A social network propagation source detection method independent of propagation model
By combining multi-dimensional feature extraction and graph neural network models with the Transformer model, the efficiency and accuracy issues of propagation source detection in social networks are solved, achieving efficient source detection in complex scenarios.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNICOM DIGITAL TECNOLOGY CO LTD
- Filing Date
- 2026-01-19
- Publication Date
- 2026-06-05
AI Technical Summary
Existing social network propagation source detection technologies suffer from insufficient efficiency and accuracy in capturing and modeling propagation snapshot time-series information, as well as in representing node topology features and community features in complex scenarios with unknown propagation models.
It employs multi-dimensional feature extraction and graph neural network models, combining the LPSI algorithm and the Transformer model. It uses graph neural networks and the Transformer self-attention mechanism to fuse features, dynamically capture the propagation source, support input of variable-length sequences, and optimize the training process through a node removal strategy.
It significantly improves the accuracy and versatility of source detection tasks in real social network scenarios, and can effectively construct features even when there is a lack of labeled data, adapting to the learning of propagation patterns in complex scenarios.
Smart Images
Figure CN122155708A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of social network propagation detection technology, and specifically to a method for detecting social network propagation sources that does not rely on propagation models. Background Technology
[0002] With the rapid growth of social networks and their penetration into all aspects of social life, detecting the source of rumors and misinformation on social networks is of great significance for public opinion control. A robust influence propagation source tracing model can quickly identify and segment high-intensity areas of topical information on social networks, rapidly construct propagation scores for large-scale social network nodes or communities, and thus accurately locate high-intensity propagation areas and even the source of propagation in the early stages of influential topics such as public opinion and rumors, thereby achieving effective control over social public opinion and the spread of rumors.
[0003] Meanwhile, the work of tracing the source of transmission is also applicable to similar epidemiological models. Based on the infection status and contact relationships of infected individuals obtained from epidemiological investigations, a good source tracing algorithm should be able to quickly narrow down and locate high-risk areas of transmission within and between communities, significantly reducing the time required to find the initial spread set or area of infectious diseases, and improving the ability to rapidly control infectious diseases.
[0004] Most existing social network propagation source detection technologies are not applicable to scenarios where the prior propagation model is unknown, and their generalization ability in complex real-world propagation scenarios is poor. Existing source detection technologies that do not rely on propagation models have a series of shortcomings in capturing and modeling propagation snapshot time-series information, representing node topological features and community features, and the scalability of multimodal representation of node label information, which affect the efficiency of source detection in real social network scenarios. Summary of the Invention
[0005] This invention aims to overcome the aforementioned shortcomings of existing technologies by providing a social network propagation source detection method that does not rely on propagation models, thereby improving the accuracy and versatility of source detection tasks in various scenarios.
[0006] The present invention provides a method for detecting social network propagation sources that does not rely on propagation models, comprising the following steps: Step S1: Based on the input of the social network graph with users as nodes and interaction relationships as edges, obtain the social network graph topology and label snapshot sequence. Extract features in three dimensions: single label propagation snapshot, overall label snapshot propagation sequence, and graph topology to obtain multi-dimensional features. Step S2: Build three independent graph neural network models, input multi-dimensional features into the independent graph neural network models for representation learning, and finally broadcast and concatenate the outputs of each module to obtain intermediate features; The structure of the graph neural network model is as follows: starting from the input layer, it sequentially connects multiple GATv2 units, fully connected layers, residual connected layers and output layers; The GATv2 unit is used to extract deep feature information of the input features; the fully connected layer is used to project the high-dimensional features output by multiple GATv2 units back to the same dimension as the original input features; the residual connection layer is used to aggregate the output and input features of the fully connected layer; and the output layer is used for the final output of the graph neural network branch.
