WO2021089013A1 - 空间图卷积网络的训练方法、电子设备及存储介质 - Google Patents

空间图卷积网络的训练方法、电子设备及存储介质 Download PDF

Info

Publication number
WO2021089013A1
WO2021089013A1 PCT/CN2020/127254 CN2020127254W WO2021089013A1 WO 2021089013 A1 WO2021089013 A1 WO 2021089013A1 CN 2020127254 W CN2020127254 W CN 2020127254W WO 2021089013 A1 WO2021089013 A1 WO 2021089013A1
Authority
WO
WIPO (PCT)
Prior art keywords
network
network structure
feature
category
graph convolutional
Prior art date
Application number
PCT/CN2020/127254
Other languages
English (en)
French (fr)
Inventor
纪超杰
吴红艳
李烨
蔡云鹏
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Publication of WO2021089013A1 publication Critical patent/WO2021089013A1/zh

Links

Images

Classifications

    • 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
    • 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
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • This application belongs to the technical field of pattern recognition, and in particular relates to a training method, electronic equipment and storage medium of a spatial graph convolutional network.
  • the classification of objects corresponding to nodes in an object-relational network and the prediction of network structure characteristics of objects corresponding to connections between nodes are two common tasks. For example, use the citation relationship between the papers to predict the subject of the paper, and to complete the missing citation relationship between the papers. For another example, the co-expression relationship between the intracellular function of the protein and the co-expression relationship between the protein loss is predicted by the co-expression of the protein in the tissue.
  • the embodiments of the present application provide a training method for a spatial graph convolutional network, a method for object classification and a prediction method for network structure characteristics of an object, an electronic device and a storage medium, which can solve the above technical problems.
  • an embodiment of the present application provides a method for training a spatial graph convolutional network, including:
  • the training data packet includes network structure characteristics of multiple objects, object attribute characteristics of each object, and label categories of some of the multiple objects;
  • the network structure characteristics of each object are The association relationship between the object and other objects; among the plurality of objects, an object with a marked category is the second object, and an object without a marked category is the first object;
  • the graph convolutional network to be trained is trained according to the training data to obtain the graph convolutional network used for object classification and object network structure attribute prediction.
  • the embodiments of the present application provide an object classification and a method for predicting the network structure characteristics of the object, including:
  • the spatial domain graph convolution network is used to process the test data to obtain the classification result of the object to be predicted and the network structure attribute prediction result of the object, and the spatial domain graph convolution network is the one described in the first aspect above
  • the graph convolutional network trained by the method is the one described in the first aspect above.
  • an embodiment of the present application provides a training device for a spatial domain graph convolutional network, including:
  • the data acquisition module is used to acquire training data; wherein the training data package includes the network structure characteristics of multiple objects, the object attribute characteristics of each object, and the label categories of some of the multiple objects;
  • the network structure feature is the association relationship between the object and other objects; among the plurality of objects, the objects with the marked category are the second objects, and the objects without the marked category are the first objects;
  • the training module is used to train the graph convolutional network to be trained according to the training data to obtain the graph convolutional network used for object classification and object network structure attribute prediction.
  • an embodiment of the present application provides a device for predicting object classification and connection relationships between objects, including:
  • the test data acquisition module is used to acquire the test data of the object to be predicted
  • the prediction module uses a spatial domain graph convolutional network to process the test data to obtain the classification result of the object to be predicted and the network structure attribute prediction result between the objects.
  • the spatial domain graph convolutional network passes through the first The graph convolutional network trained by the method described in the aspect.
  • an electronic device including:
  • an embodiment of the present application provides a computer-readable storage medium, including: the computer-readable storage medium stores a computer program, and the computer program implements the first aspect and/or the first aspect when the computer program is executed by a processor. The method steps described in the second aspect.
  • the embodiments of the present application provide a computer program product, which when the computer program product runs on an electronic device, causes the electronic device to execute the method steps described in the first aspect.
  • FIG. 1 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application.
  • FIG. 2 is a schematic diagram of an object relationship network structure provided by an embodiment of the present application.
  • FIG. 3 is a schematic flowchart of a method for training a spatial domain graph convolutional network provided by an embodiment of the present application
  • FIG. 4 is a schematic flowchart of a method for training a spatial domain graph convolutional network provided by another embodiment of the present application.
  • FIG. 5 is a schematic flowchart of a method for predicting object classification and network structure characteristics of an object provided by an embodiment of the present application
  • Fig. 6 is a schematic diagram of a spatial domain graph convolutional network training device provided by an embodiment of the present application.
  • FIG. 7 is a schematic diagram of a device for predicting object classification and network structure characteristics of an object provided by an embodiment of the present application.
  • the term “if” can be construed as “when” or “once” or “in response to determination” or “in response to detecting “.
  • the phrase “if determined” or “if detected [described condition or event]” can be interpreted as meaning “once determined” or “in response to determination” or “once detected [described condition or event]” depending on the context ]” or “in response to detection of [condition or event described]”.
  • the classification of objects corresponding to nodes in an object-relational network and the prediction of network structure characteristics of objects corresponding to connections between nodes are two common tasks. For example, use the citation relationship between the papers to predict the subject of the paper, and to complete the missing citation relationship between the papers. For another example, the co-expression relationship between the intracellular function of the protein and the loss of the protein is predicted by the co-expression of the protein in the tissue.
  • the current methods all treat the two tasks of node classification and node connection relationship prediction in isolation, and there is no method to train, learn and complete these two tasks at the same time.
  • multi-task simultaneous learning can greatly reduce the computational cost of the computer, especially for complex models such as deep learning networks.
  • the input data must contain network topology information, that is, the connection relationship between nodes, but the node classification model does not directly learn and model the connection relationship, but indirectly observes the existing The accuracy benefit brought by the connection relationship to the classification result is used for model learning. Such a single observation mode does not maximize the use of input information.
  • the existing methods do not consider the classification of nodes at all.
  • this application provides a method for training spatial graph convolutional networks, object classification and prediction of object network structure features Methods of electronic equipment and storage media.
  • Figure 1 shows an electronic device D10 provided by an embodiment of the present application, including: at least one processor D100, a memory D101, and a computer program D102 stored in the memory D101 and running on the at least one processor D100
  • the processor D100 executes the computer program D102, at least one of the training method of the spatial graph convolutional network, the object classification, and the prediction method of the network structure feature of the object provided in the embodiment of the present application is implemented.
  • the above-mentioned electronic devices may be computing devices such as desktop computers, notebooks, palmtop computer servers, server clusters, distributed servers, and cloud servers.
  • the electronic device D10 may include, but is not limited to, a processor D100 and a memory D101.
  • FIG. 1 is only an example of the electronic device D10, and does not constitute a limitation on the electronic device D10. It may include more or less components than those shown in the figure, or a combination of certain components, or different components. , For example, can also include input and output devices, network access devices, and so on.
  • the so-called processor D100 may be a central processing unit (Central Processing Unit, CPU), and the processor D100 may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), and application specific integrated circuits (Application Specific Integrated Circuits). , ASIC), ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the memory D101 may be an internal storage unit of the electronic device D10, such as a hard disk or a memory of the electronic device D10.
  • the memory D101 may also be an external storage device of the electronic device D10, for example, a plug-in hard disk equipped on the electronic device D10, a smart memory card (Smart Media Card, SMC), and a secure digital (Secure Digital, SD) card, flash card (Flash Card), etc.
  • the memory D101 may also include both an internal storage unit of the electronic device D10 and an external storage device.
  • the memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program.
  • the memory D101 can also be used to temporarily store data that has been output or will be output.
  • Figure 2 shows an object relationship network diagram provided by an embodiment of the present application.
  • the nodes in the graph are objects in the actual application scenario
  • the edges in the graph are the association relationships between the objects, which correspond to the network structure attribute characteristics of the objects
  • the numbers of the nodes in the graph are the numbers of the objects corresponding to the nodes.
  • the object relationship network diagram shown in FIG. 2 corresponds to the protein association relationship network
  • the object corresponding to the node in FIG. 2 is a protein
  • the edge in FIG. 2 corresponds to the co-expression of the two proteins in the cell.
  • the object relationship network diagram shown in Fig. 2 corresponds to the document citation relationship network
  • the node in Fig. 2 corresponds to the document
  • the edge in Fig. 2 corresponds to the citation relationship between the two documents.
  • those skilled in the art can obtain object association diagrams in different fields according to the guidelines of this application to apply to the methods provided in the embodiments of this application.
  • FIG. 3 shows a schematic flow chart of a method for training a spatial domain graph convolutional network provided by an embodiment of the present application.
  • the graph convolutional network used to obtain object classification and object network structure attribute prediction is applied to the electronic device shown in FIG. 1 above.
  • Device hereinafter referred to as server
  • the method includes step S110 and step S120, and the specific implementation principles of each step are as follows:
  • training data includes network structure characteristics of multiple objects, object attribute characteristics of each object, and label categories of some of the multiple objects; the network structure of each object The characteristic is an association relationship between the object and other objects; among the plurality of objects, an object with a marked category is the second object, and an object without a marked category is the first object.
  • the server obtains training data.
  • the training data includes the network structure characteristics of multiple objects, the object attribute characteristics of each object, and the label categories of some of the multiple objects;
  • the network structure characteristic is the association relationship between the object and other objects; among the plurality of objects, the object with the marked category is the second object, and the object without the marked category is the first object. It can be understood that, in the object relationship graph represented by the graph network structure, the object corresponds to the node in the object relationship network, and the network structure feature of the object corresponds to the connection relationship between the node and other nodes in the object relationship network.
  • the object is a protein in a protein association network; the attribute feature of the object is the subspace structure of the protein; the network structure feature of the object is that each protein and other proteins are organized
  • the label category of the part of the object is the cellular function of a part of the protein of known cellular function; the first object is the protein of unknown cellular function; the second object is the protein of known cellular function.
  • the object is a document in a document citation relationship network;
  • the attribute feature of the object is a keyword of the document title;
  • the network structure feature of the object is the citation of each document and other documents Relationship;
  • the marked category of the part of the object is a document of some known document category;
  • the first object is a document of an unknown document category, and the second object is a document of a known document category.
  • the server obtains the network structure features of multiple documents in the document citation relationship network.
