WO2024037354A1 - 一种基于图神经网络的模型训练方法及装置 - Google Patents
一种基于图神经网络的模型训练方法及装置 Download PDFInfo
- Publication number
- WO2024037354A1 WO2024037354A1 PCT/CN2023/111176 CN2023111176W WO2024037354A1 WO 2024037354 A1 WO2024037354 A1 WO 2024037354A1 CN 2023111176 W CN2023111176 W CN 2023111176W WO 2024037354 A1 WO2024037354 A1 WO 2024037354A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- training
- graph
- features
- node
- processed
- Prior art date
Links
- 238000012549 training Methods 0.000 title claims abstract description 425
- 238000000034 method Methods 0.000 title claims abstract description 82
- 238000013528 artificial neural network Methods 0.000 title claims description 8
- 230000004927 fusion Effects 0.000 claims description 49
- 238000000605 extraction Methods 0.000 claims description 36
- 230000006870 function Effects 0.000 claims description 24
- 238000010586 diagram Methods 0.000 claims description 9
- 238000004590 computer program Methods 0.000 claims description 6
- 230000008569 process Effects 0.000 abstract description 19
- 238000012545 processing Methods 0.000 abstract description 13
- 238000004364 calculation method Methods 0.000 description 8
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000003672 processing method Methods 0.000 description 2
- 230000006978 adaptation Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000003703 image analysis method Methods 0.000 description 1
- 238000007500 overflow downdraw method Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- 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
- G06F16/901—Indexing; Data structures therefor; Storage structures
- G06F16/9024—Graphs; Linked lists
-
- 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
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Definitions
- the present application relates to the field of data processing, and in particular to a model training method and device based on graph neural network.
- a model based on graph neural network can be trained to process data and obtain corresponding processing results.
- a GNN-based model can be used to process multimedia content to predict the tags of the multimedia content.
- embodiments of the present application provide a GNN-based model training method and device.
- embodiments of the present application provide a model training method based on graph neural network GNN, which is characterized in that the method includes:
- first training graph is a relationship graph corresponding to training multimedia content, or the first training graph is a relationship graph corresponding to training items;
- a GNN-based target model is trained.
- the method also includes:
- the position features, the structure features and the node features are fused to obtain fusion features, including:
- the position features, the structural features and the node features are fused to obtain the fusion features.
- obtaining the location features of the first training map based on the first training map includes:
- the first training image is input into the position feature extraction module to obtain the position feature corresponding to the target feature space.
- the dimension of the target feature space is lower than the dimension of the image space corresponding to the first training image.
- the location feature extraction module is trained as follows:
- parameters of the location feature extraction module are adjusted.
- determining the loss function based on the location features of each node of the second training graph corresponding to the target feature space and the second training graph includes:
- the loss function is obtained;
- the first node is any one of the nodes, and the loss term corresponding to the first node is determined in the following way:
- the loss term corresponding to the first node is obtained, the second node is a K-order neighbor node of the first node, and K is an integer greater than 1 or equal to 1.
- determining the first importance of the location feature includes:
- the first model is trained based on the location features corresponding to the third training image and the labels corresponding to the third training image.
- determining the structural features of the first training graph includes:
- a multi-order neighbor number sequence of the first training graph is obtained.
- the multi-order neighbor data sequence is used to indicate the number of neighbor nodes of each order included in each node in the first training graph.
- the structural features include the multi-order neighbor number sequence.
- obtaining the structural features of the first training graph based on the first training graph includes:
- a multi-order neighbor number sequence of the first training graph and a clustering feature of the first training graph are obtained, and the multi-order neighbor data sequence is used to indicate the number of neighbors in the first training graph.
- the multi-order neighbor number sequence and the clustering feature are fused to obtain the structural features of the first training graph.
- obtaining the second degree of importance of the structural feature based on the structural feature includes:
- the second model is trained based on the structural features corresponding to the fourth training graph and the labels corresponding to the fourth training graph.
- the method also includes:
- the graph to be processed is a relationship graph corresponding to the multimedia content to be processed, or the graph to be processed is a relationship graph corresponding to the items to be processed;
- Fusion of the location features of the graph to be processed, the structural features of the graph to be processed and the node features of the graph to be processed is performed to obtain the fusion features of the graph to be processed;
- inventions of the present application provide a model training device based on graph neural network GNN.
- the device includes:
- the first acquisition unit is used to acquire a first training graph and a label corresponding to the first training graph.
- the first training graph is a relationship graph corresponding to the training multimedia content, or the first training graph is a relationship graph corresponding to the training multimedia content.
- a first determination unit configured to obtain the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph based on the first training graph;
- a first fusion unit used to fuse the position features, the structure features and the node features to obtain fusion features
- a training unit configured to train a GNN-based target model according to the fused features and the label.
- the device also includes:
- a second determination unit configured to determine the first importance of the location feature and the second importance of the structural feature
- the first fusion unit is used for:
- the position features, the structural features and the node features are fused to obtain the fusion features.
- obtaining the location features of the first training map based on the first training map includes:
- the first training image is input into the position feature extraction module to obtain the position feature corresponding to the target feature space.
- the dimension of the target feature space is lower than the dimension of the image space corresponding to the first training image.
- the location feature extraction module is trained as follows:
- parameters of the location feature extraction module are adjusted.
- determining the loss function based on the location features of each node of the second training graph corresponding to the target feature space and the second training graph includes:
- the loss function is obtained;
- the first node is any one of the nodes, and the loss term corresponding to the first node is determined in the following way:
- the loss term corresponding to the first node is obtained, the second node is a K-order neighbor node of the first node, and K is an integer greater than 1 or equal to 1.
- determining the first importance of the location feature includes:
- the first model is trained based on the location features corresponding to the third training image and the labels corresponding to the third training image.
- determining the structural features of the first training graph includes:
- a multi-order neighbor number sequence of the first training graph is obtained.
- the multi-order neighbor data sequence is used to indicate the number of neighbor nodes of each order included in each node in the first training graph.
- the structural features include the multi-order neighbor number sequence.
- obtaining the structural features of the first training graph based on the first training graph includes:
- a multi-order neighbor number sequence of the first training graph and a clustering feature of the first training graph are obtained, and the multi-order neighbor data sequence is used to indicate the number of neighbors in the first training graph.
- the multi-order neighbor number sequence and the clustering feature are fused to obtain the structural features of the first training graph.
- obtaining the second degree of importance of the structural feature based on the structural feature includes:
- the second model is trained based on the structural features corresponding to the fourth training graph and the labels corresponding to the fourth training graph.
- the device also includes:
- the second acquisition unit is used to acquire a graph to be processed, where the graph to be processed is a relationship graph corresponding to the multimedia content to be processed, or the graph to be processed is a relationship graph corresponding to the item to be processed;
- a third determination unit configured to obtain the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed based on the graph to be processed;
- the second fusion unit is used to fuse the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed, to obtain the fusion features of the graph to be processed;
- the fourth determination unit is used to input the fusion features of the image to be processed into the target model to obtain the label of the image to be processed.
- embodiments of the present application provide a device, which includes a processor and a memory;
- the processor is configured to execute instructions stored in the memory, so that the device performs the method described in any one of the above first aspects.
- embodiments of the present application provide a computer-readable storage medium, including instructions that instruct a device to perform the method described in any one of the above first aspects.
- embodiments of the present application provide a computer program product.
- the computer program When the computer program When the product is run on a computer, it causes the computer to perform any of the methods described in the first aspect above.
- An embodiment of the present application provides a GNN-based model training method, which includes: obtaining a first training graph and a label corresponding to the first training graph, where the first training graph is a relationship graph corresponding to the training multimedia content, Alternatively, the first training diagram is a relationship diagram corresponding to the training items. After obtaining the first training graph, the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph can be obtained based on the first training graph, and the The position features of the first training graph, the structural features of the first training graph, and the node features of the first training graph are fused to obtain the fusion features of the first training graph.
- a GNN-based target model is trained based on the fusion features of the first training graph and the labels of the first training graph.
- the location features of the first training graph and the structural features of the first training graph are also considered. Therefore, , more effective information is used to train the target model, and accordingly, when the trained target model processes data, the processing results obtained will be more accurate.