[0007] Step S3: The intermediate features are processed through a multi-layer perceptron mixing layer to obtain fused multi-dimensional features; the structure of the multi-layer perceptron mixing layer is as follows: starting from the input layer, it is connected sequentially to the Patch Mixing layer, the Channel Mixing layer, and the output layer; The input layer processes variable-length input sequences by padding or culling; the Patch Mixing layer mixes all feature channels within each label snapshot; the Channel Mixing layer mixes each feature channel across all label snapshots; and the output layer linearly projects the mixed features onto the input dimension of the Encoder-Only Transformer model.
[0008] Step S4: Train the Encoder-Only Transformer model by combining and fusing multidimensional features.
[0009] Step S5: Use the trained Encoder-Only Transformer model for social network propagation source detection.
[0010] Preferably, the specific method for obtaining the tag snapshot sequence in step S1 is as follows: Step S11: Determine the observation period and divide the observation period into time windows, wherein the observation period is the interval from the initial state to the infection rate reaching the set threshold; Step S12: Determine and record the status of all nodes within each time window, and generate label snapshot data representing the network status within each time window; Step S13: Arrange the tag snapshot data of all time windows to form a tag snapshot sequence.
[0011] Preferably, the state of the node includes infected and uninfected; The state of the node can be directly determined by the supervision data input from the social network graph, and the state of the node is recorded as 1 or -1 depending on whether it is infected.
[0012] Preferably, the specific method for feature extraction at the single-label propagation snapshot dimension in step S1 is as follows: Step S1-1: Based on the tag snapshot sequence, calculate the incremental information and cumulative percentage information; Step S1-2: Based on the label snapshot data and cumulative percentage data, obtain their original labels, positive labels and negative labels respectively, and extract six high-order features using the LPSI algorithm; Step S1-3: Concatenate the incremental information with the higher-order features to obtain the features in the node label snapshot dimension.
[0013] Furthermore, the incremental information in step S1-1 includes a node's first infection identifier and a node's first recovery identifier, which correspond to Boolean values indicating whether node i is infected for the first time and whether it has recovered to an uninfected state for the first time in the current snapshot, respectively; the cumulative proportion is the cumulative proportion of positive and negative labels in the propagation, and the calculation method is shown in the following formula: In the formula, This represents the cumulative percentage of node i within time window number m.
[0014] Furthermore, in steps S1-2, the original labels, positive labels, and negative labels of the two types of data are as follows: the original label directly corresponds to the vector representation of the data; the positive label corresponds to the vector representation of the data where all -1s are set to 0; and the negative label corresponds to the vector representation of the data where all 1s are set to 0.
[0015] Furthermore, the LPSI algorithm described in step S1-2 is shown in the following formula: In the formula, S is the regularized Laplacian matrix after processing the adjacency matrix of the social network graph, α is the proportion parameter that controls the saliency score obtained from its neighbors and its own label state, and y is a vector that can take the incremental information and cumulative proportion information described in steps S1-1 and S1-2, as well as the original label, positive label and negative label corresponding to the label snapshot data and cumulative proportion data, respectively.
[0016] Preferably, the feature extraction of the overall propagation sequence dimension of the tag snapshot in step S1 includes: for data that provides a prior propagation model, directly using the model parameters; For data for which no prior propagation model is provided, heuristic methods are used to estimate propagation parameters.
[0017] Preferably, the feature extraction method for the graph topology dimension in step S1 is to use the classic graph algorithm and the Node2Vec algorithm.
[0018] Preferably, the multilayer perceptron hybrid layer in step S3 supports variable-length input sequences. The input is adjusted to fit different length sequences through filtering or zero-padding strategies, and the processed input is divided into M and K. o Two dimensions, where M is the length of the tag snapshot sequence, and K... o The length of the intermediate feature.
