Rumor detection method and system fusing cross-snapshot memory mechanism and structure pruning

Abstract

The application provides a rumor detection method and system fusing a cross-snapshot memory mechanism and structure pruning, and relates to the technical field of data processing. The method comprises the following steps: constructing a propagation tree based on event data; dividing the propagation tree based on a preset node timestamp to obtain a dynamic graph sequence; performing node feature random masking on each propagation snapshot to obtain enhanced node features of each propagation snapshot; optimizing the dynamic graph sequence based on the enhanced node features to obtain an optimized propagation graph sequence; performing bidirectional graph convolution coding on the optimized propagation graph sequence to obtain current snapshot node representations of each propagation snapshot; performing double-pooling operation on each current snapshot node representation to obtain graph-level representations of each current snapshot node representation; constructing a target memory bank based on the graph-level representations of each current snapshot node representation; and performing rumor classification on the target memory bank based on a classifier to obtain rumor detection results of a target event. The application can improve the accuracy of rumor detection.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a rumor detection method and system that integrates cross-snapshot memory mechanism and structural pruning. Background Technology

[0002] With the rapid development of social media platforms, information dissemination is characterized by speed, wide reach, and high user participation. While this brings convenience to the flow of social information, it also allows false information and rumors to spread rapidly.

[0003] Currently, social media rumor detection generally uses dynamic graph modeling methods. This method encodes each time-series snapshot independently, and there is a lack of explicit information exchange channels between the propagation characteristics of different time stages. It is impossible to continuously transmit key information in the early propagation characteristics to subsequent stages, and it is difficult to use the propagation characteristics of the early stage of rumor to assist in the identification and judgment of the propagation status in the middle and later stages, resulting in low accuracy of rumor detection. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, which can solve the technical problem of low accuracy in rumor detection in the prior art.

[0005] A first aspect of this invention proposes a rumor detection method that integrates cross-snapshot memory mechanisms and structural pruning, comprising:

[0006] Based on the event data of the target event collected from social media platforms, a propagation tree is constructed;

[0007] The propagation tree is segmented based on preset node timestamps to obtain a dynamic graph sequence; wherein, the dynamic graph sequence contains multiple propagation snapshots;

[0008] Randomly mask the node features of each propagation snapshot to obtain the enhanced node features of each propagation snapshot;

[0009] The dynamic graph sequence is optimized based on the enhanced node features of each propagation snapshot to obtain an optimized propagation graph sequence.

[0010] Perform bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence;

[0011] Perform double pooling on the representation of each current snapshot node to obtain a graph-level representation of each current snapshot node representation;

[0012] Construct the target memory based on the graph-level representation of each current snapshot node;

[0013] The target memory is classified into rumors based on a classifier to obtain the rumor detection results for the target event.

[0014] Optionally, constructing a propagation tree based on the event data of the target event collected from social media platforms specifically includes:

[0015] The original posts and other messages are obtained from the event data of the target event collected from social media platforms; wherein the other messages include at least comment messages, reply messages, and repost messages.

[0016] Determine the information propagation relationship between the other messages and the original post;

[0017] Based on the original post, the other messages, and the information propagation relationships, a propagation tree is constructed; wherein the root node of the propagation tree is the original post, the child nodes of the propagation tree are the other messages, and the edges of the propagation tree are the information propagation relationships.

[0018] Optionally, the step of performing a random masking of node features on each propagation snapshot to obtain enhanced node features for each propagation snapshot specifically includes:

[0019] Obtain the random mask matrix;

[0020] Based on the random mask matrix, node feature random masks are applied to each propagation snapshot to obtain enhanced node features for each propagation snapshot.

[0021] Optionally, optimizing the dynamic graph sequence based on the enhanced node features of each propagation snapshot to obtain an optimized propagation graph sequence specifically includes:

[0022] The edge importance data and node centrality data of the dynamic graph sequence are determined based on the enhanced node features of each propagation snapshot.

[0023] Based on the edge importance data and the node centrality data, a comprehensive importance score for each edge is determined;

[0024] High-contribution edges are obtained from the comprehensive importance scores of each edge; wherein the comprehensive importance score of the high-contribution edge is greater than or equal to a preset threshold.

[0025] The dynamic graph sequence is optimized based on the high-contribution edges to obtain an optimized propagation graph sequence.

[0026] Optionally, performing bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence specifically includes:

[0027] Construct a top-down propagation graph and a bottom-up propagation graph based on the optimized propagation graph sequence;

[0028] The top-down propagation graph is encoded using a bidirectional graph convolutional network to obtain a top-down node representation;

[0029] The bottom-up propagation graph is encoded using the bidirectional graph convolutional network to obtain bottom-up node representations;

[0030] The top-down node representation and the bottom-up node representation are concatenated and merged to obtain the node representation of the current snapshot of each propagation snapshot.

[0031] Optionally, the step of performing double-pooling on the representation of each current snapshot node to obtain a graph-level representation of each current snapshot node specifically includes:

[0032] Perform average pooling on each current snapshot node representation to obtain global statistical features;

[0033] Perform max pooling on each current snapshot node representation to obtain local key node features;

[0034] The global statistical features and the local key node features are concatenated and fused to obtain the graph-level representation of each current snapshot node.

