A material sea shipping network risk node analysis method based on bipartite graph embedding

Abstract

The application relates to a material sea shipping network risk node analysis method based on a bipartite graph embedding, wherein the method comprises the following steps: based on original material sea shipping network data, generating low-dimensional embedding vectors corresponding to various nodes in the sea shipping network; inputting the feature vectors and target vectors into a pre-constructed binary classification model for training to generate a link prediction model of the sea shipping network; inputting the target embedding vectors into the link prediction model to calculate the decline value of the area under the curve value corresponding to the link prediction model, and determining the risk nodes of the sea shipping network by using the target area whose decline value exceeds a preset threshold. The application obtains the embedding vectors of the nodes from the original sea shipping network, constructs a binary classification model to obtain a link prediction model of the sea shipping network, evaluates the influence degree of the corresponding area on the network evolution according to the accuracy decline of the link prediction, so that the uncertainty in the network evolution process is reduced and the management of the sea shipping network is optimized.

Description

Technical Field

[0001] This application relates to the field of intelligent prediction technology of geographic information, and in particular to a method for risk node analysis of material shipping networks based on bipartite graph embedding. Background Technology

[0002] Methods for analyzing nodes in maritime transport networks often rely on topological metrics such as centrality and degree distribution to measure node importance. However, these traditional network metrics only reflect some structural characteristics of the network and cannot delve into the potential dependencies and interactions between nodes. Especially when predicting the impact of nodes on the future evolution of maritime transport networks, existing methods have significant limitations, making it difficult to identify key nodes that play a crucial role in the direction of network evolution. Furthermore, because maritime transport networks for strategic materials often exhibit bipartite graph characteristics, many classic network analysis metrics and techniques are less applicable to this structure.

[0003] Therefore, how to fully utilize the high-dimensional and complex node relationships to identify the importance of nodes in the dynamic changes of the network, given the special characteristics of the maritime transport network for strategic materials, has become a key challenge in identifying risk nodes in the maritime transport network. Summary of the Invention

[0004] This application provides a risk node analysis method for material shipping networks based on bipartite graph embedding. This method addresses the limitations of traditional network indicators, which can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. In particular, existing methods have significant limitations when predicting the impact of nodes on the future evolution of shipping networks, making it difficult to identify key nodes that play an important role in the direction of shipping network evolution. Furthermore, since strategic material shipping networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability under this structure.

[0005] The first aspect of this application provides a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, comprising the following steps: generating low-dimensional embedding vectors corresponding to each node in the maritime transport network based on the original maritime transport network data; using the low-dimensional embedding vectors as feature variables and the edge between two nodes as target variables, inputting the feature vectors and the target vectors into a pre-constructed binary classification model for training to generate a link prediction model for the maritime transport network; randomly shuffling the embedding vectors of the target region nodes to generate target embedding vectors, and inputting the target embedding vectors into the link prediction model to calculate the decrease value of the area under the curve corresponding to the link prediction model, and using the target region where the decrease value exceeds a preset threshold to determine the risk nodes of the maritime transport network.

[0006] Optionally, in one embodiment of this application, generating low-dimensional embedding vectors corresponding to each node in the shipping network based on the original material shipping network data includes: initializing the hyperparameters of the shipping network based on the original material shipping network data to generate target hyperparameters; modeling the explicit and implicit relationships of the shipping network based on the target hyperparameters to generate modeled explicit and implicit relationships; constructing a joint optimization framework for the shipping network based on the modeled explicit and implicit relationships and the objective function, and generating low-dimensional embedding vectors corresponding to each node based on the joint optimization framework.

[0007] Optionally, in one embodiment of this application, the empirical distribution of the two nodes in the explicit relationship is as follows:

[0008]

[0009] In a bipartite network G = (U, V, E), U and V represent the sets of two types of nodes in the bipartite network, respectively, and E is the set of edges in the network. i and v j Then w represents the i-th and j-th nodes in U and V respectively. ij Then it is node u i and v j The weight of the edges between them;

[0010] The reconstruction distribution of the two nodes in the explicit relationship is as follows:

[0011]

[0012] Among them, u i v j Representing node u respectively i and v j d-dimensional embedding vector;

[0013] The KL divergence formula in the explicit relation is:

[0014]

[0015] Where log(·) refers to the logarithmic function with base e, and P(i,j) and The two nodes u are respectively i and v j The optimization objective in the explicit relation is to minimize the empirical distribution and the reconstructed distribution of O1.

[0016] The second-order proximity formula for two nodes in the implicit relation is:

[0017]

[0018] in, That is, the two nodes u in the binary network node set U. i and u j The second-order proximity of w ik For node u i and v k The weight of the edges between them. The definitions are similar;

[0019] The objective function in the implicit relation is:

[0020]

[0021] Where S represents the node sequence, D U and D V Let C represent the set of node sequences for two types of networks obtained through random walks. S (u i ) represents node u i In the context node of sequence S, P(u) c |u i That is, in u c Observed in the context of u i The probability, P(v) c |v j Similarly, maximizing O2 and O3 is the implicit relationship objective.

[0022] The calculation formula for the joint optimization framework is as follows:

[0023] max L=αlogO2+βlogO3-γO1

[0024] Here, parameters α, β, and γ are all hyperparameters used to balance different parts of the joint optimization framework.

[0025] Optionally, in one embodiment of this application, the step of using the low-dimensional embedding vector as a feature variable and the edge between two nodes as a target variable, and inputting the feature vector and the target vector into a pre-built binary classification model for training to generate a link prediction model for the maritime network, includes: concatenating the low-dimensional embedding vector to generate a feature vector for a sample, and determining the target variable based on the edge when the edge exists between the two nodes; using edges existing before a preset time point as positive samples in the training set, and edges after the preset time point as positive samples in the test set, and allocating negative samples to the training set and the test set according to a target ratio; based on the training set and the test set... In the trial set, the pre-built LightGBM model is used to perform a binary classification task of the low-dimensional embedding vector. Based on the binary classification task and the log-likelihood loss function, the target weak learner is initialized using prior information to generate initialization results. Based on the initialization results, the pseudo-residual of the regression tree is calculated, and the regression tree is fitted according to the pseudo-residual to generate fitted data. The strong learner is updated according to the fitted data to generate the final strong learner. Based on the final strong learner, the feature vector and the target vector are input into the pre-built binary classification model for training until the pre-built binary classification model meets the preset training stopping condition, thus generating the link prediction model of the maritime network.

