known, which may not be true in practical applications. This assumption limits the scope of applicability of source detection methods because in many cases, identifying the correct propagation model requires prior knowledge, which may be difficult to obtain in practice. For example, for new infectious diseases or online rumors, it becomes difficult to select an appropriate propagation model, and it is also a challenge to obtain the true values ​​of the parameters in the pre-selected underlying propagation model. Therefore, it is necessary to work on developing more flexible and adaptable information tracing methods that can perform source detection without relying on fixed propagation models.

[0008] (3) Finally, although the text content-based method has achieved certain results in rumor detection, it still has difficulties in locating the source and path of information dissemination. This method focuses more on the analysis of the content of tweets, but there are challenges in accurately locating the real source and clearly tracking the dissemination path. In addition, with the continuous development of artificial intelligence, it has become very easy to rewrite based on existing information through advanced artificial intelligence technology. Therefore, the text content-based method may find it difficult to detect information rewritten by artificial intelligence technology. Summary of the invention

[0009] An object of embodiments of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages to be described below.

[0010] The embodiments of the present invention provide a cross-platform information tracing method, device, electronic device and storage medium, which solve the problems that the above-mentioned prior art is difficult to handle cross-platform information tracing, relies on a priori propagation models and has text content limitations.

[0011] To this end, the present invention provides the following technical solutions:

[0012] In a first aspect, a cross-platform information tracing method is provided, comprising:

[0013] When tracing the source of information, users involved in the dissemination of the information are aligned across platforms, and virtual edges are added between the aligned users. The virtual edges represent the dissemination paths of the information between different platforms.

[0014] Assign node labels to users in multiple platforms, and then calculate the label value of each node to iteratively propagate label information between nodes. The calculation formula of the label value is:

[0015]

[0016] Among them, i and j represent the aligned user pair nodes between the two platforms, α is a part of the label information obtained by node i from its neighbors (0<α<1), which is used to control the weight ratio of the initial label information and the propagation information. represents the label value of node i in the t+1th iteration, Y i represents the initial label value of node i, S ij represents the probability of label propagation from node j to node i; when convergence is reached, the iteration is stopped; based on the obtained convergence result, the local peak with the highest label value is used as a potential information source node.

[0017] Preferably, in the cross-platform information tracing method, assigning node labels to users on multiple platforms includes:

[0018] 3.1) Assign positive labels to texts that are completely similar to the tweet to be traced, with an initial label value of Y i =1; the text whose similarity with the tweet to be traced exceeds the threshold is assigned a possible label, and the initial label value is Y i =0.5; negative labels are assigned to other nodes, and the initial label value is Y i =-1.

[0019] Preferably, in the cross-platform information tracing method, if a node i satisfies the following two conditions:

[0020] Node i is initially an infected node, i.e., Y i =1;

[0021] Final label value greater than the label value of its neighbors, i.e., the local peak with the highest label value;

[0022] Then the node i is identified as a potential information source node.

[0023] Preferably, in the cross-platform information tracing method, the method of performing cross-platform user alignment on users involved in the information dissemination includes:

[0024] 1.1) Encoding user node features

[0025] User information and tweet content are used as user features, which are converted into user feature vectors through a pre-trained language model. For users who have never sent tweets, a random initialization vector is used as their node feature, and the feature vector is updated with subsequent training;

[0026] 1.2) Cross-platform user alignment based on graph neural network

[0027] The network structure and the user feature vector are integrated, and the user nodes between multiple platforms are learned through the graph neural network to obtain the final feature vector representation of each user node:

[0028]

[0029] in, Represents S after integrating neighbor features r The final vector of user node i on the platform, S r Represents the network structure of a social platform, r refers to a specific social platform, Indicates v i Node in Graph S r neighbors on the network, σ represents the activation function, a and W represent the network parameters to be learned, K represents the multi-head attention mechanism, Represents the vector representation of neighbor node j;

[0030] The expression of each node is continuously updated through the guidance of the existing labels. The loss function is as follows:

[0031]

[0032] in, refers to the set of aligned user node pairs, S w and t are two different social platforms, namely, r = w and r = t, ξ is a hyperparameter, and d is the cosine similarity;

[0033] 1.3) Aligning predictions for users across platforms

[0034] The cosine similarity between each user is calculated, and the user pairs above the threshold are added with the virtual edge to align the users between different platforms, where the formula is as follows:

