CN113761286A

CN113761286A - Map embedding method and device of knowledge map and electronic equipment

Info

Publication number: CN113761286A
Application number: CN202010486192.7A
Authority: CN
Inventors: 付尧
Original assignee: Hangzhou Hikvision Digital Technology Co Ltd
Current assignee: Hangzhou Hikvision Digital Technology Co Ltd
Priority date: 2020-06-01
Filing date: 2020-06-01
Publication date: 2021-12-07
Anticipated expiration: 2040-06-01
Also published as: CN113761286B

Abstract

The embodiment of the invention provides a map embedding method and device of a knowledge map and electronic equipment. Wherein the method comprises the following steps: acquiring topological information of a first class node and a second class node, wherein the topological information is the characteristics of objects represented by the first class node, and interaction exists between the objects represented by the first class node and the second class node; respectively aiming at each type of node in the first type of node and the second type of node, carrying out graph feature extraction on the topology information of the type of node according to the topology information of the type of node to obtain deep features of the type of node; and according to the interaction information between the first class of nodes and the second class of nodes, carrying out graph feature extraction on the deep features of the first class of nodes and the deep features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes. Reduced information lost in graph embedding for heterogeneous knowledge graphs may be achieved.

Description

Map embedding method and device of knowledge map and electronic equipment

Technical Field

The invention relates to the technical field of graph data mining, in particular to a graph embedding method and device of a knowledge graph and electronic equipment.

Background

In the technical field of Graph data mining, a Knowledge Graph (knowledgegraph) is often used to describe the association relationship between different objects, nodes in the Knowledge Graph can be used to represent objects or features, and edges can be used to represent the existence of relationships between objects or features represented by connected nodes. If the knowledge graph contains object nodes for representing different types, the knowledge graph is called a heterogeneous knowledge graph, for example, one part of the nodes in the knowledge graph is used for representing users, and the other part of the nodes is used for representing articles.

The knowledge Graph is high in dimensionality and not beneficial to processing by electronic equipment, so that the knowledge Graph can be mapped to a representation with a lower dimensionality, and the process is called Graph embedding (Graph embedding) hereinafter. In the related art, the graph embedding can be completed by a random walk manner.

However, the heterogeneous knowledge graph contains nodes for representing various different types of objects and interactive information generated among different types of objects, so that the complexity of the heterogeneous knowledge graph is high, the local features of the knowledge graph can only be determined in a random walk mode, the global features of the knowledge graph are difficult to be effectively embedded into the representation, more information in the knowledge graph is lost in the obtained representation, and subsequent graph data mining is not facilitated.

Disclosure of Invention

The embodiment of the invention aims to provide a method and a device for embedding a knowledge graph and electronic equipment, so as to reduce information lost when graph embedding is carried out on a heterogeneous knowledge graph. The specific technical scheme is as follows:

in a first aspect of the embodiments of the present invention, a method for graph embedding of a knowledge graph is provided, where the knowledge graph includes multiple types of nodes, where each type of node is used to represent a same type of object, the knowledge graph includes interaction edges and association edges, the interaction edges are used to connect nodes that have interactions and different types between represented objects, and the association edges are used to connect nodes that have associations and the same types of represented objects, the method includes:

acquiring topological information of a first class node and a second class node, wherein the topological information is used for representing a topological relation between the first class nodes in the knowledge graph, and an interactive edge exists between the first class nodes and the second class nodes;

respectively aiming at each type of nodes in the first type of nodes and the second type of nodes, carrying out graph feature extraction on topological information of the type of nodes by utilizing a graph feature extraction model obtained based on deep learning to obtain deep features of the type of nodes;

and according to interaction information between the first class nodes and the second class nodes, carrying out graph feature extraction on deep features of the first class nodes and deep features of the second class nodes to obtain object representations of the first class nodes and object representations of the second class nodes, wherein the interaction information is used for representing a topological relation formed by the first class nodes, the second class nodes and interaction edges between the first class nodes and the second class nodes in the knowledge graph.

In a possible embodiment, before the performing graph feature extraction on the topology information of the class node by using the graph feature extraction model obtained based on deep learning to obtain deep features of the class node, the method further includes:

acquiring interactive characteristics between the first class nodes and the second class nodes, wherein the interactive characteristics are used for representing characteristics of interaction generated between objects represented by the first class nodes and objects represented by the second class nodes;

inputting the topology information of the first kind of nodes, the topology information of the second kind of nodes and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction, and obtaining the interacted topology information of the first kind of nodes and the interacted topology information of the second kind of nodes output by the bidirectional long and short term memory network, wherein the sequence of forward propagation in the bidirectional long and short term memory network is one of a first sequence and a second sequence, the sequence of backward propagation in the bidirectional long and short term memory network is the other one of the first sequence and the second sequence, and the first sequence is as follows: topology information and interactive characteristics of the first type of nodes and interactive characteristics of the second type of nodes, wherein the second sequence is as follows: topology information and interactive characteristics of the second type of nodes and interactive characteristics of the first type of nodes;

the method for extracting the graph feature of the topological information of the type of node by using the graph feature extraction model obtained based on deep learning to obtain the deep feature of the type of node comprises the following steps:

and carrying out graph feature extraction on the interacted topological information of the type of node by utilizing a graph feature extraction model obtained based on deep learning to obtain deep features of the type of node.

