Cold start false news detection method based on propagation path and greedy algorithm

Abstract

The invention relates to the technical field of natural language processing, in particular to a cold start false news detection method based on a propagation path and a greedy algorithm, which comprises the following steps: S1, constructing a news propagation graph, generating propagation paths based on interaction behaviors of users and news, and calculating the similarity of the propagation paths among the users; s2, based on the propagation path similarity, performing unsupervised community initialization by adopting a greedy strategy to form a user community with cohesiveness; s3, injecting a community structure as a structure prior into the graph neural network, and enhancing the propagation structure perception capability of the model in a cold start stage; and S4, carrying out true and false classification on the news events by utilizing the graph neural network with the injection structure prior, and realizing early false news detection. The method does not need to depend on a large number of labels and complete propagation structures, high detection precision and robustness can be kept in the early stage that information is sparse and communities are not formed, experiments show that the method is superior to an existing method on mainstream data sets, and the method has high practical application value.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a cold-start method for detecting fake news based on propagation paths and greedy algorithms. Background Technology

[0002] With social media becoming a core carrier of information dissemination, the early identification and containment of fake news has become an important issue for social governance and cybersecurity. Existing research indicates that fake news spreads significantly faster than real news in its early stages (such as the first 10% of the dissemination cycle). Therefore, achieving accurate detection of fake news during the cold start phase has significant practical implications.

[0003] Fake news detection methods based on propagation paths are gradually becoming a mainstream research direction. These methods are based on a key finding: fake news and real news exhibit systematic differences in their propagation topology; for example, fake news often displays characteristics such as concentrated propagation paths and high node centrality. Modeling the propagation graph constructed from user interactions such as forwarding and commenting using graph neural networks can effectively capture these structural features, thus achieving high-precision detection in the later stages of propagation.

[0004] However, in the cold start phase of news dissemination, that is, in the early stages of dissemination when user participation is scarce and the interactive structure has not yet taken shape, existing methods face significant challenges:

[0005] 1) Lack of community structure: Most graph-based detection methods rely on established user communities as contextual features, but in the early stages of propagation, the network structure is highly sparse, making it difficult to extract effective community signals;

[0006] 2) Incomplete propagation path: The number of nodes and edges in the early propagation path is limited, which makes it difficult for graph neural networks to perform effective neighborhood aggregation and feature propagation, thus failing to accurately capture abnormal diffusion patterns;

[0007] 3) High dependence on manual intervention: To alleviate the problem of insufficient labeled data in the cold start stage, some methods rely on manually setting seed nodes, heuristic community partitioning, or manual feature engineering. These methods have weak generalization ability and are difficult to adapt to dynamic and ever-changing social environments.

[0008] Existing methods generally suffer from problems such as detection lag, insufficient generalization ability, and low automation in cold-start scenarios. Faced with this technical bottleneck, there is an urgent need in the field for an innovative solution that can achieve efficient and automated fake news detection in the early stages of propagation and under sparse structural conditions. Therefore, this invention provides a cold-start fake news detection method based on propagation paths and a greedy algorithm. Summary of the Invention

[0009] This invention aims to address the limitations in detection performance of existing fake news detection methods during the cold start phase of dissemination, caused by sparse community structure, incomplete dissemination paths, and reliance on manual intervention.

[0010] This invention provides a cold-start fake news detection method based on propagation paths and a greedy algorithm. First, a user-news interaction propagation graph is constructed, and a path similarity function integrating co-occurrence frequency, temporal overlap, and structural proximity is designed. Then, an unsupervised greedy strategy is employed to iteratively aggregate similar users starting from nodes with high propagation density, forming a cohesive initial community. Finally, the community structure is injected as a structural prior into the graph neural network to enhance the model's ability to perceive sparse propagation features.

[0011] This invention requires no manual intervention or data annotation, and can automatically and robustly achieve efficient detection of fake news in the early stages of its dissemination.

[0012] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0013] A cold-start fake news detection method based on propagation path and greedy algorithm includes the following steps:

[0014] S1: Construct a news dissemination graph, generate dissemination paths based on user interaction with news, and calculate the similarity of dissemination paths between users;

[0015] S2: Based on the similarity of propagation paths, a greedy strategy is used to initialize the unsupervised community, forming a cohesive user community;

[0016] S3: Inject community structure as a structural prior into the graph neural network to enhance the model's ability to propagate structural awareness during the cold start phase;

[0017] S4: Utilize graph neural networks with injected structural priors to classify news events as true or false, thereby enabling early detection of fake news.

