Telecommunication network false information detection method and system based on graph-level learning

Abstract

The invention discloses a telecommunication network false information detection method and system based on graph-level learning. The method comprises the following steps: acquiring communication data of a communication service platform, and performing structured processing to obtain a structured propagation event data set; constructing a bidirectional sub-cascade graph based on a communication data propagation path in the structured propagation event data set; performing multi-scale feature extraction on the bidirectional sub-cascade graph based on a pre-constructed multi-scale perception false information feature extraction model to obtain a feature vector representing an information propagation event time sequence structure and a dynamic communication data propagation path; performing mean aggregation on the feature vectors to obtain global graph embedding of the target information event; and embedding the global graph into a false classifier of input communication data, and outputting a classification result of the target information event. According to the method, false information is accurately and efficiently identified from mass data based on a multi-scale perception false information feature extraction model, and false information propagation can be restrained.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information security, and more particularly to a telecommunication network false information detection method and system based on graph-level learning. BACKGROUND

[0002] With the popularization of 5G communication technology and the wide application of intelligent terminals, traditional communication methods such as short message and call are still important carriers for the spread of false information.

[0003] However, the existing false information detection technology still has significant limitations: First, the existing technology mostly uses a one-way propagation model (such as a chain structure of "sender → receiver"), which can only depict the forward diffusion process of information and ignores the reverse interaction mechanism of "receiver feedback → sender secondary propagation". For example, fraudsters often adjust their tactics based on the content of the receiver's reply (such as sending more urgent reminder messages if there is no reply). This two-way interaction feature cannot be reflected in a single view, resulting in delayed detection.

[0004] Second, false information propagation has significant time sensitivity. For example, "time-limited offer" type short messages will have a peak of forwarding within 1-2 hours after being published, and then the propagation efficiency will quickly decay. Existing graph neural network models mostly use static graph structures and do not dynamically integrate time decay factors and propagation frequency, resulting in a response delay of more than 30 minutes for sudden propagation events.

[0005] Third, false information propagation has both local cluster characteristics (such as forwarding within a family group) and global network characteristics (such as cross-regional diffusion). Existing methods either focus on local subgraphs (such as community detection) or analyze global networks (such as PageRank algorithm), and fail to achieve collaborative learning of local and global features, resulting in an identification accuracy of less than 65% for mixed propagation patterns of small-scale intensive forwarding and cross-regional diffusion.

[0006] Fourth, false information often undergoes semantic variation (such as modifying part of the content when forwarding to evade keyword filtering) during propagation, and the degree of variation is strongly correlated with the propagation influence. Existing technologies lack quantitative modeling of subgraph-level propagation influence and cannot capture the dynamic association between semantic variation and influence diffusion.

[0007] Therefore, it is an urgent problem for those skilled in the art to provide a detection scheme that can capture false information propagation characteristics from multiple aspects and solve the problems of low recognition rate and poor scene adaptability of existing technologies. SUMMARY

[0008] In view of this, the present application provides a kind of based on graph level learning's false information detection method and system of telecommunication network, by constructing bidirectional subcascade graph, time-structure fusion model and reinforcing subgraph level influence attention, comprehensively improve the identification precision and response speed of false information under complex propagation scenario.

[0009] In order to achieve the above object, the technical scheme adopted by the present application is as follows: In a first aspect, the present application provides a false information detection method for telecommunication network based on graph level learning, comprising the following steps: S1, obtain the communication data of the communication service platform, and perform structured processing to obtain a structured propagation event data set; S2, based on the communication data propagation path in the structured propagation event data set, construct a bidirectional subcascade graph related to the communication data; S3, based on a pre-constructed multi-scale perception false information feature extraction model, perform multi-scale feature extraction on the bidirectional subcascade graph to obtain a feature vector representing the time sequence structure of information propagation events and the dynamic communication data propagation path; S4, perform mean aggregation on the feature vector to obtain a global graph embedding of the target information event; input the global graph embedding into a false classifier of the communication data, and output the classification result of the target information event.