[0035] Optionally, the construction of the target memory based on the graph-level representation of each current snapshot node specifically includes:

[0036] Obtain the graph-level representation of the first propagation snapshot from the optimized propagation graph sequence;

[0037] An initialization memory is constructed based on the atlas representation of the first propagation snapshot;

[0038] Based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence, the initialization memory is updated to obtain the target memory; wherein, the other propagation snapshots are propagation snapshots in the optimized propagation graph sequence other than the first propagation snapshot.

[0039] Optionally, updating the initialization memory based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence to obtain the target memory specifically includes:

[0040] The next propagation snapshot corresponding to the first propagation snapshot is determined as the current propagation snapshot;

[0041] The initialization memory is designated as the current memory.

[0042] The current memory is determined as the query vector;

[0043] The current propagation snapshot is determined as a key vector and a value vector;

[0044] The query vector, the key vector, and the value vector are processed using a multi-head attention mechanism to obtain the current attention.

[0045] The current memory bank is updated using the current attention through a gating fusion mechanism to obtain the updated current memory bank;

[0046] If the current propagation snapshot is not the last propagation snapshot in the optimized propagation graph sequence, then the next propagation snapshot corresponding to the current propagation snapshot is determined as the new current propagation snapshot, and the new current propagation snapshot is used to perform the steps from determining the current memory as the query vector to obtaining the updated current memory;

[0047] If the current propagation snapshot is the last propagation snapshot in the optimized propagation graph sequence, then the updated current memory is determined as the target memory.

[0048] A second aspect of this invention proposes a rumor detection system that integrates cross-snapshot memory mechanism and structural pruning, comprising: a processor and a memory;

[0049] The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the rumor detection method that integrates cross-snapshot memory mechanism and structural pruning as described in the first aspect.

[0050] A third aspect of the present invention provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in the first aspect.

[0051] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0052] In this embodiment of the invention, node feature random masking enhances the model's adaptability to missing information and noise interference, avoiding misjudgments caused by incomplete data. A comprehensive importance assessment is used to adaptively prune the propagation graph, effectively eliminating redundant interactions and low-contribution nodes, reducing structural noise interference with key propagation patterns. Furthermore, bidirectional graph convolutional coding fully extracts structural features from both propagation directions, while double pooling simultaneously preserves global topological information and local key node features, making the graph-level representation more discriminative. Continuous accumulation and dynamic updating of historical propagation information through a cross-snapshot memory enables the model to capture long-range temporal dependencies, preventing the loss of important early signals in later stages. The modules work together to ensure the model can accurately identify the essential patterns of rumor propagation even in complex social media environments, thereby effectively improving the accuracy of rumor detection. Attached Figure Description

[0053] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0054] Figure 1 This is a flowchart illustrating a rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, as provided in an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of a rumor detection system that integrates cross-snapshot memory mechanism and structural pruning, provided by an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram of another rumor detection system that integrates cross-snapshot memory mechanism and structural pruning, provided by an embodiment of the present invention. Detailed Implementation

[0057] To enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the technical solutions 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, not all embodiments. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0058] The following detailed description, in conjunction with the accompanying drawings, of the rumor detection method integrating cross-snapshot memory mechanism and structural pruning provided by the present invention through specific embodiments and application scenarios, will be provided in detail.

[0059] Reference manual attached Figure 1 The diagram illustrates a flowchart of a rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, provided by an embodiment of the present invention.

[0060] This invention provides a rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, which may include the following steps:

[0061] S1: Construct a propagation tree based on the event data of the target event collected from social media platforms.

[0062] In this embodiment of the invention, for each target event: the root node represents the original posted text; the child nodes represent comments, replies, or forwarding behaviors; and the edges represent information propagation relationships.

[0063] Constructing the propagation tree: , where: V represents the set of nodes; E represents the set of propagation edges.

[0064] As an optional implementation, S1 can construct a propagation tree based on the event data of the target event collected from social media platforms in the following ways:

[0065] The original posts and other messages are obtained from the event data of the target event collected from social media platforms; wherein the other messages include at least comment messages, reply messages, and repost messages.

[0066] Determine the information propagation relationship between the other messages and the original post;

[0067] Based on the original post, the other messages, and the information propagation relationships, a propagation tree is constructed; wherein the root node of the propagation tree is the original post, the child nodes of the propagation tree are the other messages, and the edges of the propagation tree are the information propagation relationships.

[0068] This implementation method, by acquiring the original post and other messages and clarifying the information propagation relationships between them, constructs a propagation tree with the original post as the root node and other messages as child nodes. This clearly depicts the spread of rumors and user interaction patterns. The propagation tree not only fully preserves the hierarchical structure and directional information during the propagation process but also provides a standardized data foundation for subsequent temporal snapshot partitioning and graph neural network encoding. This ensures the complete transmission of propagation structure information during the modeling process, effectively supporting the model's accurate learning and recognition of propagation patterns.

[0069] S2: The propagation tree is segmented based on preset node timestamps to obtain a dynamic graph sequence.

[0070] In this embodiment of the application, the dynamic graph sequence includes multiple propagation snapshots.

[0071] The propagation tree is segmented based on node timestamps.

[0072] Suppose the propagation process is divided into K time windows:

[0073] Generate a sequence of dynamic graphs:

[0074] in: Indicates the first A snapshot of the spread. This represents the set of nodes within the i-th time window; Represents the set of propagation relationships (edges) between nodes; This represents the feature matrix of the corresponding node. Each row corresponds to the feature vector of a node, which is used to describe the text content, semantic representation or other attribute information of that node.