[0026] Optionally, in one embodiment of this application, the formulas for the feature vector and the target vector are respectively:

[0027] X uv =[u i ,v j ]

[0028]

[0029] Among them, u i v j Representing node u respectively i and v j The embedding vector, where E is the edge set of the graph;

[0030] The formulas for the positive samples in the training set and the positive samples in the test set are as follows:

[0031]

[0032] Among them, u i and v j Let E(t) represent the node, E(t) represent the set of edges at time t, and t0 represent the selected time point.

[0033] The formula for calculating the log-likelihood loss function is as follows:

[0034] L(y i ,F(x i ))=-[y i logP(y i =1|x i )+(1-y i )log(1-P(y i =1|x i ))]

[0035]

[0036] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively; i P(y) represents the model's predicted value for sample i; i =1|x i ) represents the probability that the model predicts the existence of a link in sample i; log(·) refers to the logarithmic function with base e;

[0037] The initialization formula for the target weak learner is:

[0038]

[0039] Where P(Y=1|x) is the proportion of y=1 in the training samples;

[0040] The formula for calculating the pseudo residual is:

[0041]

[0042] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i ) represents the current model's predicted value for sample i. To predict the probability that a link exists in this sample, F m-1 (x) is the learner obtained after training the (m-1)th regression tree;

[0043] The fitting formula for the regression tree is:

[0044]

[0045] Wherein, the leaf region corresponding to the m-th regression tree is R. m,j Where j = 1, 2, ..., J m J m Let m be the number of leaf nodes of the m-th regression tree;

[0046] The update formula for the strong learner is:

[0047]

[0048] Among them, c m,j The best fit for leaf node j of the m-th tree;

[0049] The final formula for generating the strong learner is:

[0050]

[0051] Where I(·) represents the indicator function, M is the total number of regression trees trained by the model, and x represents the feature vector of a sample to be predicted.

[0052] A second aspect of this application provides a risk node analysis device for a material shipping network based on bipartite graph embedding, comprising: a generation module for generating low-dimensional embedding vectors corresponding to each node in the shipping network based on original material shipping network data; a training module for using the low-dimensional embedding vectors as feature variables and the edge between two nodes as target variables, inputting the feature vectors and the target vectors into a pre-constructed binary classification model for training, thereby generating a link prediction model for the shipping network; and a determination module for randomly shuffling the embedding vectors of target region nodes to generate target embedding vectors, inputting the target embedding vectors into the link prediction model to calculate the decrease value of the area under the curve corresponding to the link prediction model, and using the target region where the decrease value exceeds a preset threshold to determine the risk nodes of the shipping network.

[0053] Optionally, in one embodiment of this application, the generation module includes: an initialization unit, configured to initialize the hyperparameters of the shipping network based on the original material shipping network data to generate target hyperparameters; a modeling unit, configured to model the explicit and implicit relationships of the shipping network based on the target hyperparameters to generate modeled explicit and implicit relationships; and a generation unit, configured to construct a joint optimization framework for the shipping network based on the modeled explicit and implicit relationships and the objective function, and generate low-dimensional embedding vectors corresponding to each node based on the joint optimization framework.

[0054] Optionally, in one embodiment of this application, the empirical distribution of the two nodes in the explicit relationship is as follows:

[0055]

[0056] In a bipartite network G = (U, V, E), U and V represent the sets of two types of nodes in the bipartite network, respectively, and E is the set of edges in the network. i and v j Then w represents the i-th and j-th nodes in U and V respectively.ij Then it is node u i and v j The weight of the edges between them;

[0057] The reconstruction distribution of the two nodes in the explicit relationship is as follows:

[0058]

[0059] Among them, u i v j Representing node u respectively i and v j d-dimensional embedding vector;

[0060] The KL divergence formula in the explicit relation is:

[0061]

[0062] Where log(·) refers to the logarithmic function with base e, and P(i,j) and The two nodes u are respectively i and v j The optimization objective in the explicit relation is to minimize the empirical distribution and the reconstructed distribution of O1.

[0063] The second-order proximity formula for two nodes in the implicit relation is:

[0064]

[0065] in, That is, the two nodes u in the binary network node set U. i and u j The second-order proximity of w ik For node u i and v k The weight of the edges between them. The definitions are similar;

[0066] The objective function in the implicit relation is:

[0067]

[0068] Where S represents the node sequence, D U and D V Let C represent the set of node sequences for two types of networks obtained through random walks. S (u i ) represents node u i In the context node of sequence S, P(u) c |u i That is, in u c Observed in the context of ui The probability, P(v) c |v j Similarly, maximizing O2 and O3 is the implicit relationship objective.

[0069] The calculation formula for the joint optimization framework is as follows:

[0070] max L=αlogO2+βlogO3-γO1

[0071] Here, parameters α, β, and γ are all hyperparameters used to balance different parts of the joint optimization framework.

[0072] Optionally, in one embodiment of this application, the training module includes: a concatenation unit, configured to concatenate the low-dimensional embedding vector to generate a feature vector of a sample, and determine the target variable based on the edge when the edge exists between the two nodes; an allocation unit, configured to use edges existing before a preset time point as positive samples in the training set, and edges existing after the preset time point as positive samples in the test set, and allocate negative samples to the training set and the test set according to a target ratio; and an execution unit, configured to perform a binary classification task of the low-dimensional embedding vector based on the training set and the test set using the pre-built binary classification model. The system performs a binary classification task and uses prior information to initialize a target weak learner based on the binary classification task and the log-likelihood loss function, generating initialization results. A fitting unit calculates the pseudo-residuals of the regression tree based on the initialization results, fits the regression tree according to the pseudo-residuals, generates fitting data, and updates the strong learner based on the fitting data to generate the final strong learner. A training unit inputs the feature vector and the target vector into the pre-built binary classification model for training, until the pre-built binary classification model meets a preset training stopping condition, generating the link prediction model for the maritime network.

[0073] Optionally, in one embodiment of this application, the formulas for the feature vector and the target vector are respectively:

[0074] X uv =[u i ,v j ]

[0075]

[0076] Among them, u i v j Representing node u respectively i and v j The embedding vector, where E is the edge set of the graph;

[0077] The formulas for the positive samples in the training set and the positive samples in the test set are as follows:

[0078]

[0079] Among them, u i and v j Let E(t) represent the node, E(t) represent the set of edges at time t, and t0 represent the selected time point.