[0010] Further, step S1 specifically includes: Obtain the SMS data, call data and interactive feedback data of the communication operator service platform in a preset time period; after account desensitization, text preprocessing and time standardization processing, generate a structured propagation event data set; The data set includes communication participating terminals, communication interaction behaviors and communication information.

[0011] Further, step S2 specifically includes: Take the communication participating terminals as nodes, generate edges according to the communication interaction behaviors, and extract node features and time stamps from the structured propagation event data set; wherein the node features include a text feature matrix and a communication behavior feature matrix; Divide the information propagation process into T time intervals according to a preset time window, construct a subcascade graph in each interval; take the information source as the root, the propagation edge as the directed edge, the terminal node as the leaf, and the feedback edge as the directed edge, respectively construct a top-down diffusion view and a bottom-up feedback view, and obtain a bidirectional subcascade graph; wherein the bidirectional subcascade graph shares the text feature matrix and the communication behavior feature matrix.

[0012] Further, in step S3, the construction process of the multi-scale perception false information feature extraction model includes: A time-aware node embedding module is constructed, which captures the time span and decay effect of communication behavior between nodes by defining a time embedding function; A time-structure-aware computing module is constructed, which optimizes the adjacency matrix of the bidirectional sub-cascaded graph using the communication behavior feature matrix, and obtains the node features of the bidirectional view by fusing time embedding and communication behavior features through a graph convolutional network. A graph embedding module for multi-view fusion is constructed, which concatenates the node features of the bidirectional views into a multi-channel feature matrix and extracts the global embedding of the sub-cascaded graph through convolution and pooling operations. A subgraph-level attention weighting module is constructed to calculate the attention weights of sub-cascaded graphs in different time windows and to perform weighted updates on the global embedding.

[0013] Furthermore, the data processing procedure of the time-aware node embedding module includes: Let the nodes in the bidirectional sub-cascade graph be... u, v The communication behavior times are respectively t 1. t 2; among which, ; Define time embedding function The temporal embedding representation at the node level is computed; it is expressed by the formula:

[0014] in, d t Embedding dimensions for time; These are the trainable frequency factor and the time decay parameter, respectively. It is a cosine function.

[0015] Furthermore, the data processing procedure of the time-structure-aware computing module includes: Adjacency matrix optimization is performed using the communication behavior feature matrix. Generate communication behavior weight matrix The adjacency matrix of the bidirectional sub-cascaded graph is optimized to obtain the optimized adjacency matrix of the diffusion view. The adjacency matrix after feedback view optimization is: ; When performing graph convolutional network computation, let the first... l -1 level nodes The hidden representation is After fusing temporal embedding and communication behavior features, the input vector of the graph convolutional network is obtained. and ; Will and The inputs are fed into a graph convolutional network, and the ReLU activation function is used to obtain the first...l Hidden node representation in layered diffusion view The node hiding representation in the feedback view ; in, For element-wise product, for t The adjacency matrix of the time-domain spread view; Let be the adjacency matrix of the feedback view at time t; This is the communication behavior weight matrix; and These are the optimized adjacency matrices for the diffusion view and the feedback view, respectively; For nodes With the target node Time embedding between them For nodes The characteristics of communication behavior; and These are the fused feature vectors of the diffusion view and the feedback view, respectively; and The diffusion view and feedback view are respectively the first The convolutional filtering matrix of the layer.

[0016] Furthermore, the data processing procedure of the graph embedding module for multi-view fusion includes: The bidirectional view node feature matrix calculated by the graph convolutional network is concatenated into a multi-channel feature matrix according to the channel dimension. Convolution operations are performed based on multi-channel feature matrices and filters to extract local communication interaction feature maps. Max pooling is performed on the convolution results, and all pooling results are concatenated to obtain the communication subgraph features, which are used as the global embedding representation of the sub-cascaded graph.

[0017] Furthermore, the data processing procedure for the subgraph-level influence attention weighting module includes: Calculate the first using the time embedding function m Time decay vector of each time window ; Based on the self-attention mechanism, and combining the features of the communication subgraph with the time decay vector, the attention weight of each sub-cascade graph is calculated; expressed by the formula:

[0018] The attention weights are multiplied by the global embedding of the sub-cascaded graph to obtain a weighted feature vector. ; in, Attention weights; These are the attention weight parameters; This is the time decay vector; for t Global embedding representation of the sub-cascade graph at time step.