[0019] Preferably, the specific structure of the Patch Mixing layer and the Channel Mixing layer in step S3 is as follows: starting from the input layer, a parallel multilayer perceptron, a normalization layer, a residual connection layer and an output layer are connected in sequence. The single multilayer perceptron includes a fully connected layer, an activation function layer, and a fully connected layer for achieving cross-snapshot information mixing within a channel or cross-channel information mixing within a node; the normalization layer is used to accelerate training convergence; the residual connection layer is used to aggregate the output and input features of the fully connected layer; and the output layer is used to adjust the output shape to meet the input shape of the next module.
[0020] Furthermore, the Patch Mixing layer comprises M independent multilayer perceptrons, each with input and output dimensions of K. o Each multilayer perceptron independently processes data from a single label snapshot across all feature channels. The transpose of the output after normalization is used as the input to the Channel Mixing layer. The Channel Mixing layer is set to K. o Each multilayer perceptron has an input and output dimension of M. Each multilayer perceptron independently processes data from one feature channel across all label snapshot sequences. The transpose of the output after normalization is used as the transpose back to the original shape as the output of the entire module.
[0021] Preferably, in step S4, the Encoder-Only Transformer model combines temporal positional encoding and feature vector centrality to construct a token sequence. During the training and prediction phases, a node elimination strategy based on random negative sampling is used to dynamically mask non-critical nodes. The scoring calculation unit is used to reorganize the output of the Transformer and obtain the source confidence score of each node through weighted average pooling. Furthermore, the execution logic of the node removal strategy is as follows: Before model training and prediction, the set of nodes to be removed is determined in advance using a random negative sampling algorithm; In the self-attention calculation of the last layer encoder, all tokens belonging to the set of nodes to be removed are located and masked. The output value of the masked tokens is directly set to -∞ so that they cannot be selected as source nodes in subsequent linear layers and pooling.
[0022] Furthermore, the formula for calculating the weighted average pooling of the scoring calculation unit is as follows: In the formula, the generation formula for the weight sequence S is: Generate an arithmetic progression weight sequence; These represent the proportion of infected nodes out of all nodes in the latest and earliest tag snapshot data, respectively; M is the length of the tag snapshot sequence. This represents the actual weight.
[0023] The innovation of this invention is: Multi-dimensional and dynamic evolution feature analysis: This invention extracts features from three independent dimensions (label propagation snapshot, overall features of the propagation sequence, and graph topology), combines LPSI algorithm enhancement, Node2Vec graph embedding, and other techniques, and achieves dynamic fusion through graph neural networks and Transformer self-attention mechanism. It overcomes the limitations of traditional methods that rely on single-dimensional or static topology analysis, significantly improving the ability to capture the dynamic evolution of the propagation source.
[0024] Optimization of node removal strategy: This invention also uses a node removal strategy in the training and prediction of the Transformer layer. This method can greatly reduce the computation of self-attention in the last Encoder layer. At the same time, in scenarios with a small amount of data, it can prevent overfitting caused by uneven distribution of source points in the training data.
[0025] Adaptability of input data length: Variable-length sequence inputs are supported in the multidimensional feature fusion module of the Multilayer Perceptron Mixer (MLP-Mixer). For longer sequence inputs, a batch of label snapshots are filtered out using a certain strategy to meet the requirements; for shorter sequence inputs, efficiency is optimized through zero-padding and skipping computations.
