Mail risk early warning system based on hierarchical graph neural network and LSTM
By constructing a hierarchical map of the 'mail city transit center-outline' and multi-source data fusion analysis, the limitations of mail risk identification in the existing technology are solved, and email risk prediction and real-time warning from a global perspective are realized, and the mail transportation process is improved. risk monitoring and control capabilities.
Patent Information
- Application Number
- CN202410936192.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-07-12
AI Technical Summary
The existing email risk identification technology lacks a global perspective and systematic risk assessment ability. The single-point screening model has limitations when facing massive and diverse emails, making it difficult to maintain efficiency and accuracy, and poses safety risks.
Based on the hierarchical graph neural network and LSTM, the hierarchical graph of the 'mail city transfer center-outline' is built. Through multi-source data fusion analysis, it realizes collaborative screening and risk warning across nodes, combining mail order data mining and X The intelligent detection technology of optical security image is used to dynamically explore potential risks of nodes.
It improves the accuracy and reliability of email risk prediction, realizes global mail risk assessment and real-time early warning, enhances risk monitoring and control capabilities during mail transportation, and reduces security risks.
Smart Images

Figure CN118897977A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of mail link risk warning, and in particular to a mail risk warning system based on hierarchical graph neural network and LSTM. Background Art
[0002] With the rise of Internet e-commerce platforms, mail transportation has developed rapidly as the main mode of goods flow. However, behind the rapid growth of mail business, risk supervision and evaluation methods are relatively lagging behind. Due to the low cost, high concealment and wide geographical coverage of mail transportation, some lawless elements may use this channel to transport illegal items, which not only threatens urban traffic and operation safety, but also becomes an unstable factor in social security management, increasing urban security risks.
[0003] At present, there are two main ways to identify mail risks in China: one is to find contraband through manual inspection and verification. Specific measures include real-name mail, inspection before delivery, and security inspection through machines. This method relies on the experience and vigilance of security inspectors to prevent potential risks by manually checking mail information and package contents. However, the efficiency and accuracy of manual inspection are often limited by personnel allocation, workload, and individual differences, and are prone to missed inspections and false inspections due to fatigue.
[0004] The other is to develop X-ray security image intelligent detection technology by using image processing and recognition technology as an auxiliary means. This method uses advanced computer vision algorithms to perform X-ray scanning and image analysis on mail packages to assist security inspectors in identifying contraband and dangerous items. Although this technology has improved detection efficiency and accuracy to a certain extent, its effectiveness is still limited by algorithm performance and equipment configuration.
[0005] The above research and applications mainly focus on mail screening at a single node, lacking a global perspective and systematic risk assessment capabilities. This single-point screening model has obvious limitations when facing massive and diverse mails. Due to the lack of collaboration and information sharing between nodes, the screening effect is often partial and one-sided. In addition, the existing screening technology may be difficult to maintain efficiency and accuracy when faced with complex and changing mail content and the ever-evolving disguise methods of contraband, resulting in certain security risks. In order to improve the effectiveness of mail risk identification, a more comprehensive and systematic solution is urgently needed that can integrate and analyze multi-node and multi-channel mail information on a global scale, and realize cross-node collaborative screening and risk warning. This requires not only technological innovation and breakthroughs, but also the consideration of real-time sharing and comprehensive utilization of information in system design and implementation, so as to establish a more sound and reliable mail risk warning system. Summary of the invention
[0006] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and to provide a mail risk warning system based on a hierarchical graph neural network and LSTM. The system introduces a hierarchical graph neural network into the field of mail link risk warning for the first time, constructs a hierarchical graph of "mail city transfer center-outlet", and dynamically mines the potential risks of nodes from a global perspective. Through this global perspective, the system can fully consider the interdependence between various nodes in the mail transfer process, effectively identify and analyze the interactive impact of upstream and downstream nodes, thereby improving the accuracy and reliability of risk prediction. At the same time, the system also combines mail order data mining technology and X-ray security inspection image intelligent detection technology, and realizes a more comprehensive mail risk assessment through the fusion analysis of multi-source data.
