Implicit relation detection method and detection system for sparse network

Abstract

The invention discloses an implicit relation detection method and detection system for a sparse network, and belongs to the technical field of electrical digital data processing, and the method comprises the following steps: obtaining embedded representation of nodes in a data set based on a graph convolutional network and relation edge sparsity; obtaining feature representation of the nodes according to the embedded representation, the meta-paths and the priorities of the meta-paths; obtaining candidate edges of the candidate node pairs; calculating the relation score of the candidate edge through the score function and the feature representation of the node in the node pair; and predicting whether the candidate edge exists or not according to the relation score. Aggregating data of nodes and relations of the sparse network to obtain embedding representation based on sparsity and feature representation based on relation edge sparsity and meta-paths; calculating candidate edge relation scores among the candidate nodes, and detecting whether an implicit relation exists or not; the complex multi-dimensional relationship between the nodes is captured through recursive propagation, the representation capability of node features is enhanced, and the accuracy of implicit relationship mining is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic digital data processing, and in particular to a method and system for detecting implicit relations of a sparse network. Background Art

[0002] With the rapid development of data networks, attribute sparse networks (sparse networks) are widely present in various scenarios, such as dark web, anonymous social networks, and low-annotation complex networks. These networks have the following common characteristics: Attribute sparsity: The explicit attributes and annotation data of nodes are usually severely missing, such as anonymous entities in dark web transactions, limited features of anonymous social networks, and missing user information on low-annotation social platforms. Structural complexity: The connection relationships between nodes are heterogeneous or dynamically changing, and implicit relationships cannot be directly obtained through explicit links. Sparse data distribution: When both attributes and relationships are sparse, existing methods find it difficult to fully model implicit associations.

[0003] Revealing these implicit relationships is of great significance for identifying key nodes of dark web transactions, discovering potential interaction patterns between anonymous users, and mining the interest behaviors of low-labeled users.

[0004] Facing the problem of mining implicit relationships in attribute-sparse networks, traditional technical solutions mainly include the following categories: Network modeling methods based on explicit dependencies: Relying on adjacency matrices and explicit relationship modeling, it is suitable for networks with complete attributes, but it is difficult to effectively model implicit relationships when node or edge attributes are missing. Embedding representation method: Generate node embeddings through neighbor feature aggregation to adapt to certain heterogeneity, but in attribute-sparse scenarios, it is difficult to fully capture effective information due to the lack of node or edge features. Network modeling methods based on deep learning: Use attention mechanisms combined with meta-paths to capture multiple types of relationships, but it is easy to overfit in sparse networks, and the propagation ability of long meta-paths is limited, making it difficult to model long-distance implicit relationships.

[0005] Therefore, it is necessary to design an implicit relationship detection method for sparse networks to mine the implicit relationships between entities or nodes in sparse networks. Summary of the invention

[0006] In view of the above technical problems existing in the prior art, the present invention provides a method and a system for detecting implicit relations in a sparse network, so as to mine implicit relations in the sparse network.

[0007] The present invention discloses a method for detecting implicit relations in a sparse network, comprising the following steps: collecting data on nodes and relations in the sparse network to obtain a data set; obtaining an embedded representation of the nodes in the data set based on a graph convolutional network and sparsity of relation edges; obtaining a feature representation of the nodes according to the embedded representation, meta-paths and their priorities; obtaining candidate edges of candidate node pairs, wherein the candidate edges are represented as implicit relations of the candidate node pairs; calculating the relation scores of the candidate edges by using a score function and the feature representations of the nodes in the node pairs; and predicting whether the candidate edges exist according to the relation scores.

[0008] Preferably, the method for determining whether a candidate edge exists includes:

[0009] If the relationship score is greater than or equal to the first threshold τ, then there is a candidate edge;

[0010] If the relationship score is less than the first threshold, there is no candidate edge.

