Visual full-process machine learning platform, control method, client and application

Through the visual full-process machine learning platform, the problems of high labor costs and complex development processes in machine learning applications are solved, and rapid iteration and efficient model construction and deployment are achieved, the overall labor costs of the enterprise are reduced, and large-scale data processing is supported.

CN114139728BActive Publication Date: 2024-11-26DIGITAL CHINA SYST INTEGRATION SERVICE
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111477541.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-06
Publication Date
2024-11-26
Estimated Expiration
2041-12-06

AI Technical Summary

Technical Problem

The existing technology has high labor costs, complex development processes and difficult to meet the needs of general scenario modeling in machine learning applications, especially in terms of rapid iteration and model deployment within the enterprise.

Method used

Provide a visual full-process machine learning platform, which builds machine learning models through drag and drop through client interface, supports built-in operators for pictures, videos, log files and container/virtual machine images, realizes interactive modeling and parameter tuning, and Provides functions such as data security guarantee, multi-department collaborative modeling and model management, which lowers the threshold for model development and deployment.

Benefits of technology

It greatly reduces the threshold for using AI tools, improves modeling efficiency, reduces labor costs, realizes the rapid implementation of machine learning algorithms and full-process management, supports data processing from tens of millions to hundreds of millions of yuan, simplifies the traditional coding process, and reduces the The labor cost of the enterprise algorithm engineer.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114139728B_ABST
    Figure CN114139728B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of data processing technology, and discloses a visual full-process machine learning platform, control method, client and application, which drags and drops built-in operators of pictures, videos, log files, backup data and container / virtual machine images on the client interface and interactively constructs a visual modeling process to obtain a machine learning model; uses the setting parameters input at the operator parameter to train the obtained machine learning model, view the results, and train again to view the results through parameter adjustment of the visual interface, and repeats the cycle; obtains the parameters of the optimal visual interface and the optimal model, and then performs visual display of pictures, videos, log files, backup data and container / virtual machine images. The visual modeling interaction process of the present invention can realize dynamic and intuitive parameter adjustment and modeling process adjustment, realizes rapid iteration of the modeling process, and realizes rapid implementation of machine learning modeling.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and in particular relates to a visual full-process machine learning platform, control method, client and application. Background Art

[0002] At present, the development and improvement of machine learning technology, especially deep learning technology, has made machine learning more and more frequently used in enterprises, especially Internet enterprises, and its application scope is becoming more and more extensive. Although the theoretical basis of various machine learning algorithms is becoming more and more perfect, and the cost of hardware equipment (CPU, GPU, disk for storing training data, etc.) required for machine learning is also getting lower and lower, there are still many other problems in the application of machine learning technology.

[0003] Machine learning is a relatively professional field, and the requirements in Internet companies are rapidly iterating. For a small requirement that requires the use of machine learning, the algorithm modeling and parameter tuning process also require the participation of algorithm engineers, and the labor cost of model development is relatively high. Moreover, after the model is developed, how to put it into practice is a big challenge for algorithm engineers who are only proficient in algorithm design and model tuning.

[0004] At present, most enterprises that use machine learning methods to reduce costs and increase efficiency in business scenarios have hired a large number of machine learning algorithm engineers to combine the professional knowledge of business personnel to model. Modeling problems in different scenarios consumes a lot of manpower and material resources. The open source platforms they use either provide simple computing resources or open source machine learning algorithms. It is difficult to meet the modeling needs of general scenarios and reduce the development costs of machine learning.

[0005] To solve the above problems, the prior art 1 patent number CN111240662A is a spark machine learning system and learning method based on task visualization dragging, which mainly proposes: a spark machine learning system based on task visualization dragging, including a process designer, a process parser and a process scheduler; the process designer is used to allow users to drag data source components, data preprocessing components, machine learning components, and output components to the design area to build a machine learning process and generate a process description language; the process parser is used to parse the machine learning flowchart constructed by the user, parse the relationship, input, and output between each component, and translate the flowchart into a set of data that can be recognized by the scheduler through the designed algorithm. The process scheduler parses the data passed by the process parser and submits the constructed machine learning process to the spark cluster for training.

[0006] In recent years, with the rise and development of big data and artificial intelligence technologies, governments at home and abroad have also made artificial intelligence a key direction for national strategic development in the next few years and formulated relevant development policies. Therefore, more and more industries and enterprises are willing to try to use artificial intelligence technology to improve the efficiency of enterprises in R&D, production, management, marketing, operation and maintenance, and guide enterprise operation decisions. For enterprises, to carry out artificial intelligence work, it is necessary to form an AI technical team, and AI technology generally includes: big data underlying technologies, such as Hadoop, Spark, etc., to provide support for the storage and efficient calculation of big data; data analysis technologies, such as statistical analysis, machine learning, natural language processing, etc., to provide support for the construction of analytical modeling solutions; analytical tool technologies, such as Python / R / Matlab / SAS, etc., to provide support for the implementation of analytical modeling scenarios; development management technologies, such as data access and storage technologies (such as databases, offline files, Kafka message queues, etc.), model deployment (scheduling, synchronous / asynchronous services, streaming services) technologies, model monitoring technologies (calculation and display of various monitoring indicators), model management technologies (unified model management, etc.), model display technologies (how to better present and apply model results in a unified manner), etc. From the above, it can be seen that for most companies, it is very difficult to form a complete AI technical team, both from a cost perspective and from an implementation perspective. Therefore, in order to better promote the widespread application of AI technology, it is necessary to create a one-stop machine learning platform based on the existing supporting machine learning tools with a platform-based mindset. It is also very difficult for professional algorithm teams within large Internet companies to implement machine learning algorithms. There are many problems in applying machine learning to actual business. For algorithm engineers, professional work such as algorithm development and model tuning may not be a problem, but they often have to solve many problems that are not related to the algorithm itself, such as the cluster environment required for model training, model deployment and release, and subsequent model online service operation and maintenance.