[0018] Furthermore, step S1 includes the following sub-steps:

[0019] S1.1: Construct a propagation graph and user propagation structure, and use a greedy algorithm to achieve community segmentation during the cold start phase, thereby assisting in the early detection of fake news. First, a news propagation graph is constructed to capture the interaction structure between users and news on social platforms. Each news event... It must contain at least one publisher and several user groups involved in the dissemination Where k is the number of users spreading the message. S1.2: Define the propagation graph as the interactions of users on social media, such as forwarding, commenting, and liking news. ,in Represents a set of users and news nodes. The edges constitute the propagation behavior. The node feature matrix is ​​denoted as... , used to represent features such as user behavior or content representation, where d is the feature dimension. To further capture the structural commonalities among users during the propagation process, Projection spreads similar images to users In this graph, the edges The weights are defined by three propagation path features:

[0020] Path co-occurrence frequency The number of times two users participate in the same propagation path;

[0021] Time window overlap rate The degree of overlap in the time periods of the interactive behaviors;

[0022] Structural adjacency The inverse of the shortest path in the propagation graph (reflecting proximity).

[0023] The user propagation path similarity function is defined as:

[0024]

[0025] Where α, β, and γ are normalized weighting parameters. This similarity function provides the structural basis for greedy community partitioning. The detection model is defined as a mapping function from news dissemination structure to true / false labels. ,in Indicates to news The corresponding propagation graph structural features include user interaction paths, community structure, and propagation similarity.

[0026] Furthermore, step S2 includes the following sub-steps:

[0027] S2.1: Community Initialization. Based on the aforementioned definition of propagation path similarity, this invention designs an unsupervised greedy aggregation strategy to complete the initial community partitioning under unlabeled conditions. Its core process is as follows:

[0028] ① Seed node selection: Select the node with the largest weighted degree from u. As the starting point of the initial seed, that is;

[0029]

[0030] ② Greedy expansion: In each iteration, select nodes from the unsplit nodes that are in the current community. The node with the highest average similarity, i.e.:

[0031]

[0032] ③Termination condition: If Or the community size reaches the maximum threshold. If the community expansion fails, the expansion will terminate and the next one will begin.

[0033] By employing a greedy strategy, the system can automatically divide multiple user communities with similar propagation behaviors in the early stages when the propagation graph is not yet complete.

[0034] S2.2: Community structure optimization. To improve the consistency and accuracy of the partitioning results, the algorithm performs the following optimizations after initialization:

[0035] Edge node reallocation: For nodes that are connected to multiple communities at the same time, their aggregation similarity is reassessed and their affiliation is adjusted.

[0036] Weak edge removal and regularization: Remove edges with weights below a threshold to suppress noise propagation path interference and improve community compactness.

[0037] Ultimately, the optimized community structure will be encoded as a label embedding or propagation prior, which will be fed into the fake news detection module (such as a convolutional classifier or structural attention network) to improve the model's judgment ability during the cold start phase.

[0038] Furthermore, step S3 includes the following sub-steps:

[0039] S3.1: Input feature modeling, for each news event This corresponds to a propagation diagram. , where the set of nodes Including forwarding users, initial publishers, and news nodes, edge set This represents the propagation path relationship between users or between users and news. Node feature matrix. Includes the following content:

[0040] ①Text representation (such as BERT encoding of news text);

[0041] ② User behavior statistics (such as forwarding frequency, account age, whether it is an institutional account);

[0042] ③ Community tag (derived from greedy community division, encoded as an embedding vector).

[0043] ④ After concatenating the above features, input them into the structure-aware graph classification model.

[0044] S3.2: Structural Prior Injection Mechanism

[0045] The greedy community partitioning provides each user node with a structural label representing the aggregation propagation similarity. This structural prior serves as an additional supervisory signal at the node level, providing the model with hints about local propagation consistency. To inject this prior, the study employs the following two methods:

[0046] Tag embedding method: Convert the one-hot encoding of each user's community into a low-dimensional structured embedding vector. , and concatenate with the original node features;

[0047] Graph structure edge weight adjustment: Enhance the edge weights between nodes within the same community, weaken the edges across communities, and achieve a biased reconstruction of the propagation graph structure.

[0048] This mechanism enables the model to perceive the "pre-structure" in the propagation graph even when the propagation is not yet fully complete, and it has a certain robustness to structural perturbations and cold-start sparsity problems.