  • the network structure feature may be a vector of the citation relationship between the document and other documents. It can also be the adjacency matrix corresponding to the network composed of multiple documents in the document citation relationship network with documents as nodes and citation relationships between documents as edges.
  • the object attribute feature of the document is the vector of the correspondence between the keywords of the document title and the keywords in the preset dictionary.
  • the dictionary is [biology, one kind, informatics, ..., prediction, ...]; in an example article, the title of the article is "MicroRNA Prediction Research in Bioinformatics", the object attribute feature of the article is [1 ,0,1, alone,1, toast]; In another example, the document title is "A trajectory prediction algorithm based on Gaussian mixture model", then the object attribute feature of the document is [0,1,0,... ...,1, alone].
  • the categories of some objects used to train graph convolutional networks are known.
  • the training data corresponding to the protein relationship network can obtain the protein attribute feature vector through the protein subspace structure dictionary and the protein subspace structure, and through the co-expression of the two proteins in the tissue As the edge of the protein relationship network, the network structure feature of the protein in the network is obtained.
  • the network structure feature can be the adjacency matrix of the protein relationship network, or it can be the object of the protein, the vector of the relationship between the protein and other proteins .
  • those skilled in the art can also obtain training data corresponding to the social relationship network, training data corresponding to the sales relationship network, and training data in other fields according to the above examples to train and process object classification in various fields and prediction of association relationships between objects.
  • Graph convolutional network can be the training data corresponding to the social relationship network, training data corresponding to the sales relationship network, and training data in other fields according to the above examples to train and process object classification in various fields and prediction of association relationships between objects.
  • the server trains the graph convolutional network to be trained based on the aforementioned training data, such as protein relation network graph data or literature reference relation network graph data, to obtain graphs for object classification and object network structure attribute prediction Convolutional network.
  • the aforementioned training data such as protein relation network graph data or literature reference relation network graph data
  • step S120 is refined, as shown in FIG. 4, including steps S121 to S125. specific:
  • the server obtains the first connection probability between objects that have no connection relationship according to the object attribute characteristics of each object, and updates the network structure characteristic of the object according to the first connection probability. It can be understood that, except for the first iteration, the object attribute characteristics in each iteration are all the object attribute characteristics updated in the previous iteration, and the object attribute characteristics adopted in the first iteration are the original object attribute characteristics obtained in step S110.
  • the first connection probability between objects is calculated according to the updated object attribute features, and the original object graph is updated according to the first connection probability Structural features, that is, the connection relationship between the updated object and other objects can also be understood as the connection relationship that complements the original object relationship graph network structure.
  • the first connection probability is not calculated in the first iteration, that is, no connection relationship completion operation between objects is performed.
  • Node identification step S122 of K cycles (convolution) produced i.e. the polymerization feature node
  • a is a linear function which has an initial parameter
  • W 1 is The dimension reduction vector
  • ⁇ is a nonlinear transformation function, such as the sigmoid function, which is used to map the calculation result to the interval [0,1] to obtain the probability value.
  • the first connection probability between two objects that have no connection relationship after obtaining the first connection probability between two objects that have no connection relationship, sort all the first probability values from large to small, and select the first 0 probability values from large to small. It is considered that there is a connection relationship between the two objects corresponding to these 0 probability values, and the network structure characteristics of the object corresponding to the probability value are updated, that is, to complete the connection relationship between the object and other objects.
  • connection probability between two objects that have no connection relationship After obtaining the first connection probability between two objects that have no connection relationship, select the object whose probability value is greater than the first threshold value, and consider that the probability value greater than the first threshold value corresponds to There is a connection relationship between the two objects, and the network structure feature of the object corresponding to the probability value is updated, that is, the connection relationship between the object and other objects is completed.
  • S122 Acquire the aggregated feature of each object according to the updated network structure feature of each object and the object attribute feature of each object, and update the object attribute feature of each object according to the aggregated feature.
  • the server obtains the aggregated characteristics of the object attribute characteristics of each object through the Graph Sampling and Aggregating (GraphSAGE) algorithm, and updates the object attribute characteristics of the object according to the aggregated characteristics.
  • GraphSAGE Graph Sampling and Aggregating
  • the server updates the attribute characteristics of each object through the Graph Attention Network (GAT) algorithm.
  • GAT Graph Attention Network
  • the application of the GraphSAGE algorithm to update the object attribute characteristics of each object is taken as an example to illustrate the convolution process in the training process of the graph convolutional neural network.
  • each node in the graph network shown in FIG. 2 is traversed. It is understandable that the access order of the nodes is not distinguished here. It is understandable that part of the nodes shown in FIG. 2 can also be selected.
  • AGGREGATE is a vector aggregation operation
  • CONCAT is a vector splicing operation
  • k-1 represents the previous cycle step
  • h k-1 represents the node feature representation generated by the previous cycle step
  • W k is a learnable parameter in the model
  • each loop step (convolution) uses different parameters.
  • S123 Calculate a second connection probability according to the object attribute characteristics and the original network structure characteristics of each object; the second connection probability is the connection between each object and other objects sampled according to the network structure characteristics of the object Probability.
  • the server calculates the second connection probability based on the object attribute characteristics of each object and the original network structure characteristics; the second connection probability is that each object is connected to the network according to the object.
  • the connection probability of other objects selected by structural feature sampling.
  • a possible implementation manner is that other objects selected according to the network structure feature of the object may be selected according to the network structure feature of each object, and the target node in the object relationship network diagram is directly connected to the object corresponding to the object. All the first hop nodes, and sampling select I Jth hop nodes that are not directly connected to the node; where I is a positive integer greater than 0, and J is a positive integer greater than 1.
  • other objects selected according to the network structure feature sampling of the object may be sampled according to the number of hops. For example, the larger the number of hops J, the smaller the number of nodes I sampled.
  • other objects selected according to the network structure feature sampling of the object may be a positive integer greater than 0 for the number of hops J, that is to say, the number of nodes of the first hop is also sampled.
  • the original network structure feature is the network structure feature obtained through step S110.
  • step S123 the selection of nodes in step S123 can be selected and adjusted by those skilled in the art according to actual needs under the teaching of this application.
  • the above method of node selection is a non-limiting exemplary description, and does not constitute a reference to this application. The limit.
  • the server obtains the predicted category of each object according to the updated object attribute characteristics of each object; according to the predicted category of the second object, the label category of the second object, and the second object
  • the connection probability, the original network structure characteristics of each object and the loss function adjust the parameters of the graph convolutional network.
  • the parameters of the graph convolutional network can be adjusted through the back propagation gradient descent method.
  • the predicted category of each object is obtained according to the updated object attribute characteristics of each object through the classifier.
  • the classifier can be a two-layer fully connected neural network or a multi-layer neural network with more than two layers. Network or other machine learning classification models.
  • the following loss function is used to calculate the reward and punishment value to adjust the parameters of the graph convolutional network
  • Loss Loss link +Loss cls
  • Loss link is the difference between the second connection probability and the original network structure feature of each object; Loss cls is the difference between the predicted category of the second object and the mark category of the second object. It is understandable that the above difference may be the absolute value of the difference, the average difference, the variance, etc., and those skilled in the art can process the determination of the difference in the loss function according to actual needs, which will not be repeated here.
  • the first loss function has a first loss function coefficient
  • the second loss function has a second loss function coefficient.
  • the coefficient of the first loss function and the coefficient of the second loss function are used to determine the Descriptive convolutional network focuses on object classification or object network structure feature prediction.
  • the coefficient for the unbiased task is ⁇
  • the coefficient for the favored task is 1- ⁇ . For example, if a protein association relationship network is known, most of the connection relationships between protein nodes are known, but the function of the protein in the cell is mostly unknown. At this time, the graph convolution network needs to be trained to focus on the node classification task. , Then the loss function is,
  • Loss ⁇ Loss link +(1- ⁇ )Loss cls
  • an annealing algorithm is used to calculate the coefficient of the first loss function and the coefficient of the second loss function according to the first reward and punishment value; or, the annealing algorithm is used according to the second reward and punishment value. Calculate the coefficient of the first loss function and the coefficient of the second loss function.
  • An annealing strategy is defined as follows, which means that as the training iteration process, temp t becomes smaller and smaller, where t is the number of training iterations of the graph convolutional network,
  • this method further adds a penalty to the phenomenon of high errors, so that the prediction error and the priority of the task that are not favored reach a balance.
  • Loss aux means that the loss is not caused by the biased training task. These two factors are combined to get: the coefficient ⁇ of the training task that is not favored and the coefficient 1- ⁇ of the training task that is favored.
  • the annealing mechanism is introduced to balance the current task focus, so as to ensure that the graph convolutional network focuses on tasks while ensuring the performance of the model during the process of learning multiple tasks.
  • the iterative training end condition may be that the preset number of iterations is reached, or the loss function may converge below the preset threshold.
  • the iterative training end condition may be that the preset number of iterations is reached, or the loss function may converge below the preset threshold.
  • Those skilled in the art can set the iteration end condition according to actual needs. If the iterative training end condition is not met, return to the step of obtaining the first connection probability between objects without a connection relationship according to the object attribute characteristics of each object to perform iterative training on the graph convolutional network; if the iterative training end condition is met, the treatment is ended Training of the trained graph convolutional network.
  • the graph convolutional network trained in the embodiment of the present application introduces a single-model multi-task graph convolutional network training method to simultaneously train the classification tasks of the objects corresponding to the nodes in the object relational network, and the inter-node
  • the network structure feature prediction task of the object corresponding to the connection relationship, and the obtained trained graph convolutional network can realize the object classification task and the object network structure feature prediction task at the same time, thereby making full use of the computing power of the computing device and improving resource utilization Rate, reducing costs.
  • the predicted network structure features are used for the connection completion of the object relational network, and the result of the connection completion is used as additional new input information to be trained together with the object classification task, which can improve the prediction of node classification. Accuracy and prediction efficiency.
  • FIG. 5 shows a method for predicting object classification and network structure features of objects provided by an embodiment of the present application, which can be implemented by the above-mentioned electronic device (hereinafter referred to as a server) shown in FIG. 1 through software/hardware.