[0019] Further, step S4 specifically includes: performing mean aggregation on the feature vector to obtain a global graph embedding of the information event; inputting the global graph embedding into a classifier to output the truth or falsehood of the information.

[0020] The feature vectors are aggregated by mean to obtain the global graph embedding. ; The global graph is embedded into the fully connected input layer, and a probability vector is output through the Softmax function. Perform binary or multi-class classification to obtain binary or multi-class classification results; The binary classification results include: real information and false information; the multi-classification results include: real information, fraudulent information, false warnings, and inaccurate interpretations.

[0021] Secondly, the present invention provides a telecommunications network fake information detection system based on graph-level learning, comprising the following units: Communication data acquisition unit: used to acquire communication data from the communication service platform and perform structured processing to obtain a structured propagation event dataset; Bidirectional sub-cascade graph construction unit: used to construct a bidirectional sub-cascade graph related to the communication data based on the communication data propagation path in the structured propagation event dataset; Multi-scale feature extraction unit: used to perform multi-scale feature extraction on the bidirectional sub-cascade graph based on a pre-built multi-scale perceptual false information feature extraction model, to obtain feature vectors representing the temporal structure of information propagation events and the dynamic communication data propagation path; Prediction result output unit: performs mean aggregation on the feature vector to obtain a global graph embedding of the target information event; embeds the global graph into a spurious classifier of the input communication data, and outputs the classification result of the target information event.

[0022] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method and system for detecting false information in telecommunications networks based on graph-level learning, which has the following beneficial effects: This invention provides a complete, bidirectional representation of the data propagation structure by constructing a bidirectional sub-cascade graph, fully simulating the real, interactive propagation process. This makes it possible to comprehensively, accurately, and dynamically characterize the complex characteristics of misinformation propagation, laying the foundation for ultimately achieving high-precision, low-latency misinformation detection.

[0023] This invention constructs a multi-scale perceptual feature extraction model for misinformation. Through a time-aware node embedding module, it addresses the problem of insufficient dynamic temporal capture, achieving fine-grained node-level time decay modeling and laying the foundation for subsequent time-series analysis. Through a time-structure-aware graph convolutional network module, it solves the problem of poor feature fusion, enhancing the model's risk perception capability and generating differentiated representations of bidirectional views. Through a multi-view fusion graph embedding representation module, it achieves feature transition from the node level to the graph level, capturing the interaction patterns of local clusters and fusing complementary information from bidirectional views. By strengthening the subgraph-level influence attention module, it addresses the problem of missing subgraph influence quantification, strengthening the role of key propagation stages and improving the model's interpretability. Through the collaborative work of these modules, multi-scale, dynamic, and refined perception of the misinformation propagation process is achieved. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 This is a flowchart of a method for detecting fake information in telecommunications networks based on graph-level learning, provided as an embodiment of the present invention.

[0026] Figure 2 This is a schematic diagram of the time-structure-aware graph convolutional network structure provided in an embodiment of the present invention.

[0027] Figure 3 This is a schematic diagram of the multi-view fusion and subgraph-level influence attention module provided in an embodiment of the present invention.

[0028] Figure 4 This is a structural diagram of a telecommunications network fake information detection system based on graph-level learning, provided as an embodiment of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1 This invention discloses a method for detecting misinformation in telecommunications networks based on graph-level learning, referring to...Figure 1 As shown, it includes the following steps: S1. Obtain communication data from the communication service platform and perform structured processing to obtain a structured propagation event dataset; S2. Construct a bidirectional sub-cascade graph related to the communication data based on the communication data propagation path in the structured propagation event dataset; S3. Based on the pre-built multi-scale perceptual false information feature extraction model, multi-scale feature extraction is performed on the bidirectional sub-cascade graph to obtain feature vectors representing the temporal structure of information propagation events and the dynamic communication data propagation path. S4. Perform mean aggregation on the feature vectors to obtain the global graph embedding of the target information event; embed the global graph into the spurious classifier of the input communication data to output the classification result of the target information event.

