CN113191482A - Heterogeneous graph neural network representation method based on element path - Google Patents

Heterogeneous graph neural network representation method based on element path Download PDF

Info

Publication number
CN113191482A
CN113191482A CN202110416061.6A CN202110416061A CN113191482A CN 113191482 A CN113191482 A CN 113191482A CN 202110416061 A CN202110416061 A CN 202110416061A CN 113191482 A CN113191482 A CN 113191482A
Authority
CN
China
Prior art keywords
path
meta
neighbor
node
heterogeneous
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
CN202110416061.6A
Other languages
Chinese (zh)
Inventor
吴国栋
范维成
汪菁瑶
涂立静
李景霞
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.)
Anhui Agricultural University AHAU
Original Assignee
Anhui Agricultural University AHAU
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 Anhui Agricultural University AHAU filed Critical Anhui Agricultural University AHAU
Priority to CN202110416061.6A priority Critical patent/CN113191482A/en
Publication of CN113191482A publication Critical patent/CN113191482A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/40—Information retrieval; Database structures therefor; File system structures therefor of multimedia data, e.g. slideshows comprising image and additional audio data
    • G06F16/45—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/95—Retrieval from the web
    • G06F16/953—Querying, e.g. by the use of web search engines
    • G06F16/9535—Search customisation based on user profiles and personalisation

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Databases & Information Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a heterogeneous graph neural network representation method based on a meta-path, which comprises the following steps: step 1, determining a plurality of meta-paths of a target node, and grouping the meta-paths according to types after sampling different types of neighbor nodes of the target node; step 2, respectively carrying out feature extraction, node initial heterogeneous content coding and feature aggregation on the neighbor nodes obtained in the step 1 to obtain heterogeneous neighbor information; step 3, respectively aggregating heterogeneous neighbor information of neighbor nodes generated in each meta-path, and obtaining corresponding embedded representation; and 4, optimizing the embedded representations in the meta paths after merging based on the attention mechanism again to generate the final embedded representation of the target node.