In a possible embodiment, the inputting the topology information of the first type of node, the topology information of the second type of node, and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction to obtain the interacted topology information of the first type of node and the interacted topology information of the second type of node output by the bidirectional long and short term memory network includes:

inputting the topology information of the first type of nodes, the topology information of the second type of nodes and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction to obtain the interacted topology information of the first type of nodes, the interacted topology information of the second type of nodes and the interacted interaction characteristics output by the bidirectional long and short term memory network;

the extracting deep features of the first type nodes and deep features of the second type nodes according to the interaction information between the first type nodes and the second type nodes to obtain object representations of the first type nodes and object representations of the second type nodes includes:

and according to the interactive features after interaction, carrying out graph feature extraction on the deep features of the first class of nodes and the deep features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

In a possible embodiment, the performing graph feature extraction on the topology information of the class node by using a graph feature extraction model obtained based on deep learning to obtain deep features of the class node includes:

and according to the node connection matrix of the type of node, carrying out graph feature extraction on the topological information of the type of node to obtain deep features of the type of node, wherein the node connection matrix is used for expressing the connection relation between each type of node and other types of nodes in the knowledge graph.

In a possible embodiment, in the interaction generated between the objects represented by the first class of nodes and the second class of nodes, the object represented by the first class of nodes is an interaction initiator;

the node connection matrix of the first type of nodes is constructed in the following way:

constructing a diagonal matrix and an adjacent matrix of the first type of nodes, wherein the diagonal matrix is used for representing the number of similar nodes connected with each node, and the adjacent matrix is used for representing the connection relation among the nodes;

and carrying out weighted average on a product matrix and an identity matrix to obtain a node connection matrix of the first type node, wherein the product matrix is the product of an inverse matrix of a diagonal matrix of the first type node and an adjacent matrix, the weight of the product matrix is in negative correlation with the diagonal matrix of the first type node, and the weight of the identity matrix is in positive correlation with the diagonal matrix of the first type node.

In a possible embodiment, in the interaction generated between the objects represented by the first class nodes and the second class nodes, the objects represented by the second class nodes are interaction acceptors;

the node connection matrix of the second type of nodes is constructed in the following way:

constructing a diagonal matrix and an adjacency matrix of the second type of nodes, wherein the diagonal matrix is used for representing the number of similar nodes connected with each node, and the adjacency matrix is used for representing the connection relation among the nodes;

and calculating the product of the evolution matrix, the adjacent matrix of the second type of nodes and the evolution matrix multiplied in sequence to obtain the node connection matrix of the second nodes, wherein the evolution matrix is obtained by evolution of the inverse matrix of the diagonal matrix of the second type of nodes.

In a possible embodiment, the performing graph feature extraction on the deep features of the first type of nodes and the deep features of the second type of nodes according to the interaction information between the first type of nodes and the second type of nodes to obtain object representations of the first type of nodes and object representations of the second type of nodes includes:

respectively aiming at each type of node in the first type of node and the second type of node, performing feature pooling on deep features of the type of node and the interaction information to obtain pooling features of the type of node;

and according to the topological relation represented by the interaction information, carrying out graph feature extraction on the pooling features of the first class of nodes and the pooling features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

In a possible embodiment, the knowledge-graph is constructed beforehand by:

for each object in a plurality of objects, constructing a knowledge sub-graph of the object, wherein the knowledge sub-graph comprises object nodes and feature nodes, each object node is used for representing one object, each feature node is used for representing one feature of the object, and the knowledge sub-graph further comprises home edges, each home edge is used for connecting one feature node and one object node and is used for representing that the object represented by the connected object node has the feature represented by the connected feature node;

aiming at each two objects with interaction in a plurality of objects, constructing an interaction graph between the two objects, wherein the interaction graph comprises object nodes and interaction edges, and the interaction edges are used for connecting nodes with different types and with interaction between the represented objects;

and fusing all the knowledge sub-pictures and all the interaction maps to obtain the knowledge map.

In a possible embodiment, said fusing all said knowledge sub-pictures and all said interaction maps to obtain a knowledge map includes:

deleting feature nodes in each knowledge sub-graph, and connecting each two object nodes connected with the same feature node by using associated edges to obtain the processed knowledge sub-graph;

and fusing all the processed knowledge sub-maps and all the interaction maps to obtain the knowledge map.

In a second aspect of the embodiments of the present invention, there is provided a graph embedding apparatus for a knowledge graph, where the knowledge graph includes multiple classes of nodes, where each class of node is used to represent a same type of object, the knowledge graph includes interactive edges and associated edges, the interactive edges are used to connect nodes that have interactions and different types between represented objects, and the associated edges are used to connect nodes that have associations and the same types of represented objects, the apparatus includes:

the system comprises a characteristic obtaining module, a characteristic obtaining module and a characteristic obtaining module, wherein the characteristic obtaining module is used for obtaining topological information of a first class node and a second class node, the topological information is used for representing a topological relation between the first class node and the second class node in the knowledge graph, and an interaction edge exists between the first class node and the second class node;

the first feature extraction module is used for extracting the graph features of the topological information of the type of nodes by utilizing a graph feature extraction model obtained based on deep learning aiming at each type of nodes in the first type of nodes and the second type of nodes respectively to obtain the deep features of the type of nodes;

and the second feature extraction module is used for extracting the graph features of the deep features of the first class nodes and the deep features of the second class nodes according to the interaction information between the first class nodes and the second class nodes to obtain the object representations of the first class nodes and the object representations of the second class nodes, wherein the interaction information is used for representing the topological relation formed by the first class nodes, the second class nodes and the interaction edges between the first class nodes and the second class nodes in the knowledge graph.