[0031] This embodiment utilizes multi-view modeling, time-aware graph neural networks, and multi-scale fusion technology to accurately depict and identify the spread of false information, making it suitable for information security protection systems in scenarios such as telecommunications operators, financial institutions, and internet platforms. The implementation steps of this embodiment include: Step 1: Communication data acquisition and preprocessing. In view of the point-to-point interaction and strong time sequence characteristics of communication data such as SMS and calls, the communication data is processed in a structured manner. Step two: Construct a bidirectional sub-cascade graph based on the characteristics of information propagation, and divide the information propagation process into time windows. Each time interval is used to construct a sub-cascaded graph within that interval. And generate a two-way view; Step 3: Construct a multi-scale perceptual fake information detection model, including time-aware node embedding representation, time-structure-aware graph convolutional network design, multi-view fusion graph embedding representation, and attention module to enhance subgraph-level influence. Step 4: Global feature aggregation and probability prediction of the propagation graph, based on the aggregation of the mean values ​​of the sub-cascaded graphs at different time windows and probability prediction.

[0032] The specific steps of this embodiment are described in detail below: Step one involves communication data acquisition, specifically including: Three types of communication data are collected: SMS data, including: text content, sender number, recipient number, sending time, SMS status (such as "delivered", "forwarded"), and whether it contains links / attachments; call data, including: caller number, called number, call start time, and call duration; and interactive feedback data, including: SMS reply content, reply time, and call callback records.

[0033] Step 1: Communication data preprocessing; specifically including: After collecting three types of communication data—SMS, call, and interactive feedback—the system first performs account anonymization, which involves irreversibly encrypting or replacing personal account information (such as mobile phone numbers and user IDs) in the communication data. Next, it performs text preprocessing, removing special characters, punctuation marks, and irrelevant noise, and then performs text segmentation. Finally, it normalizes the timestamps to facilitate the calculation of time intervals and temporal relationships, ultimately generating a structured propagation event dataset. This approach defines communication participants as nodes, communication behaviors (sending / receiving, calling / being called) as edges, and extracts specific features of the communication data, such as text, communication behavior, and time, for each node. Account anonymization, text preprocessing, and time standardization. This embodiment extracts SMS or call data corresponding to each message within a preset time period from the communication network, defines the communication participating terminals as nodes, generates edges based on communication interaction behavior, and extracts node features and timestamps; the node features include a text feature matrix. and communication behavior feature matrix The text feature matrix is ​​generated using a BERT pre-trained language model; the communication behavior feature matrix includes SMS sending frequency, call duration, number of times the same message is forwarded, cross-regional interaction frequency, reply delay time, and the number of historical violations by the terminal.

[0034] Step two involves constructing a bidirectional sub-cascade graph based on the characteristics of information propagation, which mainly includes the following parts: Node and edge construction treats each de-identified mobile phone number / terminal ID as an independent node and assigns it a unique number. In the SMS scenario, if A sends an SMS to B containing the information to be detected, a directed edge "A→B" is generated (diffusion view). If B replies to A with an SMS, a directed edge "B→A" is generated (feedback view). In the call scenario, if A initiates a call to B containing the information to be detected, a directed edge "A→B" is generated (diffusion view). If B calls back A, a directed edge "B→A" is generated (feedback view).

[0035] Considering the short-term, concentrated transmission characteristics of SMS and call data, a 30-minute time window is used to divide the information transmission cycle (e.g., 24 hours) into T=48 time intervals, with a sub-cascade graph constructed for each interval. .

[0036] Initial node features are generated by encoding SMS content using the BERT-LSTM model, with an output dimension of 512. Six quantitative features are extracted, including: SMS sending frequency in the past 24 hours, average call duration, number of times the same message is forwarded, cross-regional interaction frequency (determined based on the mobile phone number's location), reply delay time (the time difference between receiving and replying), and the number of historical violations by the terminal. After standardization, the number of dimensions is 6.

