A method for reconstructing and tracing a cross-platform false information propagation tree based on a heterogeneous propagation graph

Abstract

The application provides a cross-platform false information propagation tree reconstruction and tracing method based on a heterogeneous propagation graph. The method obtains information propagation data in multiple social media platforms and constructs a heterogeneous propagation graph containing information nodes, user nodes and domain name nodes. In view of the problem of lack of explicit forwarding records between cross-platforms, a candidate cross-platform propagation edge is constructed based on multi-modal information. For each information node, a candidate parent node set satisfying the time sequence constraint, the propagation delay constraint and the correlation constraint is constructed, the candidate propagation edge probability is calculated by using the edge feature, the node representation, the time decay weight and the cross-platform transformation cost, and the maximum directed branch structure is solved on a directed weighted candidate graph with a virtual source node introduced to obtain the most possible propagation tree or forest. The method outputs candidate source points, key propagation paths, cross-platform bridge nodes and counterfactual explanation results, and is used to provide an interpretable false information propagation evidence chain in the scene of missing forwarding relationship or cross-platform diffusion.

Description

Technical Field

[0001] This invention belongs to the field of network security and fake information detection technology, specifically involving a method for reconstructing the propagation tree of social media propagation chains and tracing the source of fake information. Background Technology

[0002] With the rapid development of social media, misinformation spreads rapidly through interactive forms such as forwarding, quoting, and commenting, posing a serious threat to social order. Propagation tree reconstruction, that is, recovering the propagation chain of "who spread it to whom" from observational data, is a key prerequisite for achieving explainable attribution. Current technologies in this field still have the following areas for improvement:

[0003] First, propagation tree reconstruction is relegated to an auxiliary task, and source localization and path reconstruction are performed independently in stages. Existing methods generally treat the propagation tree as an input feature of rumor detection classifiers rather than an independent output. For example, the COLING 2025 paper and CN118568261B only output "true / false" labels, and CN119130688A only locates the source point without reconstructing the subsequent propagation path. In addition, source localization and path reconstruction are handled independently in stages—the localization stage does not consider the integrity of the path, and the reconstruction stage relies on inaccurate source point assumptions, leading to the accumulation of errors in the two tasks, the inability to optimize collaboratively, and a lack of interpretability in the source tracing results.

[0004] Second, the utilization of multi-dimensional information in the propagation relationship needs further integration. Domain credibility is used as a static attribute and cannot be dynamically adjusted according to the number of false references; time information is only simplified to a single interval weight and is not deeply integrated at the node, edge, and attention levels; multimodal information such as text and URL is only used for node feature enhancement and not for propagation edge authenticity judgment, which limits the accuracy of path judgment.

[0005] Third, the accuracy of reconstruction under incomplete observations needs improvement, and further adaptation is required for multi-source propagation scenarios. Real data often suffers from issues such as missing forwarding information and temporal ambiguity. Existing methods (Bi-GCN, Tree-Transformer, etc.) rely on complete explicit forwarding relationships, making the reconstruction tree prone to breakage or misalignment when data is missing. Furthermore, false information often spreads in parallel from multiple sources, while existing single-source assumption methods cannot identify multiple sources or correctly attribute subsequent nodes to their respective sources.

[0006] Fourth, the cross-platform dissemination and association mechanism is still imperfect. When information is disseminated across platforms through screenshots, copy-pasting, external links, etc., there are no API-level forwarding records between platforms. Traditional methods treat events on each platform as isolated data, making it impossible to establish a cross-platform dissemination mapping and making it difficult to trace the rewriting and transfer paths.

[0007] In summary, existing technologies lack a solution that, under conditions of partial observation and multi-source interleaving, aims to reconstruct the propagation tree, integrates refined modeling and multi-dimensional discrimination, and outputs an interpretable chain of evidence. How to infer a credible propagation structure, achieve coordinated optimization of source localization and path reconstruction, and establish cross-platform propagation correlations under conditions of missing forwarding relationships are issues that require further research in this field. Summary of the Invention

[0008] This invention provides a method for reconstructing and tracing the propagation tree of misinformation across platforms based on heterogeneous propagation graphs. It aims to solve the problems in the prior art, such as the difficulty in inferring credible propagation structures and the inability to output interpretable cross-platform tracing evidence chains under conditions of partial observation, multi-source interweaving, and lack of cross-platform propagation relationships.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A cross-platform method for reconstructing and tracing the propagation tree of misinformation based on heterogeneous propagation graphs includes:

[0011] S100. Obtain dissemination event data from multiple social media platforms, including user posting records, forwarding records, comment records, citation records, timestamps, user information, information content, images, screenshots or media attachment information, URL citation records, and platform identification information;

[0012] The platform identifier information is used to indicate the platform to which the information node belongs. For cross-platform propagation scenarios, if there is no explicit forwarding interface between different platforms, potential cross-platform propagation relationships are inferred through URL redirection, external link references, account mentions, topic tags, text semantic similarity, named entity overlap, image-aware hashing, or account association relationships.

[0013] Based on data acquired from S200 and S100, a heterogeneous propagation graph is constructed. .in:

[0014] Node V includes information nodes (corresponding to posts). User nodes and domain node The initial features of information nodes are a concatenation of text features, time features, and platform attribute features; the initial features of user nodes are a concatenation of authentication status, number of followers, number of historical posts, and learnable user embeddings; the initial features of domain nodes include initial credibility. The calculation formula is:

[0015] ;

[0016] in, ICP filing status; For SSL certificate rating; This is a normalized value for the domain registration duration. Search engine indexing status; For threat intelligence scoring; this invention does not create platform nodes individually, but instead establishes platform attribute embedding and platform-to-propagation prior matrix for each platform.

[0017] Edge E includes eight types of directed edges: publishing edges Forwarding edge Comments Reference edge Domain reference edge Pay attention to the edge Time proximity Cross-platform dissemination Among them, time-nearest neighbor edges are established between any two nodes whose absolute difference in publication time is less than a preset time threshold and who have no other relationship, with a weight of . ,in For preset attenuation rate and and Use a consistent unit of time.

[0018] Cross-platform propagation edges are used to represent the potential propagation relationship between information nodes on different platforms, in order to characterize cross-platform propagation phenomena such as screenshot propagation and copy-paste propagation without explicit API forwarding records.

[0019] The direction of candidate propagation edges used in propagation tree reconstruction is uniformly set to point from information nodes with earlier publication times to information nodes with later publication times; semantic edges such as user-information and information-domain names in heterogeneous propagation graphs can have their directions set according to their relational semantics. Cross-platform propagation edges are only considered as candidate propagation relationships and do not represent definite propagation relationships. Whether they are ultimately selected into the propagation tree is determined by the propagation probability calculation of S500 and the maximum directed branch problem.