[0075] Unlike traditional dynamic graph methods, this invention retains all historical snapshot information during the propagation and evolution process, providing a foundation for subsequent cross-snapshot memory modeling.

[0076] S3: Perform random masking of node features on each propagation snapshot to obtain the enhanced node features of each propagation snapshot.

[0077] In this embodiment of the application, in order to simulate the phenomenon of missing information in real social media, the node features are randomly masked.

[0078] Generate a random mask matrix: ;

[0079] The enhanced node features are obtained: ;

[0080] Where: r is the mask rate; This represents element-wise multiplication. The node feature matrix X represents the initial feature set of all nodes in the entire propagation graph (or the current propagation snapshot). Each row corresponds to the feature vector of a node, containing text content, semantic representation, or other node attribute information. Let represent the feature vector of the i-th node, which is a component of the node feature matrix X, and has dimension F. F represents the node feature dimension, that is, the number of features contained in each node's feature vector. Node Feature Matrix , where R represents the real number space, used to indicate that each element in the matrix is ​​a continuous real value, N represents the number of nodes in the current propagation snapshot, and F represents the feature dimension corresponding to a single node. represents the feature vector of the i-th node, which is encoded by text semantic features, user attribute features, or other node-related information, and is a row element in the node feature matrix X.

[0081] By randomly masking some feature dimensions, the robustness of the model to missing and noisy data can be improved.

[0082] As an optional implementation, S3 performs node feature random masking on each propagation snapshot to obtain enhanced node features for each propagation snapshot. This can be achieved through methods such as:

[0083] Obtain the random mask matrix;

[0084] Based on the random mask matrix, node feature random masks are applied to each propagation snapshot to obtain enhanced node features for each propagation snapshot.

[0085] This implementation method selectively masks node features in each snapshot using a random mask matrix, simulating information loss caused by user deletion of posts, privacy settings, and content moderation in social media. This allows the model to learn to infer complete semantic information from incomplete features during the training phase. This mechanism forces the model to avoid over-reliance on any particular feature dimension, significantly enhancing its robustness against missing data and noise disturbances. Even in real-world applications with significant information loss, it maintains stable representation learning performance, effectively improving the robustness and generalization performance of rumor detection.

[0086] S4: Optimize the dynamic graph sequence based on the enhanced node features of each propagation snapshot to obtain an optimized propagation graph sequence.

[0087] In this embodiment of the application, structural optimization is performed on the redundant interaction relationships in the propagation graph.

[0088] First, the propagation graph after node feature masking is input into the graph attention network. For any edge in the propagation graph... Compute nodes With nodes Attention weights.

[0089] Let the node features be respectively and First, a linear transformation is performed:

[0090]

[0091] in: The weight matrix is ​​a learnable weight matrix; and These are the node feature vectors.

[0092] Then calculate the unnormalized attention coefficient:

[0093]

[0094] in For attention parameters; This indicates vector concatenation.

[0095] Finally, Softmax normalization is used:

[0096]

[0097] get , representing a node For nodes The importance of information dissemination also reflects the contribution of the edge to local semantic dissemination.

[0098] Then calculate the node PageRank centrality:

[0099] For the same propagation snapshot, the PageRank value of each node is calculated based on the propagation graph topology.

[0100] initialization:

[0101]

[0102] Subsequent iterative updates:

[0103]

[0104] in: The damping coefficient; For all pointing nodes The set of nodes; For nodes The degree of departure.

[0105] When the iteration converges, we get

[0106] That is, nodes PageRank centrality is used to measure the global influence of a node in the entire propagation network.

[0107] Construct a comprehensive importance score: ;

[0108] in: ;

[0109] Low-contribution edges are filtered out based on a threshold τ: ;

[0110] The optimized propagation graph is obtained: ;

[0111] This process reduces structural noise and computational redundancy.

[0112] As an optional implementation, S4 optimizes the dynamic graph sequence based on the enhanced node features of each propagation snapshot, and the optimized propagation graph sequence can be obtained in the following ways:

[0113] The edge importance data and node centrality data of the dynamic graph sequence are determined based on the enhanced node features of each propagation snapshot.

[0114] Based on the edge importance data and the node centrality data, a comprehensive importance score for each edge is determined;

[0115] High-contribution edges are obtained from the comprehensive importance scores of each edge; wherein the comprehensive importance score of the high-contribution edge is greater than or equal to a preset threshold.

[0116] The dynamic graph sequence is optimized based on the high-contribution edges to obtain an optimized propagation graph sequence.

[0117] This implementation method comprehensively evaluates the propagation structure from two perspectives: semantic contribution and topological influence, by integrating the importance weights of edges and the centrality of nodes. It then selects high-contribution edges with a comprehensive score exceeding a preset threshold. Compared to traditional pruning methods that rely on only a single metric, this multi-perspective evaluation mechanism avoids the accidental deletion of key propagation paths, preserving the core propagation framework while eliminating low-value interactions. The optimized propagation graph sequence not only reduces the computational burden on the graph neural network but also filters out structural noise introduced by a large number of irrelevant nodes, enabling the model to focus on key patterns that truly reflect the essence of rumor propagation, thereby significantly improving the accuracy and efficiency of subsequent encoding and detection.