[0011] Preferably, the relationship score is calculated as follows:

[0012]

[0013] Among them, P(μ,v) represents the relationship score of the candidate edge between node u and node v, h u It is represented as the feature representation of node u, h v It is represented as the feature representation of node v, W is a learnable weight matrix used to model the interaction relationship between nodes; b is a learnable bias term used to adjust the output of the scoring function; σ() is represented as the Sigmoid activation function.

[0014] Preferably, a formula for node embedding representation is:

[0015]

[0016] is the embedding representation of node u at layer k; is the embedding representation of node u at the k-1 layer; is the embedding representation of the neighbor node v at the k-1 layer, N(u) is the set of neighbor nodes of node u, ω uv is the edge weight from neighbor node v to node u, α u The sparsity weight of node u.

[0017] Preferably, another formula for node embedding representation is:

[0018]

[0019] in, is the embedding representation of node u at layer k, is the embedding representation of the neighbor node v at the k-1 layer, N(u) is the set of neighbor nodes of node u; β(e uv ) is the relationship edge e uv The sparsity weight, ω uv is the edge weight from neighbor node v to node u.

[0020] Preferably, the formula for the feature representation of a node is:

[0021]

[0022] Among them, h u It is represented as the feature representation of node u, h v Represented as the feature representation of node v, N m (u) is the set of neighbor nodes of node u on meta-path m, ω uv is the edge weight from neighbor node v to node u, ω m is the meta-path weight.

[0023] Preferably, the meta-path weight is calculated as follows:

[0024] ω m =softmax(I(m))

[0025]

[0026] I(m) is represented as the importance score of meta-path m, G(u,v) is represented as the information gain of node u and node v, L(u,v) is represented as the path length of node u and node v, node u and node v belong to meta-path m, and softmax() is represented as the normalized exponential function.

[0027] Preferably, the parameters of the sparsity of the relationship edges include the edge weights between the node and the neighboring nodes and the sparsity weights of the relationship edges.

[0028] Preferably, the method further includes obtaining a relationship score based on a prediction model:

[0029] The nodes and relationship edges of the data are modeled through a graph convolutional network and a preset meta-path priority to obtain a first prediction model;

[0030] According to the embedding representation of the node, the sparsity weight of the relationship edge and the edge weight from the neighbor node to the node are obtained;

[0031] According to the sparsity weight and the edge weight, updating the parameters of the first prediction model to obtain a second prediction model;

[0032] The data set is predicted using the second prediction model to obtain a relationship score.

[0033] The present invention also provides a detection system for implementing the above method, comprising a collection module, an aggregation module, a feature extraction module, a relationship score calculation module and a prediction module.

[0034] The acquisition module is used to collect data on nodes and relationships of the sparse network to obtain a data set;

[0035] The aggregation module is used to obtain an embedded representation of nodes in a data set based on a graph convolutional network and sparsity of relationship edges;

[0036] The feature extraction module is used to obtain a feature representation of a node according to the embedded representation, the meta-path and its priority;

[0037] The relationship score calculation module is used to obtain a set of candidate node pairs and a set of candidate edges of the candidate node pairs, and calculate the relationship scores of the candidate edges through the score function and the feature representation of the nodes in the node pairs;

[0038] The prediction module is used to predict whether the candidate edge exists according to the relationship score.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows: based on the graph convolutional network, the data of nodes and relationships of the sparse network are aggregated to obtain an embedding representation based on sparsity, as well as a feature representation based on the sparsity of the relationship edge and the meta-path; and the candidate edge relationship scores between candidate nodes are obtained through the feature representation to detect whether there is an implicit relationship between the candidate nodes; the complex multi-dimensional relationships between nodes are captured through recursive propagation, the representation ability of node features is enhanced, and the accuracy of implicit relationship mining is effectively improved, which is particularly suitable for sparse network environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a flow chart of the implicit relationship detection method for sparse networks of the present invention;

[0041] Figure 2 is a flow chart of the prediction method of the example;

[0042] Figure 3 It is a logic block diagram of the detection system of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] The present invention is further described in detail below in conjunction with the accompanying drawings:

[0045] Overview: In a sparse network, a node refers to a static object or entity, such as a user in a social network; an edge, also known as an edge relationship, refers to the relationship between two nodes.