In a possible embodiment, the apparatus further includes an information interaction module, configured to, before the graph feature extraction model obtained based on deep learning is used to perform graph feature extraction on the topology information of the class of node to obtain deep features of the class of node, obtain an interaction feature between the first class of node and the second class of node, where the interaction feature is used to represent a feature of an interaction generated between an object represented by the first class of node and an object represented by the second class of node;

the first feature extraction module performs graph feature extraction on the topological information of the type of node by using a graph feature extraction model obtained based on deep learning to obtain deep features of the type of node, and the method comprises the following steps:

In a possible embodiment, the information interaction module inputs the topology information of the first type of node, the topology information of the second type of node, and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction, so as to obtain the topology information of the first type of node after interaction and the topology information of the second type of node after interaction, which are output by the bidirectional long and short term memory network, and the information interaction module includes:

the second feature extraction module performs graph feature extraction on the deep features of the first type of nodes and the deep features of the second type of nodes according to the interaction information between the first type of nodes and the second type of nodes to obtain object representations of the first type of nodes and object representations of the second type of nodes, and the graph feature extraction module comprises the following steps:

In a possible embodiment, the performing, by the first feature extraction module, graph feature extraction on the topology information of the class node by using a graph feature extraction model obtained based on deep learning to obtain deep features of the class node includes:

In a possible embodiment, the performing, by the second feature extraction module, graph feature extraction on the deep features of the first class of nodes and the deep features of the second class of nodes according to interaction information between the first class of nodes and the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes includes:

In one possible embodiment, the apparatus further comprises a graph construction module for pre-constructing the knowledge-graph in the following manner:

In a possible embodiment, the graph construction module fuses all the knowledge sub-pictures and all the interaction maps to obtain a knowledge map, including:

In a third aspect of embodiments of the present invention, there is provided an electronic device, including:

a memory for storing a computer program;

a processor adapted to perform the method steps of any of the above first aspects when executing a program stored in the memory.

In a fourth aspect of embodiments of the present invention, a computer-readable storage medium is provided, in which a computer program is stored, which, when being executed by a processor, carries out the method steps of any one of the above-mentioned first aspects.

The embodiment of the invention has the following beneficial effects:

according to the method, the device and the electronic equipment for embedding the knowledge graph, provided by the embodiment of the invention, the topological information among the nodes of the same type and the interactive information among the nodes of different types can be extracted by utilizing graph feature extraction, so that the global information in the heterogeneous knowledge graph can be better reserved in the representation, namely, the information lost when the graph is embedded in the heterogeneous knowledge graph is reduced.

Of course, not all of the advantages described above need to be achieved at the same time in the practice of any one product or method of the invention.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts.

FIG. 1 is a schematic diagram of a knowledge graph according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart diagram of a method for graph embedding of a knowledge graph according to an embodiment of the present invention;

fig. 3 is a schematic flowchart of an information interaction method according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a network structure of a bidirectional long term and short term memory network according to an embodiment of the present invention;

FIG. 5 is a schematic flow chart of a deep feature calculation method according to an embodiment of the present invention;

FIG. 6 is a flowchart illustrating an interactive convolution method according to an embodiment of the present invention;

FIG. 7 is a schematic flow chart of a method for constructing a knowledge graph according to an embodiment of the present invention;

FIG. 8a is a schematic flow chart of a graph data mining method according to an embodiment of the present invention;

fig. 8b is a schematic structural diagram of a network for implementing end-to-end mapping of topology information to object representations according to an embodiment of the present invention;

FIG. 8c is a schematic diagram of a knowledge graph construction provided by an embodiment of the present invention;

FIG. 9a is a schematic diagram of a diagram embedding apparatus for a knowledge graph according to an embodiment of the present invention;

FIG. 9b is a schematic diagram of another example of a map embedding apparatus for a knowledge map according to an embodiment of the present invention;

FIG. 9c is a schematic diagram of another example of a map embedding apparatus for a knowledge map according to an embodiment of the present invention;

fig. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to better explain the graph embedding method of the knowledge graph provided by the embodiment of the present invention, the knowledge graph provided by the embodiment of the present invention will be explained below, and reference may be made to fig. 1, where fig. 1 is a schematic structural diagram of the knowledge graph provided by the embodiment of the present invention, where the schematic structural diagram includes:

a first class node 110, a second class node 120, a third class node 130, an interactive edge 140, and an associated edge 150.

Wherein, each first class node 110 is used for representing the same kind of object, each second class node 120 is used for representing the same kind of object, and each third class node 130 is used for representing the same kind of object. Fig. 1 is a schematic structural diagram of a knowledge graph provided in an embodiment of the present invention, and in other possible embodiments, the knowledge graph may include only two types of nodes, and may further include four or more types of nodes, which is not limited in this embodiment.

The kind of objects represented by each type of node may be different according to application scenarios, for example, the first type of node 110 may be used to represent a user, the second type of node 120 may be used to represent an item, and the third type of node 130 may be used to represent a company.