[0026] The advantages of this invention are: 1) Compared to source tracing methods based on pre-defined models, this invention adopts a Transformer architecture that does not require model assumptions. It flattens the graph structure and temporal features into a one-dimensional sequence for self-attention computation, directly learning the propagation rules through data-driven learning, which is more suitable for complex scenarios in real social networks. Compared to source tracing methods that do not require model assumptions, this invention extracts features from multiple independent dimensions, breaking through the limitations of traditional methods that rely on a single dimension or static topological analysis. It significantly improves the ability to capture the dynamic evolution of the propagation source, thus solving the problem that existing technologies cannot accurately and efficiently handle source detection tasks in real social network scenarios. 2) This invention employs a design that uses independent GNN modules to process features of different dimensions before broadcasting and concatenating them, avoiding the feature overload problem caused by directly mixing heterogeneous data. Subsequently, through a phased fusion of MLP mixing layers and adaptive Transformers (first channel-temporal mixing, then graph-sequence joint modeling), the synergistic optimization of graph structure dependency and propagation temporality is achieved. 3) This invention enhances label propagation values and heuristic propagation parameter filling through the LPSI algorithm, enabling the construction of effective features even when complete labeled data is lacking; at the same time, it combines supervised training with binary cross-entropy, enabling the model to adapt to sparse labeled scenarios in real social networks and to be quickly fine-tuned using limited labeled data. Attached Figure Description
[0027] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0028] The technical solution of the present invention will now be described with reference to the accompanying drawings.
[0029] The purpose of this invention is to detect the propagation source of social networks without relying on propagation models. The process of the method is as follows: Figure 1 As shown, it includes three steps: 1. Data collection and multi-dimensional feature extraction. 2. Graph neural network representation. 3. Sequence prediction model representation and output.
[0030] To achieve the above objectives, this invention proposes a social network propagation source detection method that does not rely on a propagation model. The specific steps are as follows: Step S1: Based on the input of the social network graph with users as nodes and interaction relationships as edges, obtain the social network graph topology and label snapshot sequence. Extract features in three dimensions: single label propagation snapshot, overall label snapshot propagation sequence, and graph topology to obtain multi-dimensional features. Step S2: Build three independent graph neural network models, input multi-dimensional features into the independent graph neural network models for representation learning, and finally broadcast and concatenate the outputs of each module to obtain intermediate features; The structure of the graph neural network model is as follows: starting from the input layer, it sequentially connects multiple GATv2 units, fully connected layers, residual connected layers and output layers; The GATv2 unit is used to extract deep feature information of the input features; the fully connected layer is used to project the high-dimensional features output by multiple GATv2 units back to the same dimension as the original input features; the residual connection layer is used to aggregate the output and input features of the fully connected layer; and the output layer is used for the final output of the graph neural network branch.
[0031] Step S3: The intermediate features are processed through a multi-layer perceptron mixing layer to obtain fused multi-dimensional features; the structure of the multi-layer perceptron mixing layer is as follows: starting from the input layer, it is connected sequentially to the Patch Mixing layer, the Channel Mixing layer, and the output layer; The input layer processes variable-length input sequences by padding or culling; the Patch Mixing layer mixes all feature channels within each label snapshot; the Channel Mixing layer mixes each feature channel across all label snapshots; and the output layer linearly projects the mixed features onto the input dimension of the Encoder-Only Transformer model.
[0032] Step S4: Train the Encoder-Only Transformer model by combining and fusing multidimensional features.
[0033] Step S5: Use the trained Encoder-Only Transformer model for social network propagation source detection.
[0034] Preferably, the specific method for obtaining the tag snapshot sequence in step S1 is as follows: Step S11: Determine the observation period and divide the observation period into time windows, wherein the observation period is the interval from the initial state to the infection rate reaching the set threshold; Step S12: Determine and record the status of all nodes within each time window, and generate label snapshot data representing the network status within each time window; Step S13: Arrange the tag snapshot data of all time windows to form a tag snapshot sequence.
[0035] Preferably, the state of the node includes infected and uninfected; The state of the node can be directly determined by the supervision data input from the social network graph, and the state of the node is recorded as 1 or -1 depending on whether it is infected.
[0036] Preferably, the specific method for feature extraction at the single-label propagation snapshot dimension in step S1 is as follows: Step S1-1: Based on the tag snapshot sequence, calculate the incremental information and cumulative percentage information; Step S1-2: Based on the label snapshot data and cumulative percentage data, obtain their original labels, positive labels and negative labels respectively, and extract six high-order features using the LPSI algorithm; Step S1-3: Concatenate the incremental information with the higher-order features to obtain the features in the node label snapshot dimension.