[0046] Existing technologies have shortcomings in mining implicit relationships in attribute sparse networks, such as strong explicit dependency, high dependency on node features, and easy overfitting of sparse data. Specifically, first, traditional methods rely too much on adjacency matrices and explicit relationship modeling, which makes traditional methods less applicable when node or edge attributes are missing; second, existing technologies are highly dependent on explicit node features, which makes it impossible to effectively model implicit relationships in sparse data scenarios; in addition, deep learning methods are prone to overfitting problems in sparse data, which affects the generalization ability of the model. Overall, existing technologies lack an effective mechanism to adapt to attribute sparse networks, resulting in poor performance in information aggregation and dissemination.

[0047] The present invention aims at the problem of mining implicit relationships of entities in attribute sparse networks, and proposes a detection method and detection system for implicit relationships that are specifically adapted to high-sparseness scenarios. By introducing a sparsity adaptive aggregation mechanism, the dependence of nodes on attributes and neighbor features is dynamically adjusted to improve the efficiency of sparse data utilization; at the same time, a high-order relationship propagation strategy is designed, and the most important propagation path is dynamically selected in combination with the meta-path priority to enhance the ability to capture implicit relationships. By solving the problem of insufficient information caused by attribute sparsity, this method can mine implicit relationships more efficiently and accurately in scenarios such as dark webs, anonymous social networks, and low-labeled social networks, thereby breaking through the limitations of existing technologies in sparse networks.

[0048] The first aspect of the present invention provides a method for detecting implicit relationships in a sparse network. Figure 1 , including the following steps:

[0049] Step 101: Collect data on nodes and relationships of a sparse network to obtain a data set.

[0050] Step 102: Based on a graph convolutional network (GCN) and the sparsity of relationship edges, an embedded representation of the nodes in the data set is obtained, wherein the embedded representation takes into account the sparsity of edges and their weights.

[0051] Graph Convolutional Network (GCN) is a deep learning model specifically designed for processing graph structured data. It can effectively extract feature information of nodes in the graph and can use the structural information of the graph (such as the connection relationship between nodes) for learning. The nodes and their connection relationships can be well matched with the nodes and their relationships in the sparse network. Embedding Representation is a representation that maps data (such as categorical data) to a high-dimensional, continuous vector space. Embedding Representation is a representation of the features of a sparse network in a vector space.

[0052] Step 103: Obtain feature representation of the node according to the embedded representation, meta-path and its priority.

[0053] A meta-path is a sequence of nodes and edges that alternate with each other, usually consisting of multiple nodes and related edges. There are many different meta-paths that can be formed between two nodes. For example, there is a chat path and a transaction path between the two nodes of user A and user B. Both meta-paths can reflect the characteristics between the two nodes. The priority of the meta-path reflects the weight of the two meta-paths in the implicit relationship.

[0054] Step 104: Obtain candidate node pair set ε candidate , and a candidate edge set of the candidate node pair set, wherein the candidate edges of the candidate edge set are represented as implicit relationships between the candidate node pairs.

[0055] The candidate node pair set has multiple candidate node pairs, and the relationship edge between two nodes of the candidate node pair constitutes a candidate edge. The candidate edge is represented as an implicit relationship between the candidate node pairs.

[0056] Step 105: Calculate the relationship score P(μ,v) of the candidate edge through the score function and the feature representation of each node in the candidate node pair.

[0057] The Sigmoid score function can be used to calculate the relationship score of the candidate edge between node μ and node v. The relationship score is also called the relationship probability.