The interaction edges 140 are used to connect nodes with different types of interaction between the objects represented by the nodes, for example, the interaction edge 140 connecting one first-type node 110 and one second-type node 120 may be used to represent the interaction between the object represented by the one first-type node 110 and the objects represented by the one second-type nodes 120. Illustratively, assuming that one first type node 110 represents user A and one second type node 120 represents item 1, if user A has an interaction with item 1, such as user A purchased item 1, user A used item 1, etc., the one first type node 110 and the one second type node 120 may be connected in the knowledge-graph using an interaction edge 140.

The associated edges 150 are used to connect nodes that are related and have the same kind of object, for example, the associated edges 150 connecting one first-type node 110 and another second-type node 110 may be used to represent that there is a relation between an object represented by the one first-type node 110 and an object represented by the another first-type node 110. For example, assuming that one first-type node 110 represents a user a and another first-type node 110 represents a user B, if the user a and the user B have an association, such as the user a and the user B belong to the same community, the user a and the user B are both young men, and social activity occurs between the user a and the user B, the one first-type node 110 and the another first-type node 110 may be connected by using an association edge 150 in the knowledge graph.

Compared with a homogeneous knowledge graph only comprising nodes representing the same kind of objects, the heterogeneous knowledge graph has nodes representing more kinds of objects and interaction edges for representing interaction information between different kinds of objects, so the complexity degree is often higher. The graph embedding method aiming at the traditional isomorphic knowledge graph, such as the graph embedding method based on random walk, is difficult to completely retain the information in the heterogeneous knowledge graph to the representation.

Based on this, an embodiment of the present invention provides a map embedding method for a knowledge graph, which may be referred to fig. 2, where fig. 2 is a schematic flow diagram of the map embedding method for a knowledge graph provided by the embodiment of the present invention, and the method may include:

s201, acquiring topology information of the first class node and the second class node.

S202, respectively aiming at each type of nodes in the first type of nodes and the second type of nodes, carrying out graph feature extraction on the topological information of the type of nodes by utilizing a graph feature extraction model obtained based on deep learning to obtain deep features of the type of nodes.

S203, according to the interaction information between the first class of nodes and the second class of nodes, carrying out graph feature extraction on the deep features of the first class of nodes and the deep features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

By adopting the embodiment, the topological information among the nodes of the same type and the interactive information among the nodes of different types can be extracted in a deep learning mode, so that the global information in the heterogeneous knowledge graph can be better retained in the representation, namely, the information lost when the graph is embedded in the heterogeneous knowledge graph is reduced.

In S201, the topology information is used to represent a topological relationship between nodes of a class in the knowledge graph. The topology information may be represented in different ways, and in one possible embodiment may be represented in a diagonal matrix and an adjacency matrix.

The diagonal matrix is used for representing the number of similar nodes connected with each node, and the adjacent matrix is used for representing the link relation among the nodes. For example, assuming that there are n first-type nodes in total, the diagonal matrix of the first-type nodes is a matrix of n × n, where the i-th row and the i-th column have elements representing the number of other first-type nodes connected to the i-th first-type node, and all elements except the diagonal elements are 0.

The adjacency matrix of the first-class nodes is a matrix of n × n, wherein elements on a diagonal line are 0, if an associated edge exists between the ith first-class node and the jth first-class node and the associated edge is a weighted edge, the element in the ith row and the jth column in the adjacency matrix is an edge weight of the associated edge between the ith first-class node and the jth first-class node, if an associated edge exists between the ith first-class node and the jth first-class node and the associated edge is a non-weighted edge, the element in the ith row and the jth column in the adjacency matrix is 1, and if no edge exists between the ith first-class node and the jth first-class node, the element in the ith row and the jth column in the adjacency matrix is 0.

According to different application scenarios, the diagonal matrix and the adjacency matrix of the first type node may be combined in different ways to represent the topological relationship between the first type nodes, for example, the inverse matrix of the diagonal matrix may be multiplied by the adjacency matrix, and the obtained product represents the topological relationship between the first type nodes. For the topological relationship between the second type nodes, reference may be made to the foregoing description of the topological relationship between the first type nodes, which is not described herein again.

It is to be understood that the first type of node and the second type of node are used herein only to distinguish two different types of nodes, and do not refer to a certain type of node. According to different application scenarios, the first class node and the second class node may be different classes of nodes, and in this embodiment, there is an interaction between objects represented by the first class node and the second class node, that is, there is an interaction edge between at least one first class node and one second class node in the knowledge graph.

In S202, the graph feature extraction model may be obtained through deep learning training in advance, and the graph feature extraction model may be different according to different application scenarios. Compared with random walk, the deep learning can extract global information more comprehensively in an end-to-end mapping mode.

In S203, the mutual information is used to represent a topological relationship between the first class node and the second class node in the knowledge-graph. For the representation manner of the topological relationship between the first-class nodes and the second-class nodes, reference may be made to the foregoing description of the topological relationship between the first-class nodes, and details are not repeated here.

Because there is interaction between the object represented by the first kind of node and the object represented by the second kind of node, the interaction characteristics of the interaction that occurs and some information that can represent the relationship between the objects represented by the first kind of node exist in the topology information of the second kind of node can be considered. Similarly, some information which can represent the relationship between the objects represented by the second type nodes exists in the interaction characteristics and the topology information of the first type nodes. Based on this, in one possible embodiment, as shown in fig. 3, the method includes:

s301, acquiring topological information of the first class node and the second class node and interaction characteristics between the first class node and the second class node.

The interactive characteristics are used for representing characteristics of interaction generated between the object represented by the first type node and the object represented by the second type node.