[0037] Furthermore, the incremental information in step S1-1 includes a node's first infection identifier and a node's first recovery identifier, which correspond to Boolean values indicating whether node i is infected for the first time and whether it has recovered to an uninfected state for the first time in the current snapshot, respectively; the cumulative proportion is the cumulative proportion of positive and negative labels in the propagation, and the calculation method is shown in the following formula: In the formula, This represents the cumulative percentage of node i within time window number m.
[0038] Furthermore, in steps S1-2, the original labels, positive labels, and negative labels of the two types of data are as follows: the original label directly corresponds to the vector representation of the data; the positive label corresponds to the vector representation of the data where all -1s are set to 0; and the negative label corresponds to the vector representation of the data where all 1s are set to 0.
[0039] Furthermore, the LPSI algorithm described in step S1-2 is shown in the following formula: In the formula, S is the regularized Laplacian matrix after processing the adjacency matrix of the social network graph, α is the proportion parameter that controls the saliency score obtained from its neighbors and its own label state, and y is a vector that can take the incremental information and cumulative proportion information described in steps S1-1 and S1-2, as well as the original label, positive label and negative label corresponding to the label snapshot data and cumulative proportion data, respectively.
[0040] Preferably, the feature extraction of the overall propagation sequence dimension of the tag snapshot in step S1 includes: for data that provides a prior propagation model, directly using the model parameters; For data for which no prior propagation model is provided, heuristic methods are used to estimate propagation parameters.
[0041] Preferably, the feature extraction method for the graph topology dimension in step S1 is to use the classic graph algorithm and the Node2Vec algorithm.
[0042] Preferably, the multilayer perceptron hybrid layer in step S3 supports variable-length input sequences. The input is adjusted to fit different length sequences through filtering or zero-padding strategies, and the processed input is divided into M and K. o Two dimensions, where M is the length of the tag snapshot sequence, and K... o The length of the intermediate feature.
[0043] Preferably, the specific structure of the Patch Mixing layer and the Channel Mixing layer in step S3 is as follows: starting from the input layer, a parallel multilayer perceptron, a normalization layer, a residual connection layer and an output layer are connected in sequence. The single multilayer perceptron includes a fully connected layer, an activation function layer, and a fully connected layer for achieving cross-snapshot information mixing within a channel or cross-channel information mixing within a node; the normalization layer is used to accelerate training convergence; the residual connection layer is used to aggregate the output and input features of the fully connected layer; and the output layer is used to adjust the output shape to meet the input shape of the next module.
[0044] Furthermore, the Patch Mixing layer comprises M independent multilayer perceptrons, each with input and output dimensions of K. o Each multilayer perceptron independently processes data from a single label snapshot across all feature channels. The transpose of the output after normalization is used as the input to the Channel Mixing layer. The Channel Mixing layer is set to K. o Each multilayer perceptron has an input and output dimension of M. Each multilayer perceptron independently processes data from one feature channel across all label snapshot sequences. The transpose of the output after normalization is used as the transpose back to the original shape as the output of the entire module.
[0045] Preferably, in step S4, the Encoder-Only Transformer model combines temporal positional encoding and feature vector centrality to construct a token sequence. During the training and prediction phases, a node elimination strategy based on random negative sampling is used to dynamically mask non-critical nodes. The scoring calculation unit is used to reorganize the output of the Transformer and obtain the source confidence score of each node through weighted average pooling. Furthermore, the execution logic of the node removal strategy is as follows: Before model training and prediction, the set of nodes to be removed is determined in advance using a random negative sampling algorithm; In the self-attention calculation of the last layer encoder, all tokens belonging to the set of nodes to be removed are located and masked. The output value of the masked tokens is directly set to -∞ so that they cannot be selected as source nodes in subsequent linear layers and pooling.