[0058] Step 106: Based on the relationship score, predict whether the candidate edge exists, that is, whether there is an implicit relationship.

[0059] More specifically, if the relationship score is greater than or equal to a first threshold τ (such as 0.5-0.8), there is a candidate edge, that is, there is an implicit relationship; if the relationship score is less than the first threshold τ, there is no candidate edge, that is, there is no implicit relationship.

[0060] Based on graph convolutional networks, the data of nodes and relationships in sparse networks are aggregated to obtain an embedding representation based on sparsity, as well as a feature representation based on relationship edge sparsity and meta-paths. The scores of candidate edge relationships between candidate nodes are obtained through feature representation to detect whether there are implicit relationships between candidate nodes. Recursive propagation is used to capture complex multi-dimensional relationships between nodes, which enhances the representation ability of node features and effectively improves the accuracy of implicit relationship mining, especially for sparse network environments.

[0061] In step 102, a sparsity adaptive aggregation mechanism can be used to dynamically adjust the information aggregation strategy according to the sparsity of nodes and edges, and generate a high-dimensional representation of the target node in combination with link characteristics (such as node degree, neighbor similarity, etc.). The node sparsity index is defined as: S(u) = 1-number of attributes / total number of attributes, with a value range of [0,1]. For edge relationships, the edge sparsity index is defined as: S(e) = 1-actual number of edges / possible number of edges.

[0062] The formula for embedding representation of a node is: (1)

[0063]

[0064] in, is the embedding representation of node u at layer k; is the embedding representation of node u at the k-1 layer; is the embedding representation of the neighbor node v at the k-1 layer, that is, the feature vector of the previous layer; N(u) is the set of neighbor nodes of node u; ω uv is the edge weight from neighbor node v to node u, indicating the influence strength of neighbor node v on node u; α u The sparsity weight of node u, ranging from [0,1], is used to balance the importance of neighbor features and node attributes: α u The closer it is to 1, the more dependent it is on neighbor features (applicable to nodes with high sparsity). u The closer to 0, the more dependent it is on its own attributes (applicable to low sparsity nodes); (1-α u ) represents the importance weight of the node attribute, and α u Complementary.

[0065] The formula is divided into two parts. The first part represents the feature aggregation from neighboring nodes, and the second part represents the attribute retention of the node itself. The two are combined through the sparsity weight α uDynamic balance. The formula aims to solve the problem of attribute sparsity, and achieves efficient information fusion by dynamically adjusting the degree of node dependence on neighbor features and its own attributes. For high-sparse nodes, more reliance is placed on their structural features (embedded information of neighbor nodes); for low-sparse nodes, more use is made of their own attribute information, thereby optimizing the embedded representation of the node. The feature information (embedded representation) of node u at the kth layer is dynamically calculated and updated according to this formula.

[0066] When the sparsity weight of the relationship edge is introduced, the feature representation update formula of node u in the kth layer is expressed as:

[0067]

[0068] in, is the embedding representation of node u at layer k, is the embedding representation of the neighbor node v at the k-1 layer, N(u) is the set of neighbor nodes of node u; β(e uv ) is the relationship edge e uv The sparsity weight is used to measure the importance of the edge. The edge with higher sparsity has lower weight, and the edge with lower sparsity has higher weight. uv Represented as the relationship edge from neighbor node v to node u; ω uv is the edge weight from neighbor node v to node u, indicating the influence strength of neighbor node v on node u.

[0069] The formula shows that the feature of node u is the feature of all its neighboring nodes. Aggregated, and each edge is classified according to its sparsity weight β(e uv ) and edge weight ω uv Adjusted the contribution size, sparsity weight β(e uv ) and edge weight ω uv It reflects the sparsity of the relationship edges. It aims to reduce the negative impact of low-quality or sparse edges on information transmission by adjusting the edge weights. By introducing edge sparsity weights, the connections between nodes are weighted, making information aggregation more accurate and reliable.