[0020] R is a finite set of edge types, defining all possible semantic relations in a heterogeneous graph. For any edge in the graph... There exists a mapping for each. Map it uniquely to One of the types.

[0021] For each edge type Set independent parameterized aggregation functions And introduce learnable edge type importance weights.

[0022] ;

[0023] in Edge type label, ; One-hot encoding for edge types; For a learnable 2-layer fully connected network, the 8-dimensional input is mapped to an 8-dimensional importance score; Normalize the scores to a probability distribution .

[0024] This enables the model to distinguish the contributions of different propagation relationships, such as direct forwarding within the platform, implicit temporal nearest neighbor propagation, and cross-platform propagation.

[0025] S300, a propagation graph built on S200, encodes nodes, edges, and platform attributes to generate edge features that integrate time, relationships, domain names, and platform information.

[0026] For nodes timestamp (Normalized to [0,1]), absolute time embedding is generated using sine / cosine position encoding. For dimensional indexes The encoding formula is:

[0027]

[0028]

[0029] in As a preset encoding dimension, this encoding can preserve the global location information of the timestamp.

[0030] For candidate edges Define the time difference (Unit: seconds). To compress the long-tailed distribution and enhance the model's perception of time intervals, a logarithmic transformation is first performed, followed by mapping to a relative time embedding via a multilayer perceptron:

[0031]

[0032]

[0033] in It is a Layer 2 fully connected network.

[0034] Timestamps are divided into discrete buckets of three granularities to capture time patterns at different scales. The minute-level bucket divides one hour into six 10-minute buckets. The hourly bucket divides a day into 24-hour buckets. The 7-day bucket divides the time window into 7 buckets. By concatenating the three embedding vectors above, a multi-granularity time vector is obtained:

[0035]

[0036] For platform attributes, construct platform embedding:

[0037]

[0038] The platform embedding includes platform type encoding, platform review strength prior, platform dissemination speed prior, and platform content form prior. The platform review strength prior characterizes the degree of intervention of the platform in false or sensitive information; the platform dissemination speed prior characterizes the speed of information spread within the platform; and the platform content form prior characterizes the content characteristics of the platform, which are mainly text, images, videos, or mixed content.

[0039] For cross-platform propagation scenarios, this invention introduces cross-platform deformation costs into edge features. This is used to quantify the degree of content distortion that occurs when information is disseminated from one platform to another. The distortion cost is defined as:

[0040]

[0041] in, For semantic similarity, To normalize text length differences, To express stylistic differences (such as the ratio of formal to informal language). The platform's prior confidence in the dissemination, These are preset weighting coefficients.

[0042] Finally, the relation type encoding, relative time embedding, multi-granularity time vector, source node domain name source feature vector, target node domain name source feature vector, source node platform embedding, and target node platform embedding are fused to generate edge features:

[0043]

[0044] in Encoding relation types, For relative time embedding, For multi-granularity time vectors, This is the domain name origin feature vector. They are nodes and nodes Embedded on the platform For cross-platform transformation costs; if a node does not reference a domain name, the corresponding domain name source feature vector is filled with a zero vector.

[0045] In this way, time and platform information are used together to participate in node representation and edge discrimination, and deformation cost is used to distinguish between rewritten and copied content and independent original content.

[0046] S400, the propagation graph structure based on S200, and the edge features of S300 are used to update the node representation through a graph neural network:

[0047] in ;

[0048] After the node representation is updated, domain name forward modulation and platform forward modulation are performed; domain name source credibility posterior update and platform propagation prior and posterior update are performed during the training or offline update phase of S600.

[0049] Perform forward domain modulation: Initialize domain credibility The initial credibility includes ICP filing status, SSL certificate level, domain registration duration, search engine indexing status, and threat intelligence score. Domain nodes are inserted into the heterogeneous propagation graph, establishing reference edges from information nodes to domain nodes. In domain forward modulation, domain node features influence the representation of information nodes referencing that domain through an attention mechanism.

[0050] ,

[0051] in, This is the representation of information nodes after domain name forward modulation. For information node representation, For learnable domain name projection matrices, For domain attention weight, This is an embedded representation of the domain name to which the information node belongs.

[0052] In platform-based forward modulation, platform attribute embedding influences the representation of information nodes belonging to that platform through an attention mechanism:

[0053] .

[0054] in, This is an embedded representation of the platform to which the information node belongs. For a learnable platform projection matrix, For platform attention weight, This represents the information node after forward modulation by the platform.

[0055] The source prediction subnetwork is a 2-layer MLP, with input nodes as follows: Output The probability of being a source of propagation During training Collaborative optimization.

[0056] S500, a node representation based on S400, constructs a set of candidate parent nodes for each information node v. Candidate edges The timing constraint, propagation delay constraint, and association constraint must be satisfied simultaneously:

[0057] Timing constraints: ;

[0058] Propagation delay constraint: ,

[0059] Between the preset minimum propagation delay and the preset maximum propagation delay. and These represent the platforms to which nodes u and v belong, respectively. When When, the platform-wide propagation delay threshold is used; when At that time, the platform uses the dynamic cross-platform propagation delay threshold from the propagation prior library:

[0060]

[0061] in, and The platform stores cross-platform propagation delay statistics in its propagation prior parameter library, and the formula for deformation cost is... (Confidence prior) refers to different fields in the same parameter library.

[0062] Association constraints include intra-platform association constraints and cross-platform association constraints.

[0063] Intra-platform association constraints: Nodes and There are forwarding, quoting, commenting, or following relationships between them; or the cosine similarity of the text content is greater than the preset similarity threshold; or they quote the same domain name and the time difference is less than the preset time threshold.

[0064] Cross-platform association constraints include: the existence of URL redirection, external link reference, account mention or topic tag correspondence; the publishing account is identified as a cross-platform associated account; the semantic similarity between information nodes on different platforms is greater than the cross-platform semantic similarity threshold; and at least one of the following is shared: named entity, image-aware hash, key phrase n-gram or URL normalization fingerprint.

[0065] For each candidate edge Represent the parent node Child node representation Candidate edge features Text semantic similarity URL structure similarity Input the edge prediction network to obtain the base propagation probability of candidate edges. :

[0066]

[0067] MLP stands for Multilayer Perceptron, and its output value, after being activated by a Sigmoid function, represents candidate edges. This represents the base probability of true propagation edges. In one embodiment, the parent node representation and child node representation are both 128-dimensional, the candidate edge features are 153-dimensional, and the text semantic similarity and URL structural similarity are each 1-dimensional, then the input dimension of the edge prediction network is 411-dimensional.

[0068] After obtaining the basic propagation probability, time decay weights and cross-platform propagation correction terms are introduced.