[0037] In this embodiment, during the construction of the bidirectional sub-cascade graph, a top-down diffusion view and a bottom-up feedback view are constructed, with the information source as the root and the propagation edge as the directed edge, and the terminal node as the leaf and the feedback edge as the directed edge. The default time window is 60 minutes, which is shortened to 30 minutes during peak propagation periods; for example, when the propagation density is >50 times / minute, the time window is adjusted to 30 minutes. This embodiment divides the information propagation process into 48 time intervals, and a sub-cascade graph is constructed within each interval. This generates a top-down diffusion view and a bottom-up feedback view; the bidirectional view shares the node feature matrix. and The diffusion view uses an adjacency matrix. This indicates that the feedback view uses an adjacency matrix. express.

[0038] Step three involves constructing a multi-scale perceptual fake information detection model, specifically including: Construct a time-aware node embedding module to incorporate relative time (the communication time span between the target node and its neighboring nodes, such as node...) Sending messages and nodes The time difference of forwarding information is embedded into the node representation and fused with node-level time decay.

[0039] A time-structure-aware computing module is constructed to optimize the adjacency matrix based on communication behavior features, highlighting the weight of high-risk communication relationships. The node features of the bidirectional view are obtained by fusing time embedding and communication behavior features based on graph convolutional networks.

[0040] A graph embedding module for multi-view fusion is constructed, which concatenates the node features of the diffusion view and the feedback view into a multi-channel feature matrix, and extracts the global embedding of the sub-cascaded graph through convolution and pooling operations.

[0041] A subgraph-level attention weighting module is constructed to calculate the attention weights of the sub-cascaded graph based on the time decay effect, and the weighted features of the sub-cascaded graph are obtained based on the attention weights.

[0042] In this embodiment, the data processing procedure for the time-aware node embedding module includes: Using response timestamps to capture the evolution of the information propagation cascade graph over time, consider the graph... Two nodes , At what time , It appears at all times, among which Compute the node-level temporal embedding representation:

[0043] in, Let the dimension be the time vector. These are the trainable frequency factor and the time decay parameter, respectively. It is a cosine function.

[0044] This embodiment utilizes trainable parameters in the function. It achieves node-level time decay effect capture, which is different from traditional graph-level decay modeling and improves the ability to express time features in a fine-grained manner.

[0045] In this embodiment, the data processing procedure for the time-structure-aware computing module includes: First, adjacency matrix optimization is performed, utilizing communication behavior characteristics to optimize the adjacency matrix and highlight the weights of high-risk communication relationships (such as high-frequency forwarding and cross-regional calls). In this embodiment, the communication behavior weight matrix is ​​set as follows: It is generated by features such as "number of times the same information is forwarded" and "frequency of cross-regional interaction" of nodes. The larger the value, the higher the propagation risk. Then the optimized adjacency matrix of the top-to-bottom view is:

[0046] Similarly, the adjacency matrix after bottom-up view optimization is:

[0047] in, For element-wise product, Let be the adjacency matrix of the spread view at time t. Let be the adjacency matrix of the feedback view at time t; The communication behavior weight matrix; the optimized adjacency matrix of the feedback view. Similarly, that is .

[0048] Secondly, graph convolutional network computation is performed, assuming the first... Layer nodes The hidden representation is After fusing temporal embeddings and communication behavior features, the input vector for the graph convolutional network is obtained:

[0049]

[0050] in, For the first Layer node representation, This indicates vector concatenation. It is a node With the target node Time embedding between them For nodes The characteristics of its communication behavior.

[0051] Will and The input is fed into a graph convolutional layer, and after passing through the ReLU activation function, the th... Hidden node representation in a top-to-bottom view:

[0052] Get the first Hidden node representation in a bottom-up view:

[0053] in, and These are the optimized adjacency matrices for the top-down view and the bottom-up view, respectively. and These are the fused feature vectors of the diffusion view and the feedback view, respectively; For the first The convolutional filtering matrix of a layered graph convolutional network is specifically adapted to the high-dimensional behavioral features of communication data, providing a bottom-up view. Similarly, the node feature matrix of the bidirectional view is finally obtained. (Diffusion view) and (Feedback view).