S302, inputting the topology information of the first node, the topology information of the second type of node and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction, and obtaining the interacted topology information of the first type of node and the interacted topology information of the second type of node output by the bidirectional long and short term memory network.

The sequence of forward propagation in the bidirectional long and short term memory network is one of the first sequence and the second sequence, and the sequence of backward propagation in the bidirectional long and short term memory network is the other one of the first sequence and the second sequence. The first sequence is: topology information of the nodes in the first column, interactive characteristics and interactive characteristics of the nodes in the second column. The second sequence is: topology information of the second type of nodes, interaction characteristics of the second type of nodes and interaction characteristics of the first type of nodes.

The network structure of the bidirectional long and short term memory network can be as shown in fig. 4, in which bold solid line arrows are used to indicate the order of forward propagation and bold dashed line arrows are used to indicate the order of backward propagation. The network structure shown in fig. 4 is only one possible network structure of the bidirectional long and short term memory network provided by the embodiment of the present invention, and in other possible embodiments, the order of forward propagation and the order of backward propagation may be opposite to the order shown in fig. 4, which is not limited by the embodiment.

And S303, respectively aiming at each type of node in the first type of node and the second type of node, carrying out graph feature extraction on the interacted topology information of the type of node according to the topology information of the type of node, and obtaining deep features of the type of node.

Taking the topology information of the first type of nodes after interaction as an example, because the topology information of the second type of nodes and the information which can be used for representing the object represented by the first type of nodes in the interaction characteristics are merged into the topology information after interaction, compared with the topology information before interaction, the topology information after interaction can better represent the object represented by the first type of nodes.

S304, according to the interaction information between the first class of nodes and the second class of nodes, carrying out graph feature extraction on the deep features of the first class of nodes and the deep features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

This step is the same as S203, and reference may be made to the related description about S203, which is not described herein again.

By selecting the embodiment, the characteristic that the bidirectional long and short term memory network can determine the incidence relation between the contexts can be fully utilized, and for each type of node, the bidirectional long and short term memory network is used for extracting the information of the relation between the objects represented by the type of node from the interactive characteristics and the topological information of other types of nodes, so that the information included in the obtained representation is further enriched.

The following description will be made of a bidirectional long-short term memory network:

the forgetting gate of the bidirectional long-short term memory network can be as follows:

f_j＝σ(W_fx_j+W_hh_j-1+b_f)

the input gates may be as follows:

in_j＝σ(W_inx_j+W_hh_j-1+b_in)

the output gates may be as follows:

o_j＝σ(W_ox_j+W_hh_j-1+b_o)

the subscript j sequentially refers to a first type node, an interactive feature and a second type node in sequence, for example, when j equals 1, the subscript j equals 2, the subscript j equals 3, the subscript j indicates the first type node, the interactive feature and the second type node, and x is_jDifferent meanings are shown according to different values of j, and x is x when j is 1_jTopology information representing nodes of the first type, x when j is 2_jRepresenting interactive features, x when j is 3_jTopology information representing nodes of the second type. h is_j-1The meaning is different according to different values of j-1, and h is different when j-1 is equal to 1_j-1Representing the topology information after interaction of the first class of nodes, h when j-1 is 2_j-1Representing the characteristics of the interaction after the interaction, h when j-1 is 3_j-1And representing the topology information of the second class of nodes after interaction. f. of_j、in_j、

c_j、c_j-1、o_jRepresenting intermediate quantities, W, in a bidirectional long-short term memory network_f、W_h、b_f、W_in、b_in、W_z、b_z、W_o、b_oAre the training parameters of the respective doors, h_fHidden state representing the forward process, h_j,fRepresents j orHidden state of the forward process of the object, e.g. h_1,fHidden state representing the forward process of the first type of node, h_2,fHidden state of forward process representing interactive features, h_3,fAnd (4) representing a hidden state of a forward process of the second type node. σ (-) represents a sigmoid activation function, tanh (-) represents a hyperbolic tangent function

Representing an inter-vector element bit-wise multiplication operation. Hidden state h of backward process_bIs calculated by_fSimilarly, the difference is only that the subscript j sequentially refers to the second-class node, the interaction, and the first-class node in order. The feature after interaction is represented by h_fAnd h_bCommon decisions, e.g. topology information h after interaction of nodes of the first kind₁Can be represented by the following formula:

h₁＝h_1,f+h_1,b。

wherein h is_1,bAnd (4) representing a hidden state of a backward process of the first type node.

In a possible embodiment, the bidirectional long and short term memory network may further output the interactive characteristics after interaction, and in this embodiment, the interactive characteristics after interaction may be used as the interaction information between the first type node and the second type node. In other words, in this embodiment, the deep features of the first type of nodes and the deep features of the second type of nodes may be subjected to graph feature extraction according to the interactive features after the interaction, so as to obtain the object representations of the first type of nodes and the second type of nodes.

A process of calculating a deep layer feature will be described below, referring to fig. 5, where fig. 5 is a schematic flow chart of a method for calculating a deep layer feature according to an embodiment of the present invention, and the method may include:

s501, constructing a node connection matrix of the node.