[0046] Furthermore, the formula for calculating the weighted average pooling of the scoring calculation unit is as follows: In the formula, the generation formula for the weight sequence S is: Generate an arithmetic progression weight sequence; These represent the proportion of infected nodes out of all nodes in the latest and earliest tag snapshot data, respectively; M is the length of the tag snapshot sequence. This represents the actual weight.
[0047] From the perspective of information processing flow, this invention can be divided into the following three parts: Part 1: Data Collection and Multidimensional Feature Extraction; To construct a tag snapshot vector with temporal information, it is first necessary to collect multiple tag propagation snapshots of the tag propagation process in the social network graph structure. Feature extraction and modeling are performed in three dimensions: individual tag propagation snapshots, the overall propagation sequence of tag snapshots, and graph topology. This series of data is the main raw data of this model.
[0048] The steps for data collection and multi-dimensional feature extraction are as follows: Step 1-1: Initial state, input the relational data of the social network to construct a social network G=(V,E) with nodes as users and edges as relations, and a sequence of label snapshots, where V represents the set of nodes and E represents the set of edges; Step 1-2-1: Using the original label data y m To obtain incremental information (v) of the label data 1stInfected v 1stRecovered ) and cumulative percentage (y acc ), using the original label data y m With cumulative value y acc This yields their corresponding positive labels, negative labels, and original labels; Step 1-2-2: Based on the original label data y m With cumulative value y accd The corresponding positive labels, negative labels, and original labels are combined with the social network graph, and a total of six higher-order features (v) are obtained through the iterative LPSI algorithm. LPSI+ v LPSI- v LPSI v LPSIacc + v LPSIacc- v LPSIacc ).
[0049] Steps 1-2-3: Concatenate the incremental information with the higher-order features to obtain the feature data of the node label snapshot dimension.
[0050] Steps 1-3: Tag snapshot of the overall propagation sequence dimensional features, i.e., propagation model features. For data inputs that provide prior propagation model parameters, the in-infection rate and out-of-infection rate are both set as the infection rate of the nodes specified by the model, while the recovery rate is the node recovery rate specified only by the SIR model; for data inputs that do not provide prior propagation model parameters, the corresponding parameters for those without prior propagation model parameters are estimated using heuristic methods. Steps 1-4: For graph topology dimension features, construct graph topology features based on a series of classic algorithms and the Node2Vec algorithm; Part Two: Graph Neural Network Representation; This invention constructs three graph neural network representation modules independently for the feature data of the above three dimensions, processes the input data of the three dimensions independently to represent graph topology information, and adjusts its shape through connection layers, broadcasting and splicing.
[0051] The steps to obtain a graph neural network representation of multi-dimensional features are as follows: Step 2-1: Initial state, input the three dimensions of features extracted in step 1; Step 2-2: Construct three graph neural network representation modules independently for the input data of the above three dimensions, and obtain their representations by passing through multiple stacked graph neural networks. Step 2-3: Transform the representation obtained in Step 2-2 into an output vector with the same dimension as the input through a fully connected layer and an activation function layer, and then perform a residual connection between the output vector and the original input features. Steps 2-4: Broadcast and stitch together the features of each dimension after residual connection according to the feature shape of the single label snapshot dimension to obtain intermediate feature output with a uniform shape.
[0052] Part Three: Sequence Prediction Model Characterization and Output; The sequence model representation consists of two parts: a multidimensional feature fusion module based on a multilayer perceptron hybrid layer (MLP-Mixer) and a self-attention computation and model prediction output based on a sequence prediction model.
