CN114124729A - Dynamic heterogeneous network representation method based on meta-path - Google Patents

Dynamic heterogeneous network representation method based on meta-path Download PDF

Info

Publication number
CN114124729A
CN114124729A CN202111393567.6A CN202111393567A CN114124729A CN 114124729 A CN114124729 A CN 114124729A CN 202111393567 A CN202111393567 A CN 202111393567A CN 114124729 A CN114124729 A CN 114124729A
Authority
CN
China
Prior art keywords
node
time
path
meta
network
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202111393567.6A
Other languages
Chinese (zh)
Inventor
谭洪胜
刘群
袁铭
王国胤
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University of Post and Telecommunications
Original Assignee
Chongqing University of Post and Telecommunications
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Chongqing University of Post and Telecommunications filed Critical Chongqing University of Post and Telecommunications
Priority to CN202111393567.6A priority Critical patent/CN114124729A/en
Publication of CN114124729A publication Critical patent/CN114124729A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention belongs to the field of graph network representation learning, and particularly relates to a dynamic heterogeneous network representation method based on a meta-path, which comprises the steps of constructing a dynamic heterogeneous network, and sampling different time weighting meta-path sequences from the network according to a time weighting meta-path; preprocessing the vector of the network node, and aggregating the information of the network node sequence of each element path after preprocessing through a GRU (generalized regression Unit); coding the time of the node sequence by adopting a relative time coding technology; acquiring deep-level characteristic information of the node sequence by adopting a Bi-GRU method, and aggregating time characteristics and structural characteristics; interacting characteristic information of different sequences at different time by using a Bi-GRU with an attention mechanism to obtain final representation of the node; the method can adapt to the node learning task and the dynamic evolution of the network under the dynamic heterogeneous network, and the downstream task can classify and cluster the nodes, thereby effectively improving the learning and representing capability of the nodes of the graph network.