[0070] In step 103, the method for obtaining the node feature representation includes the following steps:

[0071] Step 301: Obtain the importance score of the meta-path m.

[0072] The importance score of meta-path m is expressed as:

[0073]

[0074] Wherein, I(m) represents the importance score of meta-path m, G(u,v) represents the information gain of node u and node v, and L(u,v) represents the path length of node u and node v. Node u and node v belong to meta-path m. The calculation method of information gain and path length is prior art and will not be described in detail in the present invention.

[0075] Step 302: Select a meta-path according to the importance score.

[0076] For sparse networks with multiple meta-paths, shorter meta-paths with high information gain are preferred, that is, maximizing the importance score, or selecting meta-paths with large importance scores.

[0077] Step 303: Dynamically adjust the meta-path weight.

[0078] ω m =softmax(I(m))

[0079] ω m It is represented as meta-path weight and softmax() is represented as normalized exponential function.

[0080] Step 304: Aggregate the feature representations of the nodes through the meta-path.

[0081]

[0082] Among them, h u It is represented as the feature representation of node u, h v Represented as the feature representation of node v, N m (u) is the set of neighbor nodes of node u on meta-path m, ω uv is the edge weight from neighbor node v to node u, ω m is the meta-path weight.

[0083] In step 105, the relationship score is expressed as:

[0084]

[0085] Where P(μ,v) is the relationship score of the candidate edge / relationship edge between node u and node v, h u It is represented as the feature representation of node u, h v It is represented as the feature representation of node v, W is a learnable weight matrix used to model the interaction relationship between nodes; b is a learnable bias term used to adjust the output of the score function; is the Sigmoid activation function, which is used to limit the relationship score to the range of [0,1], indicating the probability of the edge. T is represented by the transposed rows and columns.

[0086] The detection method of the present invention is highly scalable, can cope with situations where node attributes are incomplete or missing, and fully utilizes link information and network structure to model implicit relationships. It fully considers the characteristics of complex environments such as the dark web, supports large-scale data processing and sparse network relationship prediction, and can be widely used in the fields of illegal transaction network tracking, crime chain identification, and intelligence analysis. It can efficiently and accurately mine hidden implicit relationships between entities in high sparsity and high anonymity scenarios, providing strong technical support for related fields.

[0087] Taking the social network of the dark web as an example, there are relationships between users such as transactions, chats, and forum messages. It is necessary to predict whether two anonymous users may have a potential transaction relationship, such as Figure 2 , prediction methods include:

[0088] Step 401: Collect data and obtain a data set.

[0089] Collect the user nodes of the network and the known relationship data between them (for example, chat frequency, transaction number, message frequency). Obtain features such as node sparsity, edge sparsity, and sparsity index.

[0090] Step 402: Aggregate heterogeneous data.

[0091] Different types of feature information (such as the sparsity of transaction paths and chat frequency) are fused to generate a comprehensive feature description. Specifically, the embedded representation of the node is obtained based on the graph convolutional network and the sparsity of the relationship edges.

[0092] Step 403: Model the nodes and relationship edges of the data through a graph convolutional network and a preset meta-path priority to obtain a first prediction model.

[0093] In one specific test, the priority / importance score of the transaction path was higher than that of the chat and forum message, respectively.

[0094] Step 404: According to the embedding representation of the node, obtain the sparsity weight of the relationship edge and the edge weight from the neighbor node to the node.

[0095] Step 405: Update the parameters of the first prediction model according to the sparsity weight and the edge weight to obtain a second prediction model.

[0096] Step 406: Use the second prediction model to predict the data set to obtain a relationship score. The second prediction model is used to output a relationship score.

[0097] Step 407: judging whether the candidate node pair has an implicit relationship according to the relationship score and the first threshold; and outputting the result.