[0053] The steps for characterizing and outputting a sequence prediction model are as follows: Step 3-1: Initial state, input the intermediate features of the multi-dimensional features; Step 3-2: The multi-dimensional feature fusion module based on the multi-layer perceptron mixing layer (MLP-Mixer) mixes the label snapshot temporal dimension and multi-channel dimension of the intermediate feature output vector to initially construct an intermediate feature output that mixes multi-channel and temporal information; Step 3-3: The self-attention module based on the sequence prediction model adopts the Encoder-Only Transformer model, which flattens the two-dimensional feature data spanned by the graph structure and the time series structure into a one-dimensional token sequence based on the original features, position features and graph topology features. This sequence is then subjected to self-attention calculation to obtain the prediction output of the one-dimensional sequence and is then reorganized into a two-dimensional result O in the form of graph batch data. Steps 3-4: The model heuristically weights the output O of the Transformer layer by the self-attention value to obtain the final source detection prediction value, and trains it based on the real data using binary cross-entropy loss.
[0054] Steps 3-5: Obtain the source detection results based on the specified recall rate and the final source detection prediction value.
[0055] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for detecting the source of social network propagation that does not rely on a propagation model, characterized in that, Includes the following steps: Step S1: Based on the input of the social network graph with users as nodes and interaction relationships as edges, obtain the social network graph topology and label snapshot sequence. Extract features in three dimensions: single label propagation snapshot, overall label snapshot propagation sequence, and graph topology to obtain multi-dimensional features. Step S2: Build three independent graph neural network models, input multi-dimensional features into the independent graph neural network models for representation learning, and finally broadcast and concatenate the outputs of each module to obtain intermediate features; The structure of the graph neural network model is as follows: starting from the input layer, it sequentially connects multiple GATv2 units, fully connected layers, residual connected layers and output layers; The GATv2 unit is used to extract deep feature information from the input features; The fully connected layer is used to project the high-dimensional features output by multiple GATv2 units back to the same dimension as the original input features; the residual connected layer is used to aggregate the output and input features of the fully connected layer; the output layer is used for the final output of this graph neural network branch. Step S3: The intermediate features are processed through a multi-layer perceptron mixing layer to obtain fused multi-dimensional features; the structure of the multi-layer perceptron mixing layer is as follows: starting from the input layer, it is connected sequentially to the Patch Mixing layer, the Channel Mixing layer, and the output layer; The input layer processes variable-length input sequences by padding or culling; the Patch Mixing layer mixes all feature channels within each label snapshot; the Channel Mixing layer mixes each feature channel across all label snapshots; and the output layer linearly projects the mixed features onto the input dimension of the Encoder-OnlyTransformer model. Step S4: Train the Encoder-Only Transformer model by combining and fusing multi-dimensional features; Step S5: Use the trained Encoder-Only Transformer model for social network propagation source detection.
2. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The specific method for obtaining the tag snapshot sequence in step S1 is as follows: Step S11: Determine the observation period and divide the observation period into time windows, wherein the observation period is the interval from the initial state to the infection rate reaching the set threshold; Step S12: Determine and record the status of all nodes within each time window, and generate labeled snapshot data representing the network status within each time window; the status of a node includes infected and uninfected, and is recorded as 1 or -1 depending on whether it is infected, which is determined by the supervision data input to the social network graph; Step S13: Arrange the tag snapshot data of all time windows to form a tag snapshot sequence.
3. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The specific method for feature extraction at the single-label propagation snapshot dimension in step S1 is as follows: Step S1-1: Based on the tag snapshot sequence, calculate the incremental information and cumulative percentage information; Step S1-2: Based on the label snapshot data and cumulative percentage data, obtain their original labels, positive labels and negative labels respectively, and extract six high-order features using the LPSI algorithm; Step S1-3: Concatenate the incremental information with the higher-order features to obtain the features in the node label snapshot dimension.
4. The method for detecting social network propagation sources that does not rely on propagation models according to claim 3, characterized in that: The incremental information in step S1-1 includes the node's first infection identifier and the node's first recovery identifier, which correspond to the Boolean values indicating whether node i was infected for the first time and whether it recovered to an uninfected state for the first time in the current snapshot, respectively; the cumulative proportion is the cumulative proportion of positive and negative labels in the propagation, and the calculation method is shown in the following formula: In the formula, This represents the cumulative percentage of node i within time window number m.