[0035]

[0036] Where d is the cosine similarity and θ is the threshold between aligned users;

[0037] Preferably, in the cross-platform information tracing method, after performing cross-platform user alignment on users involved in the information dissemination, the following steps are further included:

[0038] 2.1) Filter subgraphs on each platform based on target information and time interval

[0039] For specific information to be traced, based on the user nodes extracted from each platform and their interaction relationships, expand its first-order neighbors to obtain candidate local subgraphs r represents a platform;

[0040] 2.2) Cross-network user alignment based on local information propagation network

[0041] Applying the cross-platform aligned user data to the candidate local subgraphs In , we obtain a set of multiple platform-aligned user pairs with virtual edge information;

[0042] The set of multiple platform-aligned user pairs with virtual edge information is regarded as a local subgraph, denoted as

[0043] Then proceed to the step of assigning node labels to users.

[0044] Preferably, in the cross-platform information tracing method, the cross-platform aligned user data is applied to the local sub-graph Including: for unlabeled cross-platform alignment user data in the local sub-graph Retrain on the local subgraph or align the existing annotated cross-platform user data on the local subgraph The above reasoning results.

[0045] Preferably, in the cross-platform information tracing method, assigning node labels to users on multiple platforms and then calculating the label value of each node in iteratively propagating label information between nodes also includes the following steps:

[0046] Construct a weight matrix W on the edge set E, where W ij = 1 means there is an edge between node i and node j. The weight matrix W is further symmetrically normalized to s, where S ij represents the probability of label propagation from node j to node i

[0047] The labels are propagated across the network iteratively based on the matrix S.

[0048] In a second aspect, a cross-platform information tracing device is provided, comprising:

[0049] A cross-network user alignment module based on information propagation, which is used to perform cross-platform user alignment on users involved in the information propagation when tracing the source of the information, and add virtual edges between the aligned users, the virtual edges representing the propagation path of the information between different platforms;

[0050] The information tracing module based on the source saliency theory is used to assign node labels to users in multiple platforms, and then calculate the label value of each node to iteratively propagate label information between nodes. The calculation formula of the label value is:

[0051]

[0052] Among them, i and j represent the aligned user pair nodes between the two platforms, α is a part of the label information obtained by node i from its neighbors (0<α<1), which is used to control the weight ratio of the initial label information and the propagation information. represents the label value of node i in the t+1th iteration, Y i represents the initial label value of node i, S ij represents the probability of label propagation from node j to node i; when convergence is reached, the iteration is stopped; based on the obtained convergence result, the local peak with the highest label value is used as a potential information source node.

[0053] According to a third aspect, an electronic device is provided, 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 the instructions are executed by the at least one processor so that the at least one processor executes any one of the methods described.

[0054] According to a fourth aspect, a storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, any of the methods described above is implemented.

[0055] The embodiments of the present invention have at least the following beneficial effects:

[0056] The present invention proposes a cross-platform information tracing method. Aiming at the difficulty of information tracing on multiple social platforms, the present invention proposes a cross-network user alignment method to add "virtual edges" between different platforms, so that cross-platform information propagation behavior can be considered between multiple platforms through "virtual edges". At the same time, the method proposes to automatically identify the actual source nodes based on the source saliency theory without relying on the propagation model, thus getting rid of the limitations of the propagation model-based method. In addition, in order to reduce the computational complexity, the present invention proposes a local information propagation network construction technology based on cross-platform information propagation, which effectively reduces the complexity of information tracing and improves the efficiency of tracing.

[0057] Other advantages, objectives and features of the embodiments of the present invention will be reflected in part through the following description, and in part will be understood by those skilled in the art through study and practice of the embodiments of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1Showing aligned users between two social platforms in one embodiment of the present invention.

[0059] Figure 2 It is the overall flow chart of the framework of the present invention. DETAILED DESCRIPTION

[0060] The embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the embodiments according to the description.