The node connection matrix is used for representing the connection relation between each node of the class and other nodes of the class in the knowledge graph. As described above, the connection relationship may be represented by an angle matrix and an adjacency matrix, and therefore the node connection matrix may be calculated from a diagonal matrix and an adjacency matrix, and for convenience of description, the following formula a represents the adjacency matrix, D represents the diagonal matrix, and L represents the node connection matrix, and then the node connection matrix may be calculated according to the following formula:

L＝D^-1A

in addition, the node connection matrix of the first type node and the node connection matrix of the second type node may be the same or different, which is not limited in this embodiment. For example, in a possible embodiment, assuming that in an interaction generated between objects represented by a first class of nodes and a second class of nodes, the object represented by the first class of nodes is an interaction initiator, the object represented by the second class of nodes is an interaction acceptor, for example, the object represented by the first class of nodes is a user, and the object represented by the second class of nodes is an article, the node connection matrix of the first class of nodes may be obtained by performing weighted average on a product matrix and an identity matrix, where the product matrix is a product of an inverse matrix of a diagonal matrix of the first class of nodes and an adjacency matrix, that is, the node connection matrix of the first class of nodes may be obtained by calculating according to the following formula:

L＝(1-α)I+αD^-1A

where I is an identity matrix with the same dimensions as D and a, α is a coefficient negatively correlated to the diagonal matrix, and α has a value in the range of (0,1), which is exemplary, and in one possible embodiment, α ═ e^-D. By adopting the embodiment, the first class nodes with higher knowledge graph popularity (the number of connected similar nodes) can contribute more to the whole information propagation, so that the obtained object representation of the first class nodes can reflect the information in the knowledge graph more accurately.

In this embodiment, the node connection matrix of the second node may be obtained by calculating a product of an evolution matrix, an adjacent matrix of the second type node, and the evolution matrix multiplied in sequence, where the evolution matrix is obtained by evolution of an inverse matrix of a diagonal matrix of the second type node, that is, the node connection matrix of the second type node may be obtained by calculation according to the following formula:

L＝D^-1/2AD^-1/2

since the objects represented by the second type nodes are passive in the interaction, the information capacity of the objects represented by the second type nodes for aggregating other similar objects should be emphasized, and meanwhile, the information of the objects represented by the second type nodes are limited to be amplified. The embodiment can highlight the information capability of the object represented by the second type node for converging other similar objects, and simultaneously limit the self information of the object represented by the second type node to be amplified. Therefore, the obtained object representation of the second type node can reflect the information in the knowledge graph more accurately

And S502, according to the node connection matrix, carrying out graph feature extraction on the topological information of the type of node to obtain deep features of the type of node.

The process of extracting the graph features can be circularly performing convolution processing according to the following formula until the preset cycle number is reached:

H^l＝ELU(LH^l-1W^l)

wherein H^lRepresents the output of the first convolution process, H^l-1Represents the output of the (l-1) th convolution process, H⁰Representing topological information, W^lRepresents the parameter matrix used in the first convolution process, and ELU (. circle.) represents the activation function. With respect to W^lThe determination of (c) will be described in the following examples.

As will be described below with reference to fig. 6, fig. 6 is a schematic flowchart of an interactive convolution method according to an embodiment of the present invention, where the method includes:

s601, respectively aiming at each type of node in the first type of node and the second type of node, performing feature pooling on deep features and interaction information of the type of node to obtain pooling features of the type of node.

Taking the first type node as an example, the pooling characteristic of the first type node may be calculated according to the following formula:

wherein E is₁Representing the pooled features of the nodes of the first type,

deep features representing nodes of the first class, h₁₂Representing the mutual information. meanpooling (. cndot.) denotes the averaging of the para-positions, i.e.E₁The value of the element at any position is equal to

And h₁₂Average value of the elements at any one of the positions.

And S602, according to the topological relation represented by the interactive information, carrying out graph feature extraction on the pooling features of the first class of nodes and the pooling features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

The method for extracting the graph features of the pooled features of the first class nodes and the pooled features of the second class nodes may refer to the foregoing description about the graph feature extraction. In one possible embodiment, the convolution process may be performed circularly according to the following formula until the preset number of cycles is reached:

wherein L is₁₂A node connection matrix used for representing the incidence relation between the first class node and the second class node,

for the parameter matrix used in the first convolution processing, E^lFor the output obtained by the first convolution processing, E^l-1An output obtained for the first-1 convolution processing, and

wherein E is₁Pooling characteristics for nodes of the first type, E₂Pooling characteristics for the second class of nodes.

In order to more clearly describe the graph embedding method of the knowledge graph provided by the embodiment of the present invention, the following description is provided for the construction of the knowledge graph, and referring to fig. 7, fig. 7 is a schematic flow chart of the knowledge graph construction method provided by the embodiment of the present invention, and the schematic flow chart may include:

s701, aiming at each object in a plurality of objects, constructing a knowledge sub-map of the object.

The knowledge sub-graph further comprises attribution edges, each attribution edge is used for connecting one feature node with one object node and is used for indicating that the objects represented by the connected object nodes have the features represented by the connected feature nodes.

For example, taking two objects including a user and an item as an example, a knowledge sub-graph of the user and a knowledge sub-graph of the item may be constructed for the user and the item, respectively. Each object node in the user's knowledge sub-graph is used to represent an object, and each feature node is used to represent a feature of the user, such as place of birth, age, etc. Assuming that the object node a represents the user a and the feature node 1 represents the age group 16-18 years, if there is a home edge between the object node a and the feature node 1, the home edge may represent the age group of the user a as 16-18 years.

S702, aiming at each two objects with interaction in a plurality of objects, an interaction map between the two objects is constructed.