[0118] S5: Perform bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence.

[0119] In this embodiment of the application, a top-down propagation graph (TD) and a bottom-up propagation graph (BU) are constructed respectively.

[0120] A bidirectional graph convolutional network is used to extract the propagation representation.

[0121] Top-down encoding:

[0122] Bottom-up coding:

[0123] Top-down graph convolutional coding:

[0124] For the TD propagation graph, a graph convolutional network is used to aggregate the neighborhood information of nodes layer by layer.

[0125] For nodes , No. The layer representation is updated as follows:

[0126]

[0127] in: For nodes Neighbors in the TD graph; For the first Layer weight matrix; These are the normalization coefficients of the adjacency matrix; The ReLU activation function is used. After two layers of graph convolution, the result is...

[0128]

[0129] This indicates the propagation characteristics of a node in the direction of rumor spread.

[0130] Bottom-up graph convolutional coding:

[0131] The same structure is used for encoding the BU propagation graph.

[0132] The node update method is

[0133]

[0134] in This is for backpropagation of neighbors.

[0135] The final result is HBU = BiGCNBU(G′).

[0136] This indicates the propagation of information in the direction of user feedback.

[0137] The fusion yields the representation of the current snapshot node: .

[0138] As an optional implementation, S5 performs bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence, which may include:

[0139] Construct a top-down propagation graph and a bottom-up propagation graph based on the optimized propagation graph sequence;

[0140] The top-down propagation graph is encoded using a bidirectional graph convolutional network to obtain a top-down node representation;

[0141] The bottom-up propagation graph is encoded using the bidirectional graph convolutional network to obtain bottom-up node representations;

[0142] The top-down node representation and the bottom-up node representation are concatenated and merged to obtain the node representation of the current snapshot of each propagation snapshot.

[0143] This implementation method constructs both top-down and bottom-up propagation graphs, utilizing a bidirectional graph convolutional network to encode information in two directions: from the source to the end of the propagation and from the end back to the source. This fully exploits the bidirectional dependencies between parent and child nodes in the propagation tree. Top-down encoding effectively captures the spread pattern of rumors from the source to a wide range of users, while bottom-up encoding can perceive the reverse impact of user feedback on the source information. The node representation obtained by concatenating and fusing the two methods incorporates the structural semantics of the bidirectional propagation path. Compared to unidirectional graph neural networks, this bidirectional encoding strategy can extract richer propagation structure information, making the model's characterization of rumor propagation patterns more complete and accurate, thereby effectively improving the accuracy of subsequent detection tasks.

[0144] S6: Perform double pooling on the representation of each current snapshot node to obtain a graph-level representation of each current snapshot node representation.

[0145] Perform a dual-pooling operation on the current snapshot node representation.

[0146] Average pooling:

[0147]

[0148] Max pooling:

[0149]

[0150] in: Indicates the first The output of the layer graph convolutional network is the first Feature representation of each node (node ​​embedding); This indicates the node number, i.e., the first node in the propagation snapshot. One node; This indicates the number of layers in the graph convolutional network, such as the first layer, the second layer, etc. That is, the node features are updated once after each layer in the graph convolutional network, therefore the [layer number]... The node representation of the layer output is denoted as

[0151] By piecing together elements, a hierarchical representation can be formed:

[0152]

[0153] Get the first A graph-level representation of a snapshot.

[0154] As an optional implementation, S6 performs a double-pooling operation on the representation of each current snapshot node to obtain a graph-level representation of each current snapshot node, which may include:

[0155] Perform average pooling on each current snapshot node representation to obtain global statistical features;

[0156] Perform max pooling on each current snapshot node representation to obtain local key node features;

[0157] The global statistical features and the local key node features are concatenated and fused to obtain the graph-level representation of each current snapshot node.

[0158] In this implementation, average pooling effectively extracts the global statistical features of the propagation graph, reflecting the overall structural distribution; while max pooling accurately captures the salient features of key local nodes, highlighting important information about high-influence users or core propagation nodes. By concatenating and fusing these two types of features, the generated graph-level representation simultaneously considers information at both the overall topology and the local key nodes, overcoming the shortcomings of single pooling strategies that easily lose local key information or ignore global distribution characteristics. This results in a graph-level representation containing richer structural discrimination information, significantly enhancing the model's ability to distinguish various rumor propagation patterns, thereby effectively improving rumor detection accuracy.

[0159] S7: Construct the target memory based on the graph-level representation of each current snapshot node.

[0160] In this embodiment of the application, the memory bank is first initialized using the first propagation snapshot: ;

[0161] For subsequent snapshots:

[0162] A multi-head attention mechanism is used to enable the interaction between historical propagation information and the current propagation state.

[0163] Using a memory as a query:

[0164] The current snapshot is represented as a key and value:

[0165] Calculate attention output:

[0166] The memory bank is then updated using a gated fusion mechanism:

[0167] in: ,in, Indicates action on the historical memory bank A learnable weight matrix is ​​used to extract important features of historical propagation information; This indicates the effect on the current attention output. A learnable weight matrix is ​​used to extract important features of the current propagation state.

[0168] Final update result: ;

[0169] Repeat the process until the last snapshot.