- Figure 1 is a schematic flow chart of a GNN-based model training method provided by an embodiment of the present application
- Figure 2 is a schematic flow chart of a method for training a location feature extraction module provided by an embodiment of the present application
- Figure 3 is a schematic diagram of a process of training a target model provided by an embodiment of the present application.
- Figure 4 is a schematic structural diagram of a GNN-based model training device provided by an embodiment of the present application.
- the inventor of this application found through research that currently, when training a GNN-based model, model training can be performed based on the node features of the training graph and the labels of the training graph.
- some information in the training graph is often lost.
- part of the position information in the training graph will be lost.
- the current training method cannot distinguish two nodes that are far apart in the training graph but have the same neighborhood structure.
- some structural information in the training graph will be lost.
- the current training method cannot distinguish two nodes with the same calculation subgraph but different neighborhood structures.
- embodiments of the present application provide a GNN-based model training method and device.
- Figure 1 is a schematic flow chart of a GNN-based model training method provided by an embodiment of the present application.
- the method can be executed by a terminal device or by a server, which is not specifically limited in the embodiment of this application.
- the method may include the following steps: S101-S104.
- model training process is a process of multiple iterative calculations. Each iteration can adjust the parameters of the model, and the adjusted parameters participate in the next round of iterative calculations.
- Figure 1 takes the first training graph as an example to introduce a certain round of iteration process in training the target model based on GNN. It can be understood that there are many training graphs used to train the target model, and each training graph is processed in a similar manner when training the target model. After training on multiple training graphs, a target model with an accuracy that meets the requirements can be obtained.
- the first training graph is a relationship graph corresponding to training multimedia content, or the first training graph is a relationship graph corresponding to training items. .
- the first training graph may include multiple nodes, and the first training graph may reflect the association between the multiple nodes, for example, node 1 and node 2 in the first training graph. If there is an edge between them, it means that node 1 and node 2 are related.
- the embodiment of the present application does not specifically limit the label corresponding to the first training graph.
- the label corresponding to the first training graph is related to the data processing task of the target model. For example, if the target model is used to identify the category of multimedia content, the first training graph may be used to indicate the category of the multimedia content to be trained.
- the multimedia content mentioned in the embodiments of this application include but are not limited to text and/or images.
- the items mentioned in the embodiments of the present application may be, for example, commodities.
- S102 Obtain the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph based on the first training graph.
- an image analysis method can be used to analyze the first training image, thereby obtaining the location features of the first training image, the structural features of the first training image, and the location characteristics of the first training image. Node characteristics.
- "obtaining the location features of the first training map based on the first training map” can be implemented through a location feature extraction module.
- the first training image can be input into the location feature extraction module, thereby obtaining the location features corresponding to the first training image.
- the position feature extraction module is used to extract the position features of the graph. Therefore, after the first training image is input into the location feature extraction module, the location feature extraction module can output the location features of the first training image.
- the location features of the first training graph mentioned in the embodiments of this application may be features that can reflect the locations of nodes in the first training graph.
- the dimension of the image space corresponding to the first training graph is related to the size of the first training graph, or in other words, related to the number of nodes included in the first training graph.
- the location feature extraction module can be used to extract the location features of the image in the target feature space. In other words, after the first training image is input into the location feature extraction module, the location feature extraction module can Output the position features corresponding to the first training image in the target feature space.
- the dimension of the target feature space is lower than the dimension of the image space corresponding to the first training image.
- the target feature space may be a low-dimensional space. In this case, the amount of calculation required to obtain the location features of the first training map can be reduced, and even if the first training map is
- the training graph includes a large number of nodes, and the position feature extraction module can also obtain the position features of the first training graph.
- the location feature extraction module may be pre-trained.
- the training method of the location feature extraction module please refer to the description of Figure 2 below, which will not be described in detail here.
- the number of neighbor nodes of each order included in each node is a kind of structural information of the first training graph. Therefore, as an example, "obtaining the structural features of the first training graph based on the first training graph" can be implemented based on the first training graph to obtain multi-order neighbors of the first training graph.
- Quantity sequence for this case, the structural features include the multi-order neighbor quantity sequence.
- the multi-order neighbor data sequence is used to indicate the number of neighbor nodes of each order included in each node in the first training graph.
- node 1 is taken as an example for explanation.
- the first-order neighbor node of node 1 refers to the node directly connected to node 1;
- the second-order neighbor node of node 1 refers to the node connected to node 1 through an intermediate node, or in other words, the node separated from node 1 by one intermediate node;
- the k-order neighbor node of node 1 refers to the node connected to node 1 through (k-1) intermediate nodes, or in other words, it is separated from node 1 by (k-1) intermediate nodes. node.
- the first training graph can be traversed to obtain the multi-order neighbor number sequence of the first training graph.
- the clustering feature of the first training graph is also a kind of structural information of the first training graph. Therefore, in another example, "obtaining the structural features of the first training graph based on the first training graph" can be implemented based on the first training graph to obtain multiple features of the first training graph. The multi-order neighbor number sequence and the clustering features of the first training graph are obtained, and the multi-order neighbor number sequence and the clustering features are fused to obtain the structural features of the first training graph.
- the clustering features of the first training graph may include, for example, each node included in the first training graph corresponding to The number of node rings of each order.
- node 2 will be taken as an example for explanation.
- the k-order node ring of node 2 refers to the number of rings composed of node 2 and the k-order nodes of node 2 in the first training graph.
- the embodiments of this application do not specifically limit the specific implementation manner of fusing the multi-order neighbor number sequence and the clustering features.
- the multi-order neighbor number sequence and the clustering features can be fused.
- the above clustering features are superimposed (for example, superimposed on the feature dimension).
- the aforementioned "obtaining structural features of the first training graph based on the first training graph" can be performed by a structural feature extraction module.
- a node feature extraction module can be used to process the first training graph to obtain the third A node feature of the training graph.
- the node feature extraction module mentioned here may be, for example, a traditional node feature extraction module, which may determine the node features of the first training graph according to the traditional node feature determination method.
- the first training graph After obtaining the position features of the first training graph, the structural features of the first training graph, and the node features of the first training graph, the position features of the first training graph, the first training graph can be The structural features of the first training graph are fused with the node features of the first training graph to obtain the fusion features of the first training graph.
- the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph can be fused according to a preset feature fusion method to obtain the Fusion features of the first training image.
- the characteristics of the first training graph may also be considered.
- the first degree of importance of the positional features and the second degree of importance of the structural features of the first training graph are used to compare the positional features of the first training graph, the structural features of the first training graph and the first training graph.
- the node features of the graph are fused, so that the obtained fusion feature can combine the structural features of the first training graph and the importance of the position features of the first training graph, and correspondingly, the features with a high degree of importance can be
- more feature information is contributed, which accordingly enables the trained target model to obtain higher accuracy processing results when processing data.
- the aforementioned first degree of importance and the second degree of importance may be determined based on the data processing tasks of the target model. For example, the importance and structural characteristics of the location features corresponding to various data processing tasks may be determined in advance. correspondingly, the data processing tasks corresponding to the target model can be matched with the aforementioned "the importance of positional features and the importance of structural features corresponding to various data processing tasks", thereby obtaining the first importance and the second level To the extent.
- the first importance level may be determined based on location features of the first training graph.
- a first model can be pre-trained, and the first model is used to obtain the corresponding importance based on the input location features.
- the location features of the first training map can be input into the first model, thereby obtaining the first importance level.
- the first model may be a multilayer perceptron (MLP).
- MLP multilayer perceptron
- the first model may be trained based on the location features corresponding to the third training image and the labels corresponding to the third training image.
- the determination method can refer to the determination method of the location features of the first training map. For details, please refer to the description of the location features of the first training map above. The description will not be repeated here. .
- the second degree of importance may be determined based on structural features of the first training graph.
- a second model can be pre-trained, and the second model is used to obtain the corresponding importance based on the input structural features.
- the structural features of the first training graph can be input into the second model, thereby obtaining the second importance level.
- the second model may be an MLP.
- the second model may be trained based on the structural features corresponding to the fourth training graph and the labels corresponding to the fourth training graph.
- the determination method can refer to the determination method of the structural features of the first training graph.
- the determination method can refer to the determination method of the structural features of the first training graph.