  • the method includes steps S210 to S220.
  • the specific implementation principles of each step are as follows:
  • the test data includes network structure characteristics and object attribute characteristics of each object in the object relationship network where the object to be predicted is located.
  • the network structure feature of the object in the object relationship network where the object to be predicted is located is the relationship between each protein in the protein association relationship network where the protein to be predicted is located and other proteins in the organization.
  • Co-expression; the object attribute is characterized by the subspace structure of the protein.
  • the network structure feature of the object in the object relationship network where the object to be predicted is located is the citation relationship of each document in the document citation relationship network where the document to be predicted is located with other documents;
  • the object attribute feature is the keyword of the document title.
  • Those skilled in the art can also obtain the training data corresponding to the social relationship network, the prediction data corresponding to the sales relationship network, and the prediction data in other fields according to the above examples, and use graph convolutional networks to process the classification of objects in various fields and the prediction of association relationships between objects. task.
  • S220 Use the spatial domain graph convolutional network obtained by the method shown in FIG. 3 to process the test data to obtain the classification result of the object and the prediction result of the network structure attribute between the objects.
  • the server uses the spatial domain graph convolutional network trained by the method shown in Figure 3 to process the test data to obtain the classification result of the object and the network structure attributes between the objects forecast result.
  • the aggregated feature of each object is obtained according to the network structure feature and object attribute feature of each object, and the object attribute feature of each object is updated according to the aggregated feature; according to the network structure of each object Feature and the object attribute feature updated by each object to obtain the predicted connection probability of each object with other objects; update the network structure feature of the object according to the predicted connection probability; obtain the object attribute feature after updating the object to be predicted The predicted category of the predicted object.
  • the same object attribute feature aggregation operation as the aforementioned graph convolutional network training method is used.
  • the GraphSAGE algorithm is used to aggregate the object attribute features of the target node corresponding to each object and all neighbor nodes corresponding to the target node. Take the operation as an example.
  • the parameters refer to the description in the above embodiment,
  • the object attribute characteristics of the object are identified, and the classification result of the object to be predicted is obtained.
  • the predicted connection probability between each object (node) without a connection relationship is obtained according to the following formula, and the network structure feature of the object is updated according to the predicted connection probability, that is, the graph network connection relationship of the object to be predicted is complemented,
  • the parameters refer to the description in the above-mentioned embodiment.
  • connection relationship is to update the network structure characteristics of the object corresponding to the probability value, that is, to complete the connection relationship between the object and other objects.
  • the prediction results of the object classification and the connection relationship between the objects in the object relationship network can be obtained at the same time, thereby saving computing power of computing equipment, improving efficiency, and reducing cost.
  • FIG. 6 shows a training device for the spatial domain graph convolutional network provided by an embodiment of the present application, including:
  • the data acquisition module M110 is used to acquire training data; wherein the training data packet includes the network structure characteristics of multiple objects, the object attribute characteristics of each object, and the label category of some of the multiple objects; each object The network structure feature of is the association relationship between the object and other objects; among the multiple objects, the object with the marked category is the second object, and the object without the marked category is the first object.
  • the training module M120 is used to train the graph convolutional network to be trained according to the training data to obtain the graph convolutional network used for object classification and object network structure attribute prediction.
  • the training module M120 also includes the following sub-modules:
  • connection completion module M121 is configured to obtain a first connection probability between objects that have no connection relationship according to the object attribute characteristics of each object, and update the network structure characteristic of the object according to the first connection probability;
  • the feature aggregation module M122 is configured to obtain the aggregate feature of each object according to the updated network structure feature of each object and the object attribute feature of each object, and update the object attribute feature of each object according to the aggregate feature;
  • the sampling connection calculation module M123 calculates a second connection probability according to the object attribute characteristics of each object and the original network structure characteristics; the second connection probability is a sampling selection for each object and the network structure characteristics of the object The connection probability of other objects;
  • the parameter adjustment module M124 obtains the predicted category of each object according to the updated object attribute characteristics of each object; according to the predicted category of the second object, the label category of the second object, the second connection probability, and each object The original network structure characteristics and loss function of the adjustment of the parameters of the graph convolutional network;
  • the iterative judgment module M125 is used to judge the end condition of the iteration. If the end condition of the iteration is not satisfied, return to the step of obtaining the first connection probability between objects without connection relationship according to the object attribute characteristics of each object to iterate the graph convolutional network Training until the training of the graph convolutional network satisfies the iteration end condition, and the graph convolutional network used for object classification and object network structure attribute prediction is obtained.
  • the loss function includes a first loss function and a second loss function.
  • the parameter adjustment module M124 is configured to obtain the predicted category of each object according to the updated object attribute characteristics of each object; according to the predicted category of the second object, the label category of the second object, and the second object
  • the connection probability, the original network structure characteristics of each object, and the loss function to adjust the parameters of the graph convolutional network also include:
  • the first reward and punishment value determining module M1241 is configured to determine the first reward and punishment value according to the predicted category of the second object, the marked category of the second object, and the first loss function;
  • the second reward and punishment value determining module M1241 is configured to determine a second reward and punishment value according to the second connection probability, the original network structure characteristics of each object, and the second loss function;
  • the parameter adjustment submodule M1243 adjusts the parameters of the graph convolutional network according to the first reward and punishment value and the second reward and punishment value.
  • the parameter adjustment module M124 is further configured to determine, according to the coefficients of the first loss function and the coefficients of the second loss function, that the graph convolutional network focuses on object classification or object network Structural feature prediction.
  • the parameter adjustment module M124 is further configured to use an annealing algorithm to calculate the coefficients of the first loss function and the second loss function according to the first reward and punishment value; or,
  • An annealing algorithm is used to calculate the coefficients of the first loss function and the second loss function according to the second reward and punishment value.
  • the object is a protein in a protein association network;
  • the attribute feature of the object is the subspace structure of the protein;
  • the network structure feature of the object is the relationship between each protein and other proteins in the tissue Co-expression;
  • the label category of the part of the object is the cellular function of a part of the protein of known cellular function;
  • the first object is the protein of unknown cellular function;
  • the second object is the protein of known cellular function.
  • the object is a document in the document citation relationship network;
  • the attribute feature of the object is a keyword of the document title;
  • the network structure feature of the object is the citation relationship between each document and other documents ;
  • the mark category of the part of the object is the document of a part of the known document category;
  • the first object is the document of the unknown document category, and the second object is the document of the known document category.
  • FIG. 7 shows an object classification and connection relationship prediction device between objects provided by an embodiment of the present application, including: a test data acquisition module M210 is used to obtain test data of the object to be predicted.
  • the prediction module M220 is used to process the test data by using the spatial domain graph convolutional network obtained by the above graph convolutional network training method to obtain the classification result of the object and the prediction result of the network structure attribute between the objects.
  • test data includes network structure characteristics and object attribute characteristics of each object in the object relationship network where the object to be predicted is located;
  • the prediction module M220 is used to process the test data using the spatial domain graph convolution network to obtain the classification result of the object and the prediction result of the network structure attribute between the objects, and the spatial domain graph convolution
  • the network is a graph convolutional network trained by the method of any one of claims 1 to 6, and the prediction module M220 further includes the following sub-modules:
  • the predictive aggregation module M2201 is configured to obtain the aggregation characteristics of each object according to the network structure characteristics and the object attribute characteristics of each object, and update the object attribute characteristics of each object according to the aggregation characteristics.
  • the predictive connection module M2202 is configured to obtain the predicted connection probability of each object with other objects according to the network structure characteristics of each object and the object attribute characteristics updated by each object; and update the network structure characteristics of the object according to the predicted connection probability.
  • the prediction category module M2203 obtains the prediction category of the object to be predicted according to the updated object attribute characteristics of the object to be predicted.
  • the network structure feature of the object in the object relationship network where the object to be predicted is located is the co-expression of each protein in the tissue with other proteins in the protein association relationship network where the protein to be predicted is located;
  • the object attribute is characterized by the subspace structure of the protein.
  • the network structure feature of the object in the object relationship network where the object to be predicted is located is the citation relationship of each document with other documents in the document citation relationship network where the document to be predicted is located;
  • the attribute feature is the key word of the document title.
  • the embodiments of the present application also provide a computer-readable storage medium, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in each of the foregoing method embodiments can be realized.
  • the embodiments of the present application provide a computer program product.
  • the computer program product runs on an electronic device, the electronic device can realize the steps in the foregoing method embodiments when the electronic device is executed.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium.
  • the computer program can be stored in a computer-readable storage medium. When executed by the processor, the steps of the foregoing method embodiments can be implemented.
  • the computer program includes computer program code, and the computer program code may be in the form of source code, object code, executable file, or some intermediate forms.
  • the computer-readable medium may at least include: any entity or device capable of carrying computer program code to the photographing device/terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (Random Access Memory, RAM), electric carrier signal, telecommunications signal, and software distribution medium.
  • ROM read-only memory
  • RAM random access memory
  • electric carrier signal telecommunications signal
  • software distribution medium for example, U disk, mobile hard disk, floppy disk or CD-ROM, etc.
  • computer-readable media cannot be electrical carrier signals and telecommunication signals.
  • the disclosed apparatus/network equipment and method may be implemented in other ways.
  • the device/network device embodiments described above are only illustrative.
  • the division of the modules or units is only a logical function division, and there may be other divisions in actual implementation, such as multiple units.
  • components can be combined or integrated into another system, or some features can be omitted or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