[0080] The formula for calculating the log-likelihood loss function is as follows:

[0081] L(y i ,F(x i ))=-[y i logP(y i =1|x i )+(1-y i )log(1-P(y i =1|x i ))]

[0082]

[0083] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i P(y) represents the model's predicted value for sample i. i =1|x i ) represents the probability that the model predicts the existence of a link in sample i, and log(·) refers to the logarithmic function with base e;

[0084] The initialization formula for the target weak learner is:

[0085]

[0086] Where P(Y=1|x) is the proportion of y=1 in the training samples;

[0087] The formula for calculating the pseudo residual is:

[0088]

[0089] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i ) represents the current model's predicted value for sample i. To predict the probability that a link exists in this sample, F m-1 (x) is the learner obtained after training the (m-1)th regression tree;

[0090] The fitting formula for the regression tree is:

[0091]

[0092] Wherein, the leaf region corresponding to the m-th regression tree is R. m,j Where j = 1, 2, ..., J m J m Let m be the number of leaf nodes of the m-th regression tree;

[0093] The update formula for the strong learner is:

[0094]

[0095] Among them, c m,j The best fit for leaf node j of the m-th tree;

[0096] The final formula for generating the strong learner is:

[0097]

[0098] Where I(·) represents the indicator function, M is the total number of regression trees trained by the model, and x represents the feature vector of a sample to be predicted.

[0099] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding as described in the above embodiments.

[0100] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding.

[0101] A fifth aspect of this application provides a computer program product that stores a computer program that, when executed by a processor, implements the above-described method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding.

[0102] This application's embodiments first obtain the node embedding vectors from the original maritime transport network using the BiNE bipartite graph embedding algorithm, which serves as input to the classification model. Then, a LightGBM model is constructed to obtain a link prediction model for the maritime transport network. Next, the embedding vectors of nodes belonging to corresponding countries are randomly shuffled and input into the model. The impact of each country on network evolution is evaluated based on the decrease in link prediction accuracy, thereby reducing uncertainty in network evolution and optimizing maritime transport network management. This solves the problem that traditional network indicators can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. Especially when predicting the impact of nodes on the future evolution of the maritime transport network, existing methods have significant limitations, making it difficult to identify key nodes that play a crucial role in the direction of maritime transport network evolution. Furthermore, since strategic material maritime transport networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability in this structure.

[0103] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0104] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0105] Figure 1 This is a flowchart illustrating a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, according to an embodiment of this application.

[0106] Figure 2 This is an overall flowchart according to one embodiment of the present application;

[0107] Figure 3 A maritime network diagram based on input data according to one embodiment of this application;

[0108] Figure 4 This is a comparison diagram of the prediction results of the link prediction model before embedding vector shuffling according to an embodiment of this application and the real network.

[0109] Figure 5 This is a schematic diagram of a risk node analysis device for a material shipping network based on bipartite graph embedding, according to an embodiment of this application.

[0110] Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0111] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0112] The following describes a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, according to an embodiment of this application, with reference to the accompanying drawings. Addressing the limitations mentioned in the background art, traditional network indicators can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. Especially when predicting the impact of nodes on the future evolution of the maritime transport network, existing methods have significant limitations, making it difficult to identify key nodes that play a crucial role in the direction of the maritime transport network's evolution. Furthermore, since strategic material maritime transport networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability under this structure. This application provides a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding. In this method, the embedding vectors of nodes are first obtained from the original maritime transport network using the BiNE bipartite graph embedding algorithm and used as input to a classification model. Then, a LightGBM model is constructed to obtain a link prediction model for the maritime transport network. Next, the embedding vectors of nodes belonging to corresponding countries are randomly shuffled and input into the model. The degree of influence of the corresponding country on network evolution is evaluated based on the decrease in link prediction accuracy, thereby reducing uncertainty in the network evolution process and optimizing the management of the maritime transport network. This solves the problem that traditional network indicators can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. In particular, when predicting the impact of nodes on the future evolution of maritime networks, existing methods have significant limitations and find it difficult to identify key nodes that play an important role in the direction of maritime network evolution. In addition, since strategic material maritime networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability under this structure.

[0113] Specifically, Figure 1 This is a flowchart illustrating a risk node analysis method for a material shipping network based on bipartite graph embedding, provided in an embodiment of this application.

[0114] like Figure 1 As shown, this method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding includes the following steps:

[0115] In step S101, based on the original material shipping network data, a low-dimensional embedding vector is generated for each node in the shipping network.

[0116] It is understood that the material maritime transport network data in the embodiments of this application can be strategic material maritime transport network data.

[0117] In actual implementation, the embodiments of this application can use the BiNE (Bipartite Network Embedding) algorithm to solve for the low-dimensional embedding vectors corresponding to each node (i.e., port) in the network based on the original strategic material maritime transport network data.

[0118] Optionally, in one embodiment of this application, generating low-dimensional embedding vectors corresponding to each node in the shipping network based on the original material shipping network data includes: initializing the hyperparameters of the shipping network based on the original material shipping network data to generate target hyperparameters; modeling the explicit and implicit relationships of the shipping network based on the target hyperparameters to generate modeled explicit and implicit relationships; constructing a joint optimization framework for the shipping network based on the modeled explicit and implicit relationships and the objective function, and generating low-dimensional embedding vectors corresponding to each node based on the joint optimization framework.

[0119] Among them, such as Figure 2 As shown in the embodiments of this application, parameter initialization and data input can be performed. Based on the original strategic material maritime transport network data, as shown in Table 1, the hyperparameters of the maritime transport network are initialized to generate the target hyperparameters. The input maritime transport network data is G = (U, V, E), where U and V represent the sets of departure ports and arrival ports of the shipping route, respectively. The set of edges between two types of nodes is defined. Using u... i and v j Let i and j represent the i-th and j-th nodes in U and V, respectively, where i = 1, 2, ..., |U| and j = 1, 2, ..., |V|. The node embedding vectors are trained using the maritime network data before time t0 as input data. Table 1 shows the main hyperparameters of this application, as illustrated in Table 1:

[0120] Table 1

[0121]

[0122] Furthermore, embodiments of this application can embed nodes into a maritime network, modeling explicit and implicit relationships in the maritime network based on target hyperparameters to generate modeled explicit and implicit relationships. Explicit relationships are the edges observed in the network. Edges in a bipartite graph exist among two different types of nodes. Explicit relationships are modeled by considering the local proximity between two connected nodes. In the network space, if the connection between two nodes is strong (i.e., the edge weight is large), the probability of them co-occurring is also higher; therefore, node u... i With v jThe joint probability between them is defined as in Equation (1). In the embedding space, the inner product is used to represent the local proximity of two nodes, and the sigmoid function is used to convert the interaction values ​​into a probability space, as in Equation (2). Then, by minimizing the above empirical distribution P(i,j) and the reconstructed distribution... The difference between the distributions is used to learn the embedding vector. Here, the KL (Kullback-Leibler) divergence is used as a measure of the difference between the distributions, as shown in Equation (3).