- S104 Train a GNN-based target model based on the fusion features of the first training graph and the labels of the first training graph.
- a GNN-based target model can be trained based on the fusion features of the first training graph and the labels of the first training graph. For example, a model prediction result can be obtained based on the fusion features of the first training graph, and then the parameters of the target module can be adjusted based on the model prediction result and the label corresponding to the first training graph.
- the location features of the first training graph and the first training graph are also considered Structural characteristics, therefore, more effective information is used to train the target model.
- the trained target model processes data, the processing results obtained will be more accurate.
- Figure 2 is a schematic flowchart of a method for training a location feature extraction module provided by an embodiment of the present application.
- the method shown in Figure 2 may include, for example, the following S201-S204.
- the second training graph can be input into the position feature extraction module being trained, thereby obtaining the corresponding nodes of the second training graph output by the position feature extraction module.
- the position characteristics of the target feature space can be obtained by the position feature extraction module.
- S203 Determine a loss function based on the location features of each node of the second training graph corresponding to the target feature space and the second training graph.
- the distance between any two nodes of the second training graph in the target feature space can be determined based on the position characteristics of each node of the second training graph corresponding to the target feature space, And based on the second training graph, determine the distance between any two nodes in the second training graph, and then, based on the distance between any two nodes in the target feature space and the distance between any two nodes in the The distance in the second training graph determines the loss function.
- the loss function may be obtained based on the loss terms corresponding to each node in the second training graph. For example, the loss terms corresponding to each node can be summed to obtain the loss function.
- the first node is any one of the nodes, it can be based on the position characteristics of the first node corresponding to the target feature space, the position characteristics of the second node corresponding to the target feature space, and the first node.
- the distance between a node and the second node in the second training graph is used to obtain the loss term corresponding to the first node.
- the second node is a K-order neighbor node of the first node, and K is An integer greater than or equal to 1.
- the location characteristics of the first node corresponding to the target feature space, the location characteristics of the second node corresponding to the target feature space, and the location of the first node and the second node in the The distance in the second training graph is used to obtain the loss term corresponding to the first node.
- it can be based on the location feature of the first node corresponding to the target feature space and the second node corresponding to the target feature.
- Position characteristics of the space determine the distance between the first node and the second node in the target feature space, and further, based on the sum of the distance between the first node and the second node in the target feature space The distance between the first node and the second node in the second training graph is used to obtain the loss term corresponding to the first node.
- the embodiment of the present application does not specifically limit the specific value of K.
- the specific value of K may, for example, be determined based on actual conditions.
- the loss term corresponding to the first node can be calculated by the following formula (1).
- L los [(1-d cos (f pos (v i ),f pos (v j ))/2-(1-1/d spd (v i ,v j ))] 2
- L los is the loss term of the first node
- f pos (v i ) is the position feature of node i in the target feature space, and node i is the first node;
- f pos (v j ) is the position feature of node j in the target feature space, and node j is the second node;
- d cos (f pos (v i ), f pos (v j ) is the distance between node i and the node j in the target feature space;
- d spd (v i ,v j ) is the distance between node i and node j in the second training graph, and the value of d spd (v i ,v j ) is k.
- FIG. 3 is a schematic diagram of a process of training a target model provided by an embodiment of the present application.
- the location feature extraction module may process the first training image to obtain the location features of the first training image, and the first model may process the location features of the first training image to obtain the first importance level.
- the node feature extraction module may process the first training graph to obtain node features of the first training graph.
- the structural feature extraction module may process the first training image to obtain the structural features of the first training image
- the second model may process the structural features of the first training image to obtain the second degree of importance
- the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph are fused. , obtain the fusion features of the first training image.
- a GNN-based target model is trained based on the fusion features of the first training graph and the labels of the first training graph.
- the graph to be processed can be processed based on the target model to obtain the label of the graph to be processed.
- the following steps A1-A4 can be performed to obtain the label corresponding to the graph to be processed.
- Step A1 Obtain a graph to be processed, where the graph to be processed is a relationship graph corresponding to the multimedia content to be processed, or the graph to be processed is a relationship graph corresponding to the item to be processed.
- Step A2 Obtain the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed based on the graph to be processed.
- Step A3 Fusion of the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed, to obtain the fusion features of the graph to be processed.
- processing method of the graph to be processed is similar to the processing method of the first training graph when training the target model. Therefore, regarding steps A1-A3, the specific implementation method can refer to steps S101-S103. Herein No detailed description is given.
- Step A4 Input the fusion features of the image to be processed into the target model to obtain the label of the image to be processed.
- the fusion features of the graph to be processed can be input into the target model, thereby obtaining the label of the graph to be processed output by the target model.
- training the target model in addition to the node characteristics of the first training graph, the location characteristics of the first training graph and the structural characteristics of the first training graph are also considered, therefore, training the The target model uses more effective information.
- the trained target model processes the image to be processed, the label of the image to be processed will be more accurate.
- the embodiment of the present application also provides a device.
- the device will be introduced below with reference to the accompanying drawings.
- the device 400 may specifically include: a first acquisition unit 401, a first determination unit 402, a first fusion unit 403, and a training unit 404.
- the first acquisition unit 401 is used to acquire a first training graph and a label corresponding to the first training graph.
- the first training graph is a relationship graph corresponding to the training multimedia content, or the first training graph is a relationship graph corresponding to the training multimedia content. Relationship diagram corresponding to training items;
- the first determining unit 402 is configured to obtain the location features of the first training graph, the structural features of the first training graph, and the node features of the first training graph based on the first training graph;
- the first fusion unit 403 is used to fuse the position features, the structure features and the node features to obtain fusion features;
- a training unit 404 is configured to train a GNN-based target model according to the fused features and the label.
- the device also includes:
- a second determination unit configured to determine the first importance of the location feature, and determine the The second degree of importance of structural characteristics
- the first fusion unit 403 is used for:
- the position features, the structural features and the node features are fused to obtain the fusion features.
- obtaining the location features of the first training map based on the first training map includes:
- the first training image is input into the position feature extraction module to obtain the position feature corresponding to the target feature space.
- the dimension of the target feature space is lower than the dimension of the image space corresponding to the first training image.
- the location feature extraction module is trained as follows:
- parameters of the location feature extraction module are adjusted.
- determining the loss function based on the location features of each node of the second training graph corresponding to the target feature space and the second training graph includes:
- the loss function is obtained;
- the first node is any one of the nodes, and the loss term corresponding to the first node is determined in the following way:
- the loss term corresponding to the first node is obtained, the second node is a K-order neighbor node of the first node, and K is an integer greater than 1 or equal to 1.
- determining the first importance of the location feature includes:
- the first model is trained based on the location features corresponding to the third training image and the labels corresponding to the third training image.
- determining the structural features of the first training graph includes:
- a multi-order neighbor number sequence of the first training graph is obtained, and the multi-order neighbor number sequence is obtained.
- the order neighbor data sequence is used to indicate the number of neighbor nodes of each order included in each node in the first training graph, and the structural feature includes the multi-order neighbor number sequence.
- obtaining the structural features of the first training graph based on the first training graph includes:
- a multi-order neighbor number sequence of the first training graph and a clustering feature of the first training graph are obtained, and the multi-order neighbor data sequence is used to indicate the number of neighbors in the first training graph.
- the multi-order neighbor number sequence and the clustering feature are fused to obtain the structural features of the first training graph.
- obtaining the second degree of importance of the structural feature based on the structural feature includes:
- the second model is trained based on the structural features corresponding to the fourth training graph and the labels corresponding to the fourth training graph.
- the device also includes:
- the second acquisition unit is used to acquire a graph to be processed, where the graph to be processed is a relationship graph corresponding to the multimedia content to be processed, or the graph to be processed is a relationship graph corresponding to the item to be processed;
- a third determination unit configured to obtain the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed based on the graph to be processed;
- the second fusion unit is used to fuse the location features of the graph to be processed, the structural features of the graph to be processed, and the node features of the graph to be processed, to obtain the fusion features of the graph to be processed;
- the fourth determination unit is used to input the fusion features of the image to be processed into the target model to obtain the label of the image to be processed.