[0098] The present invention designs an adaptive aggregation mechanism based on the sparsity of nodes and relationship edges, and dynamically adjusts the degree of dependence of nodes on neighbor features and their own attributes. By reasonably allocating the weights of different information sources, the impact of attribute loss is effectively compensated, and the effect of information aggregation is optimized. A dynamic meta-path priority selection mechanism based on network structure and node features is designed, which can automatically adjust the importance of meta-paths according to the actual situation of the network. This mechanism improves the ability to capture implicit relationships and optimizes the efficiency of information propagation by reducing the propagation of redundant paths. A multi-level relationship modeling method based on graph convolutional networks (GCNs) is designed to capture complex multidimensional relationships between nodes and neighbors through recursive propagation. This method enhances the representation ability of node features, effectively improves the accuracy of implicit relationship mining, and is particularly suitable for network environments with sparse attributes.

[0099] The existing technology usually relies on the explicit attributes of nodes and complete adjacency relationships. The present invention uses a sparsity adaptive aggregation mechanism to dynamically adjust the information aggregation strategy according to the sparsity of nodes and relationship edges. It can effectively make up for the lack of information when attributes are missing, thereby improving the modeling ability in sparse networks.

[0100] Compared with the traditional method of fixedly selecting meta-paths, the dynamic meta-path priority mechanism of the present invention automatically adjusts the importance of meta-paths according to the network structure and node characteristics, reduces redundant propagation and optimizes information flow, making the capture of implicit relationships more accurate and effectively improving the adaptability to complex heterogeneous networks.

[0101] The multi-level relationship modeling method based on GCN designed by the present invention can deeply mine the multi-level interactive relationship between nodes through the combination of recursive propagation and graph convolutional network. When processing implicit relationships in complex networks, it can effectively improve the accuracy of implicit relationship mining, especially in highly sparse environments such as dark networks and low-labeled complex networks.

[0102] The method of the present invention is highly scalable and can cope with various network structures with incomplete or missing node attributes, and make up for the deficiencies through link information and the relationship between nodes. Compared with the existing methods that strongly rely on node attributes, this method shows stronger robustness in processing data sparse or anonymized environments such as dark networks and low-labeled complex networks.

[0103] The second aspect of the present invention also provides a detection system for implementing the above detection method, such as Figure 3 As shown, it includes a collection module 1, an aggregation module 2, a feature extraction module 3, a relationship score calculation module 4 and a prediction module 5.

[0104] The acquisition module 1 is used to collect data on nodes and relationships of a sparse network to obtain a data set;

[0105] The aggregation module 2 is used to obtain an embedded representation of the nodes in the data set based on a graph convolutional network (GCN) and sparsity of relationship edges;

[0106] The feature extraction module 3 is used to obtain the feature representation of the node according to the embedded representation, the meta-path and its priority;

[0107] The relationship score calculation module 4 is used to obtain a set of candidate node pairs and a set of candidate edges of the candidate node pairs, and calculate the relationship scores of the candidate edges through the score function and the feature representation of each node in the node pair;

[0108] The prediction module 5 is used to predict whether the candidate edge / implicit relationship exists according to the relationship score.

[0109] The relationship score calculation module may obtain the relationship score through the second prediction model.

[0110] This paper aims to solve the problem of implicit relationship modeling and mining in highly anonymized and data-sparse environments, and is particularly suitable for scenarios such as dark webs and anonymous social networks. It combines the advantages of graph convolutional networks (GCNs) and heterogeneous information networks, innovatively introduces sparsity adaptive aggregation mechanisms and dynamic meta-path priority mechanisms, thereby overcoming the limitations of traditional methods in attribute sparse networks.

[0111] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting implicit relations in sparse networks, characterized in that: The following steps are involved: Collect data on nodes and relationships of sparse networks to obtain a data set; Based on graph convolutional networks and sparsity of relationship edges, we obtain embedded representations of nodes in the dataset. Obtain feature representation of the node based on the embedded representation, meta-path and its priority; Obtaining candidate edges of the candidate node pairs, wherein the candidate edges are represented as implicit relationships of the candidate node pairs; Calculate the relationship score of the candidate edge through the score function and the feature representation of the node in the candidate node pair; According to the relationship score, predict whether the candidate edge exists.