[0007] To achieve the above purpose, the technical solution provided by the present invention is: an email risk warning system based on hierarchical graph neural network and LSTM, comprising:
[0008] Hierarchical graph construction module, used to construct a hierarchical graph of "mail city transfer center-outlet", abstracting the mail city transfer centers and outlets that mail passes through during transportation into nodes, constructing edges between adjacent nodes to obtain a graph structure, and obtaining a hierarchical graph structure through layered processing, so as to model the complex spatiotemporal dependency relationship between nodes at different granularities;
[0009] The mail data import module is used to import mail order data at mail city transfer centers and outlets, mail image recognition data, and historical mail data in the database, and perform preprocessing operations on the three types of data, including: deleting abnormal values and deleting redundant data to obtain preprocessed mail data;
[0010] The email feature extraction module is used to extract features from the email data preprocessed by the email data import module, use the LSTM network to encode and decode the email data, automatically learn the key features in the email data, achieve email data dimensionality reduction and feature extraction, and obtain node features;
[0011] The node feature update module is used to update the node features obtained by the email feature extraction module, and use the hierarchical graph neural network and attention mechanism to aggregate and update the messages of nodes between and within the layers, and obtain the final feature representation of the nodes in the hierarchical graph;
[0012] The node risk warning module is used to predict the probability of risk category of nodes in the hierarchical graph. The input is the final feature representation of the node obtained by the node feature update module. The node risk is classified through a neural network composed of a fully connected layer and a sigmoid activation function. The risk categories are set to five categories: low, low, medium, high, and high. When the node is in the high or high category, the system starts a safety warning prompt to remind staff to strengthen management and control.
[0013] Furthermore, the hierarchical graph construction module performs the following operations:
[0014] Let G = (G city ,G outlet ) represents a hierarchical diagram of “mail city transfer center-network”, where G city = {V c ,E c} is the mail city transit center map, and Goutlet is the outlet map;
[0015] The mail city transfer centers that the mail passes through during transportation are regarded as nodes of the mail city transfer center graph, denoted as set V c = {V i c |i=1,2,...,N c},V i c represents the central node, N c is the number of nodes in the mail city transit center graph, which is also the number of network nodes. c is the set of edges connecting the nodes in the mail city transit center graph;
[0016] Mail city transit centers have their own network maps and serve as the central nodes of their own network maps. For each central node V i c , there is a dot graph Correspondingly, and V i o and Respectively represent the node set and edge set in the network graph;
[0017] Each central node V i c Associated with a label yi = {0, 1, 2, 3, 4}, when the label value is 0, it indicates a low risk state, when the label value is 1, it indicates a relatively low risk state, when the label value is 2, it indicates a medium risk state, when the label value is 3, it indicates a relatively high risk state, and when the label value is 4, it indicates a high risk state;
[0018] Edge E in the Mail City Transit Center Graph c and the edges in the dot graph The weight is determined by the geographical similarity g between nodes s Frequency of email exchanges s The composition is expressed as follows:
[0019]
[0020] Where: xa and x b is node V a and V b Latitude, y a and b is node V a and V b Longitude, d a,b is node V a and V b The Euclidean distance between is node V a and V b The geographical similarity between Node V a and V b Frequency of email exchanges between is node V a and V b The edge weights between them, α and β are weighting coefficients.
[0021] Further, the email data import module performs the following operations:
[0022] Import mail order data, including mail barcode, collection time, recipient, organization code, customer code, sender, sender address, sender phone number, recipient, recipient address, recipient phone number and product description, and use natural language processing technology to convert text description into vectors;
[0023] Import mail image recognition data, and use the X-ray security intelligent detection algorithm and component matching algorithm to jointly determine the probability that the mail contains contraband;
[0024] Import historical email data from the database, including: the number of historically detected dangerous emails, historical blacklist phone numbers, and historical blacklist customer codes.