[0007] With the development of artificial intelligence, the demand for intelligent business scenarios has shown a blowout growth. However, the technical threshold for customizing artificial intelligence models for business scenarios is high and time-consuming. Machine learning talents with modeling experience and accumulated domain knowledge are scarce and their salaries are high. These limitations make it difficult to quickly implement artificial intelligence scenarios.

[0008] Machine learning is a relatively professional field, and in traditional companies, the demand is rapidly iterating. For a small demand that requires the use of machine learning, the algorithm modeling and parameter tuning process also requires the participation of algorithm engineers, and the labor cost of model development is relatively high. Moreover, after the model is developed, how to implement it is a big challenge for algorithm engineers who are only proficient in algorithm design and model tuning. Many companies face many problems in the process of deploying machine learning and subsequent service maintenance.

[0009] To solve the above problems, traditional AI model training often has to go through steps such as feature analysis, model selection, parameter adjustment, and evaluation. These steps take several months, and if there is no experience at all, it will take even longer.

[0010] Prior art 2CN202011188747.6 - AI model training method, model use method, computer device and storage medium - discloses an AI model training method, characterized in that it includes:

[0011] Acquire multiple sample generation models, and play chess according to the multiple sample generation models to obtain first chess game data;

[0012] Acquire second game data, and train the to-be-trained model according to the second game data and the first game data, where the second game data is real game data;

[0013] When the model to be trained converges, the model to be trained is used as the model to be evaluated, and multiple games are played with the comparison model to obtain the game results;

[0014] When the game result reaches a preset indicator, the model to be evaluated is determined to be an AI model, and the AI ​​model training is completed.

[0015] Through the above analysis, the problems and defects of the prior art are as follows:

[0016] (1) In the development of information data processing terminals in the fields of finance, education, medical care, and information industry, existing technologies require a lot of manpower and material resources to perform software data modeling for different business scenarios, which increases labor costs.

[0017] (2) Moreover, the open source platforms used in existing technologies either provide simple computing resources or open source machine learning algorithms. It is difficult to meet the modeling requirements of general scenarios, which limits the practicality of open source platforms.

[0018] (3) The design process of the open source platform in the prior art is cumbersome, and the visual display and interactive performance are poor. It cannot improve the work efficiency of the business.

[0019] The difficulty of solving the above problems and defects is:

[0020] The core points of difficulty are as follows:

[0021] So far, there is no unified standard for many things in machine learning systems, such as the storage format of data labeling results and the storage format of models.

[0022] There are too many frameworks, too many environments, and too many versions. Correspondingly, the models generated by different frameworks are also different. If you want to achieve standard model compression and release, it is also a relatively difficult task.

[0023] Going forward, machine learning is too dependent on GPUs, and the management and scheduling of GPUs is not yet mature.

[0024] Distributed training: This area is currently developing rapidly, and machine learning based on large-scale data is not yet very mature.

[0025] Differences and lack of standards are the source of pain points in machine learning. However, as machine learning develops in the engineering field, it is urgent to establish unified standards.

[0026] The science that drives progress in machine learning research is difficult and requires innovation, experimentation, and persistence;

[0027] It is also difficult to apply known machine learning models to practical work;

[0028] Machine learning engineers on the market are more expensive than ordinary software engineers, and it is very costly for companies to set up professional machine learning teams.

[0029] Choosing what kind of machine learning tools requires a thorough understanding of the advantages and disadvantages of each algorithm and model, a skill that can be acquired by learning these models. Of course, the difficulty of building this kind of knowledge exists in all areas of computer science, not just machine learning applications.

[0030] The significance of solving the above problems and defects is:

[0031] The earth-shaking digital technology revolution has brought human society into a new era of digital economy, and digital information has become a key production factor in this era. Effectively acquiring, controlling and utilizing digital information, and then reducing or eliminating uncertainty, is the key to improving the core competitiveness of enterprises. The continued survival of any enterprise will depend on a flexible, data-centric architecture that responds to the speed of change.

[0032] The one-stop machine learning platform integrates data access, data processing, data mining, and data application, helping enterprises to have independent and controllable data analysis capabilities on a large scale with low barriers, and helping enterprises make better judgments and decisions in multiple high-value and high-difficulty decision-making scenarios such as precision marketing, risk control, product optimization, customer experience, and resource allocation.

[0033] The product has high maturity and stability, good usability, and graphical, drag-and-drop modeling experience, which can quickly achieve in-depth data analysis and model building without a single line of code. This product greatly reduces the threshold for using AI tools and improves the efficiency of modeling, allowing more users in the enterprise to have the opportunity to participate in the practical process of "artificial intelligence application" development, further promoting the development of enterprise data-driven. Summary of the invention

[0034] In response to the problems existing in the prior art, the present invention provides a visual full-process machine learning platform, control method, client and application.

[0035] The present invention is implemented as follows: a visualization control method for a visualization full-process machine learning platform, the visualization control method for the visualization full-process machine learning platform comprising:

[0036] Drag and drop pictures, videos, log files, backup data, and built-in operators of containers / virtual machine images on the client interface, and interactively build a visual modeling process to obtain a machine learning model.