Description

Dynamic heterogeneous network representation method based on meta-path
Technical Field
The invention belongs to the field of graph network representation learning, and particularly relates to a dynamic heterogeneous network representation method based on a meta path.
Background
Network data is generally unstructured data, and it is difficult to mine information in the network directly using machine learning models. Network representation learning is to map high-dimensional sparse graph data to a low-dimensional space, and meanwhile, retain structural information in a network, so that low-dimensional dense structural vector representation is obtained.
At present, the study on static network representation learning and dynamic homogeneous network representation learning is relatively mature, but the dynamic heterogeneous network study which is closer to the actual network is in the beginning stage, and it is necessary to discuss and study the dynamic heterogeneous network study.
Network representation learning is to represent nodes in a network into low-dimensional vectors while preserving network structure and semantic information so as to facilitate subsequent tasks of graph mining, such as link prediction, node classification, clustering and the like. Many existing heterogeneous network representation learning methods are designed for static heterogeneous networks, and the characteristics that network space information (topological structure and attributes) changes along with time are ignored, and only spaces (structures) corresponding to different times are simply compressed together. Since the network is constantly changing over time, two nodes that have no relation in the previous second may be associated in the next second, and the establishment of links between nodes obviously changes the topology (space) of the network. Therefore, only considering the static processing mode does not conform to the evolution rule of the actual network. And the other type of dynamic homogeneous network representation learning method does not consider the difference of network nodes and links, and if the method is directly applied to a dynamic heterogeneous network, some semantic information is inevitably lost.
Disclosure of Invention
In order to enable downstream tasks to be capable of classifying, clustering, visualizing and the like butt nodes and effectively improve the learning and expression capability of nodes of a dynamic heterogeneous graph network, the invention provides a dynamic heterogeneous network expression learning method based on a meta-path, a DHNR model is constructed to perform expression learning on the network nodes, the DHNR model comprises GRUs and Bi-GRUs with attention mechanisms, and the process of acquiring the node expression vectors by the DHNR model specifically comprises the following steps:
s1: the time for establishing the link of the network node is used as the link weight and is reserved in the network to construct a dynamic heterogeneous network;
s2: sampling different time weighting element path sequences from the network according to the time weighting element paths;
s3: preprocessing a network to obtain an initial vector of each node, inputting each time weighted meta-path sequence and the vector thereof, and aggregating information of the network node sequence of each meta-path through a GRU (general packet Unit);
s4: a relative time coding technology is adopted as a time coder to code the time of the node sequence;
s5: acquiring deep-level characteristic information of the node sequence by adopting a Bi-GRU method, and aggregating time characteristics and structural characteristics;
s6: and interacting the characteristic information of different sequences at different times by using a Bi-GRU with an attention mechanism to obtain a final representation of the node.
Further, the dynamic heterogeneous network is represented as G ═ V, E, T, where V represents a node set, E represents a link set, T represents a time set on a link, and a mapping function between a node and a node type in the dynamic heterogeneous network is represented as
Figure BDA0003369174130000021
A represents a node type set; e → R, R represents the link type set, in the dynamic heterogeneous network, | A | + | R | > 2, a time link in the dynamic heterogeneous network is represented as (i, j, t) and (i, j, t) ∈ E, the time link (i, j, t) represents that the node i is connected to the node j at the moment t, | A | represents the number of the node types in the node type set, and | R | represents the number of the link types in the link type set.
Further, sampling different time-weighted meta-path sequences from the network according to the time-weighted meta-path comprises the steps of:
constructing a time weighting path, namely, obtaining a time weighting meta-path corresponding to the type meta-path for one meta-path under the guidance of a time attribute value;
acquiring a time-weighted meta-path sequence set, namely acquiring time-weighted meta-paths of each type of meta-paths under different time attribute values of T, and representing the time-weighted meta-path sequence set as
Figure BDA0003369174130000031
Wherein k is the type number of the element path; t is a time attribute value set, | T | is the number of time attributes in the time attribute value set;
Figure BDA0003369174130000032
represents a time-weighted meta-path of the kth type under the guide of the | T | time attribute value.
Further, preprocessing the vector of the network node comprises:
setting a transformation matrix for each type of nodes, so that each type of nodes is projected into the same feature space through the transformation matrix, wherein the feature space contains features in all types of node feature spaces, and the node type is phiiNode n ofiThe projection process is represented as:
Figure BDA0003369174130000033
wherein x isiIs a node niThe original characteristics of (a);
Figure BDA0003369174130000034
is a node niThe projection feature of (a);
Figure BDA0003369174130000035
is node type is phiiA transformation matrix of the node of (1).
Further, aggregating information of the sequence of network nodes of each meta-path by the GRU includes the following processes:
aggregating the different time weighted element path information of the node to form a primary vector representation, and then the node n0The polymerization process of (a) comprises:
Figure BDA0003369174130000036
Figure BDA0003369174130000037
wherein the content of the first and second substances,
Figure BDA0003369174130000038
is a node niThe projection feature of (a); phi (n)i) Is a node niNode type of niIs a node n0M +1-a hop neighbor nodes on the time weighted element path sequence;
Figure BDA0003369174130000039