[0054] Reference Figure 2 As shown, Figure 2 middle For different moments in the propagation view Node examples, These represent the values ​​learned in the (l-1)th layer, respectively. Node representation, They represent the current time. With the generation time of the three nodes Time embedding between them To integrate the feature vectors derived from temporal information and node representations, GCL is a graph convolutional layer.

[0055] In this embodiment, the data processing procedure for the graph embedding module of multi-view fusion includes: Reference Figure 3 As shown, Figure 3 middle The feature embedding of the diffuse view at time t is represented. This represents the feature embedding of the feedback view at time t. This represents the global characteristics of the entire view at time t. These represent the global features of the T views respectively. These represent the global features after adding attention weights. Predictive labels that indicate information.

[0056] In this embodiment, a multi-channel feature matrix is ​​first constructed by concatenating the node feature matrices of the bidirectional view along the channel dimension to form a multi-channel feature matrix adapted to the communication data. Channel 1 corresponds to the diffusion view feature, and channel 2 corresponds to the feedback view feature. The multi-channel feature matrix is ​​expressed by the formula:

[0057] in, For time The number of nodes in the time cascade graph, i.e. the number of communication participants within that time window; The dimension of the node feature vector; Secondly, multi-channel convolution operations are performed to capture communication interaction patterns.

[0058] Based on multi-channel input and Filters Perform convolution operations to extract local communication interaction patterns, expressed by the formula:

[0059] in, This is a feature map of local communication interactions; and These are the learnable parameter matrices, where i represents the starting position (index) of the local window on the input sequence in the current convolution operation, and i+m-1 represents the ending position (index) of the current convolution window. It is a multi-channel feature matrix. This represents the convolution operation. The ReLU activation function outputs local interactive features, such as the risk pattern of "high-frequency forwarding + cross-regional feedback".

[0060] Finally, max pooling and global feature concatenation are performed. Max pooling is applied to the feature map generated by the convolution, expressed by the following formula:

[0061] Key features in each communication interaction mode are retained, such as features corresponding to the maximum forwarding frequency and the longest call duration; the pooling results of all filters are concatenated to obtain a global embedding representation of the sub-concatenated graph, which is the core feature of communication propagation within this time window:

[0062] in, Global features generated for the filter It represents the number of filters.

[0063] In this embodiment, the data processing procedure regarding the influence of the attention module at the subgraph level includes: First, the time decay vector is calculated; for A sequence of sub-cascaded graphs within a time window , No. The time decay vector of each time window specifically characterizes the stage decay of communication propagation:

[0064] in, The first , The end time of each time window. The smaller the value, the more significant the decline in the spread of the window. For example, in the later stages of the spread of false information, the number of new recipients decreases and the influence diminishes.

[0065] Secondly, the attention weights are calculated. Based on the self-attention mechanism, and combining the features of the communication subgraph with the time decay vector, the attention weight of each sub-cascade graph is calculated. The formula is as follows:

[0066] in, Attention weights; These are the attention weight parameters; This is the time decay vector; For the final Global embedding representation of the sub-cascade graph at time step.

[0067] Finally, an attention-weighted update is performed, multiplying the attention weights by the global embedding representation of the sub-cascade graph to obtain the weighted features of the sub-cascade graph: .

[0068] Step four, the global feature aggregation and probability prediction of the propagation graph, specifically includes: Global feature aggregation, applying weighted feature sequences of sub-cascade graphs. By performing mean aggregation and integrating the communication propagation features across all time windows, a global graph embedding of the information events is obtained, representing the complete communication propagation features:

[0069] in, The updated subgraph features are weighted by attention. This indicates the number of time windows, i.e., there are Individual charts.

[0070] True / false probability prediction embeds a global graph into the input fully connected layer and outputs the true / false probability of information propagated in communication data through the Softmax function. It supports binary or multi-class classification for common types of false information in communication scenarios.

[0071] in, It is a fully connected layer. This is a probability vector, where each element corresponds to the probability of a different category. The results of binary classification include: real information and false information. The results of multi-class classification include: real information, fraudulent information, false warnings, and misinterpretations.