[0069] Introduce time-decayed attention weights to reduce the reliability of propagation over long time intervals:

[0070]

[0071] in The raw attention score output by the graph attention network. For the preset attenuation rate, The propagation time difference (in hours).

[0072] For cross-platform propagation edges, the prior weights of propagation on each platform and the cross-platform transformation cost correction term are further incorporated. The cross-platform transformation cost is defined as:

[0073]

[0074] in, For semantic similarity, Due to differences in text length, To express stylistic differences, For the platform's prior knowledge of dissemination, , Preset weights.

[0075] For cross-platform propagation edges, the corrected propagation probability, based on the base propagation probability, time decay weight, platform prior to propagation, and cross-platform transformation cost, is as follows:

[0076]

[0077] For propagation edges within the platform, the corrected propagation probability, based on the base propagation probability and time decay weight, is:

[0078]

[0079] Based on the set of candidate parent nodes and the weighted propagation probabilities, a directed weighted graph is constructed, and the propagation tree reconstruction is modeled as solving the maximum directed branch structure problem:

[0080]

[0081] To accommodate both single-source and multi-source propagation scenarios, a virtual node *r* is introduced into the candidate directed weighted graph, and virtual source edges are established from *r* to each information node *v*. The weights of these virtual source edges are output by the source prediction subnetwork. Subsequently, the Edmonds algorithm is used to solve for the maximum directed branch structure on the directed weighted graph containing the virtual source node. After solving, the virtual source node and its virtual source edges are deleted; the information nodes connected to the virtual source node are designated as candidate source nodes, and the remaining edge set constitutes one or more propagation subtrees. A source node penalty term or a source node number regularization term is introduced into the virtual source edge weights to prevent abnormal expansion of the number of candidate source nodes. When there is only one candidate source node, a single-source propagation tree is output; when there is more than one candidate source node, a multi-source propagation forest is output.

[0082] S600, posterior update of domain origin credibility, and prior-posterior update of platform propagation are performed during the model training or offline update phase (not during the normal online inference phase). Based on the propagation tree or propagation forest (hereinafter collectively referred to as "propagation structure") obtained from S500 and the auxiliary classification results, posterior update of domain origin credibility and prior-posterior update of platform propagation are performed. In domain reverse modulation, after the propagation tree reconstruction is completed, the average false probability of all information nodes referencing domain d is calculated. Update the domain credibility verification value:

[0083] ,

[0084] in For learning rate, To assess the credibility of the domain names in the previous round, The average probability of false information comes from "auxiliary category headers, fact-checking labels, or manual review results". This is used to ensure that the credibility falls within [0,1]. The posterior value is stored in the domain name source credibility parameter library for use in the next domain name forward modulation and for use in subsequent event tracing.

[0085] In platform reverse modulation, after the propagation tree reconstruction is completed, the average false probability of all information nodes on platform p is calculated. And update the platform to propagate prior and posterior values:

[0086] ,

[0087] in, Let p be the prior value of the previous round of propagation. To propagate the prior update learning rate to the platform, This represents the average false probability of all information nodes on platform p. The updated platform propagation prior value is stored in the platform propagation prior parameter library for subsequent platform forward modulation.

[0088] Outputs a visualization of the propagation structure and a cross-platform source tracing evidence chain, including: candidate source nodes and their confidence levels; several key propagation paths with the highest edge weights; a list of cross-platform bridge nodes, sorted by subtree size, average propagation probability, or number of cross-platform propagations; platform contribution analysis results, including source platform, amplification platform, and terminal platform; a list of key domain name contributions, determined based on the change in domain name credibility posterior value, the proportion cited by false information, and the weight of associated edges; key time points and event descriptions, obtained through first-order difference detection of propagation rate; cross-platform propagation stage division results, including the incubation stage within the platform, the initial cross-platform propagation stage, the multi-platform diffusion stage, and the decay stage; and counterfactual explanations, including recalculating the maximum directed branch structure and outputting the changes in source node probability, propagation scale, or cross-platform propagation probability after simulating the removal of target nodes, target domain names, target platforms, or target cross-platform propagation edges.

[0089] A further approach: The learnable parameters in S200-S500 are trained using the following method: historical event data with a fully labeled propagation tree, or propagation edge labels and source point labels generated using explicit interaction relationships, manual review results, and weakly supervised rules, minimizing the joint loss function. ,in , , The preset weight coefficients are used. After training, S100 to S600 are executed for new cross-platform propagation events to output a cross-platform joint propagation tree and an interpretable traceable evidence chain.

[0090] The beneficial effects of this invention are as follows:

[0091] 1. This invention uses propagation structure reconstruction as the direct output target, rather than merely using the propagation structure as a feature for classifying misinformation. By outputting candidate source points, key propagation paths, and cross-platform bridge nodes, it can provide a traceable and displayable propagation chain structure for tracing the source of misinformation, thereby improving the interpretability of the tracing results.

[0092] 2. This invention addresses the problem of the lack of explicit forwarding records in cross-platform scenarios by constructing candidate cross-platform propagation edges based on the order of publication, semantic similarity of text, URL normalization fingerprint, overlapping named entities, image-aware hashing, and account association. This prevents similar content on different platforms from being regarded as isolated propagation events, thereby improving the feasibility of cross-platform propagation path inference.

[0093] 3. This invention constructs a set of candidate parent nodes by means of time constraints, propagation delay constraints, intra-platform association constraints and cross-platform association constraints, and calculates the probability of candidate propagation edges by combining multi-granularity time coding, time decay weight and cross-platform deformation cost. This can reduce the risk of erroneous edge connection caused by relying solely on single time proximity or text similarity, and improve the reliability of propagation edge discrimination.

[0094] 4. This invention introduces virtual source nodes in the candidate directed weighted graph and uses the maximum directed branch solution method to generate a single-source propagation tree or a multi-source propagation forest, enabling the model to perform global path selection within the candidate edge set, avoiding the problem of propagation chain breakage or path mismatch caused by simply selecting the parent node locally.

[0095] 5. This invention generates a chain of evidence for tracing the source by analyzing domain name origin features, platform propagation priors, and counterfactual perturbations. The chain of evidence includes key domain names, key platform propagation contributions, key time points, and changes in propagation results after removing target nodes or target edges. This can help users understand the reasons for the formation of candidate source points and key propagation paths, and improve the verifiability of the tracing results. Attached Figure Description

[0096] Figure 1 The flowchart of this invention. Detailed Implementation

[0097] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. This embodiment uses simulation deduction and detailed logical proof to elaborate on the complete implementation process of the present invention. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the present invention.