the hidden state of the m +1 hop neighbor node of the target node in the sequence;
Figure BDA00033691741300000310
representing a node niThe hidden state of the layer a is output through the GRU, and a is more than 0 and less than or equal to m1+, and m is the number of sequence links; a. them+1The type of the m +1 hop neighbor of the target node; a. them+1-aIs m +1-a hop neighbor type;
Figure BDA00033691741300000311
hidden information indicating that a-1 layer hidden layer is output
Figure BDA00033691741300000312
Node n with input GRUiProjection feature of
Figure BDA00033691741300000313
Polymerization is carried out.
Further, the hidden information output by the a-1 layer hidden layer
Figure BDA00033691741300000314
Node n with input GRUiProjection feature of
Figure BDA00033691741300000315
Carrying out the polymerization, namely:
Figure BDA00033691741300000316
Figure BDA0003369174130000041
Figure BDA0003369174130000042
Figure BDA0003369174130000043
wherein A isz、Ar、AhAnd Bz、Br、BhIs a model parameter of the GRU; z is a radical ofi、riIs an update gate and a forget gate in the GRU, and the value ranges thereof are both [0,1]；
Figure BDA00033691741300000415
Representing element-by-element multiplication;
Figure BDA0003369174130000044
the new state after forgetting the input vector and the state vector at the previous moment; sigma is an activation function;
Figure BDA0003369174130000045
is the final state processed by one GRU unit.
Further, the time of the node sequences is encoded, that is, a set of fixed sine functions is defined as the time offset to encode the time of each sequence, and the encoding process includes:
Figure BDA0003369174130000046
Figure BDA0003369174130000047
Figure BDA0003369174130000048
wherein Base (t), Base (t,2i) and Base (t,2i +1) are time bias functions,
Figure BDA0003369174130000049
for the time vector at time T, T _ Linear is a Linear fine tuning function, and d is a vector dimension.
Further, aggregating the temporal and structural features comprises:
Figure BDA00033691741300000410
wherein the content of the first and second substances,
Figure BDA00033691741300000411
represents a feature obtained by polymerizing a temporal feature and a structural feature;
Figure BDA00033691741300000412
is the characteristic of the node, when i is st, it represents the time characteristic of the node, when i is RT (t), it represents the semantic characteristic of the node; mean () represents a function of averaging the array, averaging the hidden states of the bidirectional GRU;
Figure BDA00033691741300000413
representing a representation of a bi-directional GRU.
Further, the step of using the Bi-GRU with attention mechanism to interact with different sequences of feature information at different times comprises the following steps:
and (3) interacting the characteristic information at different times by using the Bi-GRU to simulate the evolution of the network, namely:
Figure BDA00033691741300000414
reference attention mechanism to capture important feature vectors, where node niAttention weight αiThe calculation formula of (a) is as follows:
Figure BDA0003369174130000051
using the calculated attention value to connect the node n0All state features of (a) are aggregated into a final vector representation of the node:
Figure BDA0003369174130000052
wherein O ═ { O ═ O1,o2,....,ok|T|Is the set of state vector sequences of Bi-GRU, ok|T|Represents the k × | T | state vectors; leak ReLU is activation function, a belongs to R1×2d'For the attention parameter, d is the representative dimension of the vector, and the specific value is determined by experiments.
Furthermore, the structural features of the sequence are obtained through GRUs, then network node semantics and time features are obtained through Bi-GRUs with an attention mechanism, in the representing training process of the network, cross entropies minimizing real values and predicted values of all marked nodes are used as loss functions, the model is optimized through a gradient descent algorithm, and the loss functions are represented as follows:
Figure BDA0003369174130000053
wherein the content of the first and second substances,
Figure BDA0003369174130000054
representing a loss function; l represents a label; c is a classifier parameter, ylIs an index set of node labels; y islIs a node tag value, ulLabel values are predicted for the node feature vectors.
The invention provides a method for representing and learning a dynamic heterogeneous network aiming at time information contained in the dynamic heterogeneous network and rich semantic information brought by different nodes and links. The timing information of the network is well captured using a bi-directional gated round robin unit with an attention mechanism. The method can generate the representation of the dynamic heterogeneous network nodes, downstream tasks can be classified and clustered by connecting nodes, and the learning and representation capabilities of the dynamic heterogeneous network nodes are effectively improved.
Drawings
FIG. 1 is a flow chart of a meta-path based dynamic heterogeneous network representation learning method according to the present invention;
FIG. 2 is a diagram of an embodiment of a meta-path based dynamic heterogeneous network representation learning method of the present invention;
FIG. 3 is a model framework diagram of a dynamic heterogeneous network representation learning method based on meta-paths according to the present invention;
fig. 4 is a schematic diagram of medium GRU polymerization.
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.
The invention provides a dynamic heterogeneous network representation method based on a meta-path, which is characterized in that a DHNR model is constructed to carry out representation learning on network nodes, the DHNR model comprises GRUs and Bi-GRUs with attention mechanisms, and the process of acquiring node representation vectors by the DHNR model specifically comprises the following steps:
s1: the time for establishing the link of the network node is used as the link weight and is reserved in the network to construct a dynamic heterogeneous network;
s2: sampling different time weighting element path sequences from the network according to the time weighting element paths;
s3: preprocessing the vector of the network node for the input sequence to obtain an initialization vector, and aggregating the information of the network node sequence of each meta-path through a GRU;
s4: a relative time coding technology is adopted as a time coder to code the time of the node sequence;
s5: acquiring deep-level characteristic information of the node sequence by adopting a Bi-GRU method, and aggregating time characteristics and structural characteristics;
