Inter-edge multi-mode fused extensible network node vector representation system

Abstract

The invention relates to the field of vector representation, and particularly discloses an extensible network node vector representation system fusing inter-edge multiple modes, which is used for solving the problem of how to efficiently generate meaningful vector representation for new nodes. According to the method, the scoring function based on the node measurement indexes is constructed to perform basic node selection as a support, and new nodes can be quickly integrated into existing vector representation under the condition that the whole model is not retrained, so that the flexibility and the practicability of the system are improved; the system is introduced into a triple in a modeling network in a complex field, and modes (symmetry, combination and the like) between edges can be fused into node vectors, so that the node characterization accuracy is enhanced.

Description

Technical Field

[0001] The present invention relates to the field of vector representation, and more specifically, to an extensible network node vector representation system integrating multiple modes between edges. Background Art

[0002] With the development of information technology, network data has become an important resource in research and application. Networks can be in various forms, including but not limited to social networks, bioinformatics networks, the Internet, corporate networks, etc. In network analysis, nodes (representing individuals or entities) and the connections between them (i.e., edges, representing relationships or interactions) constitute the basic structure of the network. In order to effectively analyze and utilize these network data, it is necessary to convert the nodes into numerical form in a meaningful way, which is the so-called node vectorization representation. Traditional node vectorization methods usually assume that the network structure is relatively stable, but in many practical scenarios, the network is dynamically changing, especially highly active platforms such as social networks, where new users (i.e., new nodes) are added very frequently. This dynamic characteristic poses a challenge to existing node vectorization technologies. Therefore, how to efficiently generate meaningful vector representations for new nodes has become an urgent problem to be solved. In order to solve the above problems, a technical solution is now provided. Summary of the invention

[0003] In order to overcome the above-mentioned defects of the prior art, the present invention provides a scalable network node vector representation system that integrates multiple modes between edges, which is used to solve the problem of how to efficiently generate meaningful vector representations for new nodes, so as to solve the problems raised in the above-mentioned background technology.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A scalable network node vector representation system integrating multi-modes between edges includes a basic mapping module and a triple task training module; the basic mapping module is used to uniformly map each point on the network to a series of basic nodes for contextual representation; the triple task training module is used to learn the topological structure information on the network in the form of triple tasks, and the specific steps are as follows:

[0006] Receive the node vector output by the basic mapping module, and convert the context encoding vectors of the head node and the tail node into the complex embedding representation of the head node and the complex embedding representation of the tail node respectively through the node real part mapping and the node imaginary part mapping. For the association relationship, obtain the corresponding vector from the relationship real part embedding and the relationship imaginary part embedding to form the relationship complex embedding representation;

[0007] The complex embedding representation of the head node is denoted as The tail node complex embedding representation is recorded as The relational plural embedding representation is denoted as In the current complex space, the distance formula between the complex embedding representation of the head node and the complex embedding representation of the tail node is:

[0008]

[0009] Where: is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node;

[0010] Initialize two M-row and p-column parameter matrices to be trained, which are divided into real and imaginary parts representing the relationship on the network in the complex space. The upper units of the neural network include the head node, the tail node, and the association relationship. The complex definition is performed to obtain the corresponding vectors for construction;

[0011] Based on the above, the training loss formula for the triplet task is determined as:

[0012]

[0013] Where: L is the triple task training loss, σ is the sigmiod function, γ is the distance adjustment parameter, the default value is 6, is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node.

[0014] As a further solution of the present invention, the basic mapping module includes a basic node selector, a node context builder and a node context encoder; the basic node selector is used to select Q basic nodes in the currently constructed neural network to form a basic node array, each basic node obtains an index number in the array, and a d-dimensional vector is randomly initialized for all basic nodes. The d-dimensional vectors of all basic nodes are arranged according to the index number to form a matrix to be trained. The basic node calculates the score through the node expression, and the first Q results are selected in descending order. The node expression is:

[0015] score=α*PPR+β*degree+γ*betweenness;

[0016] 1 = α + β + γ;

[0017] Where score is the basic node score, α is the weighting coefficient of the personalized PageRank indicator of the current neural network, PPR is the personalized PageRank indicator of the current neural network, β is the weighting coefficient of the degree centrality indicator, degree is the degree centrality indicator, γ is the weighting coefficient of the betweenness centrality indicator, and betweenness is the betweenness centrality indicator.

[0018] As a further solution of the present invention, the node context builder is used to perform matrix addition of the context feature matrix and the local basic path distance context moment, and then perform a connection operation on the rows with the local relationship context matrix to obtain the current context matrix.