[0037] Use the setting parameters entered in the operator parameters to train the obtained machine learning model, view the results, adjust the parameters through the visualization interface, train again and view the results, and repeat the cycle;

[0038] Obtain the parameters of the optimal visualization interface and the optimal model, and then visualize the images, videos, log files, backup data, and container / virtual machine images;

[0039] If the final evaluation result is not ideal, the visual process can be converted into a code process, and fine-tuned at the code level to ultimately achieve the ideal result. It can then be published as a general process and made into a proprietary industry case or a general case.

[0040] Furthermore, the dragging and dropping of pictures, videos, log files, backup data, and built-in operators of container / virtual machine images on the client interface and interactive construction of a visual modeling process to obtain a machine learning model specifically include:

[0041] Step 1: Create a visual modeling task: Get the new experiment sidebar drawer in the client interface, fill in the task name and task list;

[0042] Step 2: Interactively build the visual modeling process: Obtain the built-in machine learning modeling operators, the visual modeling canvas area, and the information that needs to be filled in for each operator and the information display area on the client interface;

[0043] Drag the operator with the mouse, enter the canvas area, the information that needs to be filled in the operator, the model operation parameters, and connect the component to the interface, and enter the feature field of the component and the operation data parameter field of the component in the component information filling area;

[0044] Step 3: View the results of visual modeling: Save the visual process drawn in the current canvas, train the visual training process, and view the results after visual modeling training.

[0045] Another object of the present invention is to provide a client, which includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the visualization control method of the visualization full-process machine learning platform.

[0046] Another object of the present invention is to provide a visualization control method of the above-mentioned visualization full-process machine learning platform for application in visualization display in the fields of finance, education, medical care, and information industries.

[0047] Another object of the present invention is to provide a financial technology case marketplace constructed by a visualization control method of the above-mentioned visualization full-process machine learning platform.

[0048] Another object of the present invention is to provide a visual full-process machine learning platform for implementing the visual control method of the visual full-process machine learning platform, wherein the visual full-process machine learning platform comprises:

[0049] Data security module, used to securely isolate user data during machine learning modeling;

[0050] Multi-department collaborative modeling module, used for collaborative work of machine learning tasks in multiple projects and collaborative modeling among multiple departments of an enterprise;

[0051] Visual drag-and-drop modeling module, which provides a visual drag-and-drop operation mode for machine learning modeling;

[0052] The operator parameter optimization module is used to optimize the operator parameters in the form of operator parameter configuration in the visualization interface;

[0053] Big data operator embedding module, used to embed big data operators such as classification, clustering, regression, and feature engineering operators in machine learning scenarios;

[0054] Operator scaling module, used for scaling up and down of dynamic resources for operator execution;

[0055] Model management module, used to manage machine learning models;

[0056] The service monitoring module is used to deploy the model online with one click for service. The service supports online prediction, offline prediction and monitoring of model services.

[0057] Furthermore, the data security assurance module includes a data management module, which is used for platform import of data from multiple sources, large file and ultra-large file processing, and object storage.

[0058] Further, the multiple source data include: local files, databases, and big data file system components;

[0059] For large file processing, the method of splitting, transferring, and merging files after uploading supports breakpoint-resume transmission of large files on a single machine.

[0060] For processing of very large files, use the docking big data file system for processing;

[0061] Object storage uses an open source protocol for object storage, and the objects include pictures, videos, log files, backup data, and container / virtual machine images.

[0062] Furthermore, the multi-department collaborative modeling module uses k8s to isolate tasks and resources for different departments or teams through namespaces.

[0063] Furthermore, the visual drag-and-drop modeling module performs one-stop, full-process processing of machine learning through visual modeling task creation, interactive construction of modeling tasks, and viewing of modeling results.

[0064] Furthermore, the operator parameter optimization module is also used to obtain a model with excellent effect by continuously adjusting parameters, and finally obtain a model prediction result with high accuracy;

[0065] The operator expansion and contraction module is also used for operators to submit big data tasks;

[0066] The model management module is also used for the management of user models and model repositories.

[0067] Combining all the above technical solutions, the advantages and positive effects of the present invention are as follows: The present invention provides a visual full-process machine learning platform, including 8 core functional modules such as user management, project engineering management, data management, data preprocessing, feature engineering, model training, model evaluation and model publishing to offline or online services. The existing technology only includes feature engineering and model training modules, lacks user role authority management in modeling, user engineering management, data source management, data security management, online prediction and offline prediction services in model deployment, and model anomaly monitoring functions.

[0068] Comparison table of prior art products and the product of the present invention:

[0069]

[0070]

[0071] The eight core functional modules provided by the present invention can achieve: The present invention provides a comprehensive data source management system for enterprises to ensure the security, isolation and stability of user data in the process of machine learning modeling. It provides collaborative operations for multi-project machine learning tasks in enterprises to meet the collaborative modeling of multiple departments in enterprises. It provides a visual drag-and-drop machine learning modeling tool, which greatly reduces the threshold for model development. It provides a form of operator parameter configuration in a visual interface to help quickly select operator parameters. It embeds dozens of common big data operators for classification, clustering, regression, and feature engineering operators. The operators are developed based on spark to support machine learning scenarios with hundreds of billions of data volumes and TB-level quantities. Operator execution based on Spark on Kubernetes can achieve dynamic resource expansion and contraction. It provides a unified model management platform to facilitate algorithm personnel to manage machine learning models. It provides a one-click deployment model to provide services online. The service supports online prediction and offline prediction and model service monitoring functions.

[0072] The present invention accelerates the above-mentioned process through a visual interactive process. Engineers can build a machine learning model training process by simply dragging and dropping built-in operators. They can also tune the model parameters by setting parameters on the page and viewing the results. Through model deployment, they can provide online machine learning model services to the outside world.

[0073] The present invention has made relevant analysis and designed and implemented a one-stop machine learning platform system. The system provides a one-stop machine learning solution, and its functions can cover the entire process of machine learning business. The present invention reduces the human cost of algorithm engineers in enterprises and breaks through the problem of insufficient artificial intelligence talents in enterprises.

[0074] Currently, there is a shortage of algorithm engineers related to artificial intelligence and their salaries are high. This system has greatly simplified the traditional machine learning process, which originally required writing code for manual feature engineering and model training, which took about several months. It only takes a few minutes to complete the process and complete the machine learning model training in a short time, reducing the workload of engineers and ultimately reducing the overall labor cost of the company's algorithm engineers.

[0075] The present invention provides support for scenarios where a single machine cannot process big data. After practical testing, the system can efficiently process data in the tens of millions or hundreds of millions. Visual modeling interaction enables rapid iteration of the modeling process. Manual modeling is difficult to implement quickly due to high technical barriers and long cycles. For complex scenarios, the visual modeling interaction process can achieve dynamic and intuitive parameter adjustment and modeling process adjustment, achieve rapid iteration of the modeling process, and achieve rapid implementation of machine learning modeling.

[0076] The present invention realizes a one-stop full-process machine learning process through the combination of the above modules, upgrading the traditional process that requires algorithm personnel to write codes and adjust parameters to a more efficient and safe visualization method, greatly improving the landing efficiency of machine learning algorithms in the industry. These module combinations provide full life cycle management services from data management, data processing, model training, model management to model reasoning, helping users to build, train and deploy models more quickly.

[0077] The platform provided by the present invention is a zero coding experience, and modeling is achieved through visual drag and drop, so business personnel can quickly get started. The modeling speed is fast. In the past, it took 3 days to implement a model by coding, but it only takes 1 day to use this platform. The effect of the present invention is mainly reflected in the rapid coverage of the entire process of machine learning. Without this platform, completing a risk control model requires 5 days of work for database engineers, 10 days of work for algorithm engineers, 5 days of work for development engineers, and 5 days of work for business personnel. Currently, using this platform only requires 10 days of work for business personnel, and business personnel can completely model and deploy by themselves, which can help enterprises save labor costs, reduce costs and increase efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a schematic diagram of a visual full-process machine learning platform provided by an embodiment of the present invention.

[0079] In the figure: 1. Data security assurance module; 2. Multi-department collaborative modeling module; 3. Visual drag-and-drop modeling module; 4. Operator parameter selection module; 5. Big data operator embedding module; 6. Operator expansion and contraction module; 7. Model management module; 8. Service monitoring module.

[0080] Figure 2This is a diagram of using k8s to isolate tasks and resources to meet the needs of multi-department collaborative modeling.

[0081] Figure 3 It is a flow chart of the visual drag-and-drop modeling method provided by an embodiment of the present invention.

[0082] Figure 4 It is a flow chart of the interactive construction method of the visual modeling process provided by an embodiment of the present invention.

[0083] Figure 5 It is a design principle diagram of prior art 3 provided by an embodiment of the present invention.

[0084] Figure 6 It is a design principle diagram of the prior art 4 provided by an embodiment of the present invention.

[0085] Figure 7 It is a flow chart of mutual conversion between visual modeling and code modeling provided by an embodiment of the present invention.

[0086] Figure 8 This is a rendering of the financial technology case market provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0088] like Figure 4 As shown, prior art 3 (https: / / cloud.tencent.com / document / product / 851 / 39401) discloses a visualization control method for a visualization machine learning platform. Figure 5 As shown, prior art 4 (https: / / segmentfault.com / a / 1190000040523935) discloses a method for building a cloud-native AI platform based on Kubernetes.

[0089] The above two prior arts only describe general technical frameworks. The embodiment of the present invention describes a platform-level visual modeling control method based on the financial field. The main content and methods of this platform can also be realized by replacing the underlying technical framework.

[0090] In terms of the details of the control method, the contents described in the prior arts 3 and 4 are mostly aimed at general scenarios, and their target customers are Internet scenario customers, while the control method described in the present invention is mostly oriented to the B-side (systems or platforms used within enterprises or merchants). The platform allocates a large part of its resources and energy to user management design, which mainly includes: authority division and resource isolation.

[0091] The visualization platform of the present invention provides a rich variety of algorithms extending from the financial field to the general AI machine learning field, which are mainly classified as follows:

[0092] -Classification operators include XGBoost multi-classification, XGBoost binary classification, logistic regression binary classification, logistic regression multi-classification, random forest binary classification, random forest multi-classification, gradient boosted tree binary classification, gradient boosted tree multi-classification, naive Yes classification, gradient boosted tree classification, and decision tree classification.

[0093] -Clustering operators include KMeans clustering and Gaussian mixture clustering;

[0094] -Regression operators include XGBoost regression, linear regression, random forest regression, and gradient boosted tree regression.

[0095] -Feature engineering includes data standardization, data normalization, type conversion, lognormal transformation, logistic transformation, labelIndexer encoding and one-hot encoding, binning, tree model (random forest), tree model (XGBoost).

[0096] -Data preprocessing includes column selection, data splitting, missing value filling, renaming, correlation coefficient, describing data characteristics, random sampling, data deduplication, data filtering, Join, and Union.

[0097] -Model evaluation includes: binary classification evaluation, multi-classification evaluation, regression evaluation, and clustering evaluation.

[0098] -Financial components include: WOE coding and scorecard operators.

[0099] -The prediction component includes: prediction operator.

[0100] Since financial scenarios often involve user attribute-related mining and involve data from multiple sources, the input data sources supported by the platform of the embodiment of the present invention are richer than those in file 1. It not only includes formatted file formats such as csv and parquet mentioned in file 1, but also supports minio, FTP, Greenplum, PostgreSQL, EDB, MySQL, DB2, Oracle, Sqlserver, CDH, Hive's big data technology stack and its multi-source heterogeneous data sources.

[0101] This invention mainly targets AI machine learning modeling needs in financial scenarios. Its versatility can be extended to other fields, such as education, medical care, information industry, etc.

[0102] The platform of the present invention not only covers AI modeling and big data technology stack in financial scenarios, but also provides reusable cases such as scoring cards and risk control based on big data for financial scenarios.

[0103] The built-in cases of the present invention are aimed at general scenarios. The financial scenarios can be gradually refined as the products are released. General cases can be gradually evolved into special cases, and the special cases can be used to supplement the platform, and finally a financial technology case market can be established.

[0104] The platform provided by the present invention supports the conversion between code and visual modeling methods, quickly adjusts parameters, and accelerates the implementation of financial technology in enterprises.

[0105] In view of the problems existing in the prior art, the present invention provides a visual full-process machine learning platform, and the present invention is described in detail below with reference to the accompanying drawings.

[0106] like Figure 1 As shown, a visual full-process machine learning platform provided by an embodiment of the present invention includes:

[0107] Data security assurance module 1 is used to provide the enterprise with a comprehensive data source management system to ensure the security, isolation and stability of user data during the machine learning modeling process.

[0108] The multi-department collaborative modeling module 2 is used to provide collaborative operations for multi-project machine learning tasks in the enterprise to meet the needs of multi-department collaborative modeling in the enterprise.

[0109] The visual drag-and-drop modeling module 3 is used to provide a visual drag-and-drop machine learning modeling tool to lower the threshold for model development.

[0110] The operator parameter optimization module 4 is used to provide a visual interface for operator parameter configuration to optimize operator parameters.

[0111] The big data operator embedding module 5 is used to embed dozens of common big data operators such as classification, clustering, regression, and feature engineering operators. The operators are developed based on Spark and support machine learning scenarios with hundreds of billions of data volumes and TB-level quantities.

[0112] Operator scaling module 6 is used for operator execution to achieve dynamic resource scaling based on Spark on Kubernetes.

[0113] Model management module 7 is used to provide a unified model management platform to facilitate algorithm personnel to manage machine learning models.

[0114] The service monitoring module 8 is used to provide the function of one-click deployment of the model to provide services online. The service supports online prediction, offline prediction and monitoring of model services.

[0115] Ordinary technicians in the industry can also use other steps to implement the visual full-process machine learning platform provided by the present invention. Figure 1 The visual full-process machine learning platform provided by the present invention is only a specific embodiment.

[0116] The visualization control method of the visualized full-process machine learning platform of the present invention provides a one-stop machine learning scenario modeling service, covering the entire process of development, launch, and evaluation of algorithm effects, including: data processing, feature engineering, sample generation, model training, model evaluation, model publishing, online prediction, and effect evaluation.

[0117] In one embodiment of the present invention, a visualization control method of a visualization full-process machine learning platform is also provided, which is applied to a client. The visualization control method of the visualization full-process machine learning platform includes:

[0118] Drag and drop pictures, videos, log files, backup data, and built-in operators of containers / virtual machine images on the client interface, and interactively build a visual modeling process to obtain a machine learning model.

[0119] Use the setting parameters entered in the operator parameters to train the machine learning model, view the results, adjust the parameters of the visualization interface to train again and view the results, and repeat the cycle; obtain the parameters of the optimal visualization interface and the optimized model, and then visualize the images, videos, log files, backup data, and container / virtual machine images.

[0120] In one embodiment of the present invention, the dragging of pictures, videos, log files, backup data, and built-in operators of containers / virtual machine images on the client interface and interactive construction of a visual modeling process to obtain a machine learning model specifically includes:

[0121] Step 1: Create a visual modeling task: Get the new experiment sidebar drawer in the client interface, fill in the task name and task list;

[0122] Step 2: Interactively build the visual modeling process: Obtain the built-in machine learning modeling operators, the visual modeling canvas area, and the information that needs to be filled in for each operator and the information display area on the client interface;

[0123] Drag the operator with the mouse, enter the canvas area, the information that needs to be filled in the operator, the model operation parameters, and connect the component to the interface, and enter the feature field of the component and the operation data parameter field of the component in the component information filling area;

[0124] Step 3: View the results of visual modeling: Save the visual process drawn in the current canvas, train the visual training process, and view the results after visual modeling training.

[0125] The technical solution of the present invention is further described below in conjunction with specific embodiments.

[0126] Embodiment 1:

[0127] In one embodiment of the present invention, the data security assurance module 1 provides an enterprise with a comprehensive data source management system to ensure the security, isolation and stability of user data during the machine learning modeling process.

[0128] In response to the common big data processing and machine learning model training needs in the fields of finance and insurance, the platform effectively manages user data through data management functions. The platform provides the function of uploading a single big data file, supports the access of multi-source heterogeneous data and big data file systems, and uses object storage internally to transfer the data for user model training, realizing low-cost and scalable user data management. The main functions implemented by the data management module are:

[0129] (1) Platform import of data from multiple sources:

[0130] Local files: csv, xls, xlsx, txt;

[0131] Database: MySQL, Oracle, SQL Server;

[0132] Big data file system components: Hive, ElasticSearch, HDFS, Hbase;

[0133] By processing the above-mentioned types of files, it covers a variety of mainstream file storage formats and effectively supports users' diverse data carriers.

[0134] (2) Large and extra-large file processing:

[0135] For large files stored on a single machine, the system supports breakpoint-resume transmission of large files on a single machine through file splitting, transmission, and merging after uploading. A single file supports a maximum of 5G. For processing of ultra-large files, the system can be connected to the big data file system, and the maximum size supported can be approximately equal to the cluster size of the big data file system.

[0136] (3) Object storage:

[0137] Object storage stores all user data objects, but unlike files, objects will no longer have a hierarchical structure in a layer structure. This platform uses MinIO, which is an object storage service based on the Apache License v2.0 open source protocol. It is compatible with the Amazon S3 cloud storage service interface and is very suitable for storing large-capacity unstructured data, such as pictures, videos, log files, backup data, and container / virtual machine images, etc., and an object file can be of any size, ranging from a few kb to a maximum of 5T. The specific advantages are: high scalability, object storage can expand the capacity of tens or even hundreds of EB, and can make full use of high-density storage;

[0138] High efficiency, flat structure, not affected by complex directory system;

[0139] No migration is required. Object storage is a horizontally scalable system. As capacity increases, data is automatically distributed to all object storage nodes based on an algorithm.

[0140] High security. Object storage usually provides data access by using HTTP to call the authentication key provided by the object storage itself.

[0141] Easy access. It not only supports HTTP(S) protocol, but also uses REST API to call and retrieve data. It also adds NFS and SMB support.

[0142] The cost is relatively low. Compared with block storage, object storage is the most cost-effective type of data storage, and when combined with cloud computing, this feature of object storage can be fully utilized.

[0143] In one embodiment of the present invention, the multi-department collaborative modeling module 2 provides collaborative operations for multi-project machine learning tasks in an enterprise to meet the needs of multi-department collaborative modeling in an enterprise.

[0144] The concept of tenants is not limited to cluster users, but can include a set of workloads consisting of computing, network, storage and other resources. In a multi-tenant cluster, it is necessary to provide as much security isolation as possible for different tenants within the cluster to avoid malicious tenants from attacking other tenants to the greatest extent possible, while ensuring that shared cluster resources are fairly distributed among tenants.

[0145] Embodiment 2:

[0146] As shown in 2. The platform provided by the present invention uses k8s to isolate tasks and resources to meet the needs of multi-department collaborative modeling. K8s logically isolates resources for different departments or teams through namespaces based on the complexity of the internal personnel structure of the enterprise, and defines the following types of business personnel:

[0147] Administrator: Has the ability to manage resources (including departments, projects, menus, roles, users, logs) and is responsible for managing the above content for all personnel on the platform.

[0148] User: Use K8s resources within the scope of permissions in the user's corresponding namespace.

[0149] The above restrictions can achieve the following three advantages:

[0150] Multi-tenant isolation: multiple users can be isolated from each other without being aware of each other, and modeling work can be performed concurrently.

[0151] Resource isolation: Make reasonable resource division in a limited environment to maximize resource utilization.

[0152] Project concurrency: supports concurrent training of different departments and multiple projects.

[0153] In one embodiment of the present invention, the visual drag-and-drop modeling module 3 provides a visual drag-and-drop machine learning modeling tool, which greatly reduces the threshold for model development.

[0154] Embodiment 3:

[0155] like Figure 3 As shown, the visual drag-and-drop modeling method of the present invention includes: one-stop and full-process processing of machine learning through visual modeling task creation, interactive construction of modeling tasks, and modeling result viewing; specifically including:

[0156] S101, Visual Modeling Task Creation:

[0157] Click the New Task button to pop up the New Experiment sidebar drawer. Fill in the task name (required) and task description (optional). Click OK to complete the new task. After the new task is created, you will enter the Visual Modeling Task List page.

[0158] S102, interactive construction of visual modeling process;

[0159] S103, view the visual modeling results.

[0160] In a preferred embodiment of the present invention, step S102 specifically includes: modeling requires custom parameter adjustment, and the modeling process includes the following steps, such as Figure 4 shown.

[0161] S201, the left side of the entire panel is the built-in machine learning modeling operator (hereinafter referred to as the operator), the middle is the visual modeling canvas area (drag and drop to connect operators), and the right side is the information that needs to be filled in for each operator and the information display area.

[0162] S202, drag the operator on the left side with the mouse, the operator will appear in the middle canvas area, and the information that needs to be filled in for the node operator will appear on the right side. Fill in the required items according to the corresponding modeling requirements.

[0163] S203, manual visualization modeling requires the user to fill in the model running parameters in the right area.

[0164] S204, drag other components to the panel again, connect the nodes from top to bottom through the connection line, and a pointing arrow will appear. Connect all components in series and parallel from top to bottom through the arrow line, which is the final running process of the model. If you need to delete the connection line, you can delete the components connected by the connection line.

[0165] S205, filling in operator information: The component information filling area is divided into parameter settings, setting the operation parameters related to the component, including: setting the feature field for entering the component; setting the operation data parameter field of the component.

[0166] Embodiment 4:

[0167] In a preferred embodiment of the present invention, step S103 of visualizing the modeling results specifically includes:

[0168] View through the buttons on the upper right side of the page, including: Save information: save the visualization process drawn in the current canvas; Start training: train the visualization training process; View results: view the results after visualization modeling training.

[0169] The platform provided by the present invention realizes visualization, zero-code model training and parameter adjustment, and visualizes the development of artificial intelligence services by providing the above-mentioned visualized machine learning experimental development environment, greatly reducing the threshold for model training and development. Rich and detailed report content, including data distribution, analysis results, model information, model comparison, model evaluation, etc., helps users easily obtain the rules and patterns hidden in business data.

[0170] Embodiment 5:

[0171] In one embodiment of the present invention, the operator parameter optimization module 4 provides a visual interface operator parameter configuration form to help quickly optimize the operator parameters. The most critical step in the entire process of machine learning is the parameter tuning of the algorithm. By continuously adjusting the parameters, a model with excellent effect can be obtained, thereby ultimately obtaining a model prediction result with higher accuracy.

[0172] After the interactive construction of the visual modeling process is completed, the platform provided by the present invention sets parameters at the operator parameters, then starts training, checks the results, and trains again through parameter adjustment on the visual interface to see whether the results are more effective. The above process is repeated to finally achieve parameter tuning and the output of the optimal model training process.

[0173] Embodiment 6:

[0174] In one embodiment of the present invention, the big data operator is embedded in the module 5; dozens of common big data operators of classification, clustering, regression, and feature engineering operators are embedded, and the operators are developed based on spark to support machine learning scenarios with hundreds of billions of data volumes and TB-level quantities. The classification operators include XGBoost multi-classification, XGBoost binary classification, logistic regression binary classification, logistic regression multi-classification, random forest binary classification, random forest multi-classification, gradient boosted tree binary classification, gradient boosted tree multi-classification, naive Yes classification, gradient boosted tree classification, and decision tree classification.

[0175] Clustering operators include KMeans clustering and Gaussian mixture clustering; regression operators include XGBoost regression, linear regression, random forest regression, and gradient boosting tree regression. Feature engineering includes data standardization, data normalization, lognormal transformation, logistic transformation, labelIndexer encoding, and one-hot encoding.

[0176] Embodiment 7:

[0177] In one embodiment of the present invention, the operator scaling module 6, the operator execution implements spark task submission based on Spark on Kubernetes. Submit directly to the API Server of K8S through spark-submit. After applying for resources, the Client starts a Kubernetes pod and runs Spark Driver. Spark Driver starts a group of Kubernetes pods. Each Kubernetes pod creates an Executor and executes the application code. After running the program code, Spark Driver cleans up the pod where the Executor is located and maintains the "Complete" state. Apache Spark is currently the most popular big data computing framework. Compared with Hadoop, it is the only choice to replace the MapReduce component. More and more companies are migrating from traditional MapReduce job scheduling to Spark. Spark has more and more supporters in the ecosystem. Of course, its outstanding internal API design makes it very easy to integrate with the existing Hadoop components (YARN / HDFS).

[0178] In the present invention, the advantages of Spark on Kubernetes are: in the application scenario of multi-tenants on a shared cloud computing platform, each user wants to have an independent and isolated application environment to reduce interference between each other's calls. For example, when making Python calls, different users may have different requirements for Python versions. Container technology can build a completely isolated, independent, and easily maintainable operating environment for different users and applications. This is something that YARN in the Hadoop era cannot meet using virtual resource allocation technology.

[0179] The same benefits of containerization in traditional software engineering apply to big data and Spark. Containers make your applications more portable, they simplify packaging of dependencies, and they enable repeatable and reliable build workflows. They reduce the overall devops load and allow you to iterate on your code faster.

[0180] In this invention, User Isolation: Deploying Spark on Kubernetes provides powerful features for free, such as multi-tenancy control using namespaces and quotas, and role-based access control for fine-grained security and data access.

[0181] Embodiment 8:

[0182] In one embodiment of the present invention, the model management module 7 provides a unified model management platform to facilitate algorithm personnel to manage machine learning models. Model management includes two functional modules: user model and model warehouse.

[0183] The main functions of the user model include:

[0184] 1) Provide users with the function of importing third-party models and guide users to correctly import their own trained models;

[0185] 2) Provide users with relevant model information to help them use the model correctly and make correct judgments on the model's calculation results;

[0186] 3) Guide users to quickly and accurately find relevant models and understand relevant information about the models and their input and output parameters;

[0187] 4) Provide relevant information for users to add new models and call models. You can view model version, model label, model framework and other information.

[0188] 5) The frameworks supported by the model include tensorflow, pytorch, spark, paddle, and sklearn;

[0189] 6) Model management platform: unifies the algorithm model format, provides support for the discovery, deployment, and switching of algorithm models, and supports partitioned models and very large models.

[0190] Embodiment 9:

[0191] In one embodiment of the present invention, the service monitoring module 8 provides a one-click deployment model to provide services online, and the service supports online prediction and offline prediction and model service monitoring functions. Specifically, it includes: supporting one-click deployment of tensorflow, pytorch, spark, paddle, sklearn models as online prediction services; supporting model service start, stop, and delete functions; supporting one-click deployment of tensorflow, pytorch, spark, paddle, sklearn models as offline prediction services; supporting model service start, stop, and delete functions; setting scheduled offline prediction scheduling tasks according to the day, week, and month; model services support monitoring of resources such as CPU, GPU, memory, storage, resource usage, and log information visualization; supporting service log display and abnormal information display, which is convenient for troubleshooting; model service container deployment in Kubernetes, load balancing according to access volume; supporting writing result data back to minio, FTP, Greenplum, PostgreSQL, EDB, MySQL, DB2, Oracle, Sqlserver, CDH, Hive.

[0192] This system is used to compare the time and accuracy of the full-process training and evaluation of the random forest multi-classification algorithm on the open source dataset iris.

[0193] This system (visual modeling and parameter adjustment) Handwritten code Hand-write code and adjust parameters Spend time 2 minutes 30 minutes 60 minutes Accuracy 96% 90% 96%

[0194] As can be seen from the table above, analysts can obtain a high accuracy rate in a short time by using the system’s visual modeling and parameter adjustment, which fully proves the system’s speed and effectiveness.

[0195] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0196] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A visualization control method for a visualization full-process machine learning platform, characterized in that: The visualization control method of the visualization full-process machine learning platform includes: Drag and drop pictures, videos, log files, backup data, and built-in operators of containers / virtual machine images on the client interface, and interactively build a visual modeling process to obtain a machine learning model. Use the setting parameters entered in the operator parameters to train the obtained machine learning model, view the results, adjust the parameters through the visualization interface, train again and view the results, and repeat the cycle; Obtain the parameters of the optimal visualization interface and the optimal model, and then visualize the images, videos, log files, backup data, and container / virtual machine images; If the final evaluation result is not ideal, the visual process can be converted into a code process, and fine-tuned at the code level to achieve the desired effect. The process can then be published as a general process and made into a proprietary industry case or a general case. The client interface is used to drag and drop pictures, videos, log files, backup data, and built-in operators of containers / virtual machine images, and interactively construct a visual modeling process to obtain a machine learning model, specifically including: Step 1: Create a visual modeling task: Get the new experiment sidebar drawer in the client interface, fill in the task name and task list; Step 2: Interactively build the visual modeling process: Obtain the built-in machine learning modeling operators, the visual modeling canvas area, and the information that needs to be filled in for each operator and the information display area on the client interface; Drag the operator with the mouse, enter the canvas area, the information that needs to be filled in the operator, the model operation parameters, and connect the component to the interface, and enter the feature field of the component and the operation data parameter field of the component in the component information filling area; Step 3: View the results of visual modeling: save the visual process drawn in the current canvas, train the visual training process, and view the results after visual modeling training in the interface; The input data source of the visualization control method includes a formatted file format, a big data technology stack, and a multi-source heterogeneous data source; The built-in cases of the visualization control method are targeted at general scenarios. The financial scenarios can be gradually refined as the products are released, and general cases can be gradually evolved into special cases, which can be used to supplement the platform and eventually establish a financial technology case market. A visualization full-process machine learning platform that implements the visualization control method of the visualization full-process machine learning platform, the visualization full-process machine learning platform comprising: Data security module, used to securely isolate user data during machine learning modeling; Multi-department collaborative modeling module, used for collaborative work of machine learning tasks in multiple projects and collaborative modeling among multiple departments of an enterprise; Visual drag-and-drop modeling module, which provides a visual drag-and-drop operation mode for machine learning modeling; The operator parameter optimization module is used to optimize the operator parameters in the form of operator parameter configuration in the visualization interface; Big data operator embedding module, used to embed big data operators such as classification, clustering, regression, and feature engineering operators in machine learning scenarios; Operator scaling module, used for scaling up and down of dynamic resources for operator execution; Model management module, used to manage machine learning models; The service monitoring module is used to deploy the model online with one click for service. The service supports online prediction, offline prediction and monitoring of model services.

2. The visualization control method of the visualization full-process machine learning platform according to claim 1, characterized in that: The data security module includes a data management module, which is used for platform import of data from multiple sources, large file and ultra-large file processing, and object storage.

3. The visualization control method of the visualization full-process machine learning platform according to claim 1, characterized in that: Data from multiple sources including: local files, databases, and big data file system components; For large file processing, the method of splitting, transferring, and merging files after uploading supports breakpoint-resume transmission of large files on a single machine. For processing of very large files, use the docking big data file system for processing; Object storage uses an open source protocol for object storage, and the objects include pictures, videos, log files, backup data, and container / virtual machine images.

4. The visualization control method of the visualization full-process machine learning platform according to claim 1, characterized in that: The multi-department collaborative modeling module uses namespaces to isolate tasks and resources for different departments or teams.

5. The visualization control method of the visualization full-process machine learning platform according to claim 1, characterized in that: The visual drag-and-drop modeling module is created through visual modeling tasks, interactively constructs modeling tasks, and views modeling results to perform one-stop, full-process processing of machine learning; The operator parameter optimization module is also used to obtain a model with excellent effect by continuously adjusting parameters, and finally obtain a model prediction result with high accuracy; The operator expansion and contraction module is also used for operators to submit big data tasks; The model management module is also used for the management of user models and model repositories.

6. A financial technology case marketplace constructed by the visualization control method of the visualization full-process machine learning platform as described in claim 1.

7. A client, characterized in that: The client includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the visualization control method of the visualization full-process machine learning platform according to claim 1.