[0123] In one embodiment of this application, the empirical distribution of the two nodes in the explicit relationship is as follows:

[0124]

[0125] In a bipartite network G = (U, V, E), U and V represent the sets of two types of nodes in the bipartite network, respectively, and E is the set of edges in the network. i and v j Then w represents the i-th and j-th nodes in U and V respectively. ij Then it is node u i and v j The weight of the edges between them;

[0126] The reconstructed distribution of the probability space in explicit relations is as follows:

[0127]

[0128] Among them, u i v j Representing node u respectively i and v j d-dimensional embedding vector;

[0129] The KL divergence formula in explicit relations is:

[0130]

[0131] Where log(·) refers to the logarithmic function with base e, and P(i,j) and Let O1 be the empirical distribution and the reconstruction distribution of the two nodes, respectively. Minimizing O1 is the optimization objective of the explicit relationship.

[0132] There is no explicit connection between two nodes of the same type, but if they are connected to the same node, they are considered to have some implicit relationship. Since there are two different types of nodes in a bipartite graph (starting and ending nodes), for two nodes of the same type, their second-order proximity is defined as shown in Equation (4). From Equation (4), two types of networks can be obtained. Random walks on these two networks can yield the required set of node sequences. Here, the minimum and maximum number of random walks for each node are limited to minT and maxT, respectively. At the same time, according to the centrality index of the nodes in the graph, nodes with higher centrality are more likely to be selected as the starting point of the random walk, so as to preserve the importance of the nodes to a certain extent. And a termination probability p is assigned to each step of the random walk so that the generated node sequence has a variable length.

[0133] After obtaining the node sequence sets of two types of nodes, the Skip-gram model is applied to these two node sequence sets to learn node embeddings. The Skip-gram model learns word embedding representations by maximizing the co-occurrence probability of the center word and its context words. For the maritime network, given a node sequence and a node, its context consists of nodes within a given window size W before and after the node in the sequence. For the two node sequence sets, the objective functions are defined as follows: Equation (5) and Equation (6). Here, the model learning rate is set to λ1.

[0134] The second-order proximity formula for two nodes in an implicit relation is:

[0135]

[0136] in, That is, two nodes u in the set of nodes U of a binary network. i and u j The second-order proximity of w ik For node u i and v k The weight of the edges between them. The definitions are similar;

[0137] The objective function in the implicit relation is:

[0138]

[0139] Where S represents the node sequence, D U and D V Let C represent the set of node sequences for two types of networks obtained through random walks. S (u i ) represents node u i In the context node of sequence S, P(u)c |u i That is, observing u in the context of uc. i The probability, P(v) c |v j Similarly, maximizing O2 and O3 is the optimization objective of the implicit relationship.

[0140] Furthermore, embodiments of this application can solve for node embedding vectors. Combining the modeling of the explicit and implicit relationships described above with their objective functions, a joint optimization framework is constructed as shown in formula (7). The joint model is optimized using the Stochastic Gradient Ascent (SGA) algorithm to obtain the embedding vectors of each node. The calculation formula for the joint optimization framework is:

[0141] max L=αlogO2+βlogO3-γO1 (7)

[0142] Here, parameters α, β, and γ are all hyperparameters used to balance different parts of the joint optimization framework.

[0143] In step S102, the low-dimensional embedding vector is used as the feature variable and the edge between two nodes is used as the target variable. The feature vector and the target vector are input into a pre-built binary classification model for training to generate a link prediction model for the maritime network.

[0144] In actual implementation, this application embodiment uses the low-dimensional embedding vectors of two nodes as feature variables and the existence of an edge between the two nodes as the target variable. The feature vectors and target vectors are input into a pre-built LightGBM binary classification model for training to generate a link prediction model for the maritime network.

[0145] Optionally, in one embodiment of this application, a link prediction model for a maritime network is generated by using a low-dimensional embedding vector as a feature variable and the edge between two nodes as a target variable, and inputting the feature vector and target vector into a pre-built binary classification model for training. This includes: concatenating the low-dimensional embedding vector to generate a feature vector for a sample, and determining the target variable based on the edge when an edge exists between the two nodes; using edges existing before a preset time point as positive samples in the training set, and edges existing after the preset time point as positive samples in the test set, and allocating negative samples to the training and test sets according to a target ratio; and then, based on the training and test sets... The test set utilizes a pre-built binary classification model to perform a binary classification task with low-dimensional embedding vectors. Based on the binary classification task and the log-likelihood loss function, a target weak learner is initialized using prior information to generate initialization results. Based on the initialization results, the pseudo-residuals of the regression tree are calculated, and the regression tree is fitted according to the pseudo-residuals to generate fitted data. The strong learner is then updated based on the fitted data to generate the final strong learner. Based on the final strong learner, the feature vector and target vector are input into the pre-built binary classification model for training until the pre-built binary classification model meets the preset training stopping condition, thus generating a link prediction model for the maritime network.

[0146] Specifically, in the embodiments of this application, input variables can be constructed by building samples in pairs of nodes. The feature variable X of a sample can be obtained by directly concatenating the embedding vectors of the two nodes. uv As shown in formula (8). The target variable Y of this sample uv Then it is node u i and v j Whether an edge exists directly is shown in formula (9), and if an edge exists between two nodes, the target variable is determined based on the edge.

[0147] X uv =[u i ,v j (8)

[0148]

[0149] Among them, u i v j Representing node u respectively i and v j The embedding vector, E is the edge set of the graph.

[0150] This application embodiment can generate training and testing sets. Positive samples for the model are directly obtained from the edge set of the graph, while negative samples are sampled from unconnected node pairs in the same number as the positive samples. Edges existing before a selected time are used as positive samples in the training set. As shown in formula (10). Edges that appear after the selected time are used as positive samples in the test set. As shown in formula (11). 80% of the negative samples are added to the training set, and the remaining 20% ​​are added to the test set.