The interaction graph includes object nodes and interaction edges, and for the interaction edges, reference may be made to the foregoing related description, which is not described herein again.

And S703, fusing all the knowledge sub-maps and all the interaction maps to obtain the knowledge map.

The feature nodes in the knowledge sub-graph are deleted aiming at each knowledge sub-graph, and each two object nodes connected with the same feature node are connected by using the associated edges to obtain the processed knowledge sub-graph. And all the processed knowledge sub-maps and all the interaction maps are fused to obtain the knowledge map.

By adopting the embodiment, the information related to various different objects can be fused to construct the knowledge graph, so that the constructed knowledge graph contains more information.

In the following, a description is given with reference to a flow of construction of a knowledge graph and graph embedding, and referring to fig. 8a, fig. 8a is a schematic flow diagram of a graph data mining method provided by an embodiment of the present invention, and the method may include:

s801, aiming at each object in a plurality of objects, constructing a knowledge sub-map of the object.

This step can be referred to the related description of the foregoing S701, and is not described herein again.

S802, aiming at each two objects with interaction in the multiple objects, an interaction map between the two objects is constructed.

This step can be referred to the related description of the aforementioned S702, and is not described herein again.

And S803, fusing all the knowledge sub-maps and all the interaction maps to obtain the knowledge map.

This step can be referred to the related description of the foregoing S703, and is not described herein again.

S804, topology information of the first type of nodes and the second type of nodes and interaction characteristics between the first type of nodes and the second type of nodes are obtained.

This step can be referred to the related description in the foregoing S201 and S301, and is not described herein again.

And S805, inputting the topology information of the first node, the topology information of the second type of node and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction to obtain the interacted topology information of the first type of node, the interacted topology information of the second type of node and the interacted interaction characteristics output by the bidirectional long and short term memory network.

For the calculation of the bidirectional long-short term memory network, referring to the above description, the training parameters of the gates in the bidirectional long-short term memory network may be preset initial values when S805 is performed for the first time.

And S806, respectively aiming at each type of the first type of nodes and the second type of nodes, carrying out graph feature extraction on the interacted topology information of the type of nodes by utilizing a graph feature extraction model obtained based on deep learning to obtain deep features of the type of nodes.

Referring to the foregoing description of S501 and S502, the parameter matrix used in the graph feature extraction when S806 is performed for the first time may be a preset initial value.

S807, respectively aiming at each type of node in the first type of node and the second type of node, performing feature pooling on the deep features of the type of node and the interactive features after interaction to obtain the pooling features of the type of node.

Reference may be made to the foregoing description of S601, which is not repeated herein.

And S808, according to the topological relation represented by the interactive information, carrying out graph feature extraction on the pooling features of the first class of nodes and the pooling features of the second class of nodes to obtain object representations of the first class of nodes and object representations of the second class of nodes.

Reference may be made to the foregoing description of S601, which is not repeated herein. The parameter matrix used in the map feature extraction at the time of the first execution of S808 may be a preset initial value.

And S809, executing subsequent downstream tasks by using the obtained representation.

The downstream tasks may be different according to different application scenarios, and may include node classification, relationship prediction, item recommendation, and the like.

And S810, constructing a loss function according to the accuracy of the downstream task execution result, updating the model parameters through the minimized loss function, and returning to execute S805 until the return times reach a preset time threshold.

The model parameters include the training parameters in the aforementioned S805, and the parameter matrix in S806 and S808. When the preset time threshold is reached, the model parameters can be considered to be optimized, and at the moment, according to the determined model parameters, the steps of S801-S808 are repeated again to obtain accurate representations of various nodes.

Wherein S805-S808 can be accomplished through the model shown in FIG. 8 b.

When the object represented by the first type node is a user and the object represented by the second type node is an article, the process of constructing the knowledge graph shown in S801-S804 can be referred to fig. 8c, and fig. 8c is a schematic diagram of a principle of constructing the knowledge graph provided by the embodiment of the present invention.

Referring to fig. 9a, fig. 9a is a schematic structural diagram of a graph embedding apparatus of a knowledge graph according to an embodiment of the present invention, where the knowledge graph includes multiple types of nodes, where each type of node is used to represent a same type of object, the knowledge graph includes interaction edges and association edges, the interaction edges are used to connect nodes that have interactions and different types between the represented objects, and the association edges are used to connect nodes that have associations and the same types of the represented objects, the apparatus includes:

a feature obtaining module 901, configured to obtain topology information of a first class of nodes and a second class of nodes, where the topology information is used to represent a topology relationship between the first class of nodes in the knowledge graph, and an interaction edge exists between the first class of nodes and the second class of nodes;

a first feature extraction module 902, configured to perform, for each of the first class of nodes and the second class of nodes, graph feature extraction on topology information of the class of nodes by using a graph feature extraction model obtained based on deep learning, so as to obtain deep features of the class of nodes;

a second feature extraction module 903, configured to perform graph feature extraction on deep features of the first type of nodes and deep features of the second type of nodes according to interaction information between the first type of nodes and the second type of nodes, so as to obtain object representations of the first type of nodes and object representations of the second type of nodes, where the interaction information is used to represent a topological relation formed by the first type of nodes, the second type of nodes, and interaction edges between the first type of nodes and the second type of nodes in the knowledge graph.