[0072] This embodiment takes into account the noise in the communication data and uses a cross-entropy loss function combined with L2 regularization to balance detection accuracy and noise resistance. The loss function is defined as follows:

[0073] in, For the number of tag categories, For model parameters, (Binary classification) or (Four categories) represent the true label values ​​of the information. For the model to the first The predicted probability of a sample. The regularization coefficient is . For parameters of Regularization term.

[0074] Example 2 This invention discloses a telecommunications network fake information detection system based on graph-level learning, referring to... Figure 4 As shown, it includes the following units: Communication data acquisition unit: used to acquire communication data from the communication service platform and perform structured processing to obtain a structured propagation event dataset; Bidirectional sub-cascade graph construction unit: used to construct a bidirectional sub-cascade graph related to the communication data based on the communication data propagation path in the structured propagation event dataset; Multi-scale feature extraction unit: used to perform multi-scale feature extraction on the bidirectional sub-cascade graph based on a pre-built multi-scale perceptual false information feature extraction model, to obtain feature vectors representing the temporal structure of information propagation events and the dynamic communication data propagation path; Prediction result output unit: performs mean aggregation on the feature vector to obtain a global graph embedding of the target information event; embeds the global graph into a spurious classifier of the input communication data, and outputs the classification result of the target information event.

[0075] The telecommunications network fake information detection system constructed in this embodiment includes: a communication data acquisition unit, used to extract SMS data, call data and corresponding behavior generation times from the communication network, and perform preprocessing; a bidirectional sub-concatenated graph construction unit, used to generate a bidirectional sub-concatenated graph and an adjacency matrix; a multi-scale feature extraction unit, used to extract multi-scale features from the bidirectional sub-concatenated graph based on a multi-scale perceptual fake information feature extraction model; and a prediction result output unit, used to perform global feature aggregation and probability prediction of the propagation graph, and output the probability of information being true or false and the judgment result.

[0076] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0077] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting misinformation in telecommunications networks based on graph-level learning, characterized in that, Includes the following steps: S1. Obtain communication data from the communication service platform and perform structured processing to obtain a structured propagation event dataset; S2. Construct a bidirectional sub-cascade graph related to the communication data based on the communication data propagation path in the structured propagation event dataset; S3. Based on the pre-constructed multi-scale perceptual false information feature extraction model, multi-scale feature extraction is performed on the bidirectional sub-cascade graph to obtain feature vectors representing the temporal structure of information propagation events and the dynamic communication data propagation path. S4. Perform mean aggregation on the feature vectors to obtain a global graph embedding of the target information event; embed the global graph into a spurious classifier of the input communication data, and output the classification result of the target information event.

2. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 1, characterized in that, Step S1 specifically includes: Acquire SMS data, call data, and interactive feedback data from a communication service platform within a preset time period; after account anonymization, text preprocessing, and time standardization, generate a structured propagation event dataset; The dataset includes communication participating terminals, communication interaction behaviors, and communication information.

3. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 2, characterized in that, Step S2 specifically includes: The communication participating terminals are used as nodes, edges are generated based on the communication interaction behavior, and node features and timestamps are extracted from the structured propagation event dataset; wherein, the node features include a text feature matrix and a communication behavior feature matrix; The information propagation process is divided into T time intervals according to a preset time window, and a sub-cascade graph is constructed in each interval. With the information source as the root and the propagation edge as the directed edge, and the terminal node as the leaf and the feedback edge as the directed edge, a top-down diffusion view and a bottom-up feedback view are constructed respectively to obtain a bidirectional sub-cascade graph. The bidirectional sub-cascade graph shares the text feature matrix and the communication behavior feature matrix.

4. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 3, characterized in that, In step S3, the construction process of the multi-scale perceptual false information feature extraction model includes: A time-aware node embedding module is constructed, which captures the time span and decay effect of communication behavior between nodes by defining a time embedding function; A time-structure-aware computing module is constructed, which optimizes the adjacency matrix of the bidirectional sub-cascaded graph using the communication behavior feature matrix, and obtains the node features of the bidirectional view by fusing time embedding and communication behavior features through a graph convolutional network. A graph embedding module for multi-view fusion is constructed, which concatenates the node features of the bidirectional views into a multi-channel feature matrix and extracts the global embedding of the sub-cascaded graph through convolution and pooling operations. A subgraph-level attention weighting module is constructed to calculate the attention weights of sub-cascaded graphs in different time windows and to perform weighted updates on the global embedding.

5. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 4, characterized in that, The data processing procedure of the time-aware node embedding module includes: Let the nodes in the bidirectional sub-cascade graph be... u , v The communication behavior times are respectively t 1. t 2; among which, ; Define time embedding function The temporal embedding representation at the node level is computed; it is expressed by the formula: in, d t Embedding dimensions for time; These are the trainable frequency factor and the time decay parameter, respectively. It is a cosine function.

6. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 5, characterized in that, The data processing procedure of the time-structure-aware computing module includes: Adjacency matrix optimization is performed using the communication behavior feature matrix. Generate communication behavior weight matrix The adjacency matrix of the bidirectional sub-cascaded graph is optimized to obtain the optimized adjacency matrix of the diffusion view. The adjacency matrix after feedback view optimization is: ; When performing graph convolutional network computation, let the first... l -1 level nodes The hidden representation is After fusing temporal embedding and communication behavior features, the input vector of the graph convolutional network is obtained. and ; Will and The inputs are fed into a graph convolutional network, and the ReLU activation function is used to obtain the first... l Hidden node representation in layered diffusion view The node hiding representation in the feedback view ; in, For element-wise product, Let be the adjacency matrix of the spread view at time t; for t The adjacency matrix of the time-feedback view; This is the communication behavior weight matrix; and These are the optimized adjacency matrices for the diffusion view and the feedback view, respectively; For nodes With the target node Time embedding between them For nodes The characteristics of communication behavior; and These are the fused feature vectors of the diffusion view and the feedback view, respectively; and The diffusion view and feedback view are respectively the first The convolutional filtering matrix of the layer.

7. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 6, characterized in that, The data processing procedure of the graph embedding module for multi-view fusion includes: The bidirectional view node feature matrix calculated by the graph convolutional network is concatenated into a multi-channel feature matrix according to the channel dimension. Convolution operations are performed based on multi-channel feature matrices and filters to extract local communication interaction feature maps. Max pooling is performed on the convolution results, and all pooling results are concatenated to obtain the communication subgraph features, which are used as the global embedding representation of the sub-cascaded graph.

8. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 7, characterized in that, The data processing procedure for the subgraph-level influence attention weighting module includes: Calculate the first using the time embedding function m Time decay vector of each time window ; Based on the self-attention mechanism, and combining the features of the communication subgraph with the time decay vector, the attention weight of each sub-cascade graph is calculated; expressed by the formula: The attention weights are multiplied by the global embedding of the sub-cascaded graph to obtain a weighted feature vector. ; in, Attention weights; These are the attention weight parameters; This is the time decay vector; for t Global embedding representation of the sub-cascade graph at time step.

9. The method for detecting false information in telecommunications networks based on graph-level learning as described in claim 7, characterized in that, Step S4 specifically includes: The feature vectors are aggregated by mean to obtain the global graph embedding. ; The global graph is embedded into the fully connected input layer, and a probability vector is output through the Softmax function. Perform binary or multi-class classification to obtain binary or multi-class classification results; The binary classification results include: real information and false information; the multi-classification results include: real information, fraudulent information, false warnings, and inaccurate interpretations.

10. A telecommunications network fake information detection system based on graph-level learning, characterized in that, Includes the following units: Communication data acquisition unit: used to acquire communication data from the communication service platform and perform structured processing to obtain a structured propagation event dataset; Bidirectional sub-cascade graph construction unit: used to construct a bidirectional sub-cascade graph related to the communication data based on the communication data propagation path in the structured propagation event dataset; Multi-scale feature extraction unit: used to perform multi-scale feature extraction on the bidirectional sub-cascade graph based on a pre-built multi-scale perceptual false information feature extraction model, to obtain feature vectors representing the temporal structure of information propagation events and the dynamic communication data propagation path; Prediction result output unit: performs mean aggregation on the feature vector to obtain a global graph embedding of the target information event; embeds the global graph into a spurious classifier of the input communication data, and outputs the classification result of the target information event.

Images