s6: and interacting the characteristic information of different sequences at different times by using a Bi-GRU with an attention mechanism to obtain a final representation of the node.
The invention provides a DHNR model, which comprises the steps of firstly dividing a neighborhood structure of a node into different subspace structures according to time, sampling sequences of all time weighting element paths for each node, integrating neighborhood information on all time weighting element path sequences of the node through a gating circulation unit (GRU), and finally learning space-time context information on the fused node sequence by using a bidirectional gating circulation unit (Bi-GRU) with an attention mechanism to obtain a final expression vector of each node. In this embodiment, as shown in fig. 1, fig. 2, and fig. 3, the structural neighborhood of each node is divided into weighted element path sequences at different times according to time; aggregating neighbor information in each temporal weighted meta-path sequence by a gated round robin unit (GRU); filtration screening was performed using a Bi-directional gated loop unit (Bi-GRU) with attention mechanism.
In the process of constructing the dynamic heterogeneous network, the time of establishing the link by the node is taken as a link weight and is reserved in the network, and the dynamic heterogeneous network is represented as G ═ V, E and T, wherein V represents a node set, E represents a link set, and T represents a time set on the link.
The mapping function between nodes and node types in the network is
Figure BDA0003369174130000071
The mapping function between links and link types is psi E → R, where A and R represent node type and link type, dynamic heterogeneousIn the network, | A | + | R | > 2, each time link (i, j, t) ∈ E indicates that node i is connected to node j at time t.
For a given node type of
Figure BDA0003369174130000072
Node of (2), weighting meta-path in time
Figure BDA0003369174130000073
Under the guidance of (2), obtaining a node n0At a time attribute value of t0A time-weighted meta-path sequence of (a); using a plurality of meta-paths P of different types1，P2,...,PkSetting T different time attribute values for each meta-path to construct different types of time weighted meta-paths, and finally obtaining a time weighted meta-path sequence set for each node
Figure BDA0003369174130000074
Nodes of the same type have different feature spaces for each type of node (e.g., node type of φi) Designing transformation matrices of a particular type
Figure BDA0003369174130000075
Thereby projecting the features of different types of nodes into the same feature space. The specific type of transformation matrix is based on the node type, and the transformation process is as follows:
Figure BDA0003369174130000076
wherein xiAnd
Figure BDA0003369174130000077
are respectively a node niThe original features and the projected features of (1).
And respectively aggregating the different time weighted element path information of each node to form a preliminary vector representation. As in fig. 4, for n0Time weighting ofAll neighbor information and self information in the meta-path sequence are aggregated, and the basic recursion process of the GRU aggregation module on the time weighting meta-path sequence is as follows:
Figure BDA0003369174130000078
Figure BDA0003369174130000079
wherein a is more than 0 and less than or equal to m +1,
Figure BDA00033691741300000710
representing a node niHidden state output by GRU, node niIs node n0And m +1-a hop neighbor nodes on the time weighted element path sequence.
Figure BDA0003369174130000081
The formula is expressed as:
Figure BDA0003369174130000082
Figure BDA0003369174130000083
Figure BDA0003369174130000084
Figure BDA0003369174130000085
wherein A isz，Ar，AhAnd Bz，Br，BhIs a parameter, zi，riIs an update gate and a forget gate, and the value ranges thereof are both [0,1]；
Figure BDA00033691741300000816
Is an element-by-element multiplication.
After m +1 times of propagation on the structural neighborhood of the time-weighted element path sequence, the node n0Weighting meta-paths in time
Figure BDA0003369174130000086
The state vector output above can be obtained by the following equation:
Figure BDA0003369174130000087
wherein
Figure BDA0003369174130000088
(d': represents a dimension).
When the relative time coding technology is used as a time coder to code the time of the node sequence, a group of fixed sine functions are defined as time offset to code the time of each sequence:
Figure BDA0003369174130000089
Figure BDA00033691741300000810
Figure BDA00033691741300000811
where t is the time attribute value of the time weighted meta-path sequence,
Figure BDA00033691741300000812
the time characteristics of the nodes are subjected to T-Linear fine tuning transformation.
Aggregating the temporal features and the structural semantics into the aforementioned node feature vector, namely:
Figure BDA00033691741300000813
wherein
Figure BDA00033691741300000814
Is a temporal feature or a semantic structural feature of the node. Taking it as the input of a bidirectional gated cyclic unit, and then applying a Mean function to the node n0All state features are averaged to obtain a time-coded vector representation of the node on a time-weighted element path sequence
Figure BDA00033691741300000815
The method comprises the following steps of interacting characteristic information of different sequences at different times by using a Bi-GRU with an attention mechanism to obtain a final representation of a node, and specifically comprising the following steps:
s61: and (3) interacting the characteristic information at different times by using the Bi-GRU to simulate the evolution of the network:
Figure BDA0003369174130000091
wherein { o1,o2,....,ok|T|Is the set of Bi-GRU state vector sequences;
s62: an attention mechanism is referenced to capture important feature vectors with an attention weight alphaiThe calculation formula of (a) is as follows:
Figure BDA0003369174130000092
wherein alpha isiHigher, oiThe more important. LeakReLU is an activation function, a belongs to R1×2d’Is an attention parameter;
s63: using the calculated attention value to connect the node n0All state features of (a) are aggregated into a final vector representation of the node:
Figure BDA0003369174130000093
wherein u is0∈Rd’Is node n0The final vector representation of (a);
s64: the cross entropy that minimizes the true and predicted values of all the marked nodes is used as a loss function:
Figure BDA0003369174130000094
where C is the classifier parameter, ylIs an indexed set of node labels. Y islAnd ulAre the node label value and the node feature vector prediction label value. And optimizing the model by a gradient descent algorithm.
Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (10)