Landscapes

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

Abstract

一种空间图卷积网络的训练方法，包括：获取训练数据（S110）；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象；以及，根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络（S120）。从而实现了可以同时处理对象关系网络的对象分类和对象的网络结构特征预测任务，节省了计算设备的算力，提高了效率。

Description

空间图卷积网络的训练方法、电子设备及存储介质 技术领域
本申请属于模式识别技术领域，尤其涉及空间图卷积网络的训练方法、电子设备及存储介质。
背景技术
在关于图网络数据处理的相关应用场景中，对象关系网络中的节点对应的对象分类，和节点间连接对应的对象的网络结构特征预测是两个较为常见的任务。例如，通过论文间的引用关系进行论文所属学科的预测，和论文间缺失的引用关系的补全。又例如，通过蛋白质间在组织内的共表达预测蛋白质的细胞内功能和蛋白质间缺失的共表达关系的补全。但是目前缺乏一种方法可以同时处理对象关系网络对应的图网络数据中的对象分类和对象网络结构特征预测任务。
发明内容
本申请实施例提供了空间图卷积网络的训练方法、对象分类和对象的网络结构特征的预测方法、电子设备及存储介质，可以解决上述技术问题。
第一方面，本申请实施例提供了一种空间图卷积网络的训练方法，包括：
获取训练数据；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象；
根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
从而实现了可以同时处理对象关系网络的对象分类和对象网络结构属性预测任务，节省了计算设备的算力，提高了效率。
第二方面，本申请实施例提供了一种对象分类和对象的网络结构特征的预测方法，包括：
获取待预测对象的测试数据；
采用空间域图卷积网络对所述测试数据进行处理，获得所述待预测对象的分类结果和对象的网络结构属性预测结果，所述空间域图卷积网络为经由上述第一方面所述的方法训练的到的图卷积网络。
第三方面，本申请实施例提供了一种空间域图卷积网络的训练装置，包括：
数据获取模块，用于获取训练数据；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象；
训练模块，用于根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
第四方面，本申请实施例提供了一种对象分类和对象间连接关系的预测装置，包括：
测试数据获取模块，用于获取待预测对象的测试数据；
预测模块，采用空间域图卷积网络对所述测试数据进行处理，获得所述待预测对象的分类结果和对象间的网络结构属性预测结果，所述空间域图卷积网络为经由上述第一方面所述的方法训练的到的图卷积网络。
第五方面，本申请实施例提供了一种电子设备，包括：
存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，所述处理器执行所述计算机程序时实现上述第一方面和/或第二方面所述的方法步骤。
第六方面，本申请实施例提供了一种计算机可读存储介质，包括：所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时实现上述第一方面和/或第二方面所述的方法步骤。
第七方面本申请实施例提供了一种计算机程序产品，当计算机程序产品在电子设备上运行时，使得电子设备执行上述第一方面所述的方法步骤。
可以理解的是，上述第二方面至第六方面的有益效果可以参见上述第一方面中的相关描述，在此不再赘述。
附图说明
为了更清楚地说明本申请实施例中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是本申请实施例提供的电子设备的结构示意图；
图2是本申请一实施例提供的对象关系网络结构示意图；
图3是本申请一实施例提供的空间域图卷积网络训练方法流程示意图；
图4是本申请另一实施例提供的空间域图卷积网络训练方法流程示意图；
图5是本申请一实施例提供的对象分类和对象的网络结构特征的预测方法的流程示意图；
图6是本申请实施例提供的空间域图卷积网络训练装置示意图；
图7是本申请实施例提供的对象分类和对象的网络结构特征的预测装置示意图。
具体实施方式
以下描述中，为了说明而不是为了限定，提出了诸如特定系统结构、技术之类的具体细节，以便透彻理解本申请实施例。然而，本领域的技术人员应当 清楚，在没有这些具体细节的其它实施例中也可以实现本申请。在其它情况中，省略对众所周知的系统、装置、电路以及方法的详细说明，以免不必要的细节妨碍本申请的描述。
应当理解，当在本申请说明书和所附权利要求书中使用时，术语“包括”指示所描述特征、整体、步骤、操作、元素和/或组件的存在，但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。
还应当理解，在本申请说明书和所附权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合，并且包括这些组合。
如在本申请说明书和所附权利要求书中所使用的那样，术语“如果”可以依据上下文被解释为“当...时”或“一旦”或“响应于确定”或“响应于检测到”。类似地，短语“如果确定”或“如果检测到[所描述条件或事件]”可以依据上下文被解释为意指“一旦确定”或“响应于确定”或“一旦检测到[所描述条件或事件]”或“响应于检测到[所描述条件或事件]”。
另外，在本申请说明书和所附权利要求书的描述中，术语“第一”、“第二”、“第三”等仅用于区分描述，而不能理解为指示或暗示相对重要性。
在本申请说明书中描述的参考“一个实施例”或“一些实施例”等意味着在本申请的一个或多个实施例中包括结合该实施例描述的特定特征、结构或特点。由此，在本说明书中的不同之处出现的语句“在一个实施例中”、“在一些实施例中”、“在其他一些实施例中”、“在另外一些实施例中”等不是必然都参考相同的实施例，而是意味着“一个或多个但不是所有的实施例”，除非是以其他方式另外特别强调。术语“包括”、“包含”、“具有”及它们的变形都意味着“包括但不限于”，除非是以其他方式另外特别强调。
在关于图网络数据处理的相关应用场景中，对象关系网络中的节点对应的对象分类，和节点间连接对应的对象的网络结构特征预测是两个较为常见的任务。例如，通过论文间的引用关系进行论文所属学科的预测，和论文间缺失的 引用关系的补全。又例如，通过蛋白质间在组织内的共表达预测蛋白质的细胞内功能和蛋白质间缺失的共表达关系。但是，目前的方法均将节点分类和节点间连接关系预测这两个任务孤立看待，没有方法能够同时训练学习和完成这两个任务。而且，也没有现有方法可以通过参数的设定，使得同一个模型可以根据参数的不同设定来切换侧重进行不同的任务。但是，往往这样的多任务同时学习可以大大减少计算机的计算成本，尤其对于深度学习网络这种复杂模型。
在图网络数据的节点分类任务中，输入数据必须包含网络拓扑结构信息，也就是节点之间的连接关系，但是节点分类模型并不直接对连接关系进行学习建模，而是通过间接观察已有的连接关系给分类结果带来的准确率收益来进行模型学习。这样单一的观察模式，并没有最大化的利用输入信息。同样，在连接预测任务中，现有方法又完全不考虑节点的类别归属。
图网络结构数据中，新的节点间的连接的引入能为节点分类和连接预测任务引入更多信息。而这个信息是能在多任务共同学习的过程中学习得到的，而目前方法都没有考虑到这点。
为解决同时处理节点分类任务和连接预测任务，以及通过连接预测任务提高节点分类任务的准确性，本申请提供了一种空间图卷积网络的训练方法、对象分类和对象的网络结构特征的预测方法电子设备及存储介质。
以下结合附图对本申请的实施例加以说明。
图1示出的是本申请实施例提供的一种电子设备D10，包括：至少一个处理器D100、存储器D101以及存储在存储器D101中并可在所述至少一个处理器D100上运行的计算机程序D102，所述处理器D100执行所述计算机程序D102时实现本申请实施例提供的空间图卷积网络的训练方法、对象分类和对象的网络结构特征的预测方法至少之一。
可以理解的是，上述电子设备，可以是桌上型计算机、笔记本、掌上电脑服务器、服务器集群、分布式服务器及云端服务器等计算设备。该电子设备D10可包括，但不仅限于，处理器D100、存储器D101。本领域技术人员可以理解， 图1仅仅是电子设备D10的举例，并不构成对电子设备D10的限定，可以包括比图示更多或更少的部件，或者组合某些部件，或者不同的部件，例如还可以包括输入输出设备、网络接入设备等。
所称处理器D100可以是中央处理单元(Central Processing Unit，CPU)，该处理器D100还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述存储器D101在一些实施例中可以是所述电子设备D10的内部存储单元，例如电子设备D10的硬盘或内存。所述存储器D101在另一些实施例中也可以是所述电子设备D10的外部存储设备，例如所述电子设备D10上配备的插接式硬盘，智能存储卡(Smart Media Card,SMC)，安全数字(Secure Digital,SD)卡，闪存卡(Flash Card)等。进一步地，所述存储器D101还可以既包括所述电子设备D10的内部存储单元也包括外部存储设备。所述存储器D101用于存储操作系统、应用程序、引导装载程序(BootLoader)、数据以及其他程序等，例如所述计算机程序的程序代码等。所述存储器D101还可以用于暂时地存储已经输出或者将要输出的数据。
为了阐述方便，以下实施例中将上述电子设备统称为服务器，可以理解的是，这并不构成对本申请的电子设备的具体限定。
图2示出的是本申请实施例提供的一个对象关系网络图。图中的节点为实际应用场景中的对象，图中的边为对象间的关联关系，对应于对象的网络结构属性特征，图中节点的编号为节点对应的对象的编号。例如，若图2表示的对象关系网络图对应于蛋白质关联关系网络，则图2中节点对应的对象是蛋白质，图2中的边对应的是两蛋白质在细胞内的共表达。又例如，若图2表示的对象关系网络图对应于文献引用关系网络，则图2中的节点对应的对象是文献，图 2中的边对应的是两文献间的引用关系。当然，本领域技术人员可以根据本申请的指引，获得不同领域的对象关联关系图以应用到本申请实施例提供的方法。
图3示出了本申请实施例提供的空间域图卷积网络的训练方法的流程示意图，用于得到对象分类和对象网络结构属性预测的图卷积网络，应用于上述图1所示的电子设备(以下称服务器)，如图3所示该方法包括步骤S110和步骤S120，各步骤的具体实现原理如下：
S110，获取训练数据；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象。
非限定性的，服务器获取训练数据，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象。可以理解的是，在图网络结构表示的对象关系图中，对象对应对象关系网络中的节点，对象的网络结构特征对应对象关系网络中节点与其他节点的连接关系。
在一个非限定性的具体示例中，所述对象为蛋白质关联关系网络中的蛋白质；所述对象属性特征为蛋白质的子空间结构；所述对象的网络结构特征为每个蛋白质和其他蛋白质在组织内的共表达；所述部分对象的标记类别为部分已知的细胞功的蛋白质的细胞功能；所述第一对象为未知细胞功能的蛋白质；所述第二对象为已知细胞功能的蛋白质。
在另一非限定性的示例中，所述对象为文献引用关系网络中的文献；所述对象属性特征为文献标题的关键词；所述对象的网络结构特征为每个文献与其他文献的引用关系；所述部分对象的标记类别为部分已知文献类别的文献；所述第一对象为未知文献类别的文献，所述第二对象为已知文献类别的文献。
以文献引用关系网络图为例，在一个非限定性的示例中，服务器获取文献 引用关系网络中多个文献的网络结构特征，所述网络结构特征可以是该文献与其他文献引用关系的向量，也可以是文献引用关系网络中的多个文献以文献为节点，以文献间存在的引用关系为边构成的网络对应的邻接矩阵。文献的对象属性特征为文献标题的关键词与预设词典中关键词对应关系的向量。例如，词典为[生物，一种，信息学，……，预测，……]；一个示例文中，献的标题为“生物信息学中的MicroRNA预测研究”则该文献的对象属性特征为[1,0,1,……,1,……]；另一示例中，文献标题为“一种基于高斯混合模型的轨迹预测算法”则该文献的对象属性特征为[0,1,0,……,1,……]。用于训练图卷积网络的部分对象的类别是已知的，例如，已知“生物信息学中的MicroRNA预测研究”属于生物信息学类别，已知“一种基于高斯混合模型的轨迹预测算法”属于计算机科学类别。则该对象在类别向量y＝[y 1,y 2,……,y n]，的相应类别的概率为1，例如，文献“生物信息学中的MicroRNA预测研究”用类别向量标识则为[0,1，0,……,0]，该类别向量中的第二个元素对应文献为生物信息学类别的概率，可以用该向量标记该文献的类别，作为该文献的标识类别数据。