[0019] As a further solution of the present invention, the specific steps of obtaining the current context matrix are:

[0020] The neighboring basic nodes are selected based on the path distance between the current node and all basic nodes. The path distance from the node to the basic node is sorted in descending order. The first k basic nodes and path distances are selected to form the basic neighbor array and basic path distance array of the node respectively. According to the basic neighbor array, the index number of each basic node is obtained, and the corresponding d-dimensional vector is queried on the basic vector matrix to form the context feature matrix under the current node;

[0021] Use the data in the basic path distance array as an index to query the basic path distance matrix to obtain the local basic path distance context matrix of the current node;

[0022] The relationship set on the in-degree or out-degree of the current node is the current relationship set. The relationship types in the current relationship set are matched with the relationship matrix to obtain the local relationship context matrix of the current node.

[0023] The context feature matrix is ​​matrix-added to the local basis path distance context moment, and then concatenated with the local relationship context matrix on the rows to obtain the current context matrix.

[0024] As a further solution of the present invention, the node context encoder is used to encode the current context matrix of the node using the Tansfromer model, and perform mean pooling on the resulting sequence to output a d-dimensional current context vector of the encoding result; in the basic mapping module, several nodes are input, and several node context encoding vectors are output.

[0025] The technical effects and advantages of the scalable network node vector representation system that integrates multiple modes between edges of the present invention are as follows: the present invention constructs a scoring function based on node metric indicators to perform basic node selection as a support, and can quickly integrate new nodes into the existing vector representation without retraining the entire model, thereby improving the flexibility and practicality of the system. In addition, by introducing triples in the modeling network in the complex domain, the patterns (symmetry, combination, etc.) between edges can be integrated into the node vector, thereby enhancing the accuracy of node representation; a scalable network node vectorization representation method is combined, which is particularly suitable for a growing and changing data environment. When a new node is added to the network, there is no need to retrain the entire neural network model from scratch. The new node can be decomposed and mapped to the existing node, and the new node can be represented by a combination of these determined old nodes; by combining the training method of performing triple tasks in the complex space, the possible symmetry, inversion, combination and other pattern relationships between edges are integrated into the information of the node representation, which helps to enhance the performance of downstream tasks such as recommendation systems and search algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of the structure of a scalable network node vector representation system integrating multiple modes between edges provided by the present invention;

[0027] Figure 2 A flow chart of the operating steps of the triple task training module provided by the present invention. DETAILED DESCRIPTION

[0028] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solution in the present invention. Obviously, the described technical solution is only a part of the present invention, not all of it. Based on the technical solution in the present invention, all other technical solutions obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0029] Figure 1 The present invention provides a structural diagram of a scalable network node vector representation system that integrates multiple edge modes. Figure 1 As shown, a scalable network node vector representation system integrating multi-modes between edges includes a basic mapping module and a triple task training module; the basic mapping module is connected to the triple task training module;

[0030] The basic mapping module is used to uniformly map each point on the network to a series of basic nodes for contextual representation;

[0031] The triple task training module is used to learn the topological structure information on the network in the form of triple tasks.

[0032] Specifically, the basic mapping module includes a basic node selector basic_node_selector, a node context builder node_context_builder and a node context encoder node_context_encoder; the basic node selector basic_node_selector is connected to the node context builder node_context_builder, and the node context builder node_context_builder is connected to the node context encoder node_context_encoder;

[0033] The basic node selector is used to select Q basic nodes basic_node in the currently constructed neural network to form a basic node array basic_node_array. Each basic node basic_node obtains the index number basic_node_idx in the array, and randomly initializes a d-dimensional vector basic_embedding for all basic nodes basic_node. All d-dimensional vectors are arranged according to the index number basic_node_idx to form a matrix to be trained basic_embedding_matrix. The basic node is scored by the node expression, and the first Q results are selected in descending order. The node expression is:

[0034] score=α*PPR+β*degree+γ*betweenness;

[0035] 1 = α + β + γ;

[0036] Where score is the basic node score, α is the weighting coefficient of the personalized PageRank indicator of the current neural network, PPR is the personalized PageRank indicator of the current neural network, β is the weighting coefficient of the degree centrality indicator, degree is the degree centrality indicator, γ is the weighting coefficient of the betweenness centrality indicator, and betweenness is the betweenness centrality indicator.