[0049] S3.3: Graph Neural Network Design

[0050] After structural prior injection, the study employs a Graph Convolutional Network (GCN) as the backbone model to learn node representations for the news dissemination graph. The model architecture is as follows:

[0051] ① Input features (Splitting original features with community priors);

[0052] ② Two-layer GCN aggregation layer:

[0053]

[0054] in This is the adjusted adjacency matrix. Its degree matrix, These are trainable weights;

[0055] ③ Graph-level aggregation uses mean pooling or attention pooling to represent all nodes;

[0056] ④ Finally, a fully connected layer is used for true / false binary classification prediction. The embedded nodes are aggregated using an attention pooling mechanism to form a news-level event representation.

[0057]

[0058] Ultimately, the news event will be represented The input is fed into a fully connected layer and a softmax classifier to predict the true or false label of the event. :

[0059] S3.4: Model Training and Optimization

[0060] The model training employs the cross-entropy loss function, where structural priors serve only as auxiliary inputs and are not involved in the supervision objective. To adapt to cold-start scenarios, only the information from the first 10% of nodes and edges is used during training, and the DropEdge data augmentation strategy is incorporated to address structural perturbations.

[0061] Furthermore, step S4 includes the following sub-steps:

[0062] S4.1: In the cold start scenario, only the first 10% of nodes and edges of the propagation path are used for model training and inference;

[0063] S4.2: Use the cross-entropy loss function for model training and combine it with the DropEdge strategy to enhance the robustness of the model;

[0064] S4.3: Evaluate the model's detection performance in the early stages of propagation using the Early-F1 metric to ensure the model has real-time response capabilities.

[0065] The beneficial effects of this invention are:

[0066] Traditional methods struggle to effectively measure user relationships during the cold start phase due to data sparsity. This invention proposes a path similarity function that integrates multi-dimensional propagation features, comprehensively considering three key indicators: path co-occurrence frequency, time window overlap rate, and structural adjacency. Co-occurrence frequency captures the macro-level collaborative patterns of users participating in the same propagation path; time overlap rate measures the synchronicity of user behavior from a micro-time sequence perspective; and structural adjacency reflects the potential influence relationships of users at the network topology level. This multi-dimensional fusion measurement method enables accurate characterization of the potential, deep-level similarity in propagation behavior among users even with extremely limited propagation paths, providing a stable and information-rich structural foundation for subsequent community partitioning and overcoming the poor robustness of single-feature measures under sparse data.

[0067] To address the issues of lack of supervised signals and stable community structure during the cold start phase, this invention abandons the traditional approach of relying on manual rules or historical labels and proposes a completely unsupervised greedy community aggregation strategy. This strategy uses the node with the highest weighted degree (i.e., the most active propagation) in the propagation graph as the initial seed, and iteratively incorporates unassigned nodes with the highest average similarity to the current community, following the principle of local optima. This process requires no prior knowledge or manual settings and can automatically and efficiently discover highly cohesive user groups in sparse networks. Compared to complex global optimization algorithms (such as spectral clustering), the greedy strategy has low computational complexity and fast convergence speed, making it highly suitable for the real-time requirements of social media platforms. The resulting community structure is tight, providing consistent contextual information for user behavior and high-quality structured input for downstream detection models.

[0068] This invention injects the community structure obtained by a greedy algorithm as prior knowledge of the graph structure into a Graph Neural Network (GCN). Specifically, this is achieved through two mechanisms: first, community label embedding, which encodes the identifier of a user's community into a low-dimensional vector and concatenates it with the original node features, enabling the model to perceive community affiliation at the node level; second, graph structure edge weight adjustment, which strengthens the connection weights between nodes within the same community and weakens the connections between different communities, thereby guiding the model to focus more on consistent patterns within communities during information propagation. This enhances the GNN's ability to represent sparse and incomplete propagation structures. The model can more effectively aggregate local neighborhood information, capture abnormal propagation signals that are difficult to detect using only original features and sparse connections, and significantly improve discriminative performance in low graph density environments.

[0069] This invention optimizes the entire process from problem definition and algorithm design to model integration, all with the goal of "early detection." It rigorously simulates a cold-start scenario by using only the first 10% (or even 5%) of the data in the propagation path during training and inference phases. The efficiency of the greedy community partitioning ensures the system can quickly respond to emerging events. The introduction of graph regularization techniques such as DropEdge further enhances the model's robustness to structural perturbations and prevents overfitting. Experimental results show that this invention achieves an average improvement of over 7% on public datasets such as FakeNewsNet and PHEME, especially on the Early-F1 metric, which measures early detection capability, significantly outperforming a series of baseline models such as Text-CNN, EANN, and UPFD-GCN. Simultaneously, its computational efficiency is comparable to mainstream graph models, and even surpasses some complex models in the early stages, demonstrating excellent engineering deployment potential and providing a reliable technical path for achieving real-time, automated, and high-precision early warning of fake news on social media platforms.