Description

Heterogeneous graph neural network representation method based on element path
Technical Field
The invention relates to the field of heterogeneous information networks, in particular to a heterogeneous graph neural network representation method based on a meta-path.
Background
Representation learning in heterogeneous information networks is to find a meaningful vector representation for each node, so as to facilitate implementation of downstream applications (such as link prediction and personalized recommendation). The heterogeneous graph contains a large amount of structural relationship information, such as unstructured content of each node. Since heterogeneous attributes or contents associated with each node need to be considered when processing a heterogeneous graph, and heterogeneous structure information composed of multiple types of nodes and edges needs to be merged, progress in heterogeneous graph processing is not as easy as in a homogeneous graph.
The existing heterogeneous information network representation method mostly uses meta-path aggregation node information, and the neighbor node information is aggregated on meta-paths with different semantics and then further integrated. The existing method for aggregating node information by meta-path faces three problems which need to be solved urgently:
(1) the neighbors of many nodes in a heterogeneous graph contain only individual types, not all.
(2) A single node in a heterogeneous graph may carry many unstructured heterogeneous contents, such as: attributes, text, or pictures.
(3) And different attributes of the neighbors in the heterogeneous graph have different importance on node embedding generation.
Disclosure of Invention
The invention aims to provide a heterogeneous graph neural network representation method based on meta-paths so as to solve the problems in the prior art.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows:
a heterogeneous graph neural network representation method based on meta-paths comprises the following steps:
step 1, determining a plurality of meta-paths of a target node, then respectively sampling different types of neighbor nodes of the target node in each meta-path by a random walk method, and grouping the neighbor nodes sampled in each meta-path according to types;
step 2, respectively carrying out feature extraction, node initial heterogeneous content coding and feature aggregation on the neighbor nodes in each meta-path obtained in the step 1, thereby respectively obtaining heterogeneous neighbor information based on types of the target node in each meta-path;
step 3, respectively aggregating heterogeneous neighbor information based on types generated in each meta-path, and aggregating the heterogeneous neighbor information based on types generated in each meta-path in the step 2 by using an attention mechanism to obtain embedded representations of all types of neighbor nodes corresponding to target nodes in the single meta-path;
and 4, optimizing the embedded representations of all types of neighbor nodes of the target node in each meta-path obtained in the step 3 after merging based on the attention mechanism, and generating the final embedded representation of the target node.
In step 1 of the present invention, the random walk method is a random walk strategy based on restart, and the sampling process in each meta path by the random walk method is as follows:
starting from the target node, iterating and randomly walking to other nodes in the corresponding unit path or returning to the target node according to probability until a neighbor node sequence with a fixed distance from the target node is successfully sampled.
In step 1 of the present invention, the process of grouping the neighbor nodes sampled in each meta-path by type is as follows:
in the single element path, for each type t of neighbor node, the first k nodes are selected from the sampled neighbor node sequence and are used as a set of type t-related neighbor sets of the target node.
In step 2 of the invention, firstly, extracting the attribute characteristics of each neighbor node in the corresponding unit path; then selecting different coding modes for a plurality of neighbor nodes in the single element path according to different attribute characteristics of the neighbor nodes for coding, and selecting the same coding mode for the neighbor nodes with the same attribute characteristics, thereby obtaining the node initial heterogeneous content codes of each neighbor node; and finally, capturing the depth feature interaction information of each neighbor node in the unit path by using a neural network LSTM, and performing feature aggregation on each neighbor node in the unit path based on the captured depth feature interaction information.
In step 3 of the invention, firstly using the heterogeneous neighbor information of each type of neighbor node of the Bi-LSTM aggregation target node in a single element path; and finally, aggregating heterogeneous neighbor information based on types in the corresponding unit path by using an attention mechanism to obtain embedded representations of all types of neighbor nodes of the target node in the corresponding unit path.
In step 4 of the method, embedded representations in each element path are combined based on an attention mechanism, and then are optimized through normalization processing.
The invention also comprises a step 5, wherein the step 5 is as follows:
and (4) inputting the final embedded representation of the target node obtained in the step (4) into a loss function as an input quantity, and obtaining an optimized model parameter of the final embedded representation of the target node through training.
The loss function in step 5 of the invention adopts a cross entropy loss function.
The method generates the target node embedded representation by sampling, coding, aggregating and other operations on the neighbors of the target node in the meta-path. Compared with the prior art, the invention has the advantages that:
1. the invention fully considers various attribute information of the nodes in the heterogeneous graph, and all the information such as attributes, texts, pictures and the like carried by the nodes are contained in the initial embedded generation representation of the nodes, so that the embedded representation is further enriched.