[0037] The node context builder is used to perform matrix addition on the context feature matrix local_basic_node_ctx_matrix and the local basic path distance context matrix local_basic_path_dist_ctx_matrix, and then perform a concat operation on the rows with the local relationship context matrix local_rel_ctx_matrix to obtain the current context matrix local_ctx_matrix.

[0038] The node_context context of a node consists of three components: k neighboring base nodes, k path distances to the neighboring base nodes, and m relationship types on the node.

[0039] If the network diameter is kg_diameter, randomly initialize a matrix to be trained, basis_path_dist_matrix, with kg_diameter rows and d columns.

[0040] If the number of relationship types in the network is M, a matrix to be trained rel_matrix with M rows and d columns is randomly initialized, and the row index of the matrix corresponds to a relationship type.

[0041] The neighboring basic nodes are selected based on the path distance path_distance between the current node node and all basic nodes basic_node. The path distance path_distance from node node to basic node basic_node is sorted in descending order, and the first k basic nodes basic_node and path distance path_distance are selected to form the basic neighbor array basis_neighborhood_array and the basic path distance array basis_path_distance_array of node node respectively. According to the basic neighbor array basis_neighborhood_array, the index number basic_node_idx of each basic node basic_node is obtained, and the corresponding d-dimensional vector basic_embedding is queried on the basic vector matrix basic_embedding_matrix to form the context feature matrix local_basic_node_ctx_matrix under the current node.

[0042] Use the data in the basic path distance array basis_path_distance_array as an index to query the basic path distance matrix basis_path_dist_matrix to obtain the local basic path distance context matrix local_basic_path_dist_ctx_matrix of the current node node.

[0043] The relationship set on the in-degree or out-degree of the current node node is the current relationship set local_rel_set. The relationship types in the current relationship set are matched with the relationship matrix rel_matrix to obtain the local relationship context matrix local_rel_ctx_matrix of the current node node.

[0044] Finally, the context feature matrix local_basic_node_ctx_matrix is ​​added to the local basic path distance context matrix local_basic_path_dist_ctx_matrix, and then concatenated with the local relationship context matrix local_rel_ctx_matrix on the rows to obtain the current context matrix local_ctx_matrix.

[0045] The node context encoder is used to encode the current context matrix local_ctx_matrix of the node node using the Tansfromer model, perform mean pooling on the resulting sequence, and output the encoded result, a d-dimensional current context vector local_ctx_embedding.

[0046] In the basic mapping module basis_model, several nodes are input, and several node context encoding vectors local_ctx_embedding are output.

[0047] Specifically, the triple task training module is used to learn the topological structure information on the network in the form of triple tasks. The specific steps are:

[0048] Initialize two M rows and p columns of parameter matrices to be trained, which are divided into the real part relation_real_embedding and the imaginary part relation_im_embedding representing the relationship on the network in the complex space. Each row vector corresponds to a relationship type.

[0049] The existing network edge units include head node head_node, tail node tail_node, and relationship _rel_edge. According to the following complex definition, the corresponding vectors are obtained to form the complex definition, which is expressed as:

[0050]

[0051] Where: is a complex vector, is the real part of the complex number, i is the imaginary unit, To map a node's context embedding to its imaginary representation, is the imaginary part of the complex number;

[0052] Through the node real part mapping map_node_real and the node imaginary part mapping map_node_im, the context encoding vectors local_ctx_embedding of the head node head_node and the tail node tail_node are converted into the complex embedding representation head_comp_embedding of the head node and the complex embedding representation tail_comp_embedding of the tail node respectively. For the association relationship rel_edge, the corresponding vectors are obtained from the real embedding relation_real_embedding and the imaginary embedding relation_im_embedding to form the complex embedding representation rel_comp_embedding of the relationship.

[0053] The head node complex embedding representation head_comp_embedding is recorded as The tail node complex embedding representation tail_comp_embedding is recorded as The relational plural embedding representation rel_comp_embedding is recorded as In the current complex space, the distance formula between the complex embedding representation of the head node and the complex embedding representation of the tail node is:

[0054]

[0055] Where: is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node;

[0056] Based on the above, the training loss formula for the triplet task is determined as:

[0057]

[0058] Where: L is the triple task training loss, σ is the sigmiod function, γ is the distance adjustment parameter, the default value is 6, is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node.