[0170] The above mechanisms enable: continuous accumulation of historical propagation information; explicit information transmission across snapshots; long-range temporal dependency modeling; and dynamic propagation evolution tracking.

[0171] As an optional implementation, S7 can construct the target memory based on the graph-level representation of each current snapshot node in the following ways:

[0172] Obtain the graph-level representation of the first propagation snapshot from the optimized propagation graph sequence;

[0173] An initialization memory is constructed based on the atlas representation of the first propagation snapshot;

[0174] Based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence, the initialization memory is updated to obtain the target memory; wherein, the other propagation snapshots are propagation snapshots in the optimized propagation graph sequence other than the first propagation snapshot.

[0175] This implementation first initializes the memory with the graph-level representation of the first propagation snapshot, establishing a historical foundation including the initial propagation state for subsequent temporal modeling. Then, through iterative updates, the graph-level representations of subsequent snapshots are gradually integrated into the memory, ensuring that key information from each stage is continuously transmitted along the timeline rather than being isolated. This mechanism allows the final target memory to incorporate evolutionary information from all time windows from the initial to the final propagation stage, effectively capturing long-range temporal dependencies in the rumor propagation process. It overcomes the shortcomings of existing methods that lose historical information due to independent snapshot encoding, making the target memory's portrayal of propagation dynamics more complete and continuous. This provides more temporally discriminative feature representations for subsequent classification, effectively improving rumor detection accuracy.

[0176] As an optional implementation, updating the initialization memory based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence to obtain the target memory can include the following methods:

[0177] The next propagation snapshot corresponding to the first propagation snapshot is determined as the current propagation snapshot;

[0178] The initialization memory is designated as the current memory.

[0179] The current memory is determined as the query vector;

[0180] The current propagation snapshot is determined as a key vector and a value vector;

[0181] The query vector, the key vector, and the value vector are processed using a multi-head attention mechanism to obtain the current attention.

[0182] The current memory bank is updated using the current attention through a gating fusion mechanism to obtain the updated current memory bank;

[0183] If the current propagation snapshot is not the last propagation snapshot in the optimized propagation graph sequence, then the next propagation snapshot corresponding to the current propagation snapshot is determined as the new current propagation snapshot, and the new current propagation snapshot is used to perform the steps from determining the current memory as the query vector to obtaining the updated current memory;

[0184] If the current propagation snapshot is the last propagation snapshot in the optimized propagation graph sequence, then the updated current memory is determined as the target memory.

[0185] This implementation uses the current memory as the query vector and the current snapshot as the key-value vector. A multi-head attention mechanism is employed to mine the relationships between historical and current information from multiple representation subspaces, effectively capturing long-range dependencies across time steps. The gating fusion mechanism adaptively learns the fusion ratio between historical memory and current attention, avoiding the loss of key information or redundant accumulation caused by simple weighting. The synergistic effect of these two mechanisms ensures that the target memory maintains accurate retention of historical evolution information during continuous updates, guaranteeing the full preservation and effective utilization of the temporal characteristics of the entire rumor propagation process. This significantly improves the model's ability to perceive and detect dynamic propagation patterns.

[0186] S8: Classify the target memory based on the classifier to obtain the rumor detection result of the target event.

[0187] In this embodiment of the application, the final memory is aggregated:

[0188] Input fully connected layer: ;in, The weight matrix represents the weight of the fully connected classification layer, which is used to map the graph-level representation of the target memory to the classification space corresponding to each rumor category. This represents the bias vector of the fully connected classification layer, which is used to correct the bias in the classification results and improve the model's classification ability.

[0189] Output categories: True Rumor, False Rumor, Unverified Rumor, and Non-Rumor, to complete rumor detection.

[0190] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0191] In this embodiment of the invention, node feature random masking enhances the model's adaptability to missing information and noise interference, avoiding misjudgments caused by incomplete data. A comprehensive importance assessment is used to adaptively prune the propagation graph, effectively eliminating redundant interactions and low-contribution nodes, reducing structural noise interference with key propagation patterns. Furthermore, bidirectional graph convolutional coding fully extracts structural features from both propagation directions, while double pooling simultaneously preserves global topological information and local key node features, making the graph-level representation more discriminative. Continuous accumulation and dynamic updating of historical propagation information through a cross-snapshot memory enables the model to capture long-range temporal dependencies, preventing the loss of important early signals in later stages. The modules work together to ensure the model can accurately identify the essential patterns of rumor propagation even in complex social media environments, thereby effectively improving the accuracy of rumor detection.

[0192] Please refer to the following: Figure 2 The rumor detection system, which integrates cross-snapshot memory mechanism and structural pruning, includes a propagation tree construction module, a dynamic graph snapshot generation module, a node feature masking module, a structural pruning module, a bidirectional graph convolutional coding module, a dual-pooling graph representation module, a cross-snapshot memory enhancement module, and a rumor classification module.

[0193] like Figure 2 As shown, the Input Rumor Propagation Tree represents the propagation structure of the event to be detected on social media. Each node in the propagation tree represents a post (including the original post, reposts, or comments), and the connections between nodes represent the information propagation relationships between users, serving as the input data for the entire model.

[0194] Snapshot Construction is a dynamic graph snapshot building module. Based on the timestamps of the propagation nodes, it divides the complete propagation tree into multiple consecutive time windows, generating multiple propagation snapshots to form a dynamic graph sequence. Its function is to discretize the continuous propagation process, making it easier for the model to learn the dynamic evolution characteristics during the propagation process.