[0070] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0071] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a schematic diagram of the model structure corresponding to the cold start fake news detection method based on propagation path and greedy algorithm of the present invention;

[0073] Figure 2 This is a detailed flowchart of the implementation process of the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example 1

[0076] like Figures 1-2 As shown in this embodiment, the cold-start fake news detection method based on propagation path and greedy algorithm includes the following steps:

[0077] S1: Download publicly available text stance datasets from the internet and conduct experiments on two benchmark datasets: FakeNewsNet and PHEME. Both datasets contain multimodal information such as news content, user propagation paths, social interaction behavior, and authenticity labels, making them suitable for propagation path modeling, structure learning, and early prediction research.

[0078] The dataset obtained above is shown in Table 1:

[0079] Table 1. Details of the FakeNewsNet and PHEME datasets

[0080]

[0081] S2: Calculation of User Propagation Path Similarity

[0082] Based on the propagation graph constructed in step S1, calculate the propagation graph between any two users. Similarity of propagation paths between them.

[0083] Path co-occurrence frequency : Count the number of identical news paths that two users jointly participated in disseminating, and then normalize the data.

[0084] Time window overlap rate Divide the time window into hours and calculate the Jaccard similarity coefficient of the time windows in which the interaction behavior of two users occurs across all news that they share.

[0085] Structural adjacency In the propagation diagram Calculate the shortest path distance d between two users and take its reciprocal. As a measure of proximity.

[0086] The overall similarity is obtained by linearly weighting the above three features:

[0087]

[0088] In this embodiment, the weighting parameter is empirically set to The similarity is then normalized to ensure a sum of 1. Based on this similarity, a user propagation similarity graph is constructed. Where nodes represent users and edge weights are corresponding values. value.

[0089] S3: Community Initialization and Optimization Based on Greedy Strategy

[0090] In similar diagrams The unsupervised greedy community partitioning algorithm is executed, and the specific steps are as follows:

[0091] S3.1: Seed Node Selection: Calculation The node with the highest weighted degree (i.e., the sum of the weights of all adjacent edges) is selected. As the seed node of the first community.

[0092] S3.2: Greedy Community Expansion: Initialize the current community. In each iteration, calculate the average similarity between all unassigned nodes and the current community: select the node with the highest average similarity. Join the community.

[0093]

[0094] S3.3: Termination condition: When the community size reaches the preset maximum value. Or the average similarity of newly added nodes is lower than the threshold. At that time, the expansion of the community will be stopped.

[0095] S3.4: Iteration and Optimization: Reselect seeds from the unassigned nodes and repeat steps S3.2-3.3 until all nodes are partitioned. Then perform community optimization: For boundary nodes connected to multiple communities simultaneously, reassign them to the community with the highest average similarity; remove nodes with weights below a threshold in the similarity graph. Weak edges are used to enhance community cohesion. This ultimately leads to a community set. .

[0096] S4: Graph Neural Network Detection Model with Structure Prior Injection

[0097] S4.1 Node Feature Construction: Feature Vector of Each User Node It is composed of three parts: a) BERT encoding of the user profile text (64 dimensions); b) user behavior statistics (such as forwarding frequency, account age, etc., 32 dimensions); c) community embedding vector (32 dimensions) obtained by mapping the one-hot encoding of its community through an embedding layer. The news node feature is the BERT encoding of its body text.

[0098] S4.2. Graph Structure Prior Injection: For the original propagation graph The adjacency matrix A is adjusted to inject community priors. If the two nodes connected by an edge belong to the same community, the edge weight is multiplied by an enhancement factor of 1.2; if they belong to different communities, it is multiplied by a suppression factor of 0.6. This yields the adjusted adjacency matrix. .

[0099] S4.3. Graph-level Representation and Classification: After two layers of GCN, an attention pooling mechanism is used to aggregate all node representations into a graph-level representation of news events. Finally, Input a fully connected layer and a Softmax classifier, and output the predicted probability of whether the news is true or false.