1. A dynamic heterogeneous network representation method based on meta-path is characterized in that a DHNR model is constructed to carry out representation learning on network nodes, the DHNR model comprises GRUs and Bi-GRUs with attention mechanisms, and the process of obtaining node representation vectors by the DHNR model specifically comprises the following steps:
s1: the time for establishing the link of the network node is used as the link weight and is reserved in the network to construct a dynamic heterogeneous network;
s2: sampling different time weighting element path sequences from the network according to the time weighting element paths;
s3: preprocessing a network to obtain an initial vector of each node, inputting each time weighted meta-path sequence and the vector thereof, and aggregating information of the network node sequence of each meta-path through a GRU (general packet Unit);
s4: a relative time coding technology is adopted as a time coder to code the time of the node sequence;
s5: acquiring deep-level characteristic information of the node sequence by adopting a Bi-GRU method, and aggregating time characteristics and structural characteristics;
s6: and interacting the characteristic information of different sequences at different times by using a Bi-GRU with an attention mechanism to obtain a final representation of the node.
2. The meta-path-based dynamic heterogeneous network representation method of claim 1, wherein the dynamic heterogeneous network is represented by G ═ V, E, T, where V represents a set of nodes, E represents a set of links, T represents a set of times on links, and a mapping function between nodes and node types in the dynamic heterogeneous network is as follows
Figure FDA0003369174120000011
A represents a node type set; e → R, R represents the link type set, in the dynamic heterogeneous network, | A | + | R | > 2, a time link in the dynamic heterogeneous network is represented as (i, j, t) and (i, j, t) ∈ E, the time link (i, j, t) represents that the node i is connected to the node j at the moment t, | A | represents the number of the node types in the node type set, and | R | represents the number of the link types in the link type set.
3. The meta-path based dynamic heterogeneous network representation method of claim 1, wherein sampling different time weighted meta-path sequences from the network according to the time weighted meta-path comprises the steps of:
constructing a time weighting path, namely, obtaining a time weighting meta-path corresponding to the type meta-path for one meta-path under the guidance of a time attribute value;
acquiring a time-weighted meta-path sequence set, namely acquiring time-weighted meta-paths of each type of meta-paths under different time attribute values of T, and representing the time-weighted meta-path sequence set as
Figure FDA0003369174120000021
Wherein k is the type number of the element path; t is a time attribute value set, | T | is the number of time attributes in the time attribute value set;
Figure FDA0003369174120000022
represents a time-weighted meta-path of the kth type under the guide of the | T | time attribute value.
4. The meta-path based dynamic heterogeneous network representation method of claim 1, wherein preprocessing the vector of the network node comprises:
setting a transformation matrix for each type of nodes, so that each type of nodes is projected into the same feature space through the transformation matrix, wherein the feature space contains features in all types of node feature spaces, and the node type is phiiNode n ofiThe projection process is represented as:
Figure FDA0003369174120000023
wherein x isiIs a node niThe original characteristics of (a);
Figure FDA0003369174120000024
is a node niThe projection feature of (a);
Figure FDA0003369174120000025
is node type is phiiA transformation matrix of the node of (1).
5. A meta-path based dynamic heterogeneous network representation method according to claim 1, wherein aggregating information of the sequence of network nodes of each meta-path through GRUs comprises the following procedures:
aggregating the different time weighted element path information of the node to form a primary vector representation, and then the node n0The polymerization process of (a) comprises:
Figure FDA0003369174120000026
Figure FDA0003369174120000027
wherein the content of the first and second substances,
Figure FDA0003369174120000028
is a node niThe projection feature of (a); phi (n)i) Is a node niNode type of niIs a node n0M +1-a hop neighbor nodes on the time weighted element path sequence;
Figure FDA0003369174120000029
the hidden state of the m +1 hop neighbor node of the target node in the sequence;
Figure FDA00033691741200000210
representing a node niThe hidden state of the a-th layer is output through the GRU, a is more than 0 and less than or equal to 1m +, and m is the number of sequence links; a. them+1The type of the m +1 hop neighbor of the target node; a. them+1-aIs m +1-a hop neighbor type;