- the device 400 is a device corresponding to the method provided by the above method embodiment, the specific implementation of each unit of the device 400 is the same concept as the above method embodiment. Therefore, the details about each unit of the device 400 For specific implementation, reference may be made to the description of the above method embodiments, which will not be described again here.
- An embodiment of the present application also provides a device, which includes a processor and a memory;
- the processor is configured to execute instructions stored in the memory, so that the device executes the GNN-based model training method described in any one of the above method embodiments.
- Embodiments of the present application provide a computer-readable storage medium, including instructions, where the instructions indicate The device executes the GNN-based model training method described in any one of the above method embodiments.
- Embodiments of the present application provide a computer program product.
- the computer program product When the computer program product is run on a computer, it causes the computer to execute the GNN-based model training method described in any one of the above method embodiments.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Databases & Information Systems (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
本申请公开了一种基于GNN的模型训练方法，该包括：获取第一训练图和第一训练图对应的标签，第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图。获得所述第一训练图之后，可以基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征，并对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。进一步地，基于所述第一训练图的融合特征和所述第一训练图的标签，训练基于GNN的目标模型。利用本方案所训练得到的目标模型在对数据进行处理时，所得到的处理结果也会更加准确。
Description
本申请要求于2022年8月18日提交中国国家知识产权局、申请号为202210994004.0、申请名称为“一种基于图神经网络的模型训练方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
本申请涉及数据处理领域，特别是涉及一种基于图神经网络的模型训练方法及装置。
可以训练基于图神经网络(graph neural network，GNN)的模型来对数据进行处理，得到对应的处理结果。例如，可以利用基于GNN的模型对多媒体内容进行处理，从而预测多媒体内容的标签。
但是，目前所训练得到的基于GNN的模型在对数据进行处理时，所得到的处理结果往往不是特别准确，因此，急需一种方案，能够解决上述问题。
发明内容
为了解决或者部分解决上述技术问题，本申请实施例提供了一种基于GNN的模型训练方法及装置。
第一方面，本申请实施例提供了一种基于图神经网络GNN的模型训练方法，其特征在于，所述方法包括：
获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图；
基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征；
对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征；
根据所述融合特征和所述标签，训练基于GNN的目标模型。
可选的，所述方法还包括：
确定所述位置特征的第一重要程度，以及确定所述结构特征的第二重要程度；
对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征，包括：
基于所述第一重要程度和所述第二重要程度，对所述位置特征、所述结构特征以及所述节点特征进行融合，得到所述融合特征。
可选的，所述基于所述第一训练图得到所述第一训练图的位置特征，包括：
将所述第一训练图输入位置特征提取模块，得到在目标特征空间对应的所述位置特征，所述目标特征空间的维度低于所述第一训练图对应的图像空间的维度。
可选的，所述位置特征提取模块，通过如下方式训练得到：
获取第二训练图；
获取所述第二训练图的各个节点对应所述目标特征空间的位置特征；
基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数；
基于所述损失函数，调整所述位置特征提取模块的参数。
可选的，所述基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数，包括：
基于所述各个节点分别对应的损失项，得到所述损失函数；其中：
第一节点为所述各个节点中的任意一个节点，所述第一节点对应的损失项，通过如下方式确定：
基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项，所述第二节点为所述第一节点的K阶邻居节点，所述K为大于1或者等于1的整数。
可选的，所述确定所述位置特征的第一重要程度，包括：
将所述位置特征输入第一模型，得到所述第一重要程度，所述第一模型用于基于输入的位置特征，得到对应的重要程度；其中：
所述第一模型是基于第三训练图对应的位置特征和所述第三训练图对应的标签训练得到的。
可选的，所述确定所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量，所述结构特征包括所述多阶邻居数量序列。
可选的，所述基于所述第一训练图得到所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列以及所述第一训练图的聚类特征，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量；
对所述多阶邻居数量序列和所述聚类特征进行融合，得到所述第一训练图的结构特征。
可选的，所述基于所述结构特征得到所述结构特征的第二重要程度，包括：
将所述结构特征输入第二模型，得到所述第二重要程度，所述第二模型用于基于输入的结构特征，得到对应的重要程度；其中：
所述第二模型，是基于第四训练图对应的结构特征和所述第四训练图对应的标签训练得到的。
可选的，所述方法还包括：
获取待处理图，所述待处理图为与待处理多媒体内容对应的关系图，或者，所述待处理图为与待处理物品对应的关系图；
基于所述待处理图得到所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征；
对所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征进行融合，得到所述待处理图的融合特征；
将所述待处理图的融合特征输入所述目标模型，得到所述待处理图的标签。
第二方面，本申请实施例提供了一种基于图神经网络GNN的模型训练装置，所述装置包括：
第一获取单元，用于获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图；
第一确定单元，用于基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征；
第一融合单元，用于对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征；
训练单元，用于根据所述融合特征和所述标签，训练基于GNN的目标模型。
可选的，所述装置还包括：
第二确定单元，用于确定所述位置特征的第一重要程度，以及确定所述结构特征的第二重要程度；
所述第一融合单元，用于：
基于所述第一重要程度和所述第二重要程度，对所述位置特征、所述结构特征以及所述节点特征进行融合，得到所述融合特征。
可选的，所述基于所述第一训练图得到所述第一训练图的位置特征，包括：
将所述第一训练图输入位置特征提取模块，得到在目标特征空间对应的所述位置特征，所述目标特征空间的维度低于所述第一训练图对应的图像空间的维度。
可选的，所述位置特征提取模块，通过如下方式训练得到：
获取第二训练图；
获取所述第二训练图的各个节点对应所述目标特征空间的位置特征；
基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数；
基于所述损失函数，调整所述位置特征提取模块的参数。
可选的，所述基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数，包括：
基于所述各个节点分别对应的损失项，得到所述损失函数；其中：
第一节点为所述各个节点中的任意一个节点，所述第一节点对应的损失项，通过如下方式确定：
基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项，所述第二节点为所述第一节点的K阶邻居节点，所述K为大于1或者等于1的整数。
可选的，所述确定所述位置特征的第一重要程度，包括：
将所述位置特征输入第一模型，得到所述第一重要程度，所述第一模型用于基于输入的位置特征，得到对应的重要程度；其中：