2. In the overall consideration of the heterogeneous graph, the method adopts the neighbors generated based on the meta-path for aggregation, the meta-path has rich semantics, the meaning emphasis points expressed by different meta-paths are different, and the meta-path can be used for aggregating the neighbor nodes with higher quality and meaning.
3. The invention adopts a plurality of element paths to carry out random walk, greatly increases the breadth and richness of node information, and divides different influence degrees of each element path by using an attention mechanism, so that the finally generated node information is more comprehensive.
Drawings
FIG. 1 is a flow chart of initial node embedding representation.
Fig. 2 is a flowchart of the same type node aggregation.
FIG. 3 is a diagram of an attention model.
Detailed Description
The invention is further illustrated with reference to the following figures and examples.
The invention relates to a heterogeneous graph neural network representation method based on a meta-path, which is characterized in that a restart type random walk strategy is utilized to select all types of neighbors of a target node based on the meta-path, and an LSTM is utilized to generate initial representation of the nodes for the neighbor nodes according to different carried information. According to different types of neighbor nodes, the Bi-LSTM is used for aggregating the node information of the same type to generate neighbor representation based on the type, and then an attention mechanism is used according to the influence degree difference of different types of nodes on a target node to generate node representation based on a single element path. Finally, a final node representation is generated among the multiple element paths by using an attention mechanism, and the generated representation is substituted into a trained loss function to optimize parameter configuration. The invention specifically comprises the following steps:
step 1, determining a plurality of meta-paths of a target node, then respectively sampling different types of neighbor nodes of the target node in each meta-path by a random walk method, and grouping the sampled neighbor nodes in each meta-path according to types.
In step 1, the random walk method is based on a restarted random walk strategy, and based on the random walk method, the neighbor nodes of the sampling target node are randomly walked in each meta-path, and the sampling process in each meta-path by the random walk method is as follows:
starting from a target node V ∈ V, wherein V is a node set in a heterogeneous graph, and iterating and randomly walking to other nodes or returning to a target node in a corresponding unit path until a neighbor node sequence with a fixed distance from the target node is successfully sampled. Meanwhile, the sampling number of the neighbor nodes of different types is limited during random walk so as to ensure that all types can be sampled.
The process of grouping the sampled neighbor nodes in each meta-path according to types in step 1 is as follows:
in the single element path, for each type t of neighbor node, the first k nodes are selected from the sampled neighbor node sequence and are used as a set of type t-related neighbor sets of the target node.
And 2, respectively carrying out feature extraction, node initial heterogeneous content coding and feature aggregation on the neighbor nodes in each meta-path obtained in the step 1, thereby respectively obtaining heterogeneous neighbor information of the neighbor nodes based on types of the target node in each meta-path. FIG. 1 shows a flow chart of initial node embedding;
the information carried by each node in the heterogeneous is not completely the same, different pre-training modes are selected according to different information carried by the nodes, for example, the information has attributes, texts and pictures, and one-hot, word2vec and CNN can be respectively used as the pre-training methods. The resulting representations are then placed into the fully-connected layers, respectively, before the LSTM is used to generate the node initial representation.
In step 2 of the invention, the attribute characteristics of each neighbor node in the corresponding unit path are extracted.
And then selecting different coding modes for a plurality of neighbor nodes in the single element path according to different attribute characteristics of the neighbor nodes for coding, and selecting the same coding mode for the neighbor nodes with the same attribute characteristics, thereby obtaining the node initial heterogeneous content codes of each neighbor node. Is provided with
Figure BDA0003025920850000041
Representing the i-th level heterogeneous content of the target node V ∈ V, where dfRepresenting the dimension of the initial encoded vector, x being different according to the attributes of the different levelsiDifferent pre-training approaches may be used.
And finally, capturing the depth feature interaction information of each neighbor node in the unit path by using a neural network LSTM, and performing feature aggregation on each neighbor node in the unit path based on the captured depth feature interaction information. The initial content encoding of the meta-path seed node v is calculated as follows:
Figure BDA0003025920850000042
in the formula (1), the first and second groups,
Figure BDA0003025920850000051
(d is the embedding dimension),
Figure BDA0003025920850000052
is provided with a parameter thetaxAll-connected neural networks of (1), CvAnd heterogeneous information of any node v in the diagram, such as the attribute, text, picture characteristic and the like of the node. The LSTM calculation is as follows:
Figure BDA0003025920850000053
Figure BDA0003025920850000054
Figure BDA0003025920850000055
Figure BDA0003025920850000056
Figure BDA0003025920850000057
Figure BDA0003025920850000058
Figure BDA0003025920850000059
is the output hidden state of the ith layer content of the node,
Figure BDA00030259208500000510