[0061] First, when the inventors conducted information tracing research, they found that the existing technologies generally ignored the importance and challenges of cross-platform information tracing, which is particularly critical in today's multi-platform social network environment. Traditional methods only focus on information tracing within a single platform, which limits the application scenarios of the algorithm. The inventors' innovation lies in expanding the focus to cross-platform tracing, making the algorithm more adaptable and comprehensive, so as to better meet the needs of real applications. Secondly, the inventors' in-depth analysis of the characteristics of information dissemination reveals the gradual weakening of methods based on text content. Because with the rapid development of artificial intelligence, new media or self-media platforms widely use artificial intelligence technology to rewrite content in information dissemination, making it difficult for traditional text content-based tracing methods to effectively detect the source of information. Then, when setting the information dissemination model, the inventors found that the method based on the propagation model usually assumes that the underlying propagation model is fixed and known, which may not be true in practical applications. It is generally difficult to use a propagation model to characterize the complex propagation mode in real social platforms, and it is difficult to set the hyperparameters in the propagation model. Especially in the scenario of cross-platform tracing, the propagation modes of various platforms are very different. For example, on Weibo and Twitter, due to the different user habits, the communication patterns on these two platforms are also very different, which further limits the applicability of the method based on the communication model.

[0062] Therefore, in response to the previously discovered problems and technical defects, the inventors conceived and designed an innovative solution that focuses on addressing the challenges of cross-platform information tracing. Based on the idea of ​​source saliency, that is, in social networks, nodes surrounded by a larger proportion of infected nodes are more likely to be information source nodes, the inventors proposed a cross-platform information tracing method based on label propagation. The method takes source saliency as the theoretical basis and attempts to automatically identify the actual source nodes without knowing the underlying propagation model. Specifically, the method first considers the characteristics of the infected area and finds that edge nodes tend to have fewer infected neighbors, while central nodes tend to have more infected neighbors, which is consistent with the propagation pattern of most social networks. In addition, in order to solve the cross-platform challenges, the inventors proposed a cross-platform information tracing method that emphasizes cross-platform user alignment in the information tracing process. By adding "virtual edges" to connect users on multiple platforms, the method allows possible propagation paths between multiple platforms to be considered. The specific process is as follows: First, in the process of information tracing detection, users involved in the information propagation are aligned across platforms. Possible propagation paths between platforms are represented by adding "virtual edges". Secondly, positive labels are assigned to user nodes that have spread the information in multiple platforms (infected nodes), and negative labels are assigned to user nodes that have not spread the information (uninfected nodes). Then, based on the probability matrix generated according to the network structure, label information is iteratively propagated between nodes. Finally, through the obtained convergence results, local peaks with the highest "infected" label values ​​can be identified as potential information source nodes. This method combines the source saliency theory and cross-platform user alignment to provide a new and efficient solution for information tracing.

[0063] The key points protected by the present invention and the corresponding technical effects are:

[0064] Key point 1, cross-network user alignment technology based on information dissemination; With the continuous development of social networks, information dissemination between multiple social platforms has become more extensive. Therefore, the present invention performs cross-network user alignment technology in the process of cross-platform information tracing. By adding "virtual edges" to connect users on multiple platforms, this method allows possible dissemination paths between multiple platforms to be considered;

[0065] Key point 2: Local information dissemination network construction technology based on cross-platform information dissemination; Due to the huge number of user nodes and complex network structure of social platforms, it is unrealistic to trace the information of all user nodes on multiple platforms. Therefore, the present invention proposes a local network construction technology based on cross-platform information dissemination to reduce the amount of information tracing calculation and effectively improve the calculation speed;

[0066] Key point 3: Information tracing technology based on source salience theory. Previous information tracing technology heavily relied on existing communication models or the accuracy of text content, while source salience theory is widely applicable to different social networks, avoiding reliance on prior knowledge and being well adapted to cross-platform information tracing scenarios.

[0067] 1. Cross-network user alignment technology based on information diffusion

[0068] In the field of cross-platform information tracing, the key nature of the core technology lies in how to effectively detect the spread of information between multiple platforms, which constitutes the most significant difference between single-platform information tracing and cross-platform information tracing. In order to accurately track the propagation path of information across platforms, the present invention introduces a cross-network user alignment technology based on information propagation. This technology makes full use of user information, tweet content and network structure on multiple platforms to accurately capture the close relationship of information propagation between users on multiple platforms. "Virtual edges" are constructed between users with close information propagation to depict the information propagation channels between multiple platforms, thereby solving the problem of cross-platform information propagation.