[0098] like Figure 1 As shown, this embodiment provides a cross-platform method for reconstructing and tracing the propagation tree of misinformation based on heterogeneous propagation graphs. It can analyze the most likely propagation path of a suspicious piece of information on a social media platform in a webpage scenario, determining its source and the process of information spreading step-by-step from the source. The specific implementation plan is as follows:

[0099] In the information tracing method described in this invention, the first stage involves learning parameters using historical data and training the model, specifically including:

[0100] Step S100: Data Acquisition and Preprocessing

[0101] For the source point labels and propagation edge labels required during the training phase, manual annotation, explicit platform interaction relationships, or weak supervision rules can be used to generate them. For specific training methods, see S601.

[0102] Since publicly available data lacks complete annotations of cross-platform propagation edges, this embodiment first describes the strategy for acquiring and preprocessing cross-platform propagation data, specifically including the following sub-steps:

[0103] S101. Training data acquisition strategy. In one embodiment, the training data may come from the public interface of a fact-checking website or platform, and the training set and test set may be divided according to the overall propagation event.

[0104] S102. Feature Extraction: Due to the inconsistent feature dimensions and units of the original data, it cannot be directly input into a graph neural network. This step extracts features from the content to be verified obtained in S101. Text features are extracted using TF-IDF vectors; timestamps are recorded in Unix format; user features include authentication status (0 or 1), number of followers (logarithmically normalized), and number of historical posts (normalized); domain names are parsed from URLs to record information such as ICP filing status, SSL certificate level, registration duration, search engine indexing status, and threat intelligence score.

[0105] To prevent omissions due to content rewriting, multimodal cross-platform content fingerprints are extracted: ① Semantic fingerprint: SimHash is used to generate a 64-bit fingerprint and extract core named entities (person names, place names, organizations), and calculate the Jaccard similarity between cross-platform posts; ② Structural fingerprint: URL subdomains and path prefixes are extracted, tracking parameters are removed, and a standardized backlink fingerprint is formed. Visual fingerprinting extracts perceptual hashes from images, and even if a screenshot is cropped or watermarked, it can still be determined to be from the same source content through hash distance.

[0106] S103. Data Normalization and Field Retention: The characteristic fields in S102 are uniformly normalized and retained, including but not limited to the platform unique identifier p, the anonymized user ID, the post timestamp, the post body, the comment list, normalized URLs appearing in the post, and image-aware hash values. Each record is accompanied by a fact-checking website verification tag indicating its authenticity. .

[0107] S104. Dataset Partitioning: In this step, all propagation events are sorted in ascending order by publication timestamp and divided into training, validation, and test sets according to a preset ratio. Each event includes the source post, all its forwarding records, and all related posts on various platforms. The entire event must belong to the same set completely and cannot be split. This ensures that all events in the test set are later than those in the training set, simulating the real-world application scenario of "predicting future events using historical data," allowing the model evaluation results to accurately reflect generalization ability.

[0108] S200, Construct heterogeneous propagation graph

[0109] This step designs feature extraction and mapping networks for three types of nodes: posts, users, and domains. Specifically, it includes the following sub-steps:

[0110] S201. Node Creation and Initialization. Based on the features extracted in S102, an "information node" is created for each post. The initial features are a concatenation of text features, time features, and platform embeddings, mapped to 128 dimensions through a linear layer. A user node is created for each user. The initial features are a concatenation of authentication status, logarithmic number of followers, number of historical posts, and learnable user embeddings, mapped to 32 dimensions through a linear layer. A "domain node" is created for each domain. The initial feature is initial credibility. It consists of splicing together its original features, totaling 16 dimensions.

[0111] Initial credibility The calculation formula is:

[0112]

[0113] in, ICP filing status; For SSL certificate rating; This is a normalized value for the domain registration duration. Search engine indexing status; Score the threat intelligence;

[0114] After the above processing, the three types of nodes obtain initial features of preset dimensions; before inputting into the graph neural network, they can be further mapped to a unified hidden dimension through a type-specific linear layer.

[0115] S202. Edge Construction: This step, based on the nodes in S201 and the relationship records in S101, explicitly constructs eight types of directed edges to differentiate the contributions of different types of propagation behaviors. A post edge points from a user node to an information node, indicating that the user posted the post, with a posting timestamp. A forwarding edge points from a forwarding user node to a forwarding post node, indicating that the user forwarded the post, with a forwarding timestamp. A comment edge points from a commenting user node to the commented post node, indicating that the user commented on the post, with a comment timestamp. A reference edge points from a forwarding post node to the forwarded post node, indicating that the forwarded post referenced the content of the original post. A domain reference edge points from an information node to a domain node, indicating that the post contains the URL of that domain. A follow edge points from a user node to a followed user node, indicating a follow relationship between users (obtainable via API). Time-nearest neighbor edges are established between any two information nodes whose absolute difference in posting time is less than a preset time threshold and which do not have other types of edges; these are bidirectional edges, and their weights are calculated using the following formula:

[0116] ,

[0117] in and The unit is minutes. The preset time decay coefficient;

[0118] Cross-platform propagation edges are defined as follows: node u belongs to platform A, and node v belongs to platform B (A ≠ B). A directed edge (earlier edge pointing to later edge) is established between u and v if and only if any of the following cross-platform associations are satisfied: external link association (u and v share the same canonical domain name); semantic association (the SimHash Hamming distance between u and v is less than 3, or the named entity Jaccard similarity is greater than 0.7); visual association (the image perceptual hash Hamming distance between u and v is less than 10); and account association (the publishers of u and v are identified as the same user across platforms (inferred from username similarity, email prefix, or mutual follow). The initial weight of the cross-platform propagation edge is set to the product of semantic similarity and time decay, serving as the original basis for subsequent Edmonds algorithm reconstruction of the cross-platform path. Cross-platform propagation edges are stored as dashed lines for easy visualization and differentiation later.

[0119] S203. Edge Type Aggregation Function Configuration. Different edge types have different propagation semantics (forwarding is direct diffusion, commenting is interactive feedback, and temporal proximity is implicit association), so they cannot be processed by the same aggregation function. Therefore, this step, based on the eight edge types in S202, can use graph attention aggregation, mean aggregation, sequence aggregation, or multilayer perceptron aggregation with independent parameters for different edge types. In one embodiment, graph attention aggregation is used for forwarding and referencing edges, sequence aggregation is used for commenting edges, multilayer perceptron aggregation is used for domain name referencing edges, and independent attention aggregation is used for cross-platform propagation edges. Simultaneously, learnable edge type importance weights are introduced. ,in To map edge-type one-hot encoding (8-dimensional) to an 8-dimensional importance score, a 2-layer network is used, enabling the model to automatically learn the importance of different behaviors such as forwarding, commenting, and quoting in the propagation process.