5. The method for detecting social network propagation sources that does not rely on propagation models according to claim 3, characterized in that: The original labels, positive labels, and negative labels of the two types of data mentioned in steps S1-2 are as follows: the original labels directly correspond to the vector representation of the data; the positive labels correspond to the vector representation of the data where all -1s are set to 0; and the negative labels correspond to the vector representation of the data where all 1s are set to 0. The LPSI algorithm described in step S1-2 is shown in the following formula: In the formula, S is the regularized Laplacian matrix after processing the adjacency matrix of the social network graph, α is the proportion parameter that controls the saliency score obtained from its neighbors and its own label state, and y is a vector that can take the incremental information and cumulative proportion information described in steps S1-1 and S1-2, as well as the original label, positive label and negative label corresponding to the label snapshot data and cumulative proportion data, respectively.
6. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The feature extraction of the overall propagation sequence dimension of the label snapshot in step S1 includes: for data that provides a prior propagation model, directly using the model parameters; For data without a prior propagation model, heuristic methods are used to estimate propagation parameters. The feature extraction method for the graph topology dimension in step S1 is to use the classic graph algorithm and the Node2Vec algorithm.
7. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The multilayer perceptron hybrid layer described in step S3 supports variable-length input sequences. By using filtering or zero-padding strategies, the input is adjusted to fit different length sequences, and the processed input is divided into M and K. o Two dimensions, where M is the length of the tag snapshot sequence, and K... o The length of the intermediate feature.
8. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The specific structures of the Patch Mixing layer and Channel Mixing layer mentioned in step S3 are as follows: starting from the input layer, the parallel multilayer perceptron, the normalization layer, the residual connection layer and the output layer are connected in sequence. The single multilayer perceptron includes a fully connected layer, an activation function layer, and a fully connected layer to achieve cross-snapshot information mixing within a channel or cross-channel information mixing within a node; the normalization layer is used to accelerate training convergence; and the residual connection layer is used to aggregate the output and input features of the fully connected layer. The output layer is used to adjust the output shape to meet the input shape of the next module; The Patch Mixing layer consists of M independent multilayer perceptrons, each with K input and output dimensions. o Each multilayer perceptron independently processes data from a single label snapshot across all feature channels. The transpose of the output after normalization is used as the input to the Channel Mixing layer. The Channel Mixing layer is set to K. o Each multilayer perceptron has an input and output dimension of M. Each multilayer perceptron independently processes data from one feature channel across all label snapshot sequences. The transpose of the output after normalization is used as the transpose back to the original shape as the output of the entire module.
9. The method for detecting social network propagation sources that does not rely on propagation models according to claim 1, characterized in that: The Encoder-Only Transformer model described in step S4 combines temporal positional encoding and feature vector centrality to construct a token sequence. During the training and prediction phases, it dynamically filters out non-critical nodes based on a node elimination strategy using random negative sampling. The scoring calculation unit is used to reshape the output of the Transformer and obtain the source confidence score of each node through weighted average pooling.
10. A method for detecting social network propagation sources that does not rely on propagation models, as described in claim 9, characterized in that: The execution logic of the node removal strategy is as follows: Before model training and prediction, the set of nodes to be removed is determined in advance using a random negative sampling algorithm; In the self-attention calculation of the last layer encoder, all tokens belonging to the set of nodes to be removed are located and masked. The output value of the masked tokens is directly set to -∞ so that they cannot be selected as source nodes in subsequent linear layers and pooling. The formula for calculating the weighted average pooling of the scoring calculation unit is as follows: In the formula, the generation formula for the weight sequence S is: Generate an arithmetic progression weight sequence; These represent the proportion of infected nodes out of all nodes in the latest and earliest tag snapshot data, respectively; M is the length of the tag snapshot sequence. This represents the actual weight.