可以理解的是，本领域技术人员可以根据上述示例获取蛋白质关系网络对应的训练数据，例如通过蛋白质子空间结构字典和蛋白质的子空间结构得到蛋白质属性特征向量，通过两蛋白质在组织内的共表达作为蛋白质关系网络的边，进而获得蛋白质在该网络中的网络结构特征，该网络结构特征可以是蛋白质关系网络的邻接矩阵，也可以是以蛋白质为对象的，该蛋白质与其他蛋白质关联关系的向量。进一步的，本领域技术人员也可以根据以上示例获取社交关系网络对应的训练数据、销售关系网络对应的训练数据，以及其他领域的训练数据，用以训练处理各领域对象分类和对象间关联关系预测的图卷积网络。
S120，根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
非限定性的，服务器根据上述训练数据，例如蛋白质关系网络图的数据或文献引用关系网络图的数据，对待训练的图卷积网络进行训练，得到用于对象 分类和对象网络结构属性预测的图卷积网络。
在一种可能的实现方式中，在图2所示的实施例的基础上，对步骤S120进行了细化，如图4所示，包括步骤S121～S125。具体的：
S121，基于所述待训练的图卷积网络，根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率，根据所述第一连接概率更新所述对象的网络结构特征。
在一个非限定性的示例中，服务器根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率，根据所述第一连接概率更新所述对象的网络结构特征。可以理解的是，除了第一次迭代每次迭代中的对象属性特征都是上一轮迭代更新的对象属性特征，第一次迭代采用的对象属性特征为S110步骤中获取的原始对象属性特征。可以理解的是，每次迭代后在步骤S110步骤获取的原始的对象图结构特征的基础上根据更新的对象属性特征计算对象间的第一连接概率，并根据第一连接概率更新原始的对象图结构特征，即更新对象与其他对象间的连接关系，也可以理解为补全原始的对象关系图网络结构的连接关系。在一种可能是实施方式中，在第一次迭代时不计算第一连接概率，也就是说不进行对象间连接关系补全操作。
在一个非限定性的具体的示例中，请一并参阅图2，如图2所示的对象关系图网络中。遍历图网络中的所有节点，选取不存在连接关系的两个节点对{(i,j),……}，通过下面公式计算每个节点对的连接概率e ij，
Figure PCTCN2020127254-appb-000001
其中，
Figure PCTCN2020127254-appb-000002
为步骤S122中第K次循环(卷积)产生的节点标识，也就是节点的聚合特征，a为一个线性函数，该函数具有初始参数，其参数通过步骤S124的反馈学习过程更新，W 1为降维向量，σ为一个非线性变换函数，例如sigmoid函数，用于将计算结果映射到[0,1]区间以获得概率值。
在一个非限定性的示例中，获取全部没有连接关系的两个对象间的第一连接概率后，对全部第一概率值由大至小进行排序，由大至小选取前O个概率值 对应的对象，认为这O个概率值对应的两个对象间有连接关系，更新该概率值对应的对象的网络结构特征，也就是补全该对象和其他对象的连接关系。
在一个非限定性的示例中，获取全部没有连接关系的两个对象间的第一连接概率后，选取概率值大于第一阈值的概率值对应的对象，认为大于第一阈值的概率值对应的两个对象间有连接关系，更新该概率值对应的对象的网络结构特征，也就是补全该对象和其他对象的连接关系。
S122，根据各个对象更新后的所述网络结构特征和各个对象的所述对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特征。
在一个非限定性的示例中，服务器通过图采样和聚合(Graph Sampling and Aggregating，GraphSAGE)算法获取各个对象的对象属性特征的聚合特征，并根据聚合特征更新对象的对象属性特征。
在一个非限定性的示例中，服务器通过图注意力网络(Graph Attention Network，GAT)算法更新各个对象的属性特征。
非限定性的，以应用GraphSAGE算法更新各个对象的所述对象属性特征为例阐述图卷积神经网络训练过程中的卷积过程。
执行K次循环过程，也就是卷积过程，每次卷积过程即为第k次卷积过程，其中K为大于等于1的整数。
在每次卷积过程中遍历图2所示的图网络中的每个节点，可以理解的是，这里不区分节点的访问顺序。可以理解的是，也可以选取图2所示节点中的部分节点。
对每个节点进行执行以下过程,直到每个节点都被访问过一次：
将当前访问的节点作为目标节点，表示为v，该节点对应的特征为x v(当k＝1时)，或h k-1 v(当k>1时)；
在S121更新的对象网络结构特征的基础上，也就是在补全的网络结构图基础上，匹配与这个目标节点v直接相连的节点，表示为N(v)；
对目标结点和所有邻居结点的特征表示，即对象的对象属性特征，进行聚合操作：
Figure PCTCN2020127254-appb-000003
其中，AGGREGATE为向量聚合操作，CONCAT为向量拼接操作，k-1表示上一个循环步，h k-1表示上一个循环步产生的节点特征表示，k＝0时，h k＝x，即使用原始节点特征表示。W k为模型中的可学习参数，每个循环步(卷积)使用不同的参数。
可以理解的是，上述以GraphSAGE算法为例进行对象(节点)的对象属性特征聚合操作，那么所有GraphSAGE算法的变形算法对节点的聚合操作均适用于本实施例，其中的算法步骤和采样步骤这里不再赘述。可以理解的是，其他空间域图卷积网络对节点特征的卷积操作也均可以用于替换本步骤的聚合(卷积)操作，这里也不再赘述。
S123，根据各个对象的所述对象属性特征和原始网络结构特征，计算第二连接概率；所述第二连接概率为每个对象与根据该对象的所述网络结构特征采样选取的其他对象的连接概率。
在一个非限定性的示例中，服务器根据各个对象的所述对象属性特征和原始的网络结构特征，计算第二连接概率；所述第二连接概率为每个对象与根据该对象的所述网络结构特征采样选取的其他对象的连接概率。
一种可能的实施方式为，根据该对象的所述网络结构特征采样选取的其他对象，可以为根据每个对象的网络结构特征，选取对象关系网络图中与该对象对应的目标节点直接连接的全部第一跳节点，以及，采样选取与该节点无直接连接的I个第J跳节点；其中I为大于0的正整数，J为大于1的正整数。
在一种可能的实施方式中，根据该对象的所述网络结构特征采样选取的其他对象，可以为根据跳数进行采样，例如，跳数J越大，采样的节点数I越小。
在一种可能的实施方式中，根据该对象的所述网络结构特征采样选取的其 他对象，可以为跳数J为大于0的正整数，也就是说对第一跳的节点数量也进行采样。
通过对目标节点外的其他节点进行采样可以保证采样的平衡和降低计算量。本领域技术人员可以根据本申请实施例的教导，在实际实施本申请实施例的技术方案时选取符合实际情况的采样方法。原始的网络结构特征即通过步骤S110获取的网络结构特征。
在一个非限定性的具体的示例中，根据更新后的对象属性特征，采用公式，
Figure PCTCN2020127254-appb-000004
计算第二连接概率，该第二连接概率为节点间连接概率。
可以理解的是，步骤S123中节点的选取，本领域技术人员可以在本申请的教导下根据实际的需要进行选取和调整，以上节点选取的方法为非限定性示例性说明，并不构成对本申请的限定。
S124，根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网的参数。
在一个非限定性的示例中，服务器根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网络的参数。非限定性的，可以通过反向传播梯度下降法调整所述图卷积网络的参数。非限定性的，通过分类器对根据更新后的各个对象的对象属性特征获得各个对象的预测类别，所述分类器可以为两层的全连接神经网络，也可以为两层以上的多层神经网络或其他机器学习分类模型。
在一个非限定性的示例中，通过下面的损失函数计算奖惩值调整所述图卷积网的参数，
Loss＝Loss link+Loss cls
其中，Loss link为第二连接概率和各个对象的原始网络结构特征的差值； Loss cls为根据所述第二对象的预测类别、所述第二对象的标记类别的差值。可以理解的是，上述差值可以为差值的绝对值、均差、方差等差值，本领域技术人员可以根据实际需要对确定损失函数中差值的处理，这里不再赘述。
在一个非限定的示例中，第一损失函数具有第一损失函数系数，第二损失函数具有第二损失函数系数，根据所述第一损失函数的系数和所述第二损失函数的系数确定所述图卷积网络偏重对象分类或对象网络结构特征预测。在一个非限定性的具体的示例中，不受偏重任务的系数为γ，受偏重任务的系数为1-γ。例如，已知一个蛋白质关联关系网络，蛋白质节点间的连接关系大部分都是已知的，但是蛋白质在细胞内的功能大部分是未知的，此时需要训练图卷积网络偏重于节点分类任务，则损失函数为，
Loss＝γLoss link+(1-γ)Loss cls
又例如，已知一个蛋白质关联关系网络，蛋白质节点间的连接关系大部分都是未知的，但是蛋白质在细胞内的功能大部分是已知的，此时需要训练图卷积网络偏重于连接预测任务，则损失函数为，
Loss＝(1-γ)Loss link+γLoss cls
在一个非限定性的示例中，采用退火算法根据所述第一奖惩值计算所述第一损失函数的系数和所述第二损失函数的系数；或，采用退火算法根据所述第二奖惩值计算所述第一损失函数的系数和所述第二损失函数的系数。在一个非限定性的具体的示例中，设置初始退火温度为temp ini，退火速率为ε。如果我们对当前的训练目标有所偏重，如更倾向于节点分类或连接预测任务，那么就需要在训练过程中加入对这两者的动态调节。因此，退火机制被引入其中，目标是使得：受偏重的任务在训练过程中随着训练的迭代过程受到递增的重视，反之，不受偏重任务所受的重视度递减。
如下定义一个退火策略，表示随着训练迭代过程，temp t越来越小，其中t为图卷积网络的训练迭代次数，
Figure PCTCN2020127254-appb-000005
同时考虑到：即使是不受偏重的任务，过大的误差也应该被避免，因此本方法又进一步对高误差现象加入了惩罚，使得不受偏重任务的预测误差和重视度达到一种平衡，
Figure PCTCN2020127254-appb-000006
Loss aux表示不受偏重的训练任务所产生的损失。综合这两个因素得到了：不受偏重的训练任务的系数γ和受偏重的训练任务的系数1-γ。
可以理解的是，通过引入退火机制来平衡当前的任务重心，从而保证图卷积网络在学习多个任务的过程中，保证模型表现的前提下，对任务有所侧重。
在执行S124之后，判定当前是否满足迭代结束条件，当前不满足迭代结束条件时，返回S121，继续执行S121～S124；当前满足迭代结束条件时，执行S125。
S125，当对所述图卷积网络的训练满足迭代结束条件，停止训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
可以理解的是，若不满足迭代训练结束条件，则返回所述根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率的步骤对所述图卷积网络进行迭代训练，直到对所述图卷积网络的训练满足迭代结束条件，停止训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
非限定性的，迭代训练结束条件可以为达到预设迭代次数，也可以为损失函数收敛到预设阈值以下。本领域技术人员可以根据实际需要设置迭代结束条件。如果未达到迭代训练结束条件则返回根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率的步骤对所述图卷积网络进行迭代训练；如果达到迭代训练结束条件则结束对待训练的图卷积网络的训练。
可以理解的是，的本申请实施例训练得到的图卷积网络，一方面引入单模型多任务的图卷积网络训练方式来同时训练对象关系网络中节点对应的对象的分类任务，和节点间连接关系对应的对象的网络结构特征预测任务，获得的经 训练的图卷积网络可以同时实现对象分类任务和对象的网络结构特征预测任务，从而充分利用了计算设备的算力，提高了资源利用率，降低了成本。另一方面，将预测得到的网络结构特征用于对象关系网络的连接补全，并将连接补全的结果作为额外的新输入信息来与对象分类任务一起协同训练，从而可以提高节点分类的预测精度和预测效率。