[0120] S300, multi-granularity time coding

[0121] To enable the model to simultaneously perceive absolute time location, propagation delay, and periodic patterns such as "day / night" and "weekday / weekend," this step performs multi-granularity time encoding, specifically including the following sub-steps:

[0122] S301, Absolute Time Embedding, based on the timestamp of S102, embeds the timestamp of each post. Normalize to the [0,1] interval:

[0123] ,

[0124] Then, sine / cosine position encoding is used to generate a 32-dimensional absolute time embedding. .

[0125] for The encoding formula is:

[0126]

[0127]

[0128] By mapping one-dimensional timestamps to a high-dimensional space through sine and cosine position coding, similar timestamps have similar coding representations, and the inner product of the coding of distant timestamps approaches zero and is approximately orthogonal, thus enabling the model to distinguish propagation events at different time points.

[0129] S302, Relative Time Embedding and Multi-Granularity Time Bucket. For each edge Calculate the time difference , Compressing the long-tailed distribution, Input a 2-layer MLP (structure: 3→64→32), output a 32-dimensional relative time embedding. Simultaneously, the timestamp is divided into three discrete buckets at three granularities: the minute-level bucket divides one hour into six 10-minute buckets, embedding 16 dimensions. The hourly bucket divides a day into 24-hour buckets, embedding 16 dimensions. The Heavenly Bucket divides the time window into 7 Heavenly Buckets, embedding 16 dimensions. Concatenate them into a 48-dimensional multi-granularity time vector The relative time embedding captures the propagation interval, the time bucket encoding captures the periodic pattern, and the three complement each other, enabling the model to distinguish between different patterns such as "fast propagation" and "delayed propagation", "weekday propagation" and "weekend propagation".

[0130] S303, Cross-platform deformation cost calculation. For cross-platform candidate edges... Calculate the cost of content transformation. :

[0131]

[0132] in, The semantic cosine similarity calculated for SBERT (Sentence-BERT) (range [0,1]); For text length differences: Normalized to [0,1]; To express stylistic differences: calculate the sum of the absolute differences in the proportions of exclamation marks, question marks, and formal words; The platform's prior confidence in the propagation (range [0,1]); =0.5, Preset weights.

[0133] S304. Edge Feature Fusion. This involves fusing relation type, relative time embedding, multi-granularity time vector, source node domain name source feature vector, target node domain name source feature vector, source node platform embedding, target node platform embedding, and cross-platform transformation cost to generate edge features.

[0134] ,

[0135] in For 8-dimensional one-hot encoding of relation types, and These are the source node domain name source feature vectors and the target node domain name source feature vectors (each with 16 dimensions). and Embedded to the platform (16 dimensions each). It is a 1-dimensional scalar. The total dimension of the edge features is 8+32+48+16+16+16+16+1=153 dimensions.

[0136] S400: Node representation learning and domain name, platform forward modulation

[0137] To enable node representations to simultaneously perceive the dynamic evolution of neighbor structure, multi-hop context, and domain reputation, this step involves node representation learning and positive modulation of domain and platform information, specifically including the following sub-steps:

[0138] S401, GNN message passing. Based on the aggregation function of S203 and the edge features of S303, a two-layer graph neural network is used for message passing to update node representations. Node The update formula is:

[0139] ,

[0140] in It is the ReLU activation function. The edge type importance weights learned in S203 This is the aggregation function for the corresponding edge type. The first layer takes the node features initialized in S201 as input, and the second layer takes the output of the first layer as input. Through message passing in the graph neural network, each node aggregates the information of its 1-hop and 2-hop neighbors, so that the node representation simultaneously contains its own features and propagation structure information.

[0141] S402, Forward Domain Name Modulation: For each referenced domain name d, obtain the domain name node representation. (16-dimensional) and current representation of information nodes (128 dimensions), calculate attention weights :

[0142] ,

[0143] in The weight matrix is ​​a learnable 144×1 matrix. This involves concatenating vectors and then updating the node representations. ,in The projection matrix is ​​128×16. If an information node references multiple domain names, the above modulation is applied sequentially. Domain name origin features are incorporated into the information node representation through an attention mechanism.

[0144] S403, Platform Forward Modulation: For each information node v of platform p, obtain the platform embedding. (16-dimensional) Calculate attention weights :

[0145] ,in The platform attention weight matrix is ​​144×1. The platform feature projection matrix is ​​128×16.

[0146] Update node representation: .

[0147] S500, Propagation Structure Restructuring

[0148] Based on the node representation of S401, a set of candidate parent nodes is constructed for each information node v. Candidate parent node u must meet the following conditions: u's publication time is earlier than v's publication time; the time difference between u and v is within a preset propagation delay range; and u and v satisfy either intra-platform association constraints or cross-platform association constraints. Intra-platform association constraints include explicit forwarding, referencing, commenting, following relationships, text similarity exceeding a preset threshold, or referencing the same domain. Cross-platform association constraints include consistent URL normalized fingerprints, overlapping named entities, text semantic similarity exceeding a preset threshold, image perceptual hash distance below a preset threshold, or account association relationships.

[0149] For each candidate edge u→v, the parent node representation, child node representation, edge features, text similarity, and URL similarity are input into the edge prediction network to obtain the basic propagation probability. Then, the probability is corrected using time decay weights, platform propagation priors, and cross-platform transformation costs. For cross-platform candidate edges, the propagation probability is further corrected by combining platform propagation priors and cross-platform transformation costs.

[0150] To accommodate both single-source and multi-source propagation scenarios, a virtual source node *r* is introduced into the candidate directed weighted graph, and virtual source edges are established from *r* to each information node. The weights of these virtual source edges are output by the source node prediction subnetwork. Subsequently, the Edmonds algorithm is used to solve for the maximum directed branch structure on the candidate directed weighted graph containing the virtual source node. After solving, the virtual source node and its virtual source edges are deleted; the information nodes connected by the virtual source edges are designated as candidate source nodes, and the remaining edge set constitutes the propagation structure. When there is only one candidate source node, a single-source propagation tree is output; when there is more than one candidate source node, a multi-source propagation forest is output.

[0151] S600, Model Training and Output

[0152] The model simultaneously optimizes source identification, propagation edge prediction, temporal consistency, and structural legality. To this end, this step defines the joint loss function and trains it, specifically including the following sub-steps:

[0153] S601. During the model training phase, if a complete propagation tree annotation is not available in historical propagation events, this invention employs a weakly supervised approach to construct training labels. Specifically, explicit forwarding, citation, replying, or commenting relationships within the platform are used as high-confidence positive sample edges; information node pairs that do not meet temporal constraints, have time intervals exceeding a preset maximum propagation delay, semantic similarity below a preset threshold, and no common URL or common entity are used as negative sample edges; initial information, earliest observable information, or high-confidence seed information from fact-checking websites or manual review results are used as source weak labels; for cross-platform propagation relationships, manually annotated or fact-checking confirmed cross-platform propagation cases are used as independent annotation data, without relying on feature rules to generate positive samples. For node pairs lacking edge labels, they are not included in the edge prediction loss calculation but only participate in unsupervised graph representation learning or candidate graph inference. Through the above methods, even in the absence of a complete manually annotated propagation tree, the edge propagation probability model and source prediction model can still be trained.