[0151]

[0152] Among them, u i and v j Let E(t) represent a node, E(t) represent the set of edges at time t, and t0 represent the selected time point.

[0153] Furthermore, the embodiments of this application can construct and train a LightGBM model. To effectively complete the training task of a binary classifier based on the above-mentioned embedding vectors, a LightGBM model based on Gradient Boosting Decision Tree (GBDT) is used. For this binary classification task, the log-likelihood loss function of formulas (12) and (13) is used. The first weak learner F0(x) is initialized using prior information according to formula (14). For building M CART regression trees, the pseudo residual (i.e., the negative gradient of the loss function) r corresponding to the m-th (m=1,2,…,M) tree is calculated using formula (15). m,i Then (x) i ,r m,i Using the training data, a regression tree is fitted with a learning rate λ2. The m-th regression tree is obtained. For the leaf node region, the best fit is calculated as shown in formula (16), and then the strong learner F is updated according to formula (17). m (x), and the final strong learner F is obtained according to formula (18) using the above process. M (x). Meanwhile, to avoid overfitting, the maximum number of leaves and the maximum depth of the tree are set to N, respectively. l With N d .

[0154] The formula for calculating the log-likelihood loss function is:

[0155] L(y i ,F(x i ))=-[y i logP(y i =1|x i )+(1-y i )log(1-P(y i =1|x i ))](12)

[0156]

[0157] Where, x i With y iLet F(x) represent the feature variable and the target variable of a single sample, respectively. i P(y) represents the model's predicted value for sample i. i =1|x i ) represents the probability that the model predicts the existence of a link in sample i, and log(·) refers to the logarithmic function with base e;

[0158] The initialization formula for the target weak learner is:

[0159]

[0160] Where P(Y=1|x) is the proportion of y=1 in the training samples;

[0161] The formula for calculating pseudo residuals is:

[0162]

[0163] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i ) represents the current model's predicted value for sample i. To predict the probability that a link exists in this sample, F m-1 (x) is the learner obtained after training the (m-1)th regression tree;

[0164] The fitting formula for the regression tree is:

[0165]

[0166] Wherein, the leaf region corresponding to the m-th regression tree is R. m,j Where j = 1, 2, ..., J m J m Let m be the number of leaf nodes of the m-th regression tree;

[0167] The update formula for a strong learner is:

[0168]

[0169] Among them, c m,j The best fit for leaf node j of the m-th tree;

[0170] The final formula for generating the strong learner is:

[0171]

[0172] Where I(·) represents the indicator function, M is the total number of regression trees trained by the model, and x represents the feature vector of a sample to be predicted.

[0173] Based on the final strong learner, the training dataset is input, and the feature vector and target vector are fed into a pre-built binary classification model for training. This process continues until the pre-built binary classification model meets a preset training stopping condition, generating a link prediction model for the maritime network. In other words, the model stops training after reaching a predetermined number of iterations following the above process. After the model training is complete, the feature variables of the test set are input into the model for forward propagation to obtain the predicted probability value. The AUC value is calculated to measure the model's ability to distinguish between positive and negative samples.

[0174] In step S103, the embedding vectors of the target area nodes are randomly shuffled to generate target embedding vectors, and the target embedding vectors are input into the link prediction model to calculate the decrease value of the area under the curve corresponding to the link prediction model. The risk nodes of the maritime network are determined by using the target areas where the decrease value exceeds a preset threshold.

[0175] It is understandable that the target region where the decline value exceeds the preset threshold can be a country where the AUC value declines beyond a certain threshold.

[0176] As one possible implementation, the embedding vectors of the target area nodes can be randomly shuffled to generate target embedding vectors. The target embedding vectors are then input into the link prediction model to calculate the area under the curve (AUC) value corresponding to the link prediction model. In the case of area nodes whose AUC value drops below a preset threshold, risk nodes in the maritime network are determined based on the area nodes.

[0177] A country is selected, and the embedding vectors of all nodes belonging to that country in the dataset are found. These embedding vectors are randomly shuffled and then input into the link prediction model for training. The output of the link prediction model with shuffled embedding vectors is calculated on the test set, and the corresponding AUC value is calculated. The AUC value of the original model is subtracted from the AUC value of the model with shuffled embedding vectors to obtain the country's impact on network evolution. Nodes in countries whose AUC value drops more than a threshold θ after shuffling are considered risk nodes.

[0178] The embodiments of this application can determine that when the accuracy of link prediction decreases significantly after the embedding vector corresponding to a certain node is shuffled, the node is considered to have a significant impact on network evolution, aiming to reduce the uncertainty in the network evolution process and optimize the management of maritime networks.

[0179] Specifically, it can be combined with Figure 3 and Figure 4 As shown, a specific embodiment is used to elaborate in detail on the working principle of a risk node analysis method for a material shipping network based on bipartite graph embedding in this application.

[0180] This application embodiment can select coal shipping network data from January 2014 to December 2023 to conduct risk node analysis of the strategic material shipping network. The network is visualized based on port geographical locations, as shown below. Figure 3 As shown, an embodiment of this application is given for data input and hyperparameter initialization, node embedding in the maritime network, link prediction model training and evaluation, and risk node analysis of the maritime network.

[0181] (I) Hyperparameter initialization and data input

[0182] Input coal shipping network data with ports as nodes from January 2014 to December 2023, with network data from January 2014 to December 2019 used to train node embeddings, specifically involving 505 ports and 4508 shipping routes. Ports can be divided into sets U and V based on departure and arrival ports. Table 2 shows the initialization hyperparameter table for this application, as shown in Table 2:

[0183] Table 2

[0184] hyperparameter name meaning hyperparameter name meaning d 128 p 0.15 W 5 <![CDATA[λ1]]> 0.008 α 0.1 M 100 β 1 <![CDATA[λ2]]> 0.1 γ 1 <![CDATA[N l ]]> 31 minT 1 <![CDATA[N d ]]> 20 maxT 32 θ 0.015 <![CDATA[t0]]> January 2020

[0185] (II) Node Embedding in Maritime Networks

[0186] 1. For the node pair corresponding to the edge in the network, use formula (1) to calculate the joint probability distribution of the two nodes. For example, for node u=1 and node v=3, their joint probability distribution in the network space is 0.0020. Determine the optimization objective of the explicit relationship according to formula (2)-(3).