[0025] Further, the email feature extraction module performs the following operations:
[0026] The mail data obtained after preprocessing by the mail data import module is integrated to obtain a three-dimensional matrix As input, represents a set of real numbers, N represents the total number of mail city transfer centers and outlets, T represents the time step, and M represents the mail data dimension;
[0027] A coding-decoding model is constructed based on the LSTM network to perform coding-decoding operations on the email data. The coding-decoding model includes an encoder, a decoder, and a fully connected layer. The encoder is composed of an LSTM network and is used to encode the three-dimensional matrix x t To do the encoding operation, that is, recursive processing, the encoder operation is expressed as:
[0028] h t ,c t =LSTM0(x t ,h t-1 ,c t-1 )
[0029] Where: h t is the tth element of the state output vector in the encoder, and is also the output of the encoder, h t-1 is the t-1th element of the state output vector in the encoder, c t and c t-1 are the t-th and t-1-th elements of the cell memory state vector in the encoder, respectively. LSTM0(·) represents the encoder;
[0030] The decoder is composed of an LSTM network, which is used to decode the output of the encoder. The operation of the decoder is expressed as:
[0031]
[0032] Where: is the tth element of the state output vector in the decoder, and is also the output of the decoder, is the t-1th element of the state output vector in the decoder, and are the t-th and t-1-th elements of the cell memory state vector in the decoder, respectively. LSTM1(·) represents the decoder;
[0033] The output of the decoder is passed through the fully connected layer, and the operation of the fully connected layer is expressed as:
[0034]
[0035] Where: is the compressed email data, which is also the node feature, d is the dimension of the compressed email data, and FC(·) is the fully connected layer operation.
[0036] Furthermore, the loss function of the encoding-decoding model is mean square error loss.
[0037] Further, the node feature updating module performs the following operations:
[0038] Define the jth row of the compressed email data obtained by the email feature extraction module to represent the initial features of the jth node in the hierarchical graph. Use the message passing mechanism of the hierarchical graph neural network to update the node features between and within the layers, so that each node contains richer global features.
[0039] For node V in the mail city transit center graph a, whose messages come from neighboring nodes and itself, the formula is as follows:
[0040]
[0041] Where: x a For node V a The characteristics of x n For node V a The characteristics of the neighbor nodes, For node V a The edge weight with neighbor nodes, M a To pass to node V a The collection of all messages;
[0042] will be passed to node V a The set of all messages M a Encoding, the formula is as follows:
[0043]
[0044] Where: m a is the message encoded vector, For encoding operations;
[0045] The message encoded vector m a For node V a The features are updated, and the formula is as follows:
[0046]
[0047] Where: is the updated node V a The features of , Att1(·) is the update function based on the attention mechanism;
[0048] Update to node V a The characteristics of the nodes in the network diagram of the central node are as follows:
[0049]
[0050] Where: The node V a is the updated node feature of node k in the network graph of the central node, m a,k The node V a is the message aggregation vector of node k in the network graph of the central node, x a,k The node V a is the original feature of node k in the network graph of the central node, is the updated node V a The feature of , Att2(·) is the update function based on the attention mechanism.
[0051] Further, the node risk warning module performs the following operations:
[0052] The node features obtained by the node feature update module are used as input, and the risk category probability prediction value of each node is obtained through a neural network composed of a fully connected layer and a sigmoid activation function. a , the calculation formula of the risk category probability prediction value is as follows:
[0053]
[0054] Where: For node V a The risk category probability prediction value, sigmoid(·) is the activation function, and FC(·) is the fully connected layer operation;
[0055] The loss function of the above neural network is cross entropy loss, and the formula is as follows:
[0056]
[0057] Where: l is the loss value, is the total number of nodes, C is the total number of risk categories, y qc is the label value of the cth category of the qth node, is the risk category probability prediction value of the cth category of the qth node;
[0058] The category with the highest probability prediction value of the risk category of the node is selected as the risk category of the node. When the node is in the high or high category, the system activates a safety warning prompt to remind staff to strengthen management and control.
[0059] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0060] 1. For the first time, a hierarchical diagram of "Mail City Transfer Center-Outlets" was constructed, which effectively reveals the dependency and hierarchical structure between various mail city transfer centers and outlets in the mail transportation process. Through this hierarchical representation method, it can not only clearly show the transmission path of mail from one node to another, but also better understand the interaction and influence between each node, and improve the overall control of the entire mail transportation link.
[0061] 2. For the first time, hierarchical graph neural network technology is introduced to predict node risks on the email link from a global perspective. By capturing the mutual influence relationship between the nodes in the email link, the system can more accurately predict potential risks and identify possible problem nodes, thereby greatly improving the accuracy of node risk prediction.