is a learnable parameter, z, f, o, c are subscripts corresponding to the forgetting gate, the input gate, the output gate and the memory cell, zi，fiAnd oiA forgetting gate vector, an input gate vector and an output gate vector of the ith layer content feature of the node are respectively.
As shown in fig. 2 and 3, step 3, aggregating the heterogeneous neighbor information of the neighbor nodes based on types generated in each meta-path, and obtaining the embedded representations of all types of neighbor nodes corresponding to the target node in the single meta-path based on the heterogeneous neighbor information aggregated in each meta-path by using an attention mechanism.
In the target node neighbor sampling process, the invention selects the neighbors with fixed length for each type of neighbors by using a restart type random walk strategy, the target node has a plurality of neighbor nodes under the same type, and the information of the plurality of neighbor nodes in the same type is aggregated by using Bi-LSTM, as shown in figure 2. For multiple types of neighbors sampled by the current meta-path, multiple types of representations are integrated by using an attention mechanism according to different importance of the multiple types of neighbors, and a target node representation of the current meta-path is generated, as shown in fig. 3.
In step 3, first use Bi-LSTM f in the unit path2 tAggregating heterogeneous neighbor information of neighbor nodes of each type t of the target node v to generate a type-based neighbor embedded representation of the target node v, wherein the calculation formula is as follows:
Figure BDA0003025920850000061
Nt(v) is the t-th type neighbor node set in the neighbor sampled by the target node v, and v' is the t-th type neighbor node instance of the target node v. f. of1(v') and f of formula (1)1(v) The definitions are the same. Here LSTM calculation formula and formulaThe formula (1) is the same.
And then, the neural network Bi-LSTM is used for aggregating heterogeneous neighbor information of all types of neighbor nodes in a corresponding single-element path, parameters in the Bi-LSTM are not shared, and the Bi-LSTM algorithm parameters used by different types of neighbor sets are different.
And finally, obtaining the embedded representation of all types of neighbor nodes of the target node in the corresponding unit path by using an attention mechanism based on the heterogeneous neighbor information aggregated in the corresponding unit path.
The calculation formula is as follows:
Figure BDA0003025920850000062
in the formula (3), OVIs a set of node types, alpha, in a meta-pathv,v,αv,tRespectively representing the attention parameters corresponding to the aggregation of the target node to the self representation and the neighbor representation, f1(v) Expressing the function, f, for equation (1)2 t(v) The function is expressed for equation (2).
To further explain the learning formula of the attention parameter, let h (v) ═ f here1(v)∪f2 t(v),t∈OV) H (v) represents a heterogeneous neighbor information set based on type of the target node, including initial representation information of the target node itself. f. ofiAnd fjFor the example in set h (v), the formula for the node thus becomes:
Figure BDA0003025920850000063
Figure BDA0003025920850000064
where LeakyReLU is the activation function, uTIs the attention parameter, αv,iWhen representing heterogeneous neighbor information of the aggregation target node v, the ith heterogeneous neighbor in the heterogeneous neighbor information set represents a corresponding attention parameter, and H (v) is the aboveA defined type-based heterogeneous neighbor information set.
And 4, as shown in fig. 3, optimizing the embedded representations of all types of neighbor nodes of the target node in each meta-path obtained in the step 3 after merging based on the attention mechanism, and generating the final embedded representation of the target node.
In step 4, P types of expressions of the target node based on the meta-paths are obtained by training P meta-paths in the meta-path set, the P meta-paths are integrated by using an attention mechanism according to different degrees of influence of different meta-paths on the target node, and a final node expression is generated, wherein the calculation process is as follows:
step 4.1, training all element path sets to obtain a target node v (not clear whether the target node or the neighbor node) which is expressed as
Figure BDA0003025920850000071
Φ1,Φ2,...,ΦpFor a given meta-path set, P is the number of meta-paths;
step 4.2, calculating the importance of each element path, wherein the calculation formula is as follows:
Figure BDA0003025920850000072
w is a weight matrix, b is a bias vector, and q is a meta-path attention vector
Step 4.3, using softmax pair
Figure BDA0003025920850000073
Carrying out normalization processing, wherein P is the number of element paths:
Figure BDA0003025920850000074
step 4.4, embedding the specific semantics generated by aggregation to obtain a final representation of the target node v, wherein the calculation formula is as follows:
Figure BDA0003025920850000075
and 5, inputting the final embedded representation of the target node obtained in the step 4 into a loss function as an input quantity, and obtaining an optimized model parameter of the final embedded representation of the target node through training.
For the semi-supervised node classification model, calculating loss by using a cross entropy loss function, and optimizing model parameters, wherein the calculation formula is as follows:
Figure BDA0003025920850000076
wherein, YLIs a set of labeled nodes, Yl,ElRespectively, the embedded representation of the label and the labeled node, and C is a parameter of the classifier.
The embodiments of the present invention are described only for the preferred embodiments of the present invention, and not for the limitation of the concept and scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the design concept of the present invention shall fall into the protection scope of the present invention, and the technical content of the present invention which is claimed is fully set forth in the claims.