[0195] Sequential / Temporal means that the dynamic graph snapshots are generated sequentially according to the time of propagation. Each snapshot represents a time stage in the propagation process, and all snapshots together describe the entire development process of the event.

[0196] GraphSnapshotDataset is a dynamic graph snapshot dataset, representing a collection of data consisting of multiple propagation snapshots. Assuming the propagation process is divided into K time windows, the dynamic graph is represented as follows:

[0197]

[0198] Each of them

[0199]

[0200] They represent: The current set of snapshot nodes; : The current snapshot edge set; : Feature matrix of the current snapshot node.

[0201] Node Feature Masking is a module that randomly masks some node features during the training phase. Its purpose is to simulate real-world social media scenarios such as user post deletion, content editing, missing data, and feature noise, thereby improving the model's robustness in environments with missing data.

[0202] Feature-level random masking randomly sets some dimensions of the node feature vectors to zero.

[0203] Robust Learning under Incomplete Graphs is a robust learning approach for incomplete propagation graphs, where the model can still learn effective representations even if the propagation graph has missing information.

[0204] Load Shedding via GAT is a load offloading (structural pruning) module based on graph attention networks. It uses graph attention networks (GAT) to calculate the importance of each edge and combines it with PageRank centrality to adaptively prune the propagation graph. The purpose is to remove redundant propagation paths, reduce computational complexity, and suppress structural noise.

[0205] Adaptive Edge Pruning dynamically determines whether to retain an edge based on its importance.

[0206] Computational Load Control reduces the number of edges involved in graph convolution calculations, thereby improving training efficiency.

[0207] GAT stands for Graph Attention Network, used to learn the importance interactions between nodes. Attention weights are calculated based on neighborhood information. Used to measure the importance of propagation between nodes.

[0208] PyG Graph Snapshot (TD / BU) is a snapshot of the PyTorch Geometric dynamic graph (bidirectional propagation graph), representing a snapshot of the dynamic graph after structural optimization.

[0209] in:

[0210] TD stands for Top-Down (top-down propagation graph);

[0211] BU stands for Bottom-Up (a graph that propagates from the bottom up).

[0212] Together, they form the BiGCN input.

[0213] Bi-Directional GCN is a bidirectional graph convolutional network that learns information propagation in two directions: the direction of rumor dissemination and the direction of user feedback. Finally, the information from both directions is fused.

[0214] TDRumorGCN (Top-Down) is a top-down propagating graph convolutional network that performs graph convolution in the direction from source node to responding node. Key learning points include: propagation and diffusion patterns; propagation hierarchy; and information diffusion paths.

[0215] BURumorGCN (Bottom-Up) is a bottom-up propagating graph convolutional network that propagates information from the reply node to the source node. Key learning areas include: user feedback; comment backflow; and propagating inverse dependencies.

[0216] Dual Pooling Readout is a dual-pooling readout module that aggregates node representations into a graph-level representation.

[0217] Simultaneously employing: Mean Pooling; Max Pooling.

[0218] Mean pooling calculates the average value of all node representations. It is used to represent the overall statistical characteristics of the propagation graph.

[0219] Max Pooling extracts the maximum activation value from all node representations. It highlights key propagation nodes.

[0220] Concatenate is a feature concatenation technique that joins the average pooling result with the max pooling result.

[0221]

[0222] The final graph-level representation is obtained.

[0223] The Global & Salient Graph Representation is a joint graph representation of the global and key nodes, which represents a propagation graph that simultaneously preserves: the overall propagation trend and the characteristics of key nodes.

[0224] Cross-Snapshot Memory Enhancement is a module that enhances memory across snapshots. Its functions are: to establish long-term memory throughout the entire propagation process; to enable information exchange between snapshots at different times; and to avoid the limitations of traditional dynamic graphs that only perform simple sequence aggregation.

[0225] The Memory Bank is a repository that stores the historical propagation status. It is continuously updated as propagation progresses.

[0226] Multi-Head Attention is a multi-head attention mechanism that calculates the correlation between historical propagation states and the current propagation state, enabling cross-snapshot information interaction.

[0227] Gate Fusion is a gated fusion mechanism that controls how much historical information is retained and how much current information is introduced. It dynamically updates the fusion memory.

[0228] FC + Softmax is a fully connected classifier.

[0229] FC stands for Fully Connected, which is responsible for feature mapping.

[0230] Softmax is responsible for converting the output into probabilities.

[0231] The classification was finally completed.

[0232] True represents true information.

[0233] False indicates false information.

[0234] Unverified indicates that the information is unverified.

[0235] Non-Rumor information is information that is not rumor.

[0236] like Figure 2As shown, this invention first inputs the rumor propagation tree of the target event and divides the propagation process into multiple time-continuous propagation snapshots based on node timestamps, constructing a dynamic graph snapshot sequence. Subsequently, random masks are applied to the node features in each propagation snapshot to simulate information gaps and noise disturbances in real social media, improving the model's robustness to incomplete data.