[0062] 3. Combine natural language processing technology with image recognition technology, make full use of mail order data and X-ray security image information, and improve the accuracy of a single node in identifying dangerous mail.
[0063] 4. The system of the present invention can provide real-time node risk warning. When potential risks are detected, it can immediately issue an alarm to security personnel, prompting them to pay attention to specific high-risk mails or nodes, assisting security personnel to reasonably allocate their attention, enhancing the risk monitoring and control capabilities of the entire mail transportation process, and effectively reducing mail security risks. BRIEF DESCRIPTION OF THE DRAWINGS
[0064] Figure 1 It is a framework diagram of the system of the present invention.
[0065] Figure 2 This is a hierarchical diagram of “mail city transfer center—outlets”; in the diagram, gray nodes represent mail city transfer centers, and white nodes represent outlets.
[0066] Figure 3 This is the structural diagram of the mail data import module.
[0067] Figure 4 This is the structure diagram of the email feature extraction module. DETAILED DESCRIPTION
[0068] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.
[0069] like Figure 1 As shown, this embodiment discloses an email risk warning system based on a hierarchical graph neural network and LSTM, which is developed using the Python language and includes:
[0070] The hierarchical graph construction module is used to construct the hierarchical graph of "mail city transfer center-network", such as Figure 2 As shown in the figure, the mail city transfer centers and outlets that the mail passes through during transportation are abstracted into nodes, and the adjacent nodes are used to build edges to obtain a graph structure. The hierarchical graph structure is obtained through layered processing to model the complex spatiotemporal dependencies between nodes at different granularities.
[0071] The mail data import module is used to import mail order data at mail city transfer centers and outlets, mail image recognition data, and historical mail data in the database, and perform preprocessing operations on the three types of data, including: deleting abnormal values and deleting redundant data to obtain preprocessed mail data;
[0072] The email feature extraction module is used to extract features from the email data preprocessed by the email data import module, use the LSTM (Long Short-Term Memory) network to encode and decode the email data, automatically learn the key features in the email data, achieve email data dimensionality reduction and feature extraction, and obtain node features;
[0073] The node feature update module is used to update the node features obtained by the email feature extraction module, and use the hierarchical graph neural network and attention mechanism to aggregate and update the messages between and within the layers of the nodes to obtain the final feature representation of the nodes in the hierarchical graph;
[0074] The node risk warning module is used to predict the probability of risk category of nodes in the hierarchical graph. The input is the final feature representation of the node obtained by the node feature update module. The node risk is classified through a neural network composed of a fully connected layer and a sigmoid activation function. The risk categories are set to five categories: low, low, medium, high, and high. When the node is in the high or high category, the system starts a safety warning prompt to remind staff to strengthen management and control.
[0075] Specifically, the hierarchical graph construction module performs the following operations:
[0076] Let G = (G city ,G outlet ) represents a hierarchical diagram of “mail city transfer center-network”, where G city = {V c ,E c} is the mail city transit center map, and Goutlet is the outlet map;
[0077] The mail city transfer centers that the mail passes through during transportation are regarded as nodes of the mail city transfer center graph, denoted as set V c = {V i c |i=1,2,...,N c},V i c represents the central node, N c is the number of nodes in the mail city transit center graph, which is also the number of network nodes. c is the set of edges connecting the nodes in the mail city transit center graph;
[0078] Mail city transit centers have their own network maps and serve as the central nodes of their own network maps. For each central node V i c , there is a dot graph Correspondingly, and V i oand Respectively represent the node set and edge set in the network graph;
[0079] Each central node V i c Associated with a label yi = {0, 1, 2, 3, 4}, when the label value is 0, it indicates a low risk state, when the label value is 1, it indicates a relatively low risk state, when the label value is 2, it indicates a medium risk state, when the label value is 3, it indicates a relatively high risk state, and when the label value is 4, it indicates a high risk state;
[0080] Edge E in the Mail City Transit Center Graph c and the edges in the dot graph The weight is determined by the geographical similarity g between nodes s Frequency of email exchanges s The composition is expressed as follows:
[0081]
[0082] Where: x a and x b is node V a and V b Latitude, y a and b is node V a and V b Longitude, d a,b is node V a and V b The Euclidean distance between is node V a and V b The geographical similarity between Node V a and V b Frequency of email exchanges between is node V a and V b The edge weights between them, α and β are weighting coefficients.