所述第一模型是基于第三训练图对应的位置特征和所述第三训练图对应的标签训练得到的。
可选的，所述确定所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量，所述结构特征包括所述多阶邻居数量序列。
可选的，所述基于所述第一训练图得到所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列以及所述第一训练图的聚类特征，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量；
对所述多阶邻居数量序列和所述聚类特征进行融合，得到所述第一训练图的结构特征。
可选的，所述基于所述结构特征得到所述结构特征的第二重要程度，包括：
将所述结构特征输入第二模型，得到所述第二重要程度，所述第二模型用于基于输入的结构特征，得到对应的重要程度；其中：
所述第二模型，是基于第四训练图对应的结构特征和所述第四训练图对应的标签训练得到的。
可选的，所述装置还包括：
第二获取单元，用于获取待处理图，所述待处理图为与待处理多媒体内容对应的关系图，或者，所述待处理图为与待处理物品对应的关系图；
第三确定单元，用于基于所述待处理图得到所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征；
第二融合单元，用于对所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征进行融合，得到所述待处理图的融合特征；
第四确定单元，用于将所述待处理图的融合特征输入所述目标模型，得到所述待处理图的标签。
第三方面，本申请实施例提供了一种设备，所述设备包括处理器和存储器；
所述处理器用于执行所述存储器中存储的指令，以使得所述设备执行以上第一方面任一项所述的方法。
第四方面，本申请实施例提供了一种计算机可读存储介质，包括指令，所述指令指示设备执行以上第一方面任一项所述的方法。
第五方面，本申请实施例提供了一种计算机程序产品，当所述计算机程序
产品在计算机上运行时，使得计算机执行以上第一方面任一项所述的方法。
与现有技术相比，本申请实施例具有以下优点：
本申请实施例提供了一种基于GNN的模型训练方法，该包括：获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图。获得所述第一训练图之后，可以基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征，并对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。进一步地，基于所述第一训练图的融合特征和所述第一训练图的标签，训练基于GNN的目标模型。在本申请实施例中，在训练所述目标模型时，除了考虑了第一训练图的节点特征之外，还考虑了第一训练图的位置特征和所述第一训练图的结构特征，因此，训练所述目标模型所使用的有效信息更多，相应的，所训练得到的目标模型在对数据进行处理时，所得到的处理结果也会更加准确。
为了更清楚地说明本申请实施例或现有技术中的技术方案，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中记载的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例提供的一种基于GNN的模型训练方法的流程示意图；
图2为本申请实施例提供的一种训练位置特征提取模块的方法的流程示意图；
图3本申请实施例提供的一种训练目标模型的过程示意图；
图4为本申请实施例提供的一种基于GNN的模型训练装置的结构示意图。
为了使本技术领域的人员更好地理解本申请方案，下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
本申请的发明人经过研究发现，目前，在训练基于GNN的模型时，可以基于训练图的节点特征和训练图的标签来进行模型训练。但是，采用这种方式，往往会丢失训练图中的部分信息。在一个示例中，采用当前的训练方式，会丢失训练图中的部分位置信息，例如，当前的训练方式，无法区分在训练图中位置较远但具有相同邻域结构的两个节点。在又一个示例中，采用当前的训练方式，会丢失训练图中的部分结构信息，例如，当前的训练方式，无法区分具有相同计算子图但有不同邻域结构的两个节点。
为了解决上述问题，本申请实施例提供了一种基于GNN的模型训练方法及装置。
下面结合附图，详细说明本申请的各种非限制性实施方式。
示例性方法
参见图1，该图为本申请实施例提供的一种基于GNN的模型训练方法的流程示意图。在本实施例中，所述方法可以由终端设备执行，也可以由服务器执行，本申请实施例不做具体限定。
在一个示例中，所述方法例如可以包括以下步骤：S101-S104。
需要说明的是，模型训练的过程是一个多次迭代计算的过程，每一次迭代都可以对模型的参数进行调整，调整后的参数参与下一轮迭代计算。
图1以第一训练图为例，对训练基于GNN的目标模型中的某一轮迭代过程进行介绍。可以理解的是，训练所述目标模型所使用的训练图有很多，在训练所述目标模型时，每个训练图的处理方式类似。在经过多个训练图训练之后，即可得到准确度符合要求的目标模型。
S101：获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图。
在本申请实施例中，所述第一训练图可以包括多个节点，该第一训练图可以体现所述多个节点之间的关联关系，例如，第一训练图中的节点1和节点2之间具备一条边，则说明节点1和节点2具备关联关系。
本申请实施例不具体限定所述第一训练图对应的标签，所述第一训练图对应的标签，与所述目标模型的数据处理任务相关。例如，所述目标模型用于识别多媒体内容的类别，则所述第一训练图可以用于指示训练多媒体内容的类别。
本申请实施例提及的多媒体内容(例如此处提及的训练多媒体内容以及下文提及的待处理多媒体内容)，包括但不限于文本和/或图像。
在一个示例中，本申请实施例中提及的物品(例如此处提及的训练物品以及下文提及的待处理物品)，例如可以是商品。
S102：基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征。
在一个示例中，可以采用图像分析方法，对所述第一训练图进行分析，从而得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征。
在一个示例中，“基于所述第一训练图得到所述第一训练图的位置特征”在具体实现时，可以通过位置特征提取模块来实现。具体地，可以将所述第一训练图输入位置特征提取模块，从而得到所述第一训练图对应的位置特征。其中，所述位置特征提取模块，用于提取图的位置特征。因此，将所述第一训练图输入所述位置特征提取模块之后，所述位置特征提取模块可以输出所述第一训练图的位置特征。本申请实施例中提及的第一训练图的位置特征，可以是能够体现所述第一训练图中的节点的位置的特征。
在一个示例中，考虑到对于第一训练图而言，第一训练图对应的图像空间的维度，与第一训练图的大小相关，或者说，与第一训练图包括的节点数量相关。所述第一训练图包括的节点数量越多，则所述图像空间的维度越高。因此，若所述第一训练图的位置特征为在所述图像空间的特征，则当所述第一训练图的节点数量较多时，确定所述第一训练图的位置特征所消耗的计算量较大，甚至，无法计算出所述第一训练图在所述图像空间的位置特征。鉴于此，在一个示例中，该位置特征提取模块，可以用于提取图在目标特征空间的位置特征，换言之，将所述第一训练图输入位置特征提取模块之后，所述位置特征提取模块可以输出第一训练图在目标特征空间对应的位置特征。
其中，所述目标特征空间的维度低于所述第一训练图对应的图像空间的维度。换言之，在一个示例中，所述目标特征空间，可以是一个低维空间，对于这种情况，可以减小计算得到所述第一训练图的位置特征的计算量，并且，即使所述第一训练图所包括的节点数量较多，所述位置特征提取模块，也能够得到所述第一训练图的位置特征。
在一个示例中，所述位置特征提取模块，可以是预先训练得到的，关于所述位置特征提取模块的训练方式，可以参考下文对于图2的描述部分，此处不做详细描述。
在一个示例中，考虑到对于第一训练图而言，其各个节点包括的各阶邻居节点的数量，是所述第一训练图的一种结构信息。因此，作为一个示例，“基于所述第一训练图得到所述第一训练图的结构特征”在具体实现时，可以基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列，对于这种情况，所述结构特征包括所述多阶邻居数量序列。其中：所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量。
关于节点的各阶邻居节点，现以节点1为例进行说明。
节点1的1阶邻居节点，指的是与节点1直连的节点；
节点1的2阶邻居节点，指的是与节点1通过一个中间节点相连的节点，或者说，是与节点1之间相隔1个中间节点的节点；
以此类推，节点1的k阶邻居节点，指的是与节点1通过(k-1)个中间节点相连的节点，或者说，是与节点1之间相隔(k-1)个中间节点的节点。
在一个示例中，可以遍历所述第一训练图，从而得到所述第一训练图的所述多阶邻居数量序列。
在一个示例中，所述第一训练图的聚类特征，也是所述第一训练图的一种结构信息。因此，在另一个示例中，“基于所述第一训练图得到所述第一训练图的结构特征”在具体实现时，可以基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列以及所述第一训练图的聚类特征，并对所述多阶邻居数量序列和所述聚类特征进行融合，得到所述第一训练图的结构特征。
关于所述第一训练图的聚类特征，需要说明的是，在一个示例中，所述第一训练图的聚类特征，例如可以包括所述第一训练图中所包括的各个节点分别对应的各阶节点环的数量。
关于节点的各阶节点换，现以节点2为例进行说明。
节点2的k阶节点环，指的是所述第一训练图中、由节点2和节点2的k阶节点所构成的环的数量。
本申请实施例不具体限定对所述多阶邻居数量序列和所述聚类特征进行融合的具体实现方式，在一个示例中，例如可以对所述多阶邻居数量序列和所
述聚类特征进行叠加(例如在特征维度上进行叠加)。
在一个示例中，前述“基于所述第一训练图得到所述第一训练图的结构特征”，可以由结构特征提取模块执行。
在一个示例中，“基于所述第一训练图得到所述第一训练图的节点特征”在具体实现时，例如可以利用节点特征提取模块对所述第一训练图进行处理，得到所述第一训练图的节点特征。此处提及的节点特征提取模块，例如可以是传统的节点特征提取模块，其可以按照传统的节点特征确定方式，确定所述第一训练图的节点特征。
S103：对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。