请参阅图5，图5示出的是本申请实施例提供的一种对象分类和对象的网络结构特征的预测方法，可由上述图1所示电子设备(以下称服务器)通过软件/硬件实现。如图5所示，该方法包括步骤S210至S220。各步骤的具体实现原理如下：
S210，获取待预测对象的测试数据。
在一个非限定性的示例中，所述测试数据包括所述待预测对象所在的对象关系网络中的各个对象的网络结构特征、对象属性特征。
非限定性的，在蛋白质关联关系网络中，所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测蛋白质所在的蛋白质关联关系网络中每个蛋白质的与其他蛋白质在组织内的共表达；所述对象属性特征为蛋白质的子空间结构。
非限定性的，在文献引用关系网络中，所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测文献所在的文献引用关系网络中每个文献的与其他文献的引用关系；所述对象属性特征为文献标题的关键词。
本领域技术人员也可以根据以上示例获取社交关系网络对应的训练数据、销售关系网络对应的预测数据，以及其他领域的预测数据，采用图卷积网络处理各领域对象分类和对象间关联关系预测的任务。
S220，采用经上述图3所示的方法得到的空间域图卷积网络对所述测试数据进行处理，获得所述对象的分类结果和对象间的网络结构属性预测结果。
在一个非限定性的示例中，服务器采用经上述图3所示的方法训练得到的空间域图卷积网络对所述测试数据进行处理，获得所述对象的分类结果和对象 间的网络结构属性预测结果。
在一个具体的非限定性的示例中，根据各个对象的网络结构特征和对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特征；根据各个对象的网络结构特征和各个对象更新的所述对象属性特征获取各个对象与其他对象的预测连接概率；根据所述预测连接概率更新所述对象的网络结构特征；根据待预测的对象更新后的对象属性特征获取待预测的对象的预测类别。
非限定性的，采用与前述图卷积网络训练方法相同的对象属性特征聚合操作，例如，以GraphSAGE算法对各个对象对应的目标结点和目标节点对应的所有邻居结点的对象属性特征进行聚合操作为例，对各个对象对应的对象属性特征采用以下公式的方法进行K次聚合操作，获得各个节点的聚合对象属性特征，再根据所述聚合特征更新各个对象的所述对象属性特征，其中的参数参考上述实施例中的阐述，
Figure PCTCN2020127254-appb-000007
非限定性的，采用图3所述方法得到的空间域图卷积网络中的分类器，例如，两层全连接神经网络，或两层以上的神经网络，或其他机器学习分类模型，对待预测对象的对象属性特征进行识别，获得该待预测对象的分类结果。
非限定性的，根据以下公式获得各个没有连接关系对象(节点)间的预测连接概率，根据所述预测连接概率更新所述对象的网络结构特征，即补全待预测对象的图网络连接关系，其中的参数参考上述实施例中的阐述。
Figure PCTCN2020127254-appb-000008
非限定性的，获取全部没有连接关系的两个对象间的预测连接概率后，对全部预测概率值由大至小进行排序，由大至小选取前Q个概率值对应的对象，判定这Q个概率值对应的两个对象间有连接关系，更新该概率值对应的对象的网络结构特征，也就是补全该对象和其他对象的连接关系。
非限定性的，获取全部没有连接关系的两个对象间的预测连接概率后，选取概率值大于预测阈值的概率值对应的对象，则判定大于第二阈值的概率值对应的两个对象间有连接关系，更新该概率值对应的对象的网络结构特征，也就是补全该对象和其他对象的连接关系。
可以理解的是，通过图5所示的方法，可以同时获得对象关系网络中对象分类和对象间连接关系的预测结果，从而节约了计算设备的算力，提高了效率，降低了成本。
应理解，上述实施例中各步骤的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
对应于上述图3以及图4所示的空间域图卷积网络的训练方法，图6示出的是本申请实施例提供的一种空间域图卷积网络的训练装置，包括：
数据获取模块M110，用于获取训练数据；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象。
训练模块M120，用于根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
其中，所述训练模块M120还包括以下子模块：
连接补全模块M121，用于根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率，根据所述第一连接概率更新所述对象的网络结构特征；
特征聚合模块M122，用于根据各个对象更新后的所述网络结构特征和各个对象的所述对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特征；
采样连接计算模块M123，根据各个对象的所述对象属性特征和原始的网络结构特征，计算第二连接概率；所述第二连接概率为每个对象与根据该对象的所述网络结构特征采样选取的其他对象的连接概率；
参数调整模块M124，根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网的参数；
迭代判断模块M125，用于判断迭代结束条件，若不满足迭代结束条件则返回根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率的步骤对所述图卷积网络进行迭代训练，直到对所述图卷积网络的训练满足迭代结束条件，得到用于对象分类和对象网络结构属性预测的图卷积网络。
在一个非限定性示例中，所述损失函数包括第一损失函数和第二损失函数。
相应的，参数调整模块M124，用于根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网的参数，还包括：
第一奖惩值确定模块M1241，用于根据所述第二对象的预测类别、所述第二对象的标记类别和所述第一损失函数确定第一奖惩值；
第二奖惩值确定模块M1241，用于根据所述第二连接概率、各个对象的原始网络结构特征和所述第二损失函数确定第二奖惩值；
参数调整子模块M1243，根据所述第一奖惩值和所述第二奖惩值调整所述图卷积网络的参数。
在一个非限定性示例中，所述参数调整模块M124还用于，根据所述第一损失函数的系数和所述第二损失函数的系数，确定所述图卷积网络偏重对象分类或对象网络结构特征预测。
在一个非限定性示例中，所述参数调整模块M124还用于，采用退火算法 根据所述第一奖惩值计算所述第一损失函数和所述第二损失函数的系数；或，
采用退火算法根据所述第二奖惩值计算所述第一损失函数和所述第二损失函数的系数。
在一个非限定性的示例中，所述对象为蛋白质关联关系网络中的蛋白质；所述对象属性特征为蛋白质的子空间结构；所述对象的网络结构特征每个蛋白质和其他蛋白质在组织内的共表达；所述部分对象的标记类别为部分已知的细胞功的蛋白质的细胞功能；所述第一对象为未知细胞功能的蛋白质；所述第二对象为已知细胞功能的蛋白质。
在一个非限定性的示例中，所述对象为文献引用关系网络中的文献；所述对象属性特征为文献标题的关键词；所述对象的网络结构特征为每个文献与其他文献的引用关系；所述部分对象的标记类别为部分已知文献类别的文献；所述第一对象为未知文献类别的文献，所述第二对象为已知文献类别的文献。
对应于上述图5所示的对象分类和对象的网络结构特征预测方法，图7示出的是本申请实施例提供的一种对象分类和对象间连接关系的预测装置，包括：测试数据获取模块M210，用于获取待预测对象的测试数据。
预测模块M220，用于采用经上述图卷积网络训练方法得到的空间域图卷积网络对所述测试数据进行处理，获得所述对象的分类结果和对象间的网络结构属性预测结果。
在一个非限定性的示例中，所述测试数据包括所述待预测对象所在的对象关系网络中的各个对象的网络结构特征、对象属性特征；
相应的，预测模块M220，用于所述采用空间域图卷积网络对所述测试数据进行处理，获得所述对象的分类结果和对象间的网络结构属性预测结果，所述空间域图卷积网络为经由权利要求1至6任一项所述的方法训练的到的图卷积网络，预测模块M220还包括以下子模块：
预测聚合模块M2201，用于根据各个对象的网络结构特征和对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特 征。
预测连接模块M2202，用于根据各个对象的网络结构特征和各个对象更新的所述对象属性特征获取各个对象与其他对象的预测连接概率；根据所述预测连接概率更新所述对象的网络结构特征。
预测类别模块M2203，根据待预测的对象更新后的对象属性特征获取待预测的对象的预测类别。
在一个非限定性的示例中，所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测蛋白质所在的蛋白质关联关系网络中每个蛋白质的与其他蛋白质在组织内的共表达；所述对象属性特征为蛋白质的子空间结构。
在一个非限定性的示例中，所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测文献所在的文献引用关系网络中每个文献的与其他文献的引用关系；所述对象属性特征为文献标题的关键词。
需要说明的是，上述图6和图7所示的装置/单元之间的信息交互、执行过程等内容，由于与本申请方法实施例基于同一构思，其具体功能及带来的技术效果，具体可参见方法实施例部分，此处不再赘述。
所属领域的技术人员可以清楚地了解到，为了描述的方便和简洁，仅以上述各功能单元、模块的划分进行举例说明，实际应用中，可以根据需要而将上述功能分配由不同的功能单元、模块完成，即将所述装置的内部结构划分成不同的功能单元或模块，以完成以上描述的全部或者部分功能。实施例中的各功能单元、模块可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中，上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。另外，各功能单元、模块的具体名称也只是为了便于相互区分，并不用于限制本申请的保护范围。上述系统中单元、模块的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介 质存储有计算机程序，所述计算机程序被处理器执行时实现可实现上述各个方法实施例中的步骤。
本申请实施例提供了一种计算机程序产品，当计算机程序产品在电子设备上运行时，使得电子设备执行时实现可实现上述各个方法实施例中的步骤。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请实现上述实施例方法中的全部或部分流程，可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一计算机可读存储介质中，该计算机程序在被处理器执行时，可实现上述各个方法实施例的步骤。其中，所述计算机程序包括计算机程序代码，所述计算机程序代码可以为源代码形式、对象代码形式、可执行文件或某些中间形式等。所述计算机可读介质至少可以包括：能够将计算机程序代码携带到拍照装置/终端设备的任何实体或装置、记录介质、计算机存储器、只读存储器(Read-Only Memory，ROM)、随机存取存储器(Random Access Memory，RAM)、电载波信号、电信信号以及软件分发介质。例如U盘、移动硬盘、磁碟或者光盘等。在某些司法管辖区，根据立法和专利实践，计算机可读介质不可以是电载波信号和电信信号。
在上述实施例中，对各个实施例的描述都各有侧重，某个实施例中没有详述或记载的部分，可以参见其它实施例的相关描述。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
在本申请所提供的实施例中，应该理解到，所揭露的装置/网络设备和方法，可以通过其它的方式实现。例如，以上所描述的装置/网络设备实施例仅仅是示意性的，例如，所述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现 时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通讯连接可以是通过一些接口，装置或单元的间接耦合或通讯连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
以上所述实施例仅用以说明本申请的技术方案，而非对其限制；尽管参照前述实施例对本申请进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本申请各实施例技术方案的精神和范围，均应包含在本申请的保护范围之内。