[0069] The specific implementation is as follows:

[0070] (1) Encoding user node features

[0071] In order to fully characterize the characteristics of users, the present invention uses user information and tweet content as user features. It is converted into a user feature vector through the Bert-base pre-trained language model, with the aim of capturing the key features of the user's posts and user information within the platform. For users who have never sent tweets, the present invention uses a random initialization vector as their node feature, and updates the feature vector with subsequent training.

[0072] (2) Cross-platform user alignment based on graph neural network

[0073] The network structure and user feature vectors are integrated, and the user nodes between multiple platforms are learned through graph neural networks to obtain the final node feature vector. This module is the core module of cross-network user alignment technology based on information propagation, which aims to integrate the network structure and user feature vectors to obtain the final vector representation of each user node:

[0074]

[0075] in, Represents S after integrating neighbor features r The final vector of user node i on the platform, S r Represents the network structure of a social platform, r refers to a specific social platform, Indicates v i Node in Graph Sr neighbors on the network, σ represents the activation function, a and W represent the network parameters to be learned, K represents the multi-head attention mechanism, Represents the vector representation of neighbor node j. The expression of each node is continuously updated through the guidance of existing labels. The loss function is as follows:

[0076]

[0077] in, refers to the set of aligned user node pairs, S w and S t are two different social platforms, namely r = w and r = t, and ξ is a hyperparameter that we want to maximize the margin. The intuitive description of the above formula is: we want the vector representations of users aligned on the two platforms to be as similar as possible The vector representations of users who are not aligned are as far apart as possible. Where d is a distance characterization function, such as cosine similarity.

[0078] (3) Predictions for aligned users across platforms

[0079] After learning the feature representation of each user in the second stage, the present invention aligns the users between different platforms and adds virtual edges. Specifically, the cosine similarity between each user is calculated, and virtual edges are added to user pairs above the threshold. The formula is as follows:

[0080]

[0081] Where d is the cosine similarity and θ is the threshold for aligning users.

[0082] Through the process of this module, the present invention can add "virtual edges" between platforms. Aligned users have similar network structures and posting content, and are similar users on different platforms. This means that the two users on the two platforms are likely to spread each other's information and become the medium for information dissemination between the two platforms. Figure 1 shown.

[0083] 2. Local information dissemination network construction technology based on cross-platform information dissemination

[0084] Since social networks are usually large in scale, it is usually not practical to trace the information of the entire network, which requires a lot of computing resources and time. Therefore, the present invention proposes a local information propagation network construction technology for cross-platform information propagation, aiming to extract candidate subgraphs and reduce the computational scale.

[0085] (1) Filter sub-graphs on each platform based on target information and time interval

[0086] For specific information to be traced, the present invention extracts the detected involved nodes and their follow-up relationships, like-comment relationships and other interactive relationships as edges on various platforms. In order to further explore potential dissemination users, the present invention continues to expand the first-order neighbors of the subgraph to obtain candidate subgraphs r represents a certain platform. Since it starts from the target node of the specific message to be traced, the construction of the subgraph can be completed within limited time and computational complexity.

[0087] (2) Cross-network user alignment based on local information diffusion network

[0088] After the cross-network user alignment model is trained on the existing annotated aligned user data or the annotated data in other information propagation channels (i.e., technology 1, cross-network user alignment technology based on information propagation), the present invention applies it to the candidate subgraph For cross-network scenarios, the input of the user alignment model is the candidate subgraphs of the two platforms, and the model output is the "virtual edges" between all the two subgraphs, that is, the set of cross-platform aligned user pairs. The two nodes connected by the virtual edge represent the social network accounts of a target person in the real world on the two platforms. There may be multiple "virtual edges", that is, there are multiple groups of cross-platform aligned users. Through the above model prediction process, cross-network user alignment can be achieved in specific scenarios of information to be traced.

[0089] By constructing a local information propagation network, the present invention can effectively reduce the computational complexity and computational resources of user alignment. Note that the cross-network user alignment model in this step can be based on the local subgraph currently extracted by the present invention. You can also train on other subgraphs and labeled data and directly train on the current subgraph. The former will have higher accuracy, and the latter can significantly reduce the demand for labeled data. In short, through the current steps, the present invention can obtain candidate subgraphs with "virtual edges", and each subgraph Due to the existence of virtual edges, it can be regarded as a large subgraph, denoted as

[0090] 3. Information Tracing Technology Based on Source Salience Theory

[0091] With the help of the first two key technologies, the present invention can obtain a candidate subgraph for target information propagation The subgraph includes multiple platforms, and each platform is connected by aligning the "virtual edges" of users. In order to solve the reliance on the propagation model in the prior art, the present invention proposes a cross-platform information tracing technology based on the source saliency theory.