[0237] Subsequently, a structural pruning module based on graph attention networks and PageRank centrality is introduced to evaluate the importance of redundant edges in the propagation graph and adaptively delete low-contribution propagation paths, resulting in a dynamically optimized graph snapshot. The optimized propagation snapshots are used to construct top-down and bottom-up propagation graphs, respectively, and then input into a bidirectional graph convolutional network for encoding, extracting node representations from the propagation direction and the user feedback direction.

[0238] Furthermore, average pooling and max pooling are performed on the node representation of each propagation snapshot, and the results of the two pooling methods are concatenated to obtain the graph-level representation of the current propagation snapshot. Based on this, the present invention proposes a cross-snapshot memory enhancement module, which initializes the memory bank with the first propagation snapshot and uses a multi-head attention mechanism to realize information interaction between the current propagation state and historical memory. Then, combined with a gating fusion strategy, the memory bank is continuously updated to achieve continuous accumulation of propagation information and long-term temporal dependency modeling.

[0239] Finally, the updated target memory is input into a fully connected classifier, and the probability of the target event belonging to one of four categories—real information, false information, unverified information, or non-rumor information—is output through Softmax to complete the rumor detection.

[0240] Advantages of the overall framework of this invention

[0241] It has stronger dynamic propagation modeling capabilities. Through dynamic graph snapshot construction and cross-snapshot memory enhancement mechanisms, it achieves continuous modeling of the entire propagation process, which can capture long-term temporal dependencies more accurately compared to traditional independent snapshot encoding methods.

[0242] It exhibits higher robustness. The node feature masking mechanism simulates information gaps in real social media scenarios, improving the model's detection stability under conditions of incomplete node features and high text noise.

[0243] The propagation structure is more refined. The structural pruning mechanism based on graph attention networks and PageRank centrality can effectively remove redundant propagation edges, reduce computational complexity, and reduce the interference of structural noise on the model.

[0244] The graph-level representation is more comprehensive. The dual-pooling readout module integrates the results of average pooling and max pooling, taking into account both the overall topological features of the propagation graph and the information of key propagation nodes, thereby improving the propagation representation capability.

[0245] Historical information can be continuously accumulated. The cross-snapshot memory enhancement module uses a memory bank, multi-head attention, and gating fusion mechanism to continuously retain and dynamically update historical propagation information, avoiding the problem of historical information loss in traditional dynamic graph methods.

[0246] Overall detection performance is superior. This invention organically combines node feature enhancement, structural optimization, bidirectional propagation modeling, multi-granularity graph representation, and cross-snapshot memory enhancement, improving the accuracy, robustness, and generalization ability of rumor detection in complex social media environments while ensuring computational efficiency.

[0247] Specifically, the system first uses social media propagation data of the target event as input. The propagation tree construction module organizes the original posts, comments, replies, and forwards into a propagation tree structure according to their propagation relationships. The dynamic graph snapshot generation module segments the propagation tree based on node timestamps, generating a GraphSnapshotDataset dynamic graph snapshot sequence arranged in chronological order. Subsequently, the node feature masking module performs random masking at the feature level using Feature-level Random Masking to simulate information loss scenarios and enhance model robustness. The structure pruning module adopts an adaptive edge pruning strategy based on graph attention networks, filtering out low-contribution connections by calculating the importance score of each edge to control computational load. The bidirectional graph convolutional encoding module constructs two propagation directions, one top-down and one bottom-up, and performs bidirectional encoding using TDRumorGCN and BURumorGCN respectively to capture multi-level dependencies in the propagation structure. The dual-pooling graph representation module concatenates and fuses the global statistical features extracted by average pooling and max pooling in parallel with local key node features to obtain multi-granular graph-level representations of each snapshot. The cross-snapshot memory enhancement module, based on a memory bank and multi-head attention mechanism, achieves continuous accumulation and dynamic updating of historical dissemination information through gating fusion, and explicitly models the temporal evolution process. Finally, the rumor classification results are output through a fully connected layer and a Softmax classifier, and the detected categories include four types: True (real information), False (false information), Unverified (unverified information), and Non-Rumor (non-rumor information).

[0248] The modules described above are connected sequentially according to the data flow order, with the output of the previous module serving as the input of the next module, together forming a complete dynamic rumor detection system.

[0249] Reference manual attached Figure 3 The diagram shows a schematic of another rumor detection system that integrates cross-snapshot memory mechanism and structural pruning, provided by an embodiment of the present invention.

[0250] This invention provides a rumor detection system 30 that integrates cross-snapshot memory mechanism and structural pruning, including: a processor 301 and a memory 302;

[0251] The memory 302 stores programs or instructions that can run on the processor 301. When the program or instructions are executed by the processor 301, they implement the steps of the above-described rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, and can achieve the same technical effect. To avoid repetition, the present invention will not elaborate further.

[0252] It should be understood that the processor 301 in this embodiment of the invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0253] It should also be understood that the memory 302 in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct memory bus RAM (DR RAM).

[0254] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0255] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0256] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0257] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0258] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0259] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0260] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0261] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0262] This invention provides a readable storage medium comprising: storing a program or instructions on the readable storage medium, wherein when the program or instructions are executed by a processor, the program or instructions implement the steps of the above-described rumor detection method that integrates cross-snapshot memory mechanism and structural pruning, and can achieve the same technical effect. To avoid repetition, this invention will not elaborate further.

[0263] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any changes 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 protection scope of the present invention.