[0187] 2. Based on formulas (4)-(6), the optimization objectives of the implicit relations of the two sets are obtained.

[0188] 3. According to formula (7), the joint optimization framework is solved using stochastic gradient ascent to obtain the embedding vectors corresponding to each node. For example, for node u = 1, its embedding vector is: u1 = [0.2250, 0.2197, 0.2084, 0.3207, 0.2802, ...] (128 dimensions)

[0189] (III) Training and Evaluation of Link Prediction Model

[0190] 1. According to formulas (8)-(9), all input variables are obtained. For example, for node u=1 and node v=2, the feature variable of this sample is X. uv = [0.2250, 0.2197, ..., 0.3327, 0.7812, ...] (256 dimensions), with the objective variable being Y. uv =0.

[0191] 2. Using the edges of the coal shipping network from January 2014 to December 2019 as positive samples in the training set, and the edges from January 2020 to December 2023 as positive samples in the test set, the same number of negative samples as the positive samples are sampled from the non-existent edges, and then added to the training set and the test set at a ratio of 4:1.

[0192] 3. Initialize the model parameters, then train the LightGBM model using the above test set and perform predictions on the test set. Taking node u=1 and node v=2 as an example, the link prediction probability is obtained as follows: The AUC value of the model is obtained based on the true probability and predicted probability of all samples. For example, the AUC value of the link prediction model in this embodiment is 0.8048.

[0193] (iv) Risk Node Analysis of Maritime Networks

[0194] 1. Randomly shuffle the embedding vectors by country. For example, for country CN, the port nodes belonging to it are v = 2, 5, 10, ... Randomly shuffle the embedding vectors corresponding to these nodes. For example, the embedding vector for v = 2 is:

[0195] v2=[0.3327,0.7812,0.3021,0.4129,0.1802,…]

[0196] Randomly shuffled as follows:

[0197] v'2=[0.1802,0.3327,0.4129,0.7812,0.3021,…]

[0198] The shuffled embedding vectors are then used as input variables for the model.

[0199] 2. Calculate the AUC value of the link prediction model after shuffling the embedding vectors. For example, the AUC value of the link prediction model after shuffling the embedding vectors of all port nodes belonging to CN is 0.7749. The AUC value decreased by 0.0299, exceeding the given threshold of 0.015. Therefore, CN is considered to have a significant impact on the evolution of the maritime network and is regarded as a risk node.

[0200] Test Analysis: Compare the predicted coal shipping network links before shuffling the embedding vectors with the actual observed data (e.g., Figure 3 As shown in the figure, the two are highly consistent. Furthermore, the accuracy, precision, and recall of this link prediction model are calculated to be 0.7365, 0.7278, and 0.8088, respectively, indicating that the prediction model has relatively high accuracy. Since subsequent risk node analyses are all based on this link prediction model, the results of the risk node analysis can be considered to have high reliability and practicality.

[0201] The proposed method for risk node analysis in maritime transport networks based on bipartite graph embedding, according to an embodiment of this application, first obtains the embedding vectors of nodes from the original maritime transport network using the BiNE bipartite graph embedding algorithm, which serves as input to a classification model. Then, a LightGBM model is constructed to obtain a link prediction model for the maritime transport network. Next, the embedding vectors of nodes belonging to corresponding countries are randomly shuffled and input into the model. The impact of each country on network evolution is evaluated based on the decrease in link prediction accuracy, thereby reducing uncertainty in network evolution and optimizing maritime transport network management. This solves the problem that traditional network indicators can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. Especially when predicting the impact of nodes on the future evolution of the maritime transport network, existing methods have significant limitations, making it difficult to identify key nodes that play a crucial role in the direction of maritime transport network evolution. Furthermore, since strategic material maritime transport networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability in this structure.

[0202] Next, referring to the accompanying drawings, a risk node analysis device for a material shipping network based on bipartite graph embedding is described according to an embodiment of this application.

[0203] Figure 5 This is a schematic diagram of the structure of a risk node analysis device for a material shipping network based on bipartite graph embedding, according to an embodiment of this application.

[0204] like Figure 5 As shown, the risk node analysis device 10 for the material shipping network based on bipartite graph embedding includes: a generation module 100, a training module 200, and a determination module 300.

[0205] Specifically, the generation module 100 is used to generate low-dimensional embedding vectors corresponding to each node in the shipping network based on the original shipping network data.

[0206] Training module 200 is used to train a pre-built binary classification model by using low-dimensional embedding vectors as feature variables and the edge between two nodes as target variables, in order to generate a link prediction model for the maritime network.

[0207] The determination module 300 is used to randomly shuffle the embedding vectors of the target area nodes to generate target embedding vectors, and input the target embedding vectors into the link prediction model to calculate the decrease value of the area under the curve corresponding to the link prediction model, and use the target area with the decrease value exceeding the preset threshold to determine the risk nodes of the maritime network.

[0208] Optionally, in one embodiment of this application, the generation module 100 includes: an initialization unit, a modeling unit, and a generation unit.

[0209] The initialization unit is used to initialize the hyperparameters of the shipping network based on the original material shipping network data in order to generate the target hyperparameters.

[0210] Modeling units are used to model the explicit and implicit relationships of a maritime network based on target hyperparameters, generating the modeled explicit and implicit relationships.

[0211] The generation unit is used to construct a joint optimization framework for the maritime network based on the modeled explicit and implicit relationships and the objective function, and to generate low-dimensional embedding vectors for each node based on the joint optimization framework.