[0154] S602, Joint Training Model: Based on the propagation structure reconstruction results of S500 and the reverse modulation of S603, a joint loss function is defined:

[0155] ,

[0156] in Cross-entropy loss for source point classification, The binary cross-entropy loss for edge prediction. Time-based penalty loss , loss due to in-degree constraint Take this place =0.5、 =0.3、 =0.2.

[0157] Backpropagation is performed using the Adam optimizer (learning rate 0.001) to update all learnable parameters in the model (including GAT parameters, LSTM parameters, MLP parameters, etc.). , , wait).

[0158] S603. After each training round, collect the set of all information nodes referencing domain name d. The probability of each node v being classified as false information is obtained from the model classification head. (Sigmoid output), calculate the average spurious probability:

[0159] ,

[0160] Then update the domain credibility:

[0161] Learning rate It is 0.1.

[0162] Updated Stored in the domain name source credibility parameter library, it will be used as a new parameter in the next training iteration. Forward modulation is used for S402. Through reverse modulation, the credibility of a domain name and the authenticity of the content referencing that domain name form a mutually verifying closed loop—the posterior value of the credibility of a domain name source cited by a large amount of false information decreases, while the posterior value of the credibility of a domain name source cited by true information increases or remains unchanged, thus forming an accumulative and evolvable domain name source credibility parameter system.

[0163] Similarly, collect all information nodes of platform p. The computing platform propagates prior and posterior values:

[0164] ,

[0165] in, The updated platform propagation prior and posterior values ​​are stored in the platform propagation prior parameter library. This is performed after each training epoch, updating the parameters accordingly. It will be used for the positive modulation of S403 in the next training round.

[0166] This step is not executed during the inference phase.

[0167] Phase Two: Model Reasoning and Application. The hundreds digits 1-6 in the following step numbers correspond to modules S100-S600 in the invention content.

[0168] S151. Obtain a target news article to be traced and its dissemination records on various platforms, and perform the same feature extraction operations as S101-S103 on the obtained data: extract text features (128 dimensions), timestamps (Unix format), user features (authentication status, logarithm of followers, number of historical posts), initial credibility features of the domain name, and multimodal fingerprint extraction. Use the exact same feature extraction process as in the training phase to ensure consistent data distribution.

[0169] S152: This step loads the best model parameters saved in S602 and sets the model to evaluation mode (disabling Dropout, using global statistics with BatchNorm), thereby reusing the propagation rules learned during the training phase.

[0170] S251-S253: Constructing a heterogeneous propagation graph of the events to be traced: Based on the data in S151 and the model in S152, repeat S201 to S203 to construct a heterogeneous propagation graph (three types of nodes and eight types of edges) of the events to be traced. Unlike the training phase, no edges of interest are constructed at this stage (if no API is available), but temporal nearest neighbor edges and cross-platform propagation edges are still constructed. By reusing the same graph construction logic, a heterogeneous propagation graph of the events to be traced is obtained that is consistent with the graph structure in the training phase.

[0171] S351-S353, Time Encoding of the Events to be Sourced: Based on the propagation graph of S251, repeat S301 to S303 to generate 153-dimensional edge features of the events to be sourced. By reusing the same temporal encoding logic, 153-dimensional side features with the same distribution as those in the training phase are obtained.

[0172] S451-S452, Node Representation Learning and Forward Modulation: Based on the model parameters of S152 and the edge features of S351, GNN message passing (S401) and forward modulation (S402, S403) are performed. Unlike the training phase, backward modulation (S603) is not performed; instead, the converged posterior values ​​of domain name source credibility and platform propagation prior and posterior values ​​are trained using S602. By calling the trained GNN for forward propagation, each node obtains a 128-dimensional representation containing structural, temporal, content, and domain name credibility information.

[0173] S551-S554: Reconstructing the propagation tree of the event to be traced and locating the source: Based on the node representation in S451, constructing the candidate parent node set and calculating the propagation probability of candidate edges are performed, and a virtual source node r is introduced into the candidate directed weighted graph. Then, the Edmonds algorithm is used to solve the maximum directed branch structure. After deleting the virtual source node and its virtual source edges, candidate source points and their corresponding propagation structures are obtained. If there is only one candidate source point, a single-source propagation tree is output; if there are more than one candidate source point, a multi-source propagation forest is output. Through the above reasoning process, the most likely source point and propagation path are inferred from the observed data.

[0174] S651. This step, based on the propagation structure of S551, outputs a multi-dimensional evidence chain. Starting from each candidate source node or the root node of the propagation subtree, it outputs... And the platform on which it resides; starting from each candidate source node or the root node of the propagation subtree, perform a depth-first traversal of the corresponding propagation subtree, and calculate the weight of each root-to-leaf path. Output K critical propagation paths in descending order, each path including a node sequence and the propagation time interval for each edge; identify the starting node of all cross-platform propagation edges and output them in descending order of subtree size. Perform platform propagation role classification, calculate the number of source nodes, the number of new nodes, and the out-degree growth rate of each platform, and output the source platform, amplification platform, and terminal platform.

[0175] Calculate the change in credibility for each domain name:

[0176] ,

[0177] Output the K key domain names in descending order, along with their initial credibility, subsequent credibility, and the proportion of false information referencing that domain name;

[0178] Calculate the changes in propagation priors for each platform:

[0179] ,

[0180] Key platforms are output in descending order, along with initial propagation priors, posterior propagation priors, and the proportion of false information on the platform, to reflect the evolutionary contribution of the platform's credibility.

[0181] Calculate the propagation rate First-order difference (number of new nodes added within a unit time window, window width 1 hour, step size 10 minutes) The peak time t* exceeding the threshold τ is detected as the key time point, and the timestamp and the list of newly added nodes at that time are output. Based on the distribution of cross-platform propagation edges on the time axis, four stages are defined. The target node X is temporarily removed from the propagation structure, and the source node confidence is recalculated. Output If the change If the threshold is exceeded, the output "Node X is a key propagation node" is generated. All evidence chains are assembled into JSON format and output, and a visualization is generated simultaneously.

[0182] For domain-level counterfactual interpretation, temporarily remove the target domain name node from the graph and re-execute steps S401 to S500 (only for the affected subgraph where the information node referencing the domain name is located), calculate the change in the representation of each information node and the change in the source node confidence after the propagation structure recalculation. .