Figure FDA00033691741200000211
hidden information indicating that a-1 layer hidden layer is output
Figure FDA00033691741200000212
Node n with input GRUiProjection feature of
Figure FDA0003369174120000031
Polymerization is carried out.
6. The meta-path based dynamic heterogeneous network representation method according to claim 5, wherein the hidden information outputted from the a-1 layer hidden layer
Figure FDA0003369174120000032
Node n with input GRUiProjection feature of
Figure FDA0003369174120000033
Carrying out the polymerization, namely:
Figure FDA0003369174120000034
Figure FDA0003369174120000035
Figure FDA0003369174120000036
Figure FDA0003369174120000037
wherein A isz、Ar、AhAnd Bz、Br、BhIs a model parameter of the GRU; z is a radical ofi、riIs an update gate and a forget gate in the GRU, and the value ranges thereof are both [0,1]；
Figure FDA00033691741200000318
Representing element-by-element multiplication;
Figure FDA0003369174120000038
for the input vector and the previous timeThe state vector carries out a new state after forgetting; sigma is an activation function;
Figure FDA0003369174120000039
is the final state processed by one GRU unit.
7. A meta-path based dynamic heterogeneous network representation method according to claim 1, wherein the time of the node sequences is encoded, that is, a fixed set of sine functions is defined as time offsets to encode the time of each sequence, and the encoding process comprises:
Figure FDA00033691741200000310
Figure FDA00033691741200000311
Figure FDA00033691741200000312
wherein Base (t), Base (t,2i) and Base (t,2i +1) are time bias functions,
Figure FDA00033691741200000313
for the time vector at time T, T _ Linear is a Linear fine tuning function, and d is a vector dimension.
8. The meta-path based dynamic heterogeneous network representation method of claim 1, wherein aggregating temporal features and structural features comprises:
Figure FDA00033691741200000314
wherein the content of the first and second substances,
Figure FDA00033691741200000315
represents a feature obtained by polymerizing a temporal feature and a structural feature;
Figure FDA00033691741200000316
is the characteristic of the node, when i is st, it represents the time characteristic of the node, when i is RT (t), it represents the semantic characteristic of the node; mean () represents a function of averaging the array, averaging the hidden states of the bidirectional GRU;
Figure FDA00033691741200000317
representing a representation of a bi-directional GRU.
9. The meta-path-based dynamic heterogeneous network representation method of claim 1, wherein the step of using Bi-GRUs with attention mechanism to interact with different sequences of feature information at different times comprises the steps of:
and (3) interacting the characteristic information at different times by using the Bi-GRU to simulate the evolution of the network, namely:
Figure FDA0003369174120000041
reference attention mechanism to capture important feature vectors, where node niAttention weight αiThe calculation formula of (a) is as follows:
Figure FDA0003369174120000042
using the calculated attention value to connect the node n0All state features of (a) are aggregated into a final vector representation of the node:
Figure FDA0003369174120000043
wherein O ═ { O ═ O1,o2,....,ok|T|Is the set of state vector sequences of Bi-GRU, ok|T|Represents the k × T state vectors; LeakReLU is an activation function, a belongs to R1×2d'For the attention parameter, d is the representative dimension of the vector, and the specific value is determined by experiments.
10. The meta-path-based dynamic heterogeneous network representation method according to claim 1, wherein after the structural features of the sequence are obtained through the GRU, the semantic and temporal features of the network nodes are obtained through the Bi-GRU with attention mechanism, in the representation training process of the network, the cross entropy that minimizes the true and predicted values of all the labeled nodes is adopted as a loss function, and the model is optimized through a gradient descent algorithm, and the loss function is represented as:
Figure FDA0003369174120000044
wherein the content of the first and second substances,
Figure FDA0003369174120000045
representing a loss function; l represents a label; c is a classifier parameter, ylIs an index set of node labels; y islIs a node tag value, ulLabel values are predicted for the node feature vectors.
CN202111393567.6A 2021-11-23 2021-11-23 Dynamic heterogeneous network representation method based on meta-path Pending CN114124729A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111393567.6A CN114124729A (en) 2021-11-23 2021-11-23 Dynamic heterogeneous network representation method based on meta-path