[0083] Specifically, Figure 3 As shown, the mail data import module performs the following operations:
[0084] Import mail order data, including mail barcode, collection time, recipient, organization code, customer code, sender, sender address, sender phone number, recipient, recipient address, recipient phone number and product description, and use natural language processing technology to convert text description into vectors;
[0085] Import mail image recognition data, and use the X-ray security intelligent detection algorithm and component matching algorithm to jointly determine the probability that the mail contains contraband;
[0086] Import historical email data from the database, including: the number of historically detected dangerous emails, historical blacklist phone numbers, and historical blacklist customer codes.
[0087] Specifically, Figure 4 As shown, the email feature extraction module performs the following operations:
[0088] The mail data obtained after preprocessing by the mail data import module is integrated to obtain a three-dimensional matrix As input, represents a set of real numbers, N represents the total number of mail city transfer centers and outlets, T represents the time step, and M represents the mail data dimension;
[0089] A coding-decoding model is constructed based on the LSTM network, and its loss function is the mean square error loss, which is used to encode and decode the email data. The coding-decoding model includes an encoder, a decoder and a fully connected layer. The encoder is composed of an LSTM network, which is used to encode the three-dimensional matrix x t To do the encoding operation, that is, recursive processing, the encoder operation is expressed as:
[0090] h t ,c t =LSTM0(x t ,h t-1 ,c t-1 )
[0091] Where: h t is the tth element of the state output vector in the encoder, and is also the output of the encoder, h t-1 is the t-1th element of the state output vector in the encoder, c t and c t-1 are the t-th and t-1-th elements of the cell memory state vector in the encoder, respectively. LSTM0(·) represents the encoder;
[0092] The decoder is composed of an LSTM network, which is used to decode the output of the encoder. The operation of the decoder is expressed as:
[0093]
[0094] Where: is the tth element of the state output vector in the decoder, and is also the output of the decoder, is the t-1th element of the state output vector in the decoder, and are the t-th and t-1-th elements of the cell memory state vector in the decoder, respectively. LSTM1(·) represents the decoder;
[0095] The output of the decoder is passed through the fully connected layer, and the operation of the fully connected layer is expressed as:
[0096]
[0097] Where: is the compressed email data, which is also the node feature, d is the dimension of the compressed email data, and FC(·) is the fully connected layer operation.
[0098] Specifically, the node feature updating module performs the following operations:
[0099] Define the jth row of the compressed email data obtained by the email feature extraction module to represent the initial features of the jth node in the hierarchical graph. Use the message passing mechanism of the hierarchical graph neural network to update the node features between and within the layers, so that each node contains richer global features.
[0100] Take node V in the mail city transfer center graph a For example, node V a The messages come from neighboring nodes and itself, and the formula is as follows:
[0101]
[0102] Where: x a For node V a The characteristics of x n For node V a The characteristics of the neighbor nodes, For node V a The edge weight with neighbor nodes, M a To pass to node V a The collection of all messages;
[0103] will be passed to node V a The set of all messages M a Encoding, the formula is as follows:
[0104]
[0105] Where: m a is the message encoded vector, For encoding operations;
[0106] The message encoded vector m a For node V a The features are updated, and the formula is as follows:
[0107]
[0108] Where: is the updated node V a The features of , Att1(·) is the update function based on the attention mechanism;
[0109] Update to node V a The characteristics of the nodes in the network diagram of the central node are as follows:
[0110]
[0111] Where: The node V a is the updated node feature of node k in the network graph of the central node, m a,k The node V a is the message aggregation vector of node k in the network graph of the central node, x a,k The node V a is the original feature of node k in the network graph of the central node, is the updated node V a The feature of , Att2(·) is the update function based on the attention mechanism.