[0212] Optionally, in one embodiment of this application, the empirical distribution of two nodes in an explicit relationship is as follows:

[0213]

[0214] In a bipartite network G = (U, V, E), U and V represent the sets of two types of nodes in the bipartite network, respectively, and E is the set of edges in the network. i and v j Then w represents the i-th and j-th nodes in U and V respectively. ij Then it is node u i and v j The weight of the edges between them;

[0215] The reconstruction distribution of two nodes in an explicit relation is as follows:

[0216]

[0217] Among them, u i v j Representing node u respectively i and v j d-dimensional embedding vector;

[0218] The KL divergence formula in explicit relations is:

[0219]

[0220] Where log(·) refers to the logarithmic function with base e, and P(i,j) and These are the two nodes u mentioned above. i and v j Minimizing O1 from the empirical distribution and the reconstructed distribution is the optimization objective in the explicit relation;

[0221] The second-order proximity formula for two nodes in an implicit relation is:

[0222]

[0223] in, That is, two nodes u in the set of nodes U of a binary network. i and u j The second-order proximity of w ik For node u i and v k The weight of the edges between them. The definitions are similar;

[0224] The objective function in the implicit relation is:

[0225]

[0226] Where S represents the node sequence, D U and D V Let C represent the set of node sequences of two types of networks obtained through random walks. S (u i ) represents node u i In the context node of sequence S, P(u) c |u i That is, in u c Observed in the context of u i The probability, P(v) c |v j Similarly, maximizing O2 and O3 is the implicit relational objective;

[0227] The calculation formula for the joint optimization framework is as follows:

[0228] max L=αlogO2+βlogO3-γO1

[0229] Here, parameters α, β, and γ are all hyperparameters used to balance different parts of the joint optimization framework.

[0230] Optionally, in one embodiment of this application, the training module 200 includes: a splicing unit, an allocation unit, an execution unit, and a fitting unit.

[0231] The splicing unit is used to splice low-dimensional embedding vectors to generate a feature vector of a sample, and to determine the target variable based on the edge when there is an edge between two nodes.

[0232] The allocation unit is used to use edges that exist before a preset time point as positive samples in the training set and edges that exist after a preset time point as positive samples in the test set, and to allocate negative samples to the training set and the test set according to a target ratio.

[0233] The execution unit is used to perform a binary classification task with low-dimensional embedding vectors based on the training set and the test set, and to initialize the target weak learner with prior information based on the binary classification task and the log-likelihood loss function, and generate the initialization result.

[0234] The fitting unit is used to calculate the pseudo-residuals of the regression tree based on the initialization results, fit the regression tree according to the pseudo-residuals, generate fitting data, and update the strong learner according to the fitting data to generate the final strong learner.

[0235] The training unit is used to train a pre-built binary classification model based on the final strong learner by inputting the feature vector and the target vector into the model until the pre-built binary classification model meets the preset training stopping condition, thereby generating a link prediction model for the maritime network.

[0236] Optionally, in one embodiment of this application, the formulas for the feature vector and the target vector are as follows:

[0237] X uv =[u i ,v j ]

[0238]

[0239] Among them, u i v j Representing node u respectively i and v j The embedding vector, where E is the edge set of the graph;

[0240] The formulas for positive samples in the training set and positive samples in the test set are as follows:

[0241]

[0242] Among them, u i and v j Let E(t) represent a node, E(t) represent the set of edges at time t, and t0 represent the selected time point.

[0243] The formula for calculating the log-likelihood loss function is:

[0244] L(y i ,F(x i ))=-[y i logP(y i =1|x i )+(1-y i )log(1-P(y i =1|x i ))]

[0245]

[0246] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i P(y) represents the model's predicted value for sample i. i =1|x i ) represents the probability that the model predicts the existence of a link in sample i, and log(·) refers to the logarithmic function with base e;

[0247] The initialization formula for the target weak learner is:

[0248]

[0249] Where P(Y=1|x) is the proportion of y=1 in the training samples;

[0250] The formula for calculating pseudo residuals is:

[0251]

[0252] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i ) represents the current model's predicted value for sample i. To predict the probability that a link exists in this sample, F m-1 (x) is the learner obtained after training the (m-1)th regression tree;

[0253] The fitting formula for the regression tree is:

[0254]

[0255] Wherein, the leaf region corresponding to the m-th regression tree is R. m,j Where j = 1, 2, ..., J m J m Let m be the number of leaf nodes of the m-th regression tree;

[0256] The update formula for a strong learner is:

[0257]

[0258] Among them, c m,j The best fit for leaf node j of the m-th tree;

[0259] The final formula for generating the strong learner is:

[0260]

[0261] Where I(·) represents the indicator function, M is the total number of regression trees trained by the model, and x represents the feature vector of a sample to be predicted.

[0262] It should be noted that the foregoing explanation of the embodiment of the risk node analysis method for the material shipping network based on bipartite graph embedding also applies to the risk node analysis device for the material shipping network based on bipartite graph embedding in this embodiment, and will not be repeated here.

[0263] The risk node analysis device for maritime transport networks based on bipartite graph embedding proposed in this application first obtains the embedding vectors of nodes from the original maritime transport network using the BiNE bipartite graph embedding algorithm, which serves as input to the classification model. Then, a LightGBM model is constructed to obtain the link prediction model for the maritime transport network. Next, the embedding vectors of nodes belonging to corresponding countries are randomly shuffled and input into the model. The impact of each country on network evolution is evaluated based on the decrease in link prediction accuracy, thereby reducing uncertainty in network evolution and optimizing maritime transport network management. This solves the problem that traditional network indicators can only reflect some structural features of the network and cannot deeply explore the potential dependencies and interactions between nodes. Especially when predicting the impact of nodes on the future evolution of the maritime transport network, existing methods have significant limitations, making it difficult to identify key nodes that play an important role in the direction of maritime transport network evolution. Furthermore, since strategic material maritime transport networks often exhibit bipartite graph characteristics, many classic network analysis indicators and techniques have poor applicability in this structure.

[0264] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0265] The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0266] When the processor 602 executes the program, it implements the risk node analysis method for a material shipping network based on bipartite graph embedding provided in the above embodiments.

[0267] Furthermore, electronic devices also include:

[0268] Communication interface 603 is used for communication between memory 601 and processor 602.

[0269] The memory 601 is used to store computer programs that can run on the processor 602.

[0270] The memory 601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0271] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0272] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0273] The processor 602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0274] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for analyzing risk nodes in a material shipping network based on bipartite graph embedding.

[0275] This application also provides a computer program product storing a computer program that, when executed by a processor, implements the above-described method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding.

[0276] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0277] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0278] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0279] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0280] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0281] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0282] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0283] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, characterized in that, Includes the following steps: Based on the original maritime shipping network data, low-dimensional embedding vectors are generated for each node in the shipping network. Using the low-dimensional embedding vector as the feature variable and the edge between two nodes as the target variable, the feature vector and the target vector are input into a pre-built binary classification model for training to generate the link prediction model of the maritime network. The embedding vectors of the target region nodes are randomly shuffled to generate target embedding vectors. These target embedding vectors are then input into the link prediction model to calculate the decrease in the area under the curve corresponding to the link prediction model. Target regions where the decrease exceeds a preset threshold are used to identify risk nodes in the maritime network.