得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征之后，可以对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。
在一个示例中，可以按照预设的特征融合方式，对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。
在又一个示例中，在所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合时，还可以考虑所述第一训练图的位置特征的第一重要程度以及所述第一训练图的结构特征的第二重要程度，来对所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征进行融合，从而使得所得到的融合特征，能够结合所述第一训练图的结构特征的和所述第一训练图的位置特征的重要程度，相应的，使得重要程度高的特征在训练目标模型时，贡献更多的特征信息，相应的，能够使得训练得到的目标模型在对数据进行处理时，所得到的的处理结果的准确度更高。
在一个示例中，前述第一重要程度和所述第二重要程度可以是基于目标模型的数据处理任务确定的，例如，可以预先确定各种数据处理任务分别对应的位置特征的重要程度和结构特征的重要程度，相应的，可以将所述目标模型对应的数据处理任务与前述“各种数据处理任务分别对应的位置特征的重要程度和结构特征的重要程度”进行匹配，从而得到所述第一重要程度和所述第二重
要程度。
在又一个示例中，可以基于所述第一训练图的位置特征确定所述第一重要程度。作为一个示例，可以预先训练第一模型，所述第一模型用于基于输入的位置特征，得到对应的重要程度。对于这种情况，可以将所述第一训练图的位置特征输入所述第一模型，从而得到所述第一重要程度。
本申请实施例不具体限定所述第一模型，作为一个示例，所述第一模型例如可以是多层感知机(multilayer perceptron，MLP)。
在一个示例中，所述第一模型，例如可以是基于第三训练图对应的位置特征和所述第三训练图对应的标签训练得到的。
关于所述第三训练图，可以参考上文对于第一训练图的描述部分，此处不做详述。
关于所述第三训练图的位置特征，其确定方式可以参考第一训练图的位置特征的确定方式，具体可参考上文对于第一训练图的位置特征的描述部分，此处不做重复描述。
关于所述第三训练图的标签，可以参考上文对于第一训练图的标签的描述部分，此处不做详述。
在另一个示例中，可以基于所述第一训练图的结构特征确定所述第二重要程度。作为一个示例，可以预先训练第二模型，所述第二模型用于基于输入的结构特征，得到对应的重要程度。对于这种情况，可以将所述第一训练图的结构特征输入所述第二模型，从而得到所述第二重要程度。
本申请实施例不具体限定所述第二模型，作为一个示例，所述第二模型例如可以是MLP。
在一个示例中，所述第二模型，例如可以是基于第四训练图对应的结构特征和所述第四训练图对应的标签训练得到的。
关于所述第四训练图，可以参考上文对于第一训练图的描述部分，此处不做详述。
关于所述第四训练图的结构特征，其确定方式可以参考第一训练图的结构特征的确定方式，具体可参考上文对于第一训练图的结构特征的描述部分，此处不做重复描述。
关于所述第四训练图的标签，可以参考上文对于第一训练图的标签的描述
部分，此处不做详述。
S104：根据所述第一训练图的融合特征和所述第一训练图的标签，训练基于GNN的目标模型。
得到所述第一训练图的融合特征之后，可以基于所述第一训练图的融合特征和所述第一训练图的标签，训练基于GNN的目标模型。例如，可以基于所述第一训练图的融合特征，得到模型预测结果，而后，基于所述模型预测结果和所述第一训练图对应的标签，调整所述目标模块的参数。
通过以上描述可知，在本申请实施例中，在训练所述目标模型时，除了考虑了第一训练图的节点特征之外，还考虑了第一训练图的位置特征和所述第一训练图的结构特征，因此，训练所述目标模型所使用的有效信息更多，相应的，所训练得到的目标模型在对数据进行处理时，所得到的处理结果也会更加准确。
接下来，结合图2，对前述位置特征提取模块的训练方法进行介绍。参见图2，该图为本申请实施例提供的一种训练位置特征提取模块的方法的流程示意图。
图2所示的方法，例如可以包括如下S201-S204。
S201：获取第二训练图。
关于所述第二训练图，可以参考上文对于第一训练图的描述部分，此处不做重复描述。
S202：获取所述第二训练图的各个节点对应所述目标特征空间的位置特征。
获取所述第二训练图之后，可以将所述第二训练图输入正在训练的所述位置特征提取模块中，从而得到所述位置特征提取模块所输出的所述第二训练图的各个节点对应所述目标特征空间的位置特征。
S203：基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数。
S204：基于所述损失函数，调整所述位置特征提取模块的参数。
S203在具体实现时，例如可以基于所述第二训练图的各个节点对应所述目标特征空间的位置特征，确定所述第二训练图在任意两个节点在所述目标特征空间中的距离，并根据所述第二训练图，确定任意两个节点在所述第二训练图中的距离，而后，基于任意两个节点在所述目标特征空间中的距离和该任意两个节点在所述第二训练图中的距离，确定损失函数。
在一个示例中，考虑到对于第二训练图而言，若所述第二训练图包括的节点数量众多，则对于任意两个节点，均计算该两个节点在所述第二训练图中的距离以及在目标特征空间中的距离，则确定所述损失函数的计算量则比较大。为了减少确定所述损失函数的计算量，S203在具体实现时，例如可以根据第二训练图中各个节点分别对应的损失项，得到所述损失函数。例如，可以对所述各个节点分别对应的损失项进行求和，从而得到所述损失函数。
其中：
第一节点为所述各个节点中的任意一个节点，则可以基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项，所述第二节点为所述第一节点的K阶邻居节点，所述K为大于1或者等于1的整数。
采用这种情况，对于第一节点而言，无需计算第一节点和所述第二训练图中各个节点分别在第二训练图以及在所述目标特征空间中的距离。由于第二节点是第一节点的K阶邻居节点，因此，第一节点和第二节点在第二训练图中的距离为K。因此，采用本方案，计算第一节点和第一节点的K阶邻居节点在所述目标特征空间的距离即可，有效减少了确定所述损失函数的计算量。
在一个示例中，“基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项”在具体实现时，例如可以基于所述第一节点对应所述目标特征空间的位置特征和第二节点对应所述目标特征空间的位置特征，确定所述第一节点和所述第二节点在所述目标特征空间的距离，进一步地，基于所述第一节点和所述第二节点在所述目标特征空间的距离和所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项。
本申请实施例不具体限定所述K的具体取值，所述K的具体取值例如可以根据实际情况确定。
在一个示例中，第一节点对应的损失项可以通过如下公式(1)计算得到。
Llos＝[(1-dcos(fpos(vi),fpos(vj))/2-(1-1/dspd(vi,vj))]2 公式(1)
Llos＝[(1-dcos(fpos(vi),fpos(vj))/2-(1-1/dspd(vi,vj))]2 公式(1)
在公式(1)中：
Llos为第一节点的损失项；
fpos(vi)为节点i在目标特征空间的位置特征，节点i为第一节点；
fpos(vj)为节点j在目标特征空间的位置特征，节点j为第二节点；
dcos(fpos(vi),fpos(vj)为节点i和所述节点j在目标特征空间的距离；
dspd(vi,vj)为节点i和节点j在第二训练图中的距离，dspd(vi,vj)的值为k。
接下来，结合图3，对所述目标模型的训练过程进行介绍。
参见图3，该图为本申请实施例提供的一种训练目标模型的过程示意图。
如图3所示：
可以由位置特征提取模块对第一训练图进行处理，得到第一训练图的位置特征，并由第一模型对所述第一训练图的位置特征进行处理，得到第一重要程度。
另外，可以由节点特征提取模块对所述第一训练图进行处理，得到第一训练图的节点特征。
另外，可以由结构特征提取模块对第一训练图进行处理，得到第一训练图的结构特征，并由第二模型对所述第一训练图的结构特征进行处理，得到第二重要程度。
而后，基于所述第一重要程度和所述第二重要程度，对所述第一训练图的位置特征、所述第一训练图的结构特征、以及所述第一训练图的节点特征进行融合，得到所述第一训练图的融合特征。
进一步地，基于所述第一训练图的融合特征和所述第一训练图的标签，训练基于GNN的目标模型。
在一个示例中，训练得到所述目标模型之后，可以基于所述目标模型对待处理图进行处理，得到所述待处理图的标签。在一个示例中，可以执行如下步骤A1-A4，从而得到待处理图对应的标签。
步骤A1：获取待处理图，所述待处理图为与待处理多媒体内容对应的关系图，或者，所述待处理图为与待处理物品对应的关系图。
步骤A2：基于所述待处理图得到所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征。
步骤A3：对所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征进行融合，得到所述待处理图的融合特征。
需要说明的是，对待处理图的处理方式，与在训练目标模型时对第一训练图的处理方式是类似的，因此，关于步骤A1-A3，其具体实现方式可以参考步骤S101-S103，此处不做详细描述。
步骤A4：将所述待处理图的融合特征输入所述目标模型，得到所述待处理图的标签。
得到待处理图的融合特征之后，可以将所述待处理图的融合特征输入所述目标模型，从而得到所述目标模型输出的所述待处理图的标签。
正是由于在训练所述目标模型时，除了考虑了第一训练图的节点特征之外，还考虑了第一训练图的位置特征和所述第一训练图的结构特征，因此，训练所述目标模型所使用的有效信息更多，相应的，所训练得到的目标模型在对待处理图进行处理时，所得到的待处理图的标签也会更加准确。
示例性设备
基于以上实施例提供的方法，本申请实施例还提供了一种装置，以下结合附图介绍该装置。
参见图4，该图为本申请实施例提供的一种基于GNN的模型训练装置的结构示意图。所述装置400例如可以具体包括：第一获取单元401、第一确定单元402、第一融合单元403以及训练单元404。
第一获取单元401，用于获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图；
第一确定单元402，用于基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征；
第一融合单元403，用于对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征；
训练单元404，用于根据所述融合特征和所述标签，训练基于GNN的目标模型。