Claims (11)

  1. 一种空间域图卷积网络的训练方法，其特征在于，包括：
    获取训练数据；其中，所述训练数据包多个对象的网络结构特征、每个对象的对象属性特征、以及所述多个对象中部分对象的标记类别；每个对象的所述网络结构特征为该对象与其他对象间的关联关系；所述多个对象中具有标记类别的对象为第二对象，不具有标记类别的对象为第一对象；
    根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
  2. 如权利要求1所述的方法，其特征在于，所述根据所述训练数据对待训练的图卷积网络进行训练，得到用于对象分类和对象网络结构属性预测的图卷积网络，包括：
    基于所述待训练的图卷积网络，根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率，根据所述第一连接概率更新所述对象的网络结构特征；
    根据各个对象更新后的所述网络结构特征和各个对象的所述对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特征；
    根据各个对象的所述对象属性特征和原始网络结构特征，计算第二连接概率；所述第二连接概率为每个对象与根据该对象的所述网络结构特征采样选取的其他对象的连接概率；
    根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网的参数；
    返回所述根据各个对象的对象属性特征获取没有连接关系的对象间的第一连接概率的步骤对所述图卷积网络进行迭代训练，直到对所述图卷积网络的训 练满足迭代结束条件，停止训练，得到用于对象分类和对象网络结构属性预测的图卷积网络。
  3. 如权利要求2所述的方法，其特征在于，所述损失函数包括第一损失函数和第二损失函数；
    相应的，根据更新后的各个对象的对象属性特征获得各个对象的预测类别；根据所述第二对象的预测类别、所述第二对象的标记类别、所述第二连接概率、各个对象的原始网络结构特征和损失函数调整所述图卷积网的参数，包括：
    根据所述第二对象的预测类别、所述第二对象的标记类别和所述第一损失函数确定第一奖惩值；
    根据所述第二连接概率、所述各个对象的原始网络结构特征和所述第二损失函数确定第二奖惩值；
    根据所述第一奖惩值和所述第二奖惩值调整所述图卷积网络的参数。
  4. 如权利要求3所述的方法，其特征在于，所述根据所述第二对象的预测类别、所述第二对象的标记类别和所述第一损失函数确定第一奖惩值前，还包括：
    根据所述第一损失函数的系数和所述第二损失函数的系数，确定所述图卷积网络偏重对象分类或对象网络结构特征预测。
  5. 如权利要求4所述的方法，其特征在于，还包括：
    采用退火算法根据所述第一奖惩值计算所述第一损失函数的系数和所述第二损失函数的系数；或，
    采用退火算法根据所述第二奖惩值计算所述第一损失函数的系数和所述第二损失函数的系数。
  6. 如权利要求1至5任一项所述的方法，其特征在于，
    所述对象为蛋白质关联关系网络中的蛋白质；所述对象属性特征为蛋白质的子空间结构；所述对象的网络结构特征为每个蛋白质与其他蛋白质在组织内的共表达；所述部分对象的标记类别为部分已知的细胞功的蛋白质的细胞功能； 所述第一对象为未知细胞功能的蛋白质；所述第二对象为已知细胞功能的蛋白质；或，
    所述对象为文献引用关系网络中的文献；所述对象属性特征为文献标题的关键词；所述对象的网络结构特征为每个文献与其他文献的引用关系；所述部分对象的标记类别为部分已知类别的文献的文献类别；所述第一对象为未知文献类别的文献，所述第二对象为已知文献类别的文献。
  7. 一种对象分类和对象的网络结构特征的预测方法，其特征在于，包括：
    获取待预测对象的测试数据；
    采用空间域图卷积网络对所述测试数据进行处理，获得所述待预测对象的分类结果和对象的网络结构特征预测结果，所述空间域图卷积网络为经由权利要求1至6任一项所述的方法训练的到的图卷积网络。
  8. 如权利要求7所述的方法，其特征在于，所述测试数据包括所述待预测对象所在的对象关系网络中的各个对象的网络结构特征、对象属性特征；
    相应的，所述采用空间域图卷积网络对所述测试数据进行处理，获得所述对象的分类结果和对象间的网络结构属性预测结果，所述空间域图卷积网络为经由权利要求1至6任一项所述的方法训练的到的图卷积网络，包括：
    根据各个对象的所述网络结构特征和所述对象属性特征获取各个对象的聚合特征，根据所述聚合特征更新各个对象的所述对象属性特征；
    根据各个对象的所述网络结构特征和各个对象更新的所述对象属性特征获取各个对象与其他对象的预测连接概率；根据所述预测连接概率更新所述对象的网络结构特征；
    根据所述待预测的对象更新后的对象属性特征获取所述待预测的对象的预测类别。
  9. 如权利要求8所述的方法，其特征在于，
    所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测蛋白质所在的蛋白质关联关系网络中每个蛋白质的与其他蛋白质在组织内的共表 达；所述对象属性特征为蛋白质的子空间结构；或，
    所述待预测对象所在的对象关系网络中对象的网络结构特征为待预测文献所在的文献引用关系网络中每个文献的与其他文献的引用关系；所述对象属性特征为文献标题的关键词。
  10. 一种电子设备，包括存储器、处理器以及存储在所述存储器中并可在所述处理器上运行的计算机程序，其特征在于，所述处理器执行所述计算机程序时实现如权利要求1至6，和/或7至9任一项所述的方法。
  11. 一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现如权利要求1至6，和/或7至9任一项所述的方法。
PCT/CN2020/127254 2019-11-06 2020-11-06 空间图卷积网络的训练方法、电子设备及存储介质 WO2021089013A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911075406.5 2019-11-06
CN201911075406.5A CN111079780B (zh) 2019-11-06 2019-11-06 空间图卷积网络的训练方法、电子设备及存储介质