Claims (8)

1. A heterogeneous graph neural network representation method based on meta-path is characterized in that: the method comprises the following steps:
step 1, determining a plurality of meta-paths of a target node, then respectively sampling different types of neighbor nodes of the target node in each meta-path by a random walk method, and grouping the neighbor nodes sampled in each meta-path according to types;
step 2, respectively carrying out feature extraction, node initial heterogeneous content coding and feature aggregation on the neighbor nodes in each meta-path obtained in the step 1, thereby respectively obtaining heterogeneous neighbor information based on types of the target node in each meta-path;
step 3, respectively aggregating heterogeneous neighbor information based on types generated in each meta-path, and aggregating the heterogeneous neighbor information based on types generated in each meta-path in the step 2 by using an attention mechanism to obtain embedded representations of all types of neighbor nodes corresponding to target nodes in the single meta-path;
and 4, optimizing the embedded representations of all types of neighbor nodes of the target node in each meta-path obtained in the step 3 after merging based on the attention mechanism, and generating the final embedded representation of the target node.
2. The meta-path based heterogeneous graph neural network representation method of claim 1, wherein: in step 1, the random walk method is a random walk strategy based on restart, and the sampling process in each element path by the random walk method is as follows:
starting from the target node, iterating and randomly walking to other nodes in the corresponding unit path or returning to the target node according to probability until a neighbor node sequence with a fixed distance from the target node is successfully sampled.
3. The meta-path based heterogeneous graph neural network representation method of claim 1 or 2, wherein: in step 1, the process of grouping the sampled neighbor nodes in each meta-path according to types is as follows:
in the single element path, for each type t of neighbor node, the first k nodes are selected from the sampled neighbor node sequence and are used as a set of type t-related neighbor sets of the target node.
4. The meta-path based heterogeneous graph neural network representation method of claim 1, wherein: in step 2, firstly, extracting the attribute characteristics of each neighbor node in the corresponding unit path; then selecting different coding modes for a plurality of neighbor nodes in the single element path according to different attribute characteristics of the neighbor nodes for coding, and selecting the same coding mode for the neighbor nodes with the same attribute characteristics, thereby obtaining the node initial heterogeneous content codes of each neighbor node; and finally, capturing the depth feature interaction information of each neighbor node in the unit path by using a neural network LSTM, and performing feature aggregation on each neighbor node in the unit path based on the captured depth feature interaction information.
5. The meta-path based heterogeneous graph neural network representation method of claim 1, wherein: in the step 3, firstly using the heterogeneous neighbor information of each type of neighbor node of the Bi-LSTM aggregation target node in the single element path; and finally, aggregating heterogeneous neighbor information based on types in the corresponding unit path by using an attention mechanism to obtain embedded representations of all types of neighbor nodes of the target node in the corresponding unit path.
6. The meta-path based heterogeneous graph neural network representation method of claim 1, wherein: and 4, merging the embedded representations in each element path based on an attention mechanism, and then optimizing through normalization processing.
7. The meta-path based heterogeneous graph neural network representation method of claim 1, wherein: further comprising step 5, step 5 is as follows:
and (4) inputting the final embedded representation of the target node obtained in the step (4) into a loss function as an input quantity, and obtaining an optimized model parameter of the final embedded representation of the target node through training.
8. The meta-path based heterogeneous graph neural network representation method of claim 7, wherein: the loss function in the step 5 adopts a cross entropy loss function.
CN202110416061.6A 2021-04-19 2021-04-19 Heterogeneous graph neural network representation method based on element path Pending CN113191482A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110416061.6A CN113191482A (en) 2021-04-19 2021-04-19 Heterogeneous graph neural network representation method based on element path