[0092] (1) Assigning node labels

[0093] For candidate local subgraphs For user nodes on the source, the present invention assigns different labels to each tweet: the text that is completely similar to the tweet to be traced is assigned a positive label (Y i =1), and the text whose similarity with the tweet to be traced exceeds the threshold is assigned a possible label (Y i =0.5), assign negative labels to other nodes (Y i =-1). The label information describes the participation of each node in the past propagation process.

[0094] (2) Label Propagation Algorithm

[0095] The label status of the node describes the participation of each user node in the target information propagation process. Before starting label propagation, the present invention should establish a weight matrix, which determines the probability of label propagation between nodes. The present invention constructs a weight matrix W on the edge set E, where W ij = 1 means there is an edge between node i and node j. The matrix is ​​further symmetrically normalized to S, where S ij represents the probability of label propagation from node j to node i.

[0096] Based on the matrix S, the present invention iteratively propagates labels on the network. In each iteration t, each node obtains a small portion of label information from its neighborhood and retains some label information Y of its initial state i Therefore, the label value of node i at time t+1 is as follows:

[0097]

[0098] Among them, α is a part of the label information obtained by node i from its neighbors (0<α<1), which is used to control the weight ratio of initial label information and propagation information. represents the label value of node i in the t+1th iteration, Y i represents the initial label value of node i. When convergence is reached, the present invention can stop this iteration. Here, "convergence" means that the label value of the node does not change in several consecutive iterations of label propagation.

[0099] (3) Source node identification

[0100] Assume that at the end of the above label propagation process, the label vector Finally converges to If a node i satisfies the following two conditions, then the node i is identified as a source node: 1) Node i is originally an infected node, i.e., Y i=1; 2) its final label value Greater than the label value of its neighbors.

[0101] The first condition reflects the fact that infected nodes are more likely to be sources than uninfected nodes. Although this choice may miss some recovered source nodes under certain propagation models (e.g., the SIR model), it avoids confusion about nodes that have never been infected. The second condition ensures that the detected sources are local maxima in the label propagation results, which is consistent with the main idea of ​​source saliency. Therefore, these local maxima are determined by the node infection status (label) and the network structure.

[0102] An embodiment of the present invention also provides an electronic device. The electronic device includes: at least one processor, and a memory connected to the at least one processor in communication, wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the method described. Specifically, the above-mentioned memory and processor are connected together via a bus, and can be a general memory and processor, which is not specifically limited here. When the processor runs the computer program stored in the memory, it can perform the various operations and functions described in the embodiments of the present invention.

[0103] The embodiment of the present invention further provides a storage medium on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented. The specific implementation can be found in the method embodiment, which will not be described in detail here.

[0104] Although the embodiments of the present invention are disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the embodiments of the present invention. For those familiar with the art, other modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scopes, the embodiments of the present invention are not limited to the specific details and the legends shown and described here.

Claims

1. A cross-platform information tracing method, characterized in that: include: When tracing the source of information, users involved in the dissemination of the information are aligned across platforms, and virtual edges are added between the aligned users. The virtual edges represent the dissemination paths of the information between different platforms. Assign node labels to users in multiple platforms, and then calculate the label value of each node to iteratively propagate label information between nodes. The calculation formula of the label value is: Among them, i and j represent the aligned user pair nodes between the two platforms, α is a part of the label information obtained by node i from its neighbors (0<α<1), which is used to control the weight ratio of the initial label information and the propagation information. represents the label value of node i in the t+1th iteration, Y i represents the initial label value of node i, S ij represents the probability of label propagation from node j to node i; when convergence is reached, the iteration is stopped; based on the obtained convergence result, the local peak with the highest label value is used as a potential information source node.