可选的，所述装置还包括：
第二确定单元，用于确定所述位置特征的第一重要程度，以及确定所述结
构特征的第二重要程度；
所述第一融合单元403，用于：
基于所述第一重要程度和所述第二重要程度，对所述位置特征、所述结构特征以及所述节点特征进行融合，得到所述融合特征。
可选的，所述基于所述第一训练图得到所述第一训练图的位置特征，包括：
将所述第一训练图输入位置特征提取模块，得到在目标特征空间对应的所述位置特征，所述目标特征空间的维度低于所述第一训练图对应的图像空间的维度。
可选的，所述位置特征提取模块，通过如下方式训练得到：
获取第二训练图；
获取所述第二训练图的各个节点对应所述目标特征空间的位置特征；
基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数；
基于所述损失函数，调整所述位置特征提取模块的参数。
可选的，所述基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数，包括：
基于所述各个节点分别对应的损失项，得到所述损失函数；其中：
第一节点为所述各个节点中的任意一个节点，所述第一节点对应的损失项，通过如下方式确定：
基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项，所述第二节点为所述第一节点的K阶邻居节点，所述K为大于1或者等于1的整数。
可选的，所述确定所述位置特征的第一重要程度，包括：
将所述位置特征输入第一模型，得到所述第一重要程度，所述第一模型用于基于输入的位置特征，得到对应的重要程度；其中：
所述第一模型是基于第三训练图对应的位置特征和所述第三训练图对应的标签训练得到的。
可选的，所述确定所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列，所述多
阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量，所述结构特征包括所述多阶邻居数量序列。
可选的，所述基于所述第一训练图得到所述第一训练图的结构特征，包括：
基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列以及所述第一训练图的聚类特征，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量；
对所述多阶邻居数量序列和所述聚类特征进行融合，得到所述第一训练图的结构特征。
可选的，所述基于所述结构特征得到所述结构特征的第二重要程度，包括：
将所述结构特征输入第二模型，得到所述第二重要程度，所述第二模型用于基于输入的结构特征，得到对应的重要程度；其中：
所述第二模型，是基于第四训练图对应的结构特征和所述第四训练图对应的标签训练得到的。
可选的，所述装置还包括：
第二获取单元，用于获取待处理图，所述待处理图为与待处理多媒体内容对应的关系图，或者，所述待处理图为与待处理物品对应的关系图；
第三确定单元，用于基于所述待处理图得到所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征；
第二融合单元，用于对所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征进行融合，得到所述待处理图的融合特征；
第四确定单元，用于将所述待处理图的融合特征输入所述目标模型，得到所述待处理图的标签。
由于所述装置400是与以上方法实施例提供的方法对应的装置，所述装置400的各个单元的具体实现，均与以上方法实施例为同一构思，因此，关于所述装置400的各个单元的具体实现，可以参考以上方法实施例的描述部分，此处不再赘述。
本申请实施例还提供了一种设备，所述设备包括处理器和存储器；
所述处理器用于执行所述存储器中存储的指令，以使得所述设备执行以上方法实施例任一项所述的基于GNN的模型训练方法。
本申请实施例提供了一种计算机可读存储介质，包括指令，所述指令指示
设备执行以上方法实施例任一项所述的基于GNN的模型训练方法。
本申请实施例提供了一种计算机程序产品，当所述计算机程序产品在计算机上运行时，使得计算机执行以上方法实施例任一项所述的基于GNN的模型训练方法。
本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本申请的其它实施方案。本申请旨在涵盖本申请的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本申请的一般性原理并包括本公开未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本申请的真正范围和精神由下面的权利要求指出。
应当理解的是，本申请并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本申请的范围仅由所附的权利要求来限制。
以上所述仅为本申请的较佳实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本申请的保护范围之内。
Claims (14)
- 一种基于图神经网络GNN的模型训练方法，其特征在于，所述方法包括：获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图；基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征；对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征；根据所述融合特征和所述标签，训练基于GNN的目标模型。
- 根据权利要求1所述的方法，其特征在于，所述方法还包括：确定所述位置特征的第一重要程度，以及确定所述结构特征的第二重要程度；对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征，包括：基于所述第一重要程度和所述第二重要程度，对所述位置特征、所述结构特征以及所述节点特征进行融合，得到所述融合特征。
- 根据权利要求1所述的方法，其特征在于，所述基于所述第一训练图得到所述第一训练图的位置特征，包括：将所述第一训练图输入位置特征提取模块，得到在目标特征空间对应的所述位置特征，所述目标特征空间的维度低于所述第一训练图对应的图像空间的维度。
- 根据权利要求3所述的方法，其特征在于，所述位置特征提取模块，通过如下方式训练得到：获取第二训练图；获取所述第二训练图的各个节点对应所述目标特征空间的位置特征；基于所述第二训练图的各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数；基于所述损失函数，调整所述位置特征提取模块的参数。
- 根据权利要求4所述的方法，其特征在于，所述基于所述第二训练图的 各个节点对应所述目标特征空间的位置特征、以及所述第二训练图，确定损失函数，包括：基于所述各个节点分别对应的损失项，得到所述损失函数；其中：第一节点为所述各个节点中的任意一个节点，所述第一节点对应的损失项，通过如下方式确定：基于所述第一节点对应所述目标特征空间的位置特征、第二节点对应所述目标特征空间的位置特征、以及所述第一节点和所述第二节点在所述第二训练图中的距离，得到所述第一节点对应的损失项，所述第二节点为所述第一节点的K阶邻居节点，所述K为大于1或者等于1的整数。
- 根据权利要求2所述的方法，其特征在于，所述确定所述位置特征的第一重要程度，包括：将所述位置特征输入第一模型，得到所述第一重要程度，所述第一模型用于基于输入的位置特征，得到对应的重要程度；其中：所述第一模型是基于第三训练图对应的位置特征和所述第三训练图对应的标签训练得到的。
- 根据权利要求1所述的方法，其特征在于，所述确定所述第一训练图的结构特征，包括：基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量，所述结构特征包括所述多阶邻居数量序列。
- 根据权利要求1所述的方法，其特征在于，所述基于所述第一训练图得到所述第一训练图的结构特征，包括：基于所述第一训练图，得到所述第一训练图的多阶邻居数量序列以及所述第一训练图的聚类特征，所述多阶邻居数据序列用于指示所述第一训练图中各个节点分别包括的各阶邻居节点的数量；对所述多阶邻居数量序列和所述聚类特征进行融合，得到所述第一训练图的结构特征。
- 根据权利要求2所述的方法，其特征在于，所述基于所述结构特征得到所述结构特征的第二重要程度，包括：将所述结构特征输入第二模型，得到所述第二重要程度，所述第二模型用 于基于输入的结构特征，得到对应的重要程度；其中：所述第二模型，是基于第四训练图对应的结构特征和所述第四训练图对应的标签训练得到的。
- 根据权利要求1所述的方法，其特征在于，所述方法还包括：获取待处理图，所述待处理图为与待处理多媒体内容对应的关系图，或者，所述待处理图为与待处理物品对应的关系图；基于所述待处理图得到所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征；对所述待处理图的位置特征、所述待处理图的结构特征以及所述待处理图的节点特征进行融合，得到所述待处理图的融合特征；将所述待处理图的融合特征输入所述目标模型，得到所述待处理图的标签。
- 一种基于图神经网络GNN的模型训练装置，其特征在于，所述装置包括：第一获取单元，用于获取第一训练图和所述第一训练图对应的标签，所述第一训练图为与训练多媒体内容对应的关系图，或者，所述第一训练图为与训练物品对应的关系图；第一确定单元，用于基于所述第一训练图得到所述第一训练图的位置特征、所述第一训练图的结构特征以及所述第一训练图的节点特征；第一融合单元，用于对所述位置特征、所述结构特征以及所述节点特征进行融合，得到融合特征；训练单元，用于根据所述融合特征和所述标签，训练基于GNN的目标模型。
- 一种设备，其特征在于，所述设备包括处理器和存储器；所述处理器用于执行所述存储器中存储的指令，以使得所述设备执行如权利要求1至10中任一项所述的方法。
- 一种计算机可读存储介质，其特征在于，包括指令，所述指令指示设备执行如权利要求1至10中任一项所述的方法。
- 一种计算机程序产品，其特征在于，当所述计算机程序产品在计算机上运行时，使得计算机执行如权利要求1至10中任一项所述的方法。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210994004.0 | 2022-08-18 | ||
| CN202210994004.0A CN115221976B (zh) | 2022-08-18 | 2022-08-18 | 一种基于图神经网络的模型训练方法及装置 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2024037354A1 true WO2024037354A1 (zh) | 2024-02-22 |