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110416061.6A CN113191482A (en) 2021-04-19 2021-04-19 Heterogeneous graph neural network representation method based on element path

Publications (1)

Publication Number Publication Date
CN113191482A true CN113191482A (en) 2021-07-30

Family

ID=76977312

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110416061.6A Pending CN113191482A (en) 2021-04-19 2021-04-19 Heterogeneous graph neural network representation method based on element path

Country Status (1)

Country Link
CN (1) CN113191482A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114124729A (en) * 2021-11-23 2022-03-01 重庆邮电大学 Dynamic heterogeneous network representation method based on meta-path
CN114528221A (en) * 2022-02-24 2022-05-24 北京航空航天大学 Software defect prediction method based on heterogeneous graph neural network

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114124729A (en) * 2021-11-23 2022-03-01 重庆邮电大学 Dynamic heterogeneous network representation method based on meta-path
CN114528221A (en) * 2022-02-24 2022-05-24 北京航空航天大学 Software defect prediction method based on heterogeneous graph neural network
CN114528221B (en) * 2022-02-24 2023-04-07 北京航空航天大学 Software defect prediction method based on heterogeneous graph neural network

Similar Documents

Publication Publication Date Title
CN109299342B (en) Cross-modal retrieval method based on cycle generation type countermeasure network
CN110619081B (en) News pushing method based on interactive graph neural network
CN109857871B (en) User relationship discovery method based on social network mass contextual data
CN112765477B (en) Information processing method and device, information recommendation method and device, electronic equipment and storage medium
Zheleva et al. Higher-order graphical models for classification in social and affiliation networks
CN110781319B (en) Common semantic representation and search method and device for cross-media big data
CN113191482A (en) Heterogeneous graph neural network representation method based on element path
CN110795641A (en) Network rumor propagation control method based on representation learning
CN112182511B (en) Complex semantic enhanced heterogeneous information network representation learning method and device
CN112559764A (en) Content recommendation method based on domain knowledge graph
CN112487200B (en) Improved deep recommendation method containing multi-side information and multi-task learning
CN115964568A (en) Personalized recommendation method based on edge cache
CN114612761A (en) Network architecture searching method for image recognition
CN112488316A (en) Event intention reasoning method, device, equipment and storage medium
CN110347853B (en) Image hash code generation method based on recurrent neural network
Li et al. Variational graph autoencoder with adversarial mutual information learning for network representation learning
CN116090504A (en) Training method and device for graphic neural network model, classifying method and computing equipment
Liang et al. Generative ai-driven semantic communication networks: Architecture, technologies and applications
CN117237140A (en) Social network influence maximization method fusing graph convolution neural network and transducer
Cai et al. RI-GCN: Review-aware interactive graph convolutional network for review-based item recommendation
CN117009674A (en) Cloud native API recommendation method integrating data enhancement and contrast learning
CN116662656A (en) Movie recommendation method based on collaborative enhancement and graph annotation intention neural network
CN117171447A (en) Online interest group recommendation method based on self-attention and contrast learning
CN117033754A (en) Model processing method, device, equipment and storage medium for pushing resources
CN112685614B (en) Social media robot group rapid detection method

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
RJ01 Rejection of invention patent application after publication
RJ01 Rejection of invention patent application after publication

Application publication date: 20210730