[0183] Example of deduction

[0184] Suppose the event to be traced contains the following observation data: Node A (Platform X, t=0) posts a message with the URL "example.com / news" and the text "A major emergency has occurred in a certain place"; Node B (Platform Y, time t=7200s) posts the text "Breaking! A major event in a certain place", without a URL, but contains the same set of named entities, and the SimHash Hamming distance is 2; Node C (Platform Y, t=9000s) forwards B and comments "This is terrible"; Node D (Platform Z, t=86400s) posts a screenshot of A (OCR extracted text similarity 0.88), with the additional comment "It's being circulated online"; Node E (Platform Y, t=7300s) has the text "A major event in a certain place" but different named entities, does not share a URL, and has a semantic similarity of only 0.41 with A.

[0185] The processing flow in this embodiment is as follows:

[0186] Mapping phase: A and B satisfy cross-platform semantic association, establish a cross-platform edge A→B; A and D satisfy visual association, establish a cross-platform edge A→D. Calculate deformation cost: ≈0.145, The values ​​≈0.13 are all relatively small, indicating reliable cross-platform association. In the candidate parent node set, B's candidate parent node is A, C's candidate parent node is B, and D's candidate parent node is A. E did not enter A's propagation subtree because it did not meet the cross-platform association constraint and had high transformation costs. Propagation probability calculation (illustrated value): ; ; The propagation tree consisting of A→B→C and A→D is obtained by solving the Edmonds algorithm.

[0187] The evidence chain output is as follows: the source is A (platform X, confidence level 0.92); the cross-platform bridge node is A; the originating platform is X, the amplifying platform is Y, and the terminal platform is Z; the cross-platform propagation stage is divided into the incubation period within the platform (t=0-600s), the first cross-platform propagation to Y (t=7200s), the diffusion within platform Y (t=7200-9000s), and the cross-platform propagation to Z (t=86400s).

[0188] This example illustrates that, in cross-platform screenshot propagation scenarios lacking explicit forwarding records, the present invention can infer the most likely cross-platform propagation path and output an interpretable chain of evidence based on multimodal fingerprint matching, deformation cost quantification, and global directed branch optimization.

[0189] Through the coordinated efforts of the above steps, this invention trains a source tracing model using historical datasets and applies it to the reconstruction of new event propagation trees. Even in observation scenarios where forwarding relationships are missing, timestamps contain noise, or multi-source propagation is intertwined, it can still infer the propagation path of false information and output interpretable source tracing results. Those skilled in the art can adjust the parameter settings according to actual application scenarios without departing from the scope of protection of this invention.

[0190] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features.

Claims

1. A cross-platform method for reconstructing and tracing the propagation tree of misinformation based on heterogeneous propagation graphs, characterized in that, Specifically, the steps include the following: S100. Obtain dissemination event data on social media platforms, including user posting records, forwarding records, comment records, citation records, timestamps, user information, information content, images or screenshots, URL citation records, and platform identification information; Construct a heterogeneous propagation graph based on S200 and S100 data. Among them: nodes Including information nodes User nodes and domain node ;side Includes eight types of directed edges: publishing edges Forwarding edge Comments Reference edge Domain reference edge Pay attention to the edge Time proximity Cross-platform dissemination R is a finite set of edge types; S300 and a heterogeneous propagation graph built on S200 are used to encode the temporal features of nodes and edges. Relationship type encoding, relative time embedding, multi-granularity time vectors, domain name origin feature vectors, platform embedding, and cross-platform transformation costs are fused to generate edge features. ; S400, based on the propagation graph structure of S200 and the edge features of S300, updates the node representation through a graph neural network and performs domain name forward modulation and platform forward modulation; the domain name forward modulation is that the domain name node features influence the representation of information nodes referencing the domain name through an attention mechanism; the platform forward modulation is that the platform attribute embedding influences the representation of information nodes belonging to the platform through an attention mechanism. S500 and S400-based node representation: For each information node, construct a set of candidate parent nodes, calculate the probability that the candidate edge is a real propagation edge and introduce time decay attention weights. Based on the set of candidate parent nodes and the weighted propagation probability, construct a directed weighted graph. Model the propagation tree reconstruction as solving the maximum directed spanning tree problem. Use the Edmonds algorithm to solve the propagation tree. The node with an in-degree of 0 in the propagation tree is the candidate source node. S600, based on the propagation tree or propagation forest obtained from S500, performs domain name reverse modulation and platform reverse modulation, updates the domain name credibility posterior value and the platform propagation prior and posterior values; and outputs the visual structure of the propagation tree or propagation forest and the traceability evidence chain.

2. The method for reconstructing and tracing the cross-platform propagation tree of false information based on heterogeneous propagation graphs according to claim 1, characterized in that, S200 specifically includes the following steps: S201. Create information nodes, user nodes, and domain name nodes, and initialize the characteristics of each node; initial trust level of domain name nodes. The calculation formula is: ； in, ICP filing status; For SSL certificate rating; This is a normalized value for the domain registration duration. Search engine indexing status; Score the threat intelligence; S202. Construct eight types of directed edges, where time-nearest neighbor edges are established between any two nodes whose absolute difference in publication time is less than a preset time threshold and have no other relationship, with the following weights: ； in For preset attenuation rate and and Use a consistent unit of time; The conditions for establishing a cross-platform propagation edge are: node Belongs to platform A, node If the user belongs to platform B and A≠B, and satisfies at least one of the following associations: sharing the same normalized domain name, SimHash Hamming distance less than a preset threshold or named entity Jaccard similarity greater than a preset threshold, image perceptual hash Hamming distance less than a preset threshold, or the publisher is identified as the same user across platforms; and the propagation edge direction is from the node with earlier publication time to the node with later publication time. S203, For each edge type Set independent parameterized aggregation functions And introduce learnable edge type importance weights: 。 3. The method for cross-platform false information propagation tree reconstruction and source tracing based on heterogeneous propagation graphs according to claim 1, characterized in that, S300 includes the following sub-steps: S301, For nodes timestamp Absolute time embedding is generated using sine / cosine position encoding. , For the preset encoding dimension, for the dimension index The encoding formula is: ； ； S302, For candidate edges Define the time difference First, perform a logarithmic transformation. The long-tailed distribution is compressed and then mapped to a relative time embedding using a multilayer perceptron. ； in It is a 2-layer fully connected network that outputs a 32-dimensional relative time embedding. Simultaneously, the timestamp is divided into three discrete buckets at three granularities: the minute-level bucket divides one hour into six 10-minute buckets. The hourly bucket divides a day into 24-hour buckets. The 7-day bucket divides the time window into 7 buckets. By concatenating the three embedding vectors above, a multi-granularity time vector is obtained: ； For platform attributes, construct platform embedding: ； Among them, platform embedding includes platform type encoding, platform review strength prior, platform dissemination speed prior, and platform content form prior; S303, For cross-platform candidate edges Calculate the cost of content transformation. : ； in, For semantic similarity; For text length differences: To express stylistic differences: This refers to the platform's prior confidence level in the dissemination of information. , Preset weights; S304. The relationship type, relative time embedding, multi-granularity time vector, source node domain name source feature vector, target node domain name source feature vector, source node platform embedding, target node platform embedding, and cross-platform transformation cost are fused to generate edge features: ； in , For relative time embedding, For multi-granularity time vectors, This is a feature vector representing the origin of a domain name. and They are nodes and nodes Embedded within the platform; Cost of cross-platform transformation.