In a possible embodiment, as shown in fig. 9b, the apparatus further includes an information interaction module 904, configured to obtain an interaction feature between the first type node and the second type node before the deep feature of the type node is obtained by performing graph feature extraction on the topology information of the type node by using the graph feature extraction model obtained based on deep learning, where the interaction feature is used to represent a feature of an interaction generated between an object represented by the first type node and an object represented by the second type node;

the first feature extraction module 902 performs graph feature extraction on the topology information of the type of node by using a graph feature extraction model obtained based on deep learning, so as to obtain deep features of the type of node, including:

In a possible embodiment, the information interaction module 904 inputs the topology information of the first type of node, the topology information of the second type of node, and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction, so as to obtain the topology information of the first type of node after interaction and the topology information of the second type of node after interaction, which are output by the bidirectional long and short term memory network, and includes:

the second feature extraction module 903 performs graph feature extraction on the deep features of the first type of node and the deep features of the second type of node according to the interaction information between the first type of node and the second type of node, so as to obtain object representations of the first type of node and object representations of the second type of node, including:

In a possible embodiment, the first feature extraction module 902 performs graph feature extraction on the topology information of the class node by using a graph feature extraction model obtained based on deep learning, so as to obtain deep features of the class node, including:

In a possible embodiment, the second feature extraction module 903 performs graph feature extraction on the deep features of the first type of node and the deep features of the second type of node according to the interaction information between the first type of node and the second type of node, so as to obtain the object representation of the first type of node and the object representation of the second type of node, including:

In one possible embodiment, as shown in fig. 9c, the apparatus further comprises a graph construction module 905 for pre-constructing the knowledge-graph in the following manner:

In a possible embodiment, the graph building module 905 fuses all the knowledge sub-pictures and all the interaction maps to obtain a knowledge map, including:

An embodiment of the present invention further provides an electronic device, as shown in fig. 10, including:

a memory 1001 for storing a computer program;

the processor 1002 is configured to implement the following steps when executing the program stored in the memory 1001:

In a possible embodiment, the knowledge-graph is constructed beforehand by:

The Memory mentioned in the above electronic device may include a Random Access Memory (RAM) or a Non-Volatile Memory (NVM), such as at least one disk Memory. Optionally, the memory may also be at least one memory device located remotely from the processor.

The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components.

In yet another embodiment of the present invention, a computer-readable storage medium is further provided, in which a computer program is stored, and the computer program, when executed by a processor, implements the steps of the graph embedding method of any one of the above-mentioned knowledge-graphs.

In yet another embodiment, a computer program product containing instructions is provided, which when run on a computer, causes the computer to perform the graph embedding method of any of the above-described embodiments.

In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, cause the processes or functions described in accordance with the embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website site, computer, server, or data center to another website site, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

It is noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

All the embodiments in the present specification are described in a related manner, and the same and similar parts among the embodiments may be referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, as for the apparatus, the electronic device, the computer-readable storage medium, and the computer program product, since they are substantially similar to the method embodiments, the description is relatively simple, and in relation to the description, reference may be made to part of the description of the method embodiments.

The above description is only for the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention shall fall within the protection scope of the present invention.

Claims

1. A graph embedding method of a knowledge graph, wherein the knowledge graph comprises a plurality of types of nodes, each type of node is used for representing the same type of object, the knowledge graph comprises interactive edges and associated edges, the interactive edges are used for connecting nodes which have interaction and different types among the represented objects, and the associated edges are used for connecting nodes which have association and the same type among the represented objects, the method comprises the following steps:

2. The method according to claim 1, wherein before the graph feature extraction is performed on the topology information of the class node by using the graph feature extraction model obtained based on deep learning to obtain deep features of the class node, the method further comprises:

3. The method according to claim 2, wherein the inputting the topology information of the first kind of nodes, the topology information of the second kind of nodes, and the interaction characteristics into a preset bidirectional long and short term memory network for information interaction to obtain the interacted topology information of the first kind of nodes and the interacted topology information of the second kind of nodes output by the bidirectional long and short term memory network comprises:

4. The method according to claim 1, wherein the extracting the topology information of the class node by using the graph feature extraction model obtained based on deep learning to obtain the deep features of the class node comprises:

5. The method according to claim 4, wherein in the interaction generated between the objects represented by the first kind of nodes and the second kind of nodes, the object represented by the first kind of nodes is an interaction initiator;

6. The method according to claim 4, wherein in the interaction generated between the objects represented by the first class node and the second class node, the object represented by the second class node is an interaction receiver;

7. The method according to claim 1, wherein the extracting deep features of the first type nodes and deep features of the second type nodes according to interaction information between the first type nodes and the second type nodes to obtain object representations of the first type nodes and object representations of the second type nodes comprises:

8. The method according to any one of claims 1 to 7, wherein the knowledge-graph is constructed in advance by:

9. The method according to claim 8, wherein said fusing all said knowledge-sub-pictures and all said interaction maps to obtain a knowledge-map comprises:

10. An apparatus for graph embedding of a knowledge graph, wherein the knowledge graph comprises a plurality of types of nodes, each type of node is used for representing a same type of object, the knowledge graph comprises interaction edges and association edges, the interaction edges are used for connecting nodes which are represented by objects and have different types, and the association edges are used for connecting nodes which are represented by objects and have associated types and the same type, the apparatus comprises:

11. An electronic device, comprising:

a memory for storing a computer program;

a processor for implementing the method steps of any of claims 1-9 when executing a program stored in the memory.

12. A computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, which computer program, when being executed by a processor, carries out the method steps of any one of the claims 1-9.