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111393567.6A CN114124729A (en) 2021-11-23 2021-11-23 Dynamic heterogeneous network representation method based on meta-path

Publications (1)

Publication Number Publication Date
CN114124729A true CN114124729A (en) 2022-03-01

Family

ID=80439929

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111393567.6A Pending CN114124729A (en) 2021-11-23 2021-11-23 Dynamic heterogeneous network representation method based on meta-path

Country Status (1)

Country Link
CN (1) CN114124729A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117131938A (en) * 2023-10-26 2023-11-28 合肥工业大学 Dynamic implicit relation mining method and system based on graph deep learning

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110555050A (en) * 2018-03-30 2019-12-10 华东师范大学 heterogeneous network node representation learning method based on meta-path
WO2020199524A1 (en) * 2019-04-02 2020-10-08 长安大学 Method for matching ride-sharing travellers based on network representation learning
CN113191482A (en) * 2021-04-19 2021-07-30 安徽农业大学 Heterogeneous graph neural network representation method based on element path

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110555050A (en) * 2018-03-30 2019-12-10 华东师范大学 heterogeneous network node representation learning method based on meta-path
WO2020199524A1 (en) * 2019-04-02 2020-10-08 长安大学 Method for matching ride-sharing travellers based on network representation learning
CN113191482A (en) * 2021-04-19 2021-07-30 安徽农业大学 Heterogeneous graph neural network representation method based on element path

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
袁铭: "基于变分推断和元路径分解的异质网络表示方法", 《上海交通大学学报》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117131938A (en) * 2023-10-26 2023-11-28 合肥工业大学 Dynamic implicit relation mining method and system based on graph deep learning
CN117131938B (en) * 2023-10-26 2024-01-19 合肥工业大学 Dynamic implicit relation mining method and system based on graph deep learning

Similar Documents

Publication Publication Date Title
Tan et al. RelativeNAS: Relative neural architecture search via slow-fast learning
CN114816997B (en) Defect prediction method based on graph neural network and bidirectional GRU feature extraction
CN108549692A (en) The method that sparse multivariate logistic regression model under Spark frames classifies to text emotion
CN110851654A (en) Industrial equipment fault detection and classification method based on tensor data dimension reduction
CN113688253B (en) Hierarchical perception temporal knowledge graph representation learning method
CN111310068A (en) Social network node classification method based on dynamic graph
Chen et al. Generative dynamic link prediction
Yang et al. Mutualnet: Adaptive convnet via mutual learning from different model configurations
CN113554100A (en) Web service classification method for enhancing attention network of special composition picture
CN114124729A (en) Dynamic heterogeneous network representation method based on meta-path
CN116186390A (en) Hypergraph-fused contrast learning session recommendation method
CN115983341A (en) Node classification method based on relation aggregation hypergraph
Le et al. GCRINT: network traffic imputation using graph convolutional recurrent neural network
CN116306780B (en) Dynamic graph link generation method
CN112396166A (en) Graph convolution neural network training method and device based on mixed granularity aggregator
CN117056226A (en) Cross-project software defect number prediction method based on transfer learning
CN110020379B (en) Link prediction method based on deep dynamic network embedded representation model
CN113537613B (en) Temporal network prediction method for die body perception
CN113360772B (en) Interpretable recommendation model training method and device
Manoju et al. Conductivity based agglomerative spectral clustering for community detection
CN114792384A (en) Graph classification method and system integrating high-order structure embedding and composite pooling
CN114254214A (en) Traffic prediction method and system based on space-time hierarchical network
Chen et al. Image Recognition Based on Convolution Neural Network with Small Data Set
Ko et al. Deep model compression and inference speedup of sum–product networks on tensor trains
He et al. Overview of Key Performance Indicator Anomaly Detection

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination