A visualization management system for graph neural networks

Abstract

The application provides a kind of visual management system for graph neural network, including graph dataset analysis and visualization module, model structure building and training module, log management module and parameter optimization module. By standardizing the storage standard of graph dataset, and using visualization library to visualize the graph dataset, the differences of graph dataset are compared. Through component-based construction of graph neural network model, flexible calling, customization, training and sharing are realized, and development steps are simplified. By recording the operation of graph neural network model, the parameters of trained model are recorded and a log is established. The log is stored in a standardized manner and compared and analyzed visually. On this basis, the model parameters are optimized and the results are recorded. The optimized model parameters can be called at any time. Application programming interfaces are constructed to connect each module, and the graph dataset and the optimized graph neural network model can be called and displayed at any time according to user needs, training tasks or model evaluation are implemented, and the development efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph neural network, and particularly relates to a visual management system for graph neural network. BACKGROUND

[0002] In the present stage of scientific research, there are a large amount of graph structure data, such as social networks, traffic networks, criminal organization, protein networks, etc. The graph data is more close to the network topology structure relationship in the real scene, and contains more abundant information. In recent years, the graph neural network (GNN) model has developed rapidly, which can comprehensively mine the topology structure information and node attribute information of the graph data, and has many successful application examples in financial anti-fraud, social networks, user recommendation, biomedicine and other fields.

[0003] The training process of the graph neural network model generally includes loading a graph structure data set, constructing a graph neural network model, training the graph neural network model, training and analyzing the graph neural network model log, and adjusting the parameters of the graph neural network to train an optimal graph neural network model for a certain data set. The existing technology lacks a scheme capable of persistent storage and management, and therefore, there is an urgent need for a management system capable of effectively storing graph data and managing graph neural network models to simplify the training and evaluation process. SUMMARY

[0004] The embodiment of the present application provides a visual management system for graph neural network to eliminate or improve one or more defects in the prior art, and solve the problem that the prior art cannot efficiently manage graph data sets and call graph neural network for training analysis.

[0005] The technical scheme of the present application is as follows:

[0006] In one aspect, the present application provides a visual management system for graph neural network, comprising:

[0007] A graph data set analysis and visualization module is configured to obtain one or more graph data sets uploaded by a user and store them according to a set storage standard, generate a graph data set information table, and visualize each graph data set based on a preset visualization library;

[0008] A model structure building and training module includes an intra-layer module, an inter-layer module and a model parameter module, which is configured to construct one or more graph neural network models in a block manner, and to train a specified graph neural network model by calling a specified graph data set and generate a model information table;

[0009] A log management module is configured to record logs of uploading, modifying, deleting, training and calling of each graph neural network model, each log records parameter values of the graph neural network model after corresponding operation, and a log information table is generated according to each log for management and visual analysis;

[0010] A parameter tuning module is configured to tune parameters of each graph neural network model and generate a tuning information table.

[0011] The graph dataset analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module are connected through an application programming interface.

[0012] In some embodiments, the graph dataset analysis and visualization module divides each graph dataset into a training set, a test set and a validation set according to a set proportion.

[0013] In some embodiments, the graph dataset analysis and visualization module is further configured to extract graph structure feature information of each graph dataset using NetworkX and DGL and save the information in a MongoDB database.

[0014] In some embodiments, the graph dataset analysis and visualization module is further configured to use Echarts to visually display the graph structure feature information of each graph dataset through a table, a node degree distribution line chart and / or a graph partial node connection force chart.

[0015] In some embodiments, the graph dataset analysis and visualization module is further configured to obtain a difference part of the table, the node degree distribution line chart and / or the graph partial node connection force chart corresponding to the graph structure feature information of each graph dataset, and perform highlight display.

[0016] In some embodiments, the model structure building and training module is configured to manage each graph neural network in a componentized manner, and record at least a model name of each graph neural network, a graph dataset related to each graph neural network, module information contained in each graph neural network, parameters of each graph neural network and training result indicators of each graph neural network.

[0017] The module information includes one or more submodule information constituting each graph neural network, an association structure between the submodules, and default values and parameter ranges of the submodules.

[0018] In some embodiments, the model structure building and training module is further configured to:

[0019] Receive a user's parameter setting request for a first specified graph neural network model, and store and call the user's preset parameters through a Json shared file to train the first specified graph neural network model based on the preset parameters.

[0020] In some embodiments, the log management module is further configured to:

[0021] Obtain one or more parameter items to be analyzed selected by a user for a second specified graph neural network model;

[0022] By searching the log information table and the tuning information table, find the training batches in which each parameter item to be analyzed in the second specified graph neural network model is inconsistent while the rest of the parameter items are consistent, obtain the model training effect of the corresponding training batches, and compare and evaluate the influence of each parameter item to be analyzed on the training effect of the second specified graph neural network model.

[0023] In some embodiments, the log management module is further configured to:

[0024] Query the log information table to split the key-value pair information of the default parameter value of each parameter in each log in the database; if there is a missing model parameter in the log, fill in the missing parameter value according to the corresponding key-value pair information, update the log, and update the log.

[0025] And, the log information table is tuned and visualized according to the set rules.

[0026] In some embodiments, the parameter tuning module is further configured to generate a tuning target curve according to the tuning information table for visual presentation.

[0027] The present application has at least the following advantages:

[0028] The visualization management system for graph neural networks of the present application standardizes the storage standard of graph data sets through the graph data set analysis and visualization module, and uses a visualization library to visually display the graph data sets and compare the differences between the graph data sets. Through the modularization of the model structure building and training module, the graph neural network model is flexibly called, customized, trained and shared, simplifying the development steps. The log management module records the operation of the graph neural network model, records the model parameters after training, and establishes a log. The log is stored in a standardized manner and is visually compared and analyzed. On this basis, the model is parameter tuned and the results are recorded. The optimized model parameters can be called at any time. An application programming interface is constructed to connect each module. According to user needs, the graph data set and the optimized graph neural network model can be called and displayed at any time, and training tasks or model evaluation can be implemented, thereby improving the development efficiency.

[0029] Additional advantages, objects, and features of the application will be set forth in part by the description that follows, and in part will become apparent to those skilled in the art upon examination of same, or can be learned by practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.

[0030] It will be understood by those within the art that the objects and advantages of the application can be met by other embodiments that are not specifically described herein. Indeed, various embodiments of the application can be practiced or carried out in various ways. Furthermore, any BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings, which are incorporated in and form a part of the specification, illustrate the present application and together with the description, serve to explain the principles of the application. In the drawings:

[0032] Figure 1 Structure diagram of the visualization management system for graph neural network according to an embodiment of the present application.

[0033] Figure 2 System architecture diagram of the visualization management system for graph neural network according to an embodiment of the present application.

[0034] Figure 3 System technology flow diagram of the visualization management system for graph neural network according to an embodiment of the present application.

[0035] Figure 4 Log uploading flow diagram in the visualization management system for graph neural network according to an embodiment of the present application.

[0036] Figure 5 Usage flow diagram of the log management module in the visualization management system for graph neural network according to an embodiment of the present application.

[0037] Figure 6 Parameter tuning flow diagram of the parameter tuning module in the visualization management system for graph neural network according to an embodiment of the present application. DETAILED DESCRIPTION

[0038] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the embodiments and drawings. Herein, the illustrative embodiments of the present application and their descriptions are used to explain the present application, but are not used as limitations to the present application.

[0039] It should also be noted that, in order to avoid the present application being obscured by unnecessary details, only the structures and / or processing steps closely related to the solutions according to the present application are shown in the drawings, and other details not closely related to the present application are omitted.

[0040] It should be emphasized that the term "comprising" as used herein is to be interpreted as "including", "containing", or "consisting" but not excluding one or more additional features, elements, steps or components.

[0041] Here, it should also be noted that the term "connected" as used herein, unless otherwise specified, can mean not only direct connection, but also indirect connection in the presence of an intermediate.

[0042] At present, the graph structure dataset management system cannot support users to upload custom datasets, and does not support comparison between two datasets. The construction and training of graph neural network models are generally programmed and built in special code programming software. Graph dataset analysis, model building, log analysis, automatic parameter tuning and other operations are processed separately, and users need to go to other platforms for analysis or code implementation. This way, the user's operation and training process of the graph neural network model is not continuous, and the collaboration of various modules cannot help users better research.

[0043] And the existing graph neural network training platform lacks log management for user model training, and cannot provide log saving and log management services. Users can only manually save log files and write scripts to analyze relevant information. The existing deep learning log visualization system can only visualize log information after it is deployed, and can only visualize the log of a single model. It does not distinguish between graph neural network model structure and tuning parameters, and does not support comparison of differences between multiple model logs.

[0044] After the user analyzes the log of the graph neural network model, he may choose to manually adjust the parameters and retrain the model or select specific parameters for automatic parameter tuning of the model. The existing automatic parameter tuning function can only optimize parameters in a black box. If interrupted in the middle, it can only start over, and the optimization time of the neural network model is relatively long. Such automatic parameter tuning of the graph neural network model is not only time-consuming, but also has poor stability.

[0045] Therefore, the shortcomings of the prior art can be summarized as follows:

[0046] 1) It does not support automatic visualization of user-defined graph structure features of datasets, and cannot visualize and compare the differences between graph dataset structures in log visualization analysis.

[0047] 2) Prior art only supports the user to upload the log file and then visualize the parameter results of the relevant log, does not support continuous model log visualization, cannot structurally distinguish the differences between the model, dataset, structural parameter, training parameter and the like, and does not support the screening of the differences of multiple parameter combinations, such as taking multiple parameters of different models as a group to compare the influence of different combinations on model training, and the existing system does not support the grouping visualization function of model logs.

[0048] 3) In the existing system, the user cannot quickly perform automatic parameter tuning of the relevant model after analyzing the log parameter influence, and the existing automatic parameter tuning of the deep neural network model cannot be visualized and monitored, and the tuning speed is slow.

[0049] 4) The log management and model training parts of the existing management system are separated, and the two parts cannot interact, so that the user is low in efficiency and poor in continuity in the process of researching and developing the model combination to form the optimal parameters.

[0050] Therefore, the present application provides a visual management system for a graph neural network, as shown in Figure 1 , which comprises a graph dataset analysis and visualization module, a model structure building and training module, a log management module and a parameter tuning module, and the graph dataset analysis and visualization module, the model structure building and training module, the log management module and the parameter tuning module are connected through an application programming interface.

[0051] The graph dataset analysis and visualization module is used to obtain one or more graph datasets uploaded by the user and store them according to the set storage standard, generate a graph dataset information table and visually display each graph dataset based on a preset visualization library.

[0052] The model structure building and training module comprises an intra-layer module, an inter-layer module and a model parameter module, which are used to construct one or more graph neural network models in a block manner, and to train the specified graph neural network model by calling the specified graph dataset and generate a model information table.

[0053] The log management module is used to record the uploading, modifying, deleting, training and calling logs of each graph neural network model, and the parameter values of the graph neural network model after each operation are recorded, and a log information table is generated based on each log for management and visual analysis.

[0054] The parameter tuning module is used to tune the parameters of each graph neural network model and generate a tuning information table.

[0055] In this embodiment, the system architecture composed of each module can be described as five levels, as shown in Figure 2 , which are hardware layer, persistence layer, model layer, backend layer and frontend layer.

[0056] In the hardware layer, the graph dataset analysis and visualization module, the model structure building and training module, the log management module, and the parameter tuning module can adopt a computer processor or a single-chip microcomputer or other electronic devices capable of storing and executing computer programs. In actual application, a central processing unit (CPU) can be adopted. Since the training of the graph neural network model is involved, the model structure building and training module can adopt a graphics processing unit (GPU).

[0057] In the data storage and management layer, the related data information can be stored based on a NoSQL database such as MongoDB in the embodiment. Specifically, the graph dataset information table generated by the graph dataset analysis and visualization module is stored to store the structural feature information of the graph dataset; the model information table generated by the model structure building and training module is stored to store the structural definition information of the model; the log information table of the log management module is stored to store the training log information related to the model; and the tuning information table generated by the parameter tuning module is stored to store the information of the automatic tuning process.

[0058] In the model layer, the existing graph neural network development technology library can be implemented, such as the DGL (Deep Graph Library) library, NetworkX, or Pytorch, including the graph structure information extraction module (i.e., the graph dataset analysis and visualization module), the graph neural network model training module (i.e., the model structure building and training module), and the automatic parameter tuning tool (i.e., the parameter tuning module). In the graph structure information extraction module, the user uploads the dataset, and the graph is constructed through DGL and NetworkX. The graph structure-related features are automatically extracted and stored in the graph data information table.

[0059] The graph neural network model training module implements the conventional graph convolutional neural network (GCN), graph attention neural network (GAT), GraphSage, etc. In addition, the GNN design space is modularized and componentized into three parts: intra-layer module, inter-layer module, and model parameter module. The user can design and combine different sub-modules, set the parameter space, generate, and train a new graph neural network model.

[0060] The automatic parameter tuning tool starts the automatic parameter tuning of the graph neural network model based on the automatic parameter tuning tool package such as the Hyperopt parameter tuning package combined with the system-implemented graph neural network model, uses Netstat to monitor the tuning process in real time, and stores the tuning state information into the tuning information table.

[0061] In the backend layer, the open framework based on python configures the route of the response front end based on the resource locator, such as Django framework: use the view function (views) to define the function of the relevant function return response view of the route of the resource locator, and use the template to design the front end HTML web page template.

[0062] The code function can flexibly add, delete, modify and query the data in the MongoDB database in Python. The backend layer can realize the message passing of the related parameters in the form of Json sharing file, and use Popen in Subprocess to start a new process to train the graph neural network model and monitor the returned process number. The front and back end interact with data through Ajax.

[0063] The front end layer can use HTML, BootStrap, JavaScript, JQuery, etc. to complete the design of the interface. Jupyter Notebook and IFrame are used to embed the system interface into the code writing of Jupyter Notebook, so that users can use the system flexibly and conveniently.

[0064] The front end layer mainly designs seven interfaces:

[0065] First, the graph dataset upload / structure visualization interface is used to adapt the graph dataset analysis and visualization module, and realizes the addition, deletion, modification and query of the graph dataset, and the visualization of the extracted graph dataset structure features.

[0066] Second, the graph dataset comparison interface is used to adapt the graph dataset analysis and visualization module, and realizes the comparison of the differences and the calculation of the similarity of the structure features of different graph datasets.

[0067] Third, the model information management interface is used to adapt the model structure building and training module, and realizes the definition and structure management of the graph neural network model structure information.

[0068] Fourth, the model training interface is used to adapt the model structure building and training module, and realizes the rapid combination of the graph neural network model and the model training.

[0069] Fifth, the log information management analysis interface is used to manage the log management module, and realizes the persistent and standardized storage of the model training log information, and supports the quick filtering of log parameter query, log grouping and log comparison.

[0070] Sixth, the automatic tuning monitoring interface is used to manage the parameter tuning module, and supports the start of automatic parameter tuning and real-time monitoring of the tuning.

[0071] Seventh, the Jupyter Notebook interface is used for customizing the application programming interface (API) interface, and an IFrame is used to embed the system into code writing, and the implementation of the graph dataset training interface, the model structure definition interface, the model training and log management interface, and the automatic tuning interface are opened, which corresponds to the user's regular process of researching graph neural network models.

[0072] The graph neural network-oriented visualization management system described in the embodiment is explained below based on the development process of the graph neural network model, and the specific process includes: graph dataset analysis and visualization→graph neural network model structure definition training→model training log management and visualization analysis→model automatic parameter tuning and monitoring. The system constructed above embeds the related API into the Jupyter Notebook through the IFrame, and directly calls the corresponding system interface API in code writing.

[0073] The execution steps of the system include:

[0074] The first step is to visualize the graph structure characteristics of the graph dataset, so that the user can visualize the graph data structure characteristics before developing the model, analyze and compare the differences between different graph data structure characteristics, and also visualize and analyze the influence of different dataset graph structure characteristics on the log during log visualization analysis.

[0075] The second step is to define the graph neural network model structure, which defines the graph neural network model used for training and also defines the log management structure for the corresponding graph neural network model structure characteristics.

[0076] The third step is log visualization management, which is based on the model structure defined in the second step to construct a structured log management, supports comparison of parameters of different graph neural network models, and supports parameter grouping function comparison.

[0077] The fourth step is automatic parameter tuning, which enables the user to quickly perform automatic parameter tuning after analyzing the graph neural network model parameters, and performs visual automatic parameter tuning monitoring based on the log management function, and uses the pre-stored and managed historical tuning information table to speed up the automatic parameter tuning process.

[0078] The above four steps embed a Jupyter Notebook in the system to build a fast interface of the function, so that the user can conveniently and quickly call the related function during the code development process of the graph neural network model, improve the interactivity with the user, and improve the development research efficiency.

[0079] Specifically, as Figure 3As shown, the graph dataset analysis and visualization module implements graph dataset analysis and visualization functions, the front-end layer of the system supports user uploading of graph datasets, the back-end layer stores the user-uploaded datasets persistently and normatively in a specified file and manages the path normatively, and automatically divides the training set, test set, and validation set. Specifically, the graph dataset analysis and visualization module divides each graph dataset into the training set, test set, and validation set according to a set proportion.

[0080] The graph dataset analysis and visualization module constructs graphs based on NetworkX and DGL (Deep Graph Library is a framework for deep learning on graphs) from the graph datasets, extracts the graph structure feature information of each graph dataset, and saves the information into a graph data information table in MongoDB.

[0081] The graph dataset analysis and visualization module also uses Echarts to visualize the graph structure feature information of each graph dataset through tables, node degree distribution line graphs, and / or graph partial node connection force diagrams. Differences between two graph datasets are automatically compared and highlighted to show the differences between the dataset features on the front end.

[0082] As shown, Figure 3 The model structure building and training module is used to implement model structure building and training. The back-end layer of the system can build various conventional graph neural network models GCN, GAT, GraphSage, etc. based on DGL. Comprehensive analysis of the model structure space of the graph neural network (GNN) model can modularize and componentize the GNN model, including three main modules: intra-layer module, inter-layer module, and model parameter module, support users to combine each module to generate various variants of the GNN model, and thus find a GNN model suitable for their own graph dataset. In some embodiments, the model structure building and training module is used to componentize and manage each graph neural network, at least recording the model name of each graph neural network, the graph dataset related to each graph neural network, the module information contained in each graph neural network, the parameters of each graph neural network, and the training result indicators of each graph neural network; wherein the module information includes one or more sub-module information that constitutes each graph neural network, the association structure between each sub-module, and the default values and parameter ranges of each sub-module. The model structure building and training module is also used to: receive a user's parameter setting request for a first specified graph neural network model, and store and call the user's preset parameters through a Json shared file to train the first specified graph neural network model based on the preset parameters.

[0083] The model structure building and training module builds a model information table to manage the feature information related to the model. The table divides the model into: model name (Model Name), model related graph data set (DataName), model submodule information (Modules, which records the submodule information constituting the graph neural network model, wherein each submodule contains default value and parameter value range), parameter information (Parameters, which records the parameters related to the model, and each parameter contains default value and value range), model output training result index (Result), and other possible information of the model (Others).

[0084] The model structure building and training module manages the structured and standardized models through the model information table. When a user implements a new model or adds a new module component to an existing model, the user can upload the model module definition through the model definition and modification interface provided by the front-end system. The system automatically updates the model module to the model information table for management and synchronizes to the subsequent log management function and automatic tuning function.

[0085] During the training of the GNN model, the system provides the GNN modularization interface on the front end, and the user can directly set the parameters on the front end to directly start a custom GNN model training. The system transmits the parameters input from the front end through the Json shared file, starts the training of the corresponding model according to the received parameters through subprocess, and automatically saves the logs generated by the model training in the log information table.

[0086] As shown in Figure 3 The log management module implements log management, and realizes the persistent management and visual analysis of the model training logs in the log management module. The log management module mainly includes four functions: log uploading, multi-model log keyword filtering visualization, log grouping, and log comparison. The system cooperates with the model information table and the log information table to manage and analyze the model. The log information table stores the logs of the multi-graph neural network model, and each piece of data records a training log data.

[0087] In some embodiments, the log management module is further configured to: query the log information table to split the key-value pair information of the default parameter value of each parameter in each log in the database; if there is a missing model parameter in the log, fill in the default parameter value according to the corresponding key-value pair information, update the log, and optimize and visualize the log information table according to the set rules. Specifically, in the log management module, the log uploading function is as shown in Figure 4As shown, first, the Init method is called to connect the database MongoDB, and after connecting the database, the UpLoadLog method is called to upload the string type log information. The system automatically splits the key-value pair information of the log parameters, queries the parameter defined in the model information table according to the ModelName in each log, and if some parameter information is missing in the uploaded log, the corresponding default parameter value in the default model information table is automatically filled in. If not missing, the corresponding parameter value of the extracted log is filled in. Finally, the complete model log information is stored in the log information table. In this way, each log corresponds to the parameters of the corresponding model in the model information table, so as to standardize the management of the log. The system also develops a log uploading toolkit. When the user trains a custom model in the local compilation environment, the user can call the toolkit in the code to upload the model training log to the log information table in real time.

[0088] The system realizes the visualization of log information, as shown in the following figure: Figure 5 As shown, first, the model information table is given to select the model list to view, and the user can select the model to view in the front-end layer. If the user selects a single model, the system will query the corresponding model parameter information in the model information table and return; if multiple models are selected, the system will query the parameters of the corresponding models in the model information table, take the intersection of the parameters common to multiple models and return, which can include five groups of parameters: DataName, Modules, Parameters, Results and Others.

[0089] After returning the related parameters of the single or multiple models selected by the user, the user selects the log parameter that wants to view in these parameters, and the system queries the corresponding training log in the log information table according to the user's selected parameter, returns the log information table corresponding to the query parameter for visualization, so that the user can accurately and conveniently focus on the log parameter content of his own research, and removes redundant information. After the front-end displays the summary parameter information of the model viewed by the user, the user can click on a log to directly view the detailed information of the log.

[0090] In some embodiments, the log management module is also configured to: obtain one or more to-be-analyzed parameter items selected by a user for a second specified graph neural network model. By searching the log information table and the tuning information table, the training batches in which the to-be-analyzed parameter items of the second specified graph neural network model are inconsistent and the remaining parameter items are consistent are searched based on the control variable, the model training effect of the corresponding training batch is obtained, and the influence relationship of each to-be-analyzed parameter item on the training effect of the second specified graph neural network model is compared and evaluated.

[0091] Specifically, the system supports user selection of parameters to group logs. After the user selects a parameter to be analyzed, the system groups the logs according to the model parameters recorded in the model information table and the trained model parameters recorded in the log information table, divides the logs with the same values of parameters other than the parameter to be analyzed into the same group, and distinguishes different groups by different colors in the front-end layer to quickly view the influence of the change of the parameter or parameters to be analyzed on the model results. When the user selects multiple logs for comparison, the system jumps to a log comparison interface, analyzes the differences between the selected multiple logs, gives the differences in specific parameters in the logs and the specific information of each log data, and highlights the parameter values of the differences. The user can click to view the differences between the log results and the model parameters of the optimal log results.

[0092] As shown in Figure 3 , the parameter tuning module is used for automatic tuning of the parameters of the graph neural network model. An automatic parameter tuning function of the graph neural network model is designed and developed on the basis of the Hyperopt toolkit, as shown in Figure 6 , the parameters of a certain model in the model information table are divided into two categories: tuning parameters and default parameters. The tuning parameters are parameters that need to be tuned by the tuning algorithm, and the user can customize the parameter tuning range. The default parameters are parameters that do not need to be tuned, and the parameter values support user-defined parameter values or directly use the default parameter values in the model information table.

[0093] When the user starts the parameter tuning function, a tuning log for this tuning process is generated and saved to the table. The system records and monitors the tuning process. During tuning, the system generates a model parameter combination according to a value selected from the tuning range of each tuning parameter combined with the default parameters, starts a model training, and checks from the log information table whether there is a model log that has been trained before starting the model training. If there is, the result information of the log is directly called to speed up the tuning, otherwise the model with the parameter combination is trained and the log result is saved.

[0094] During the automatic parameter tuning process, the training log of each model tuning is stored and visualized in the front-end interface. The user can view the tuning target curve generated during each tuning training process and the detailed information of the related model parameter log at any time.

[0095] Further, the embodiment also develops an API by embedding a Jupyter Notebook, embeds a Jupyter Notebook compiling interface in a system interface, and realizes three interface interfaces, i.e., a dataset interface interface, a model structure definition interface interface, and a log management and automatic parameter tuning interface interface, through an IFrame. When a user compiles and realizes a custom model, the user can call the interface when writing a Jupyter Notebook code, wherein the dataset interface is used for visualizing dataset uploading and viewing graph data structure features; the model structure definition interface is used for visualizing definition of a network model structure and starting model training; the model training and log management are used for visualizing and analyzing historical model training logs in model training; and the automatic parameter tuning interface is used for automatic parameter tuning to realize an optimal parameter combination model. Finally, the management and analysis of a graph neural network model and logs are realized while writing a code.

[0096] In some embodiments, the parameter tuning module is further configured to generate a tuning target curve according to the tuning information table and perform visual presentation.

[0097] The application constructs an integrated and interfaced system platform of graph structure data structure visualization, graph neural model module combined training, real-time multi-model multi-log management and efficient analysis and comparison. Compared with the previous single model development and training of a graph neural network model, the application combines graph data visualization, model training log management and model tuning into the development of the graph neural network model, provides a platform for users and realizes a convenient interface, so that the users can quickly call the functions of each module during development, and the development efficiency is improved.

[0098] Through multi-model training log analysis of the graph neural network model, the graph neural network model is realized structured management, and the model is represented through the structured module designed by the system. Based on the model structured management, the model training log is realized standardized and persistent management. The application supports management of multiple model logs, quickly compares the similarities and differences between the training log results of specific modules of multiple models, realizes visual analysis and comparison, and provides convenience for user model analysis and development.

[0099] The application combines model structured management and log structured management, realizes automatic tuning monitoring and visual monitoring, speeds up the tuning process through historical log query, and solves the disadvantages of unmonitorable and slow automatic parameter tuning process of the previous deep neural network model.

[0100] In summary, the visualization management system for the graph neural network has the advantages that the storage standard of the graph data set is specified through the graph data set analysis and visualization module, the graph data set is visually displayed by using the visualization library, and the differences of the graph data sets are compared. The graph neural network model is constructed and trained by the model structure building and training module, the model is flexibly called, customized, trained and shared, the development steps are simplified, the operation on the graph neural network model is recorded by the log management module, the model parameters after training are recorded and the log is established, the log is stored in a standardized manner and is visually compared and analyzed, the model parameters after optimization are recorded on the basis of parameter optimization of the model, and the optimized model parameters are called at any time. The application programming interface is constructed to connect the modules, the graph data set and the optimized graph neural network model are called and displayed at any time according to the user demand, the training task or the model evaluation is implemented, and the development efficiency is improved.

[0101] Those of ordinary skill in the art will appreciate that the various exemplary components, systems and methods described herein can be implemented in hardware, software or a combination of both. Whether the implementation is in hardware or software depends on the particular application and design constraints imposed on the overall system. Skilled artisans can implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application. When implemented in hardware, the hardware can comprise, for example, an electronic circuit, an Application Specific Integrated Circuit (ASIC), a suitable firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the application are the program or code segments to perform the necessary tasks. The program or code segments can be stored in a machine-readable medium, or transmitted by a carrier wave in a data signal over a transmission medium or a communication link. A "machine-readable medium" includes any medium that can store or transfer information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. The code segments can be downloaded via a computer network, such as the Internet, an intranet, etc.

[0102] It should also be noted that the exemplary embodiments mentioned in the present application describe some methods or systems based on a series of steps or devices. However, the present application is not limited to the order of the above steps, that is, the steps can be performed in the order mentioned in the embodiments, or in an order different from the embodiments, or several steps can be performed simultaneously.

[0103] In the present application, the features described and / or exemplified for one embodiment can be used in the same way or in an analogous way in one or more other embodiments, and / or in combination with or instead of features of other embodiments.

[0104] The above merely provides the preferred embodiments of the present application, but not for limiting the present application. For the person skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A visualization management system for graph neural networks, the system comprising: Comprise: a graph dataset analysis and visualization module for obtaining one or more graph datasets uploaded by a user and storing them according to set storage standards, generating a graph dataset information table, and visualizing each graph dataset based on a preset visualization library; and extracting the graph structure feature information of each graph dataset using NetworkX and DGL and saving it in a MongoDB database, visualizing the graph structure feature information of each graph dataset using Echarts through tables, node degree distribution line graphs, and / or graph partial node connection force diagrams; obtaining the difference between the corresponding tables, node degree distribution line graphs, and / or graph partial node connection force diagrams of the graph structure feature information of each graph dataset, and highlighting the differences; a model structure building and training module, including an intra-layer module, an inter-layer module, and a model parameter module, for block-based construction of one or more graph neural network models, and for calling a specified graph dataset to train a specified graph neural network model and generating a model information table; and for component-based management of each graph neural network model, recording at least the model name of each graph neural network, the graph dataset related to each graph neural network, the module information included in each graph neural network, the parameters of each graph neural network, and the training result indicators of each graph neural network; wherein the module information includes one or more sub-module information that makes up each graph neural network, the association structure between each sub-module, and the default values and parameter ranges of each sub-module; a log management module for recording the upload, modification, deletion, training, and calling logs of each graph neural network model, the parameters of the graph neural network model after each operation, and generating a log information table for management and visual analysis based on each log; a parameter tuning module for parameter tuning of each graph neural network model and generating a tuning information table; The graph dataset analysis and visualization module, the model structure building and training module, the log management module, and the parameter tuning module are connected through an application programming interface.

2. The graph neural network oriented visualization management system according to claim 1, wherein, The graph dataset analysis and visualization module divides each graph dataset into a training set, a test set, and a validation set according to a set proportion.

3. The graph neural network oriented visualization management system according to claim 1, wherein, The model structure building and training module is also used to: receive a user's parameter setting request for a first specified graph neural network model, and store and call the user's preset parameters through a Json shared file to train the first specified graph neural network model based on the preset parameters.

4. The graph neural network oriented visualization management system according to claim 1, wherein, The log management module is also used to: obtain one or more parameter items selected by a user for a second specified graph neural network model; find the training batches in which each of the parameter items to be analyzed is inconsistent while the rest of the parameter items are consistent in the second specified graph neural network model based on the control variable method by searching the log information table and the tuning information table, obtain the model training effect of the corresponding training batches, and compare and evaluate the influence of each parameter item to be analyzed on the training effect of the second specified graph neural network model.

5. The graph neural network oriented visualization management system according to claim 1, wherein, The log management module is also used to: query the log information table to split the key-value pair information of each parameter default parameter value in the database; if there is a missing model parameter in the log, the corresponding key-value pair information is called to fill in the default parameter value, and the log is updated; and, the log information table is optimized and visualized according to the set rules.

6. The graph neural network oriented visualization management system according to claim 1, wherein, The parameter optimization module is also used for: According to the optimization information table, the optimization target curve is generated and visualized.

Images