[0059] The embodiment of the present invention constructs a scoring function based on node metric indicators to perform basic node selection as a support, and can quickly integrate new nodes into the existing vector representation without retraining the entire model, thereby improving the flexibility and practicality of the system. In addition, by introducing triples in the modeling network in the complex domain, the patterns (symmetry, combination, etc.) between edges can be integrated into the node vector to enhance the accuracy of node representation; a scalable network node vectorization representation method is combined, which is particularly suitable for a growing and changing data environment. When a new node is added to the network, there is no need to retrain the entire neural network model from scratch. The new node can be decomposed and mapped to the existing node, and the new node can be represented by a combination of these determined old nodes; by combining the training method of performing triple tasks in the complex space, the possible symmetry, inversion, combination and other pattern relationships between the edges are integrated into the node representation information, which helps to enhance the performance of downstream tasks such as recommendation systems and search algorithms.

[0060] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0061] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A scalable network node vector representation system integrating multi-modes between edges, including a basic mapping module and a triple task training module; characterized in that: The basic mapping module is used to uniformly map each point on the network to a series of basic nodes for context representation; the triple task training module is used to learn the topological structure information on the network in the form of triple tasks. The specific steps are as follows: Receive the node vector output by the basic mapping module, and convert the context encoding vectors of the head node and the tail node into the complex embedding representation of the head node and the complex embedding representation of the tail node respectively through the node real part mapping and the node imaginary part mapping. For the association relationship, obtain the corresponding vector from the relationship real part embedding and the relationship imaginary part embedding to form the relationship complex embedding representation; The complex embedding representation of the head node is denoted as The tail node complex embedding representation is recorded as The relational plural embedding representation is denoted as In the current complex space, the distance formula between the complex embedding representation of the head node and the complex embedding representation of the tail node is: Where: is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node; Initialize two M-row and p-column parameter matrices to be trained, which are divided into real and imaginary parts representing the relationship on the network in the complex space. The upper units of the neural network include the head node, the tail node, and the association relationship. The complex definition is performed to obtain the corresponding vectors for construction; Based on the above, the training loss formula for the triplet task is determined as: Where: L is the triple task training loss, σ is the sigmiod function, γ is the distance adjustment parameter, the default value is 6, is the distance between the complex embedding representation of the head node and the complex embedding representation of the tail node.

2. The scalable network node vector representation system integrating multi-modes between edges according to claim 1 is characterized in that: The basic mapping module includes a basic node selector, a node context builder, and a node context encoder; the basic node selector is used to select Q basic nodes in the currently constructed neural network to form a basic node array, each basic node obtains an index number in the array, and a d-dimensional vector is randomly initialized for all basic nodes. The d-dimensional vectors of all basic nodes are arranged according to the index number to form a matrix to be trained. The basic node calculates the score through the node expression, and the first Q results are selected in descending order. The node expression is: score=α*PPR+β*degree+γ*betweenness; 1 = α + β + γ; Where score is the basic node score, α is the weighting coefficient of the personalized PageRank indicator of the current neural network, PPR is the personalized PageRank indicator of the current neural network, β is the weighting coefficient of the degree centrality indicator, degree is the degree centrality indicator, γ is the weighting coefficient of the betweenness centrality indicator, and betweenness is the betweenness centrality indicator.

3. The scalable network node vector representation system integrating multi-modes between edges according to claim 2 is characterized in that: The node context builder is used to perform matrix addition of the context feature matrix and the local base path distance context moment, and then perform a row concatenation operation with the local relationship context matrix to obtain the current context matrix.

4. The scalable network node vector representation system integrating multi-modes between edges according to claim 3 is characterized in that: The specific steps for obtaining the current context matrix are: The neighboring basic nodes are selected based on the path distance between the current node and all basic nodes. The path distance from the node to the basic node is sorted in descending order. The first k basic nodes and path distances are selected to form the basic neighbor array and basic path distance array of the node respectively. According to the basic neighbor array, the index number of each basic node is obtained, and the corresponding d-dimensional vector is queried on the basic vector matrix to form the context feature matrix under the current node; Use the data in the basic path distance array as an index to query the basic path distance matrix to obtain the local basic path distance context matrix of the current node; The relationship set on the in-degree or out-degree of the current node is the current relationship set. The relationship types in the current relationship set are matched with the relationship matrix to obtain the local relationship context matrix of the current node. The context feature matrix is ​​matrix-added to the local basis path distance context moment, and then concatenated with the local relationship context matrix on the rows to obtain the current context matrix.

5. The scalable network node vector representation system integrating multi-modes between edges according to claim 2, characterized in that: The node context encoder is used to encode the current context matrix of the node using the Tansfromer model, and perform mean pooling on the resulting sequence to output the encoded result as a d-dimensional current context vector; In the basic mapping module, several nodes are input and several node context encoding vectors are output.

Images