[0100] S5: Model Training and Cold Start Evaluation

[0101] To simulate a cold start scenario, only the earliest 10% of nodes and edges in each news propagation path are used to construct subgraphs during training and testing. Cross-entropy loss is employed as the loss function. To enhance robustness, the DropEdge technique is applied to the adjacency matrix during training, randomly discarding 10% of the edges. Five-fold cross-validation results on the FakeNewsNet dataset show that the proposed method achieves an F1 score of 79.8% in the cold start setting, significantly outperforming the baseline.

[0102] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0103] (1) Comparative experiment

[0104] The results of this experiment were compared with other detection methods on two datasets. Specifically, the experimental results for the six benchmark position detection methods, Text-CNN, BiGRU, EANN, UPFD-GCN, PSA-GT, and SAN, are shown in Tables 3 and 4.

[0105] Table 3: Comparative experimental results on the FakeNewsNet dataset (mean ± standard deviation)

[0106]

[0107] Table 4: Comparative experimental results on the PHEME dataset (mean ± standard deviation)

[0108]

[0109] As can be seen from the results in Tables 3 and 4, the method proposed in this invention is significantly better than the existing comparative model on both datasets, especially in the Early-F1 index, where it is improved by an average of 3 to 14 percentage points.

[0110] (2) Ablation test

[0111] To better verify the effectiveness of the model of this invention, an ablation experiment was conducted in this embodiment. To further verify the effectiveness of each key module in the proposed method, a series of ablation experiments were designed to analyze the performance changes of the model after removing or replacing some modules. Specifically, the contributions of greedy initialization and structural prior injection to the overall detection performance were evaluated from two perspectives.

[0112] This embodiment constructs the following three variant models:

[0113] Ours w / o Greedy: Removes the greedy algorithm and adopts a random initialization of the community structure, that is, users are randomly assigned to a fixed number of communities;

[0114] Ours w / o Structural Prior: Removes the structural prior injection module, i.e., does not use community embedding vectors, and only retains the original text and behavioral features;

[0115] Ours w / o Both: Simultaneously removes greedy initialization and structural priors, retaining only the basic propagation graph construction and the original feature input.

[0116] The three models mentioned above were compared with the complete model Ours-GreedyGNN under the same cold start experimental settings.

[0117] Table 4 shows the ablation experiment results on the FakeNewsNet dataset (with Early-F1 as the core metric):

[0118]

[0119] The experimental results above show that:

[0120] After removing the greedy initialization module (without Greedy), the model performance dropped significantly, indicating that a reasonable division of the community structure during the cold start phase plays a crucial role in propagation path modeling.

[0121] Removing the structural prior injection module (without Structural Prior) further reduces Early-F1, indicating that community embedding vectors effectively improve the model's ability to perceive sparse propagation features.

[0122] The performance degradation was most significant when both modules were removed simultaneously (w / o Both), with Early-F1 decreasing by 6.6 percentage points, verifying the synergistic enhancement effect of the two modules proposed in this invention.

[0123] Furthermore, repeated experiments at different time windows (first 5%, 10%, 15%, and 20% of propagation data) showed that structure-guided propagation has a general advantage in modeling early propagation.

[0124] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A cold start method for detecting fake news based on propagation paths, characterized in that, Includes the following steps: S1: Construct a news dissemination graph, generate dissemination paths based on user interaction with news, and calculate the similarity of dissemination paths between users; S2: Based on the similarity of propagation paths, a greedy strategy is used to initialize the unsupervised community, forming a cohesive user community; S3: Inject community structure as a structural prior into the graph neural network to enhance the model's ability to propagate structural awareness during the cold start phase; S4: Utilize graph neural networks with injected structural priors to classify news events as true or false, thereby enabling early detection of fake news.

2. The cold start method for detecting fake news based on propagation paths as described in claim 1, characterized in that: S1 includes the following sub-steps: S1.1: Construct a propagation graph and user propagation structure, and use a greedy algorithm to achieve community segmentation in the cold start phase, thereby assisting in the early detection of fake news; first, construct a news propagation graph to capture the interaction structure between users and news on social platforms; each news event It must contain at least one publisher and several user groups involved in the dissemination Where k is the number of users spreading the message; S1.2: Define the propagation graph as the interactions of users on social media, such as forwarding, commenting, and liking news. ,in Represents a set of users and news nodes. The edges constituted by the propagation behavior; The node feature matrix is ​​denoted as , used to represent features such as user behavior or content representation, where d is the feature dimension; To further capture the structural commonalities among users during the dissemination process, Projection to spread similar images to users In this graph, the edges The weights are defined by three propagation path features: Path co-occurrence frequency The number of times two users participate in the same propagation path; Time window overlap rate The degree of overlap in the time periods of the interactive behaviors; Structural adjacency The reciprocal of the shortest path in the propagation graph; The user propagation path similarity function is defined as: Where α, β, and γ are normalized weighting parameters; this similarity function provides the structural basis for greedy community partitioning; the detection model is defined as a mapping function from news dissemination structure to true / false labels. ,in Indicates to news The corresponding propagation graph structural features include user interaction paths, community structure, and propagation similarity.