2. The cross-platform information tracing method according to claim 1, characterized in that: Assign node labels to users on multiple platforms, including: 3.1) Assign positive labels to texts that are completely similar to the tweet to be traced, with an initial label value of Y i =1; the text whose similarity with the tweet to be traced exceeds the threshold is assigned a possible label, and the initial label value is Y i =0.5; negative labels are assigned to other nodes, and the initial label value is Y i =-1.

3. The cross-platform information tracing method according to claim 2, characterized in that: If a node i satisfies the following two conditions: Node i is initially an infected node, i.e., Y i =1; Final label value greater than the label value of its neighbors, i.e., the local peak with the highest label value; Then the node i is identified as a potential information source node.

4. The cross-platform information tracing method according to claim 1, characterized in that: Methods for aligning users involved in the information dissemination across platforms include: 1.1) Encoding user node features User information and tweet content are used as user features, which are converted into user feature vectors through a pre-trained language model. For users who have never sent tweets, random initialization vectors are used as their node features, and the feature vectors are updated with subsequent training. 1.2) Cross-platform user alignment based on graph neural network The network structure and the user feature vector are integrated, and the user nodes between multiple platforms are learned through the graph neural network to obtain the final feature vector representation of each user node: in, Represents S after integrating neighbor features r The final vector of user node i on the platform, S r Represents the network structure of a social platform, r refers to a specific social platform, Indicates v i Node in Graph S r neighbors on the network, σ represents the activation function, a and W represent the network parameters to be learned, K represents the multi-head attention mechanism, Represents the vector representation of neighbor node j; The expression of each node is continuously updated through the guidance of the existing labels. The loss function is as follows: in, refers to the set of aligned user node pairs, S w and S t are two different social platforms, namely, r = w and r = t, ξ is a hyperparameter, and d is the cosine similarity; 1.3) Aligning predictions for users across platforms The cosine similarity between each user is calculated, and the user pairs above the threshold are added with the virtual edge to align the users between different platforms, where the formula is as follows: Where d is the cosine similarity and θ is the threshold for aligning users.

5. The cross-platform information tracing method according to any one of claims 1 to 4, characterized in that: After performing cross-platform user alignment on users involved in the information dissemination, the following steps are also included: 2.1) Filter subgraphs on each platform based on target information and time interval For specific information to be traced, based on the user nodes extracted from each platform and their interaction relationships, expand its first-order neighbors to obtain candidate local subgraphs r represents a platform; 2.2) Cross-network user alignment based on local information propagation network Applying the cross-platform aligned user data to the candidate local subgraphs In , we obtain a set of multiple platform-aligned user pairs with virtual edge information; The set of multiple platform-aligned user pairs with virtual edge information is regarded as a local subgraph, denoted as Then proceed to the step of assigning node labels to users.

6. The cross-platform information tracing method according to claim 5, characterized in that: Apply the cross-platform alignment user data to the local sub-graph Including: for unlabeled cross-platform alignment user data in the local sub-graph Retrain on the local subgraph or align the existing annotated cross-platform user data on the local subgraph The above reasoning results.

7. The cross-platform information tracing method according to claim 5, characterized in that: Assigning node labels to users in multiple platforms and then calculating the label value of each node to iteratively propagate label information between nodes also includes the following steps: Construct a weight matrix W on the edge set E, where W ij = 1 means there is an edge between node i and node j. The weight matrix W is further symmetrically normalized to S, where S ij It represents the probability of label propagation from node j to node i and iteratively propagates labels on the network based on the matrix S.

8. A cross-platform information tracing device, characterized in that: include: A cross-network user alignment module based on information propagation, which is used to perform cross-platform user alignment on users involved in the information propagation when tracing the source of the information, and add virtual edges between the aligned users, the virtual edges representing the propagation path of the information between different platforms; The information tracing module based on the source saliency theory is used to assign node labels to users in multiple platforms, and then calculate the label value of each node to iteratively propagate label information between nodes. The calculation formula of the label value is: Among them, i and j represent the aligned user pair nodes between the two platforms, α is a part of the label information obtained by node i from its neighbors (0<α<1), which is used to control the weight ratio of the initial label information and the propagation information. represents the label value of node i in the t+1th iteration, Y i represents the initial label value of node i, S ij represents the probability of label propagation from node j to node i; when convergence is reached, the iteration is stopped; based on the obtained convergence result, the local peak with the highest label value is used as a potential information source node.

9. An electronic device, characterized in that: include: 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 performs the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images