2. The method according to claim 1, characterized in that, The process of generating low-dimensional embedding vectors for each node in the shipping network based on the original shipping network data includes: Based on the original material shipping network data, the hyperparameters of the shipping network are initialized to generate the target hyperparameters; Based on the target hyperparameters, the explicit and implicit relationships of the maritime network are modeled to generate the modeled explicit and implicit relationships. Based on the modeled explicit and implicit relationships and the objective function, a joint optimization framework for the maritime network is constructed, and low-dimensional embedding vectors corresponding to each node are generated according to the joint optimization framework.

3. The method according to claim 2, characterized in that, The empirical distribution of the two nodes in the explicit relationship is as follows: In a bipartite network G = (U, V, E), U and V represent the sets of two types of nodes in the bipartite network, respectively, and E is the set of edges in the network. i and v j Then w represents the i-th and j-th nodes in U and V respectively. ij Then it is node u i and v j The weight of the edges between them; The reconstruction distribution of the two nodes in the explicit relationship is as follows: Among them, u i v j Representing node u respectively i and v j d-dimensional embedding vector; The KL divergence formula in the explicit relation is: Where log(·) refers to the logarithmic function with base e, and P(i,j) and These are the two nodes u respectively. i and v j The optimization objective in the explicit relationship is to minimize the empirical distribution and the reconstructed distribution of O1. The second-order proximity formula for two nodes in the implicit relation is: in, That is, the two nodes u in the binary network node set U. i and u j The second-order proximity of w ik For node u i and v k The weight of the edges between them. The definitions are similar; The objective function in the implicit relation is: Where S represents the node sequence, D U and D V Let C represent the set of node sequences of two types of networks obtained through random walks. S (u i ) represents node u i In the context node of sequence S, P(u) c |u i That is, in u c Observed in the context of u i The probability, P(v) c |v j Similarly, maximizing O2 and O3 is the implicit relationship objective. The calculation formula for the joint optimization framework is as follows: max L=αlog O2+βlog O3-γO1 Here, parameters α, β, and γ are all hyperparameters used to balance different parts of the joint optimization framework.

4. The method according to claim 1, characterized in that, The step of using the low-dimensional embedding vector as the feature variable and the edge between two nodes as the target variable, inputting the feature vector and the target vector into a pre-built binary classification model for training to generate the link prediction model of the maritime network, includes: The low-dimensional embedding vector is concatenated to generate a feature vector for a sample, and the target variable is determined based on the edge between the two nodes if the edge exists. Edges existing before a preset time point are used as positive samples in the training set, and edges existing after the preset time point are used as positive samples in the test set. Negative samples are allocated to the training set and the test set according to a target ratio. Based on the training set and the test set, the binary classification task of the low-dimensional embedding vector is performed using the pre-built binary classification model, and the target weak learner is initialized using prior information based on the binary classification task and the log-likelihood loss function to generate the initialization result. Based on the initialization results, the pseudo residuals of the regression tree are calculated, and the regression tree is fitted according to the pseudo residuals to generate fitted data. The strong learner is then updated according to the fitted data to generate the final strong learner. Based on the final strong learner, the feature vector and the target vector are input into the pre-built binary classification model for training until the pre-built binary classification model meets the preset training stopping condition, thereby generating the link prediction model of the maritime network.

5. The method according to claim 4, characterized in that, The formulas for the feature vector and the target vector are as follows: X uv =[u i ,v j ] Among them, u i v j Representing node u respectively i and v j The embedding vector, where E is the edge set of the graph; The formulas for the positive samples in the training set and the positive samples in the test set are as follows: Among them, u i and v j Let E(t) represent the node, E(t) represent the set of edges at time t, and t0 represent the selected time point. The formula for calculating the log-likelihood loss function is as follows: L(y i ,F(x i ))=-[y i logP(y i =1|x i )+(1-y i )log(1-P(y i =1|x i ))] Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i P(y) represents the model's predicted value for sample i. i =1|x i ) represents the probability that the model predicts the existence of a link in sample i, and log(·) refers to the logarithmic function with base e; The initialization formula for the target weak learner is: Where P(Y=1|x) is the proportion of y=1 in the training samples; The formula for calculating the pseudo residual is: Where, x i With y i Let F(x) represent the feature variable and the target variable of a single sample, respectively. i ) represents the current model's predicted value for sample i. To predict the probability that a link exists in this sample, F m-1 (x) is the learner obtained after training the (m-1)th regression tree; The fitting formula for the regression tree is: Wherein, the leaf region corresponding to the m-th regression tree is R. m,j Where j = 1, 2, ..., J m J m Let be the number of leaf nodes of the m-th regression tree; The update formula for the strong learner is: Among them, c m,j The best fit for leaf node j of the m-th tree; The final formula for generating the strong learner is: Where I(·) represents the indicator function, M is the total number of regression trees trained by the model, and x represents the feature vector of a sample to be predicted.

6. A risk node analysis device for a maritime transport network based on bipartite graph embedding, characterized in that, include: The generation module is used to generate low-dimensional embedding vectors for each node in the shipping network based on the original shipping network data. The training module is used to train a pre-built binary classification model by using the low-dimensional embedding vector as the feature variable and the edge between two nodes as the target variable, so as to generate a link prediction model for the maritime network. The determination module is used to randomly shuffle the embedding vectors of the target area nodes to generate target embedding vectors, and input the target embedding vectors into the link prediction model to calculate the decrease value of the area under the curve corresponding to the link prediction model, and use the target area where the decrease value exceeds a preset threshold to determine the risk nodes of the maritime network.

7. The apparatus according to claim 6, characterized in that, The generation module includes: An initialization unit is used to initialize the hyperparameters of the shipping network based on the original material shipping network data, so as to generate target hyperparameters; The modeling unit is used to model the explicit and implicit relationships of the maritime network based on the target hyperparameters, so as to generate the modeled explicit and implicit relationships; The generation unit is used to construct a joint optimization framework for the maritime network based on the modeled explicit and implicit relationships and the objective function, and to generate low-dimensional embedding vectors corresponding to each node based on the joint optimization framework.

8. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding as described in any one of claims 1-5.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, as described in any one of claims 1-5.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement a method for analyzing risk nodes in a maritime transport network based on bipartite graph embedding, as described in any one of claims 1-5.

Images