4. The method for cross-platform false information propagation tree reconstruction and source tracing based on heterogeneous propagation graphs according to claim 1, characterized in that, S400 includes the following sub-steps: S401. Update the node representation using a graph neural network. The update formula is: ； S402. Perform forward domain name modulation: Insert domain name nodes into the heterogeneous propagation graph, establish reference edges from information nodes to domain name nodes, and influence the representation of information nodes referencing that domain name through an attention mechanism: ； in, This is the representation of information nodes after domain name forward modulation. For information node representation, For learnable domain name projection matrices, For domain attention weight, An embedded representation of the domain name to which the information node belongs; S403, Performing Platform Forward Modulation: Platform attribute embedding influences the representation of information nodes belonging to that platform through an attention mechanism: ； in, This is an embedded representation of the platform to which the information node belongs. For a learnable platform projection matrix, For platform attention weight, This represents the information node after forward modulation by the platform.

5. The method for cross-platform false information propagation tree reconstruction and source tracing based on heterogeneous propagation graphs according to claim 1, characterized in that, S500 includes the following sub-steps: S501. Construct a set of candidate parent nodes and candidate edges. The following constraints must be met simultaneously: Timing constraints: ; Propagation delay constraint: ; and The platforms to which nodes u and v belong are respectively. When, the platform-wide propagation delay threshold is used; when At that time, the platform uses the dynamic cross-platform propagation delay threshold from the propagation prior library: ； and This refers to the cross-platform propagation delay statistics stored in the platform's propagation prior parameter library. Association constraints: nodes and There are forwarding, quoting, commenting, or following relationships between them; or the cosine similarity of the text content is greater than the preset similarity threshold; or they quote the same domain name and the time difference is less than the preset time threshold. Cross-platform association constraints include: the existence of URL redirection, external link reference, account mention or topic tag correspondence; the publishing account is identified as a cross-platform associated account; the semantic similarity between information nodes on different platforms is greater than the cross-platform semantic similarity threshold; and at least one of the following is shared: named entity, image-aware hash, key phrase n-gram or URL normalization fingerprint. S502, For each candidate edge The probability that it is a real propagation edge is calculated using a multilayer perceptron. : ； in and Represented as nodes, Cosine similarity of the text content. Cosine similarity to the URL structure is used, and time-decayed attention weights are introduced: ； in The raw attention score output by the graph attention network. For the preset attenuation rate, The propagation time difference; for cross-platform propagation edges, the corrected propagation probability is obtained based on the base propagation probability, time decay weight, platform prior to propagation, and cross-platform transformation cost: ； For propagation edges within the platform, the corrected propagation probability, based on the base propagation probability and time decay weight, is: ； S503. Based on the set of candidate parent nodes and the weighted propagation probabilities, construct a directed weighted graph with the goal of maximizing the sum of the weighted probabilities of all edges in the tree: ； The solution process uses the Edmonds algorithm to solve the maximum directed branch structure on a directed weighted graph containing virtual source nodes. The weight of the virtual source edge pointing from the virtual source node to each information node is output by the source prediction subnetwork. The source prediction subnetwork takes the node representation of the information node as input and outputs the probability that the node is the propagation source. S504. After solving, delete the virtual source node and its virtual source edge. The information node connected to the virtual source node is used as the candidate source node. The remaining edge set forms one or more propagation subtrees. If there is only one candidate source node, output a single-source propagation tree. If there is more than one candidate source node, output a multi-source propagation forest.

6. The method for cross-platform disinformation propagation tree reconstruction and source tracing based on heterogeneous propagation graphs according to claim 1, characterized in that, The source tracing evidence chain in S600 includes: candidate source nodes and their confidence levels, and the edge with the highest weight. Changes in the posterior value of key propagation paths and domain name credibility List of key domain name contributions and list of key platform contributions to changes in prior and posterior values ​​of platform propagation, through propagation rate First-order difference The key time points and event descriptions obtained from the detection, along with the counterfactual explanation—that is, the change in the source probability after simulating the removal of node X. .

7. The method for cross-platform false information propagation tree reconstruction and source tracing based on heterogeneous propagation graphs according to claim 1, characterized in that, The specific implementation of domain name reverse modulation and platform reverse modulation in S600 includes: S601, Domain Name Reverse Modulation: Based on the propagation tree or propagation forest obtained in S500, calculate the referenced domain name. Average false probability of all information nodes Update the domain credibility verification value: ， in, For learning rate, To assess the credibility of the domain names in the previous round, The average false probability of all information nodes referencing domain d. This is used to ensure that the credibility falls within the [0,1] interval; S602, Perform platform reverse modulation: Calculate the average false probability p_fake_avg(p) of all information nodes on platform p, and update the platform propagation prior and posterior values: ， in, This is the prior value of the previous round of propagation for platform p. To propagate the prior update learning rate to the platform, Let be the average false probability of all information nodes on platform p.

8. A method for reconstructing and tracing the cross-platform propagation tree of false information based on heterogeneous propagation graphs according to any one of claims 1-6, characterized in that, This also includes the model training phase: utilizing historical event data with fully labeled propagation trees, or propagation edge labels and source node labels generated using explicit interaction relationships, manual review results, and weakly supervised rules, to minimize the joint loss function. ;in Cross-entropy loss for source point classification, Binary cross-entropy loss for edge prediction; For time-series penalty loss, Loss is due to in-degree constraint. , , Learnable parameters from S200 to S500 are trained using preset weight coefficients.

9. A method for reconstructing and tracing the cross-platform propagation tree of false information based on heterogeneous propagation graphs according to any one of claims 1-7, characterized in that, The counterfactual interpretation also includes domain-level counterfactual interpretation: temporarily remove the target domain name node from the heterogeneous propagation graph, re-execute the node representation update and propagation tree reconstruction steps from S400 to S500, calculate the change in the representation of each information node referencing the domain name and the change in the source point confidence after the propagation tree is recalculated, and output the degree of influence of the domain name on the tracing results.

Images