Claims

1. A rumor detection method integrating cross-snapshot memory mechanism and structural pruning, characterized in that, include: Based on the event data of the target event collected from social media platforms, a propagation tree is constructed; The propagation tree is segmented based on preset node timestamps to obtain a dynamic graph sequence; wherein, the dynamic graph sequence contains multiple propagation snapshots; Randomly mask the node features of each propagation snapshot to obtain the enhanced node features of each propagation snapshot; The dynamic graph sequence is optimized based on the enhanced node features of each propagation snapshot to obtain an optimized propagation graph sequence. Perform bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence; Perform double pooling on the representation of each current snapshot node to obtain a graph-level representation of each current snapshot node representation; Construct the target memory based on the graph-level representation of each current snapshot node; The target memory is classified into rumors based on a classifier to obtain the rumor detection results for the target event.

2. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in claim 1, characterized in that, The construction of a propagation tree based on event data of target events collected from social media platforms specifically includes: The original posts and other messages are obtained from the event data of the target event collected from social media platforms; wherein the other messages include at least comment messages, reply messages and repost messages; Determine the information propagation relationship between the other messages and the original post; Based on the original post, the other messages, and the information propagation relationships, a propagation tree is constructed; wherein the root node of the propagation tree is the original post, the child nodes of the propagation tree are the other messages, and the edges of the propagation tree are the information propagation relationships.

3. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in claim 1, characterized in that, The step of performing random masking of node features on each propagation snapshot to obtain enhanced node features for each propagation snapshot specifically includes: Obtain the random mask matrix; Based on the random mask matrix, node feature random masks are applied to each propagation snapshot to obtain enhanced node features for each propagation snapshot.

4. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in claim 1, characterized in that, The optimization of the dynamic graph sequence based on the enhanced node features of each propagation snapshot to obtain an optimized propagation graph sequence specifically includes: The edge importance data and node centrality data of the dynamic graph sequence are determined based on the enhanced node features of each propagation snapshot. Based on the edge importance data and the node centrality data, a comprehensive importance score for each edge is determined; High-contribution edges are obtained from the comprehensive importance scores of each edge; wherein the comprehensive importance score of the high-contribution edge is greater than or equal to a preset threshold. The dynamic graph sequence is optimized based on the high-contribution edges to obtain an optimized propagation graph sequence.

5. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in claim 1, characterized in that, The step of performing bidirectional graph convolutional encoding on the optimized propagation graph sequence to obtain the current snapshot node representation of each propagation snapshot contained in the optimized propagation graph sequence specifically includes: Construct a top-down propagation graph and a bottom-up propagation graph based on the optimized propagation graph sequence; The top-down propagation graph is encoded using a bidirectional graph convolutional network to obtain a top-down node representation; The bottom-up propagation graph is encoded using the bidirectional graph convolutional network to obtain bottom-up node representations; The top-down node representation and the bottom-up node representation are concatenated and merged to obtain the node representation of the current snapshot of each propagation snapshot.

6. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning as described in claim 1, characterized in that, The process of performing double pooling on the representations of each current snapshot node to obtain a graph-level representation of each current snapshot node specifically includes: Perform average pooling on each current snapshot node representation to obtain global statistical features; Perform max pooling on each current snapshot node representation to obtain local key node features; The global statistical features and the local key node features are concatenated and fused to obtain the graph-level representation of each current snapshot node.

7. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning according to any one of claims 1 to 6, characterized in that, The construction of the target memory based on the graph-level representation of each current snapshot node specifically includes: Obtain the graph-level representation of the first propagation snapshot from the optimized propagation graph sequence; An initialization memory is constructed based on the atlas representation of the first propagation snapshot; Based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence, the initialization memory is updated to obtain the target memory; wherein, the other propagation snapshots are propagation snapshots in the optimized propagation graph sequence other than the first propagation snapshot.

8. The rumor detection method integrating cross-snapshot memory mechanism and structural pruning according to claim 7, characterized in that, The process of updating the initialization memory based on the graph-level representation of the initialization memory and other propagation snapshots in the optimized propagation graph sequence to obtain the target memory specifically includes: The next propagation snapshot corresponding to the first propagation snapshot is determined as the current propagation snapshot; The initialization memory is designated as the current memory. The current memory is determined as the query vector; The current propagation snapshot is determined as a key vector and a value vector; The query vector, the key vector, and the value vector are processed using a multi-head attention mechanism to obtain the current attention. The current memory bank is updated using the current attention through a gating fusion mechanism to obtain the updated current memory bank; If the current propagation snapshot is not the last propagation snapshot in the optimized propagation graph sequence, then the next propagation snapshot corresponding to the current propagation snapshot is determined as the new current propagation snapshot, and the new current propagation snapshot is used to perform the steps from determining the current memory as the query vector to obtaining the updated current memory; If the current propagation snapshot is the last propagation snapshot in the optimized propagation graph sequence, then the updated current memory is determined as the target memory.

9. A rumor detection system integrating cross-snapshot memory mechanism and structural pruning, characterized in that, include: Processor and memory; The memory stores programs or instructions that can run on the processor, which, when executed by the processor, implement the steps of the rumor detection method that integrates cross-snapshot memory mechanism and structural pruning as described in any one of claims 1 to 8.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the rumor detection method that integrates cross-snapshot memory mechanism and structural pruning as described in any one of claims 1 to 8.

Images