3. The cold start method for detecting fake news based on propagation paths as described in claim 1, characterized in that: S2 includes the following sub-steps: S2.1: Community initialization. Based on the definition of similarity of propagation paths, an unsupervised greedy aggregation strategy is set to complete the initial community division under unlabeled conditions. Through the greedy aggregation strategy, the system automatically divides multiple user communities with similar propagation behaviors in the early stage when the propagation graph is not yet complete. S2.2: Community structure optimization. To improve the consistency and accuracy of the partitioning results, the algorithm performs the following optimizations after initialization: Edge node reallocation: For nodes that are connected to multiple communities at the same time, their aggregation similarity is reassessed and their affiliation is adjusted. Weak edge removal and regularization: Remove edges with weights below a threshold to suppress noise propagation path interference and improve community compactness; Ultimately, the optimized community structure will be encoded as a tag embedding or propagation prior, which will be fed into the fake news detection module as input to improve the model's judgment ability during the cold start phase.

4. The cold start method for detecting fake news based on propagation paths as described in claim 3, characterized in that: The greedy aggregation strategy in step S2.1 specifically includes the following sub-steps: S2.1.1: Seed node selection: Select the node with the largest weighted degree from u. As the starting point of the initial seed, that is; S2.1.2: Greedy Expansion: In each iteration, select nodes from the unpartitioned nodes that are in the current community. The node with the highest average similarity, i.e.: S2.1.3: Termination Condition: If Or the community size reaches the maximum threshold. If the community expansion fails, the expansion will terminate and the next one will begin.

5. The cold start method for detecting fake news based on propagation paths as described in claim 1, characterized in that: S3 includes the following sub-steps: S3.1: Input feature modeling, for each news event This corresponds to a propagation diagram. , where the set of nodes Including forwarding users, initial publishers, and news nodes, edge set Represents the propagation path relationship between users or between users and news; node feature matrix This includes: text representation, user behavior statistics, and community tags. The above features are concatenated and then input into a structure-aware graph classification model. S3.2: Structural Prior Injection Mechanism The greedy community partitioning provides each user node with a structural label for aggregated propagation similarity. This structural prior serves as an additional supervisory signal at the node level, providing the model with hints about local propagation consistency. To inject this prior, the study employs the following two methods: Tag embedding method: Convert the one-hot encoding of each user's community into a low-dimensional structured embedding vector. , and concatenate with the original node features; Graph structure edge weight adjustment: Enhance the edge weights between nodes within the same community, weaken the edges across communities, and achieve a biased reconstruction of the propagation graph structure; S3.3: Graph Neural Network Design After structural prior injection, a Graph Convolutional Network (GCN) is used as the backbone model to learn node representations for the news dissemination graph; the model architecture includes: Input features ; Two GCN aggregation layers: in This is the adjusted adjacency matrix. Its degree matrix, These are trainable weights; Graph-level aggregation uses mean pooling or attention pooling to aggregate all nodes; Finally, a fully connected layer is used for true / false binary classification prediction; the embedded nodes are aggregated through an attention pooling mechanism to form a news-level event representation. Ultimately, the news event will be represented The input is fed into a fully connected layer and a softmax classifier to predict the true or false label of the event. ; S3.4: Model Training and Optimization The model training uses the cross-entropy loss function, where the structural prior is only used as an auxiliary input and does not participate in the supervision objective. To adapt to the cold start scenario, only the information of the first 10% of nodes and edges is used during the training process, and the DropEdge data augmentation strategy is added to the structural perturbation.

6. The cold start method for detecting fake news based on propagation paths as described in claim 1, characterized in that: S4 includes the following sub-steps: S4.1: In the cold start scenario, only the first 10% of nodes and edges of the propagation path are used for model training and inference; S4.2: The model is trained using the cross-entropy loss function and combined with the DropEdge strategy to enhance the robustness of the model; S4.3: Evaluate the model's detection performance in the early stages of propagation using the Early-F1 metric to ensure the model has real-time response capabilities.

Images