Publications (1)

Publication Number Publication Date
WO2021089013A1 true WO2021089013A1 (zh) 2021-05-14

Family

ID=70310660

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/127254 WO2021089013A1 (zh) 2019-11-06 2020-11-06 空间图卷积网络的训练方法、电子设备及存储介质

Country Status (2)

Country Link
CN (1) CN111079780B (zh)
WO (1) WO2021089013A1 (zh)

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111079780B (zh) * 2019-11-06 2023-06-23 中国科学院深圳先进技术研究院 空间图卷积网络的训练方法、电子设备及存储介质
CN112183299B (zh) * 2020-09-23 2024-02-09 成都佳华物链云科技有限公司 行人属性预测方法、装置、电子设备及存储介质
CN112562339B (zh) * 2020-12-09 2022-01-04 福州大学 基于多任务图卷积网络的鲁棒性交通流预测方法
CN112967231B (zh) * 2021-02-05 2022-11-15 五邑大学 焊接质量检测方法及其装置、计算机可读存储介质
CN113033669B (zh) * 2021-03-29 2024-05-14 东北大学 一种基于可学习的特征图过滤与图注意力网络的视觉场景识别方法
CN112966114B (zh) * 2021-04-10 2023-08-15 北京工商大学 基于对称图卷积神经网络的文献分类方法和装置
CN114169466B (zh) * 2021-12-24 2023-07-07 马上消费金融股份有限公司 图数据处理、文章分类、流量预测方法、装置、设备及存储介质
CN114444665A (zh) * 2022-02-02 2022-05-06 上海图灵智算量子科技有限公司 基于图卷积神经网络的伊辛求解器及实现伊辛模型的方法
CN115995024A (zh) * 2023-03-22 2023-04-21 成都理工大学 基于类图神经网络的图像分类方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109214719A (zh) * 2018-11-02 2019-01-15 广东电网有限责任公司 一种基于人工智能的营销稽查分析的系统和方法
US20190108639A1 (en) * 2017-10-09 2019-04-11 The Board Of Trustees Of The Leland Stanford Junior University Systems and Methods for Semantic Segmentation of 3D Point Clouds
CN110009093A (zh) * 2018-12-07 2019-07-12 阿里巴巴集团控股有限公司 用于分析关系网络图的神经网络系统和方法
CN110378543A (zh) * 2018-04-12 2019-10-25 百度在线网络技术（北京）有限公司 离职风险预测方法、装置、计算机设备和存储介质
CN111079780A (zh) * 2019-11-06 2020-04-28 中国科学院深圳先进技术研究院 空间图卷积网络的训练方法、电子设备及存储介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108648095A (zh) * 2018-05-10 2018-10-12 浙江工业大学 一种基于图卷积网络梯度的节点信息隐藏方法
CN110069726B (zh) * 2019-04-26 2021-06-29 福州大学 一种适用于DBLP与arXiv的文献网络间锚链接关系预测方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190108639A1 (en) * 2017-10-09 2019-04-11 The Board Of Trustees Of The Leland Stanford Junior University Systems and Methods for Semantic Segmentation of 3D Point Clouds
CN110378543A (zh) * 2018-04-12 2019-10-25 百度在线网络技术（北京）有限公司 离职风险预测方法、装置、计算机设备和存储介质
CN109214719A (zh) * 2018-11-02 2019-01-15 广东电网有限责任公司 一种基于人工智能的营销稽查分析的系统和方法
CN110009093A (zh) * 2018-12-07 2019-07-12 阿里巴巴集团控股有限公司 用于分析关系网络图的神经网络系统和方法
CN111079780A (zh) * 2019-11-06 2020-04-28 中国科学院深圳先进技术研究院 空间图卷积网络的训练方法、电子设备及存储介质

Also Published As

Publication number Publication date
CN111079780A (zh) 2020-04-28
CN111079780B (zh) 2023-06-23

Similar Documents

Publication Publication Date Title
WO2021089013A1 (zh) 空间图卷积网络的训练方法、电子设备及存储介质
WO2018099084A1 (zh) 一种神经网络模型训练方法、装置、芯片和系统
WO2020007138A1 (zh) 一种事件识别的方法、模型训练的方法、设备及存储介质
WO2020098606A1 (zh) 节点分类方法、模型训练方法、装置、设备及存储介质
CN112862093B (zh) 一种图神经网络训练方法及装置
WO2020224106A1 (zh) 基于神经网络的文本分类方法、系统及计算机设备
CN110135681B (zh) 风险用户识别方法、装置、可读存储介质及终端设备
WO2021089012A1 (zh) 图网络模型的节点分类方法、装置及终端设备
WO2021169301A1 (zh) 一种选取样本图像的方法、装置、存储介质和服务器
WO2022252458A1 (zh) 一种分类模型训练方法、装置、设备及介质
CN113128671B (zh) 一种基于多模态机器学习的服务需求动态预测方法及系统
CN116596095B (zh) 基于机器学习的碳排放量预测模型的训练方法及装置
WO2023217127A1 (zh) 一种因果关系确定方法及相关设备
EP4343616A1 (en) Image classification method, model training method, device, storage medium, and computer program
CN111159481B (zh) 图数据的边预测方法、装置及终端设备
CN111178196B (zh) 一种细胞分类的方法、装置及设备
CN110598869B (zh) 基于序列模型的分类方法、装置、电子设备
WO2021253938A1 (zh) 一种神经网络的训练方法、视频识别方法及装置
CN114254738A (zh) 双层演化的动态图卷积神经网络模型构建方法及应用
CN114528990A (zh) 一种神经网络搜索方法及系统
WO2022252694A1 (zh) 神经网络优化方法及其装置
US20230186150A1 (en) Hyperparameter selection using budget-aware bayesian optimization
CN114970732A (zh) 分类模型的后验校准方法、装置、计算机设备及介质
JP6993250B2 (ja) コンテンツ特徴量抽出装置、方法、及びプログラム
US11609936B2 (en) Graph data processing method, device, and computer program product

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20885526

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20885526

Country of ref document: EP

Kind code of ref document: A1