[0112] Specifically, the node risk warning module performs the following operations:
[0113] The node features obtained by the node feature update module are used as input, and the risk category probability prediction value of each node is obtained through a neural network composed of a fully connected layer and a sigmoid activation function. a For example, node V a The formula for calculating the predicted value of the risk category probability is as follows:
[0114]
[0115] Where: For node V a The risk category probability prediction value, sigmoid(·) is the activation function, and FC(·) is the fully connected layer operation;
[0116] The loss function of the above neural network is cross entropy loss, and the formula is as follows:
[0117]
[0118] Where: l is the loss value, is the total number of nodes, C is the total number of risk categories, y qc is the label value of the cth category of the qth node, is the risk category probability prediction value of the cth category of the qth node;
[0119] The category with the highest probability prediction value of the risk category of the node is selected as the risk category of the node. When the node is in the high or high category, the system activates a safety warning prompt to remind staff to strengthen management and control.
[0120] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.
Claims
1. The email risk warning system based on hierarchical graph neural network and LSTM is characterized by: include: Hierarchical graph construction module, used to construct a hierarchical graph of "mail city transfer center-outlet", abstracting the mail city transfer centers and outlets that mail passes through during transportation into nodes, constructing edges between adjacent nodes to obtain a graph structure, and obtaining a hierarchical graph structure through layered processing, so as to model the complex spatiotemporal dependency relationship between nodes at different granularities; The mail data import module is used to import mail order data at mail city transfer centers and outlets, mail image recognition data, and historical mail data in the database, and perform preprocessing operations on the three types of data, including: deleting abnormal values and deleting redundant data to obtain preprocessed mail data; The email feature extraction module is used to extract features from the email data preprocessed by the email data import module, use the LSTM network to encode and decode the email data, automatically learn the key features in the email data, achieve email data dimensionality reduction and feature extraction, and obtain node features; The node feature update module is used to update the node features obtained by the email feature extraction module, and use the hierarchical graph neural network and attention mechanism to aggregate and update the messages of nodes between and within the layers, and obtain the final feature representation of the nodes in the hierarchical graph; The node risk warning module is used to predict the probability of risk category of nodes in the hierarchical graph. The input is the final feature representation of the node obtained by the node feature update module. The node risk is classified through a neural network composed of a fully connected layer and a sigmoid activation function. The risk categories are set to five categories: low, low, medium, high, and high. When the node is in the high or high category, the system starts a safety warning prompt to remind staff to strengthen management and control.
2. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 1 is characterized in that: The hierarchical graph building module performs the following operations: Let G = (G city ,G outlet ) represents a hierarchical diagram of "mail city transfer center-network", where G city = {V c ,E c } is the mail city transit center map, and Goutlet is the outlet map; The mail city transfer centers that the mail passes through during transportation are regarded as nodes of the mail city transfer center graph, denoted as set V c = {V i c |i=1,2,...,N c },V i c represents the central node, N c is the number of nodes in the mail city transit center graph, which is also the number of network nodes. c is the set of edges connecting the nodes in the mail city transit center graph; Mail city transit centers have their own network maps and serve as the central nodes of their own network maps. For each central node V i c , there is a dot graph Correspondingly, and V i o and Respectively represent the node set and edge set in the network graph; Each central node V i c Associated with a label yi = {0, 1, 2, 3, 4}, when the label value is 0, it indicates a low risk state, when the label value is 1, it indicates a relatively low risk state, when the label value is 2, it indicates a medium risk state, when the label value is 3, it indicates a relatively high risk state, and when the label value is 4, it indicates a high risk state; Edge E in the Mail City Transit Center Graph c and the edges in the dot graph The weight is determined by the geographical similarity g between nodes s Frequency of email exchanges s The composition is expressed as follows: Where: x a and x b is node V a and V b Latitude, y a and b is node V a and V b Longitude, d a,b is node V a and V b The Euclidean distance between is node V a and V b The geographical similarity between Node V a and V b Frequency of email exchanges between is node V a and V b The edge weights between them, α and β are weighting coefficients.
3. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 2 is characterized in that: The mail data import module performs the following operations: Import mail order data, including mail barcode, collection time, recipient, organization code, customer code, sender, sender address, sender phone number, recipient, recipient address, recipient phone number and product description, and use natural language processing technology to convert text description into vectors; Import mail image recognition data, and use the X-ray security intelligent detection algorithm and component matching algorithm to jointly determine the probability that the mail contains contraband; Import historical email data from the database, including: the number of historically detected dangerous emails, historical blacklist phone numbers, and historical blacklist customer codes.
4. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 3 is characterized in that: The email feature extraction module performs the following operations: The mail data obtained after preprocessing by the mail data import module is integrated to obtain a three-dimensional matrix As input, represents a set of real numbers, N represents the total number of mail city transfer centers and outlets, T represents the time step, and M represents the mail data dimension; A coding-decoding model is constructed based on the LSTM network to perform coding-decoding operations on the email data. The coding-decoding model includes an encoder, a decoder, and a fully connected layer. The encoder is composed of an LSTM network and is used to encode the three-dimensional matrix x t To do the encoding operation, that is, recursive processing, the encoder operation is expressed as: h t ,c t =LSTM0(x t ,h t-1 ,c t-1 ) Where: h t is the tth element of the state output vector in the encoder, and is also the output of the encoder, h t-1 is the t-1th element of the state output vector in the encoder, c t and c t-1 are the t-th and t-1-th elements of the cell memory state vector in the encoder, respectively. LSTM0(·) represents the encoder; The decoder is composed of an LSTM network, which is used to decode the output of the encoder. The operation of the decoder is expressed as: Where: is the tth element of the state output vector in the decoder, and is also the output of the decoder, is the t-1th element of the state output vector in the decoder, and are the t-th and t-1-th elements of the cell memory state vector in the decoder, respectively. LSTM1(·) represents the decoder; The output of the decoder is passed through the fully connected layer, and the operation of the fully connected layer is expressed as: Where: is the compressed email data, which is also the node feature, d is the dimension of the compressed email data, and FC(·) is the fully connected layer operation.
5. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 4 is characterized in that: The loss function of the encoding-decoding model is mean square error loss.
6. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 5 is characterized in that: The node feature update module performs the following operations: Define the jth row of the compressed email data obtained by the email feature extraction module to represent the initial features of the jth node in the hierarchical graph. Use the message passing mechanism of the hierarchical graph neural network to update the node features between and within the layers, so that each node contains richer global features. For node V in the mail city transit center graph a , whose messages come from neighboring nodes and itself, the formula is as follows: Where: x a For node V a The characteristics of x n For node V a The characteristics of the neighbor nodes, For node V a The edge weight with neighbor nodes, M a To pass to node V a The collection of all messages; will be passed to node V a The set of all messages M a Encoding, the formula is as follows: Where: m a is the message encoded vector, For encoding operations; The message encoded vector m a For node V a The features are updated, and the formula is as follows: Where: is the updated node V a The features of , Att1(·) is the update function based on the attention mechanism; Update to node V a The characteristics of the nodes in the network diagram of the central node are as follows: Where: The node V a is the updated node feature of node k in the network graph of the central node, m a,k The node V a is the message aggregation vector of node k in the network graph of the central node, x a,k The node V a is the original feature of node k in the network graph of the central node, is the updated node V a The feature of , Att2(·) is the update function based on the attention mechanism.
7. The email risk early warning system based on hierarchical graph neural network and LSTM according to claim 6 is characterized in that: The node risk warning module performs the following operations: The node features obtained by the node feature update module are used as input, and the risk category probability prediction value of each node is obtained through a neural network composed of a fully connected layer and a sigmoid activation function. a , the calculation formula of the risk category probability prediction value is as follows: Where: For node V a The risk category probability prediction value, sigmoid(·) is the activation function, and FC(·) is the fully connected layer operation; The loss function of the above neural network is cross entropy loss, and the formula is as follows: Where: l is the loss value, is the total number of nodes, C is the total number of risk categories, y qc is the label value of the cth category of the qth node, is the risk category probability prediction value of the cth category of the qth node; The category with the highest probability prediction value of the risk category of the node is selected as the risk category of the node. When the node is in the high or high category, the system activates a safety warning prompt to remind staff to strengthen management and control.
Citation Information
Patent Citations
Text classification model training method, text classification method and related device
CN113869458A
Encoder training method and device and storage medium
CN114418069A
Implicit discourse relation identification method fusing parameter validity and relation label level semantic priori
CN116822526A
Industrial chain risk assessment system combining hierarchical graph neural network and long and short term memory
CN118313668A
Text sentiment analysis method based on multi-level graph pooling
US20220405480A1