2. The implicit relationship detection method according to claim 1, characterized in that: Methods for determining whether a candidate edge exists include: If the relationship score is greater than or equal to the first threshold, the candidate edge exists; If the relationship score is less than the first threshold, the candidate edge does not exist.

3. The implicit relationship detection method according to claim 1, characterized in that: The relationship score is calculated as: Among them, P(μ,v) represents the relationship score of the candidate edge between node u and node v, h u It is represented as the feature representation of node u, h v It is represented as the feature representation of node v, W is a learnable weight matrix used to model the interaction relationship between nodes; b is a learnable bias term used to adjust the output of the scoring function; σ() is represented as the Sigmoid activation function.

4. The implicit relationship detection method according to claim 1, characterized in that: The formula for the node embedding representation is: is the embedding representation of node u at layer k; is the embedding representation of node u at the k-1 layer; is the embedding representation of the neighbor node v at the k-1 layer, N(u) is the set of neighbor nodes of node u, ω uv is the edge weight from neighbor node v to node u, α u The sparsity weight of node u.

5. The implicit relationship detection method according to claim 1, characterized in that: The formula for the node embedding representation is: in, is the embedding representation of node u at layer k, is the embedding representation of the neighbor node v at the k-1 layer, N(u) is the set of neighbor nodes of node u; β(e uv ) is the relationship edge e uv The sparsity weight, ω uv is the edge weight from neighbor node v to node u.

6. The implicit relationship detection method according to claim 1, characterized in that: The formula for node feature representation is: Among them, h u It is represented as the feature representation of node u, h v Represented as the feature representation of node v, N m (u) is the set of neighbor nodes of node u on meta-path m, ω uv is the edge weight from neighbor node v to node u, ω m is the meta-path weight.

7. The implicit relationship detection method according to claim 6, characterized in that: The calculation method of meta path weight is: ω m =softmax(I(m)) I(m) is represented as the importance score of meta-path m, G(u,v) is represented as the information gain of node u and node v, L(u,v) is represented as the path length of node u and node v, node u and node v belong to meta-path m, and softmax() is represented as the normalized exponential function.

8. The implicit relationship detection method according to claim 1, characterized in that: The parameters of relationship edge sparsity include the edge weight between the node and the neighboring node and the sparsity weight of the relationship edge.

9. The implicit relationship detection method according to claim 1, characterized in that: Also included are methods based on obtaining relationship scores through predictive models: The nodes and relationship edges of the data are modeled through a graph convolutional network and a preset meta-path priority to obtain a first prediction model; According to the embedding representation of the node, the sparsity weight of the relationship edge and the edge weight from the neighbor node to the node are obtained; According to the sparsity weight and the edge weight, updating the parameters of the first prediction model to obtain a second prediction model; The data set is predicted using the second prediction model to obtain a relationship score.

10. A detection system, characterized in that: Used to implement the implicit relationship detection method according to any one of claims 1 to 9, characterized in that it includes a collection module, an aggregation module, a feature extraction module, a relationship score calculation module and a prediction module, The acquisition module is used to collect data on nodes and relationships of the sparse network to obtain a data set; The aggregation module is used to obtain an embedded representation of nodes in a data set based on a graph convolutional network and sparsity of relationship edges; The feature extraction module is used to obtain a feature representation of a node according to the embedded representation, the meta-path and its priority; The relationship score calculation module is used to obtain a set of candidate node pairs and a set of candidate edges of the candidate node pairs, and calculate the relationship scores of the candidate edges through the score function and the feature representation of the nodes in the candidate node pairs; The prediction module is used to predict whether the candidate edge exists according to the relationship score.

Images