Family
ID=83616467
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2023/111176 WO2024037354A1 (zh) | 2022-08-18 | 2023-08-04 | 一种基于图神经网络的模型训练方法及装置 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN115221976B (zh) |
| WO (1) | WO2024037354A1 (zh) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115221976B (zh) * | 2022-08-18 | 2024-05-24 | 抖音视界有限公司 | 一种基于图神经网络的模型训练方法及装置 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111985622A (zh) * | 2020-08-25 | 2020-11-24 | 支付宝(杭州)信息技术有限公司 | 一种图神经网络训练方法和系统 |
| CN114265954A (zh) * | 2021-12-02 | 2022-04-01 | 西安电子科技大学 | 基于位置与结构信息的图表示学习方法 |
| CN114510966A (zh) * | 2022-01-14 | 2022-05-17 | 电子科技大学 | 一种基于图神经网络的端到端大脑因果网络构建方法 |
| CN115221976A (zh) * | 2022-08-18 | 2022-10-21 | 抖音视界有限公司 | 一种基于图神经网络的模型训练方法及装置 |
Family Cites Families (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110502659B (zh) * | 2019-08-23 | 2022-07-15 | 深圳市商汤科技有限公司 | 图像特征提取及网络的训练方法、装置和设备 |
| CN111400560A (zh) * | 2020-03-10 | 2020-07-10 | 支付宝(杭州)信息技术有限公司 | 一种基于异构图神经网络模型进行预测的方法和系统 |
| US20220138502A1 (en) * | 2020-08-25 | 2022-05-05 | Alipay (Hangzhou) Information Technology Co., Ltd. | Graph neural network training methods and systems |
| US11227190B1 (en) * | 2021-06-29 | 2022-01-18 | Alipay (Hangzhou) Information Technology Co., Ltd. | Graph neural network training methods and systems |
| US20220101103A1 (en) * | 2020-09-25 | 2022-03-31 | Royal Bank Of Canada | System and method for structure learning for graph neural networks |
| CN112734034A (zh) * | 2020-12-31 | 2021-04-30 | 平安科技（深圳）有限公司 | 模型训练方法、调用方法、装置、计算机设备和存储介质 |
| CN112766498B (zh) * | 2021-01-29 | 2022-11-22 | 北京达佳互联信息技术有限公司 | 模型训练方法及装置 |
| CN112766500B (zh) * | 2021-02-07 | 2022-05-17 | 支付宝(杭州)信息技术有限公司 | 图神经网络的训练方法及装置 |
| CN112861791B (zh) * | 2021-03-11 | 2022-08-23 | 河北工业大学 | 一种结合图神经网络和多特征融合的唇语识别方法 |
| CN112802013B (zh) * | 2021-03-12 | 2022-12-02 | 清华大学 | 基于图神经网络与多任务学习的脑疾病检测方法和装置 |
| CN113869353A (zh) * | 2021-08-16 | 2021-12-31 | 深延科技(北京)有限公司 | 模型训练方法、老虎关键点检测方法及相关装置 |
| CN114139684A (zh) * | 2021-12-02 | 2022-03-04 | 脸萌有限公司 | 图神经网络生成方法、装置、系统、介质以及电子设备 |
| CN114580654A (zh) * | 2022-02-23 | 2022-06-03 | 亿景智联(北京)科技有限公司 | 基于联邦学习的图神经网络模型训练方法以及装置 |
| CN114782468A (zh) * | 2022-04-29 | 2022-07-22 | 深圳开立生物医疗科技股份有限公司 | 图像分割模型训练方法、图像处理方法及相关装置 |
-
2022
- 2022-08-18 CN CN202210994004.0A patent/CN115221976B/zh active Active
-
2023
- 2023-08-04 WO PCT/CN2023/111176 patent/WO2024037354A1/zh unknown
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111985622A (zh) * | 2020-08-25 | 2020-11-24 | 支付宝(杭州)信息技术有限公司 | 一种图神经网络训练方法和系统 |
| CN114265954A (zh) * | 2021-12-02 | 2022-04-01 | 西安电子科技大学 | 基于位置与结构信息的图表示学习方法 |
| CN114510966A (zh) * | 2022-01-14 | 2022-05-17 | 电子科技大学 | 一种基于图神经网络的端到端大脑因果网络构建方法 |
| CN115221976A (zh) * | 2022-08-18 | 2022-10-21 | 抖音视界有限公司 | 一种基于图神经网络的模型训练方法及装置 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115221976B (zh) | 2024-05-24 |
| CN115221976A (zh) | 2022-10-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US20240046306A1 (en) | Methods and apparatus for detection of spam publication | |
| WO2024037354A1 (zh) | 一种基于图神经网络的模型训练方法及装置 | |
| CN108205581B (zh) | 生成数字媒介环境中的紧凑视频特征表示 | |
| US20210182912A1 (en) | Method and apparatus for labeling data | |
| US9311644B2 (en) | Item listing categorization system | |
| US20170177725A1 (en) | Developing An Item Data Model For An Item | |
| CN111160410B (zh) | 一种物体检测方法和装置 | |
| WO2019212570A1 (en) | Transition regularized matrix factorization for sequential recommendation | |
| KR102151756B1 (ko) | 작업자 관리 및 중개 방법 및 장치 | |
| CN111930858A (zh) | 一种异质信息网络的表示学习方法、装置及电子设备 | |
| US20140214844A1 (en) | Multiple classification models in a pipeline | |
| CN112949832B (zh) | 一种网络结构搜索方法、装置、电子设备及存储介质 | |
| KR102393517B1 (ko) | 풀필먼트 서비스를 제공하는 방법 및 그를 위한 서비스 시스템 | |
| CN113343085A (zh) | 一种信息推荐方法、装置、存储介质及电子设备 | |
| WO2024074146A1 (zh) | 多媒体数据处理方法、装置、设备及存储介质 | |
| US20230351221A1 (en) | Knowledge graph reasoning method and apparatus, model training method and apparatus, and computer device | |
| KR102363581B1 (ko) | 풀필먼트 서비스를 제공하는 방법 및 그를 위한 서비스 시스템 | |
| US20230359503A1 (en) | Distributed Actor-Based Information System and Method | |
| US20230245486A1 (en) | Methods, systems, articles of manufacture and apparatus to improve tagging accuracy | |
| KR102363580B1 (ko) | 풀필먼트 서비스를 제공하는 방법 및 그를 위한 서비스 시스템 | |
| JP6974222B2 (ja) | 検索装置、検索方法および検索プログラム | |
| TWI604322B (zh) | 解決方案搜尋系統及解決方案搜尋系統的操作方法 | |
| Liu et al. | Prototype-based Cross-Modal Object Tracking | |
| CN115827732A (zh) | 目标领域信息搜索方法、模型的训练方法及装置 | |
| CN115438767A (zh) | 神经网络模型的配置确定方法、装置、设备及存储介质 |
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: 23854258 Country of ref document: EP Kind code of ref document: A1 |