Digital feature derivation method and system in model life cycle

Abstract

The invention relates to a digital feature derivation method and system in a model life cycle, and the method comprises the steps: obtaining time sequence data, continuously adding incremental data to the head of a sequence, and removing the data at the tail of the time sequence; performing feature statistics on the time sequence data through a preset feature statistics operator; and receiving data input of a data supply side through a preset instant computing engine, performing real-time processing on the input data based on the metadata configuration and the scene rule configuration, and outputting real-time feature data. According to the method, through a time sequence increment updating mode, repeated data processing and calculation are remarkably reduced, resource consumption is reduced, and the production speed of the feature data is increased. The real-time calculation engine serves as a comprehensive supplement of the pre-calculation features, and the instantaneity of the feature data and the service delivery efficiency are further improved. Through a sound model evaluation system, the model effect can be tracked in real time and accurately evaluated, and the model performance is further improved.

Description

Technical Field

[0001] The present invention relates to the field of big data processing technology, and in particular to a method and system for deriving digital features in a model life cycle. Background Art

[0002] In the current field of machine learning and big data processing, the model launch process has formed a set of standardized steps, including problem definition, sample exploration, exploratory data analysis, feature data collection and feature production construction, model training, and model deployment and release. This series of steps covers the entire process from understanding business needs to the actual application of the model, which involves both core issues in the field of machine learning and key links in big data processing. With the continuous advancement of big data processing frameworks, the operational threshold for enterprises to process massive data sets has been greatly reduced, allowing model engineers to independently complete all work from problem definition to model deployment and release.

[0003] At present, model engineers mainly rely on the capabilities provided by the big data processing framework in terms of feature data collection and feature production construction. By writing data processing scripts or using data processing tools, they clean, transform, and aggregate the original data to generate feature data for model training. At the same time, during the model training phase, engineers will use various machine learning algorithms and frameworks to train feature data to obtain a model with excellent performance. Finally, during the model deployment and release phase, the trained model is integrated into the business system to achieve online prediction and decision-making of the model.

[0004] However, with the continuous application of upper-level algorithm models and the accumulation of underlying feature data, existing technologies have gradually exposed a series of problems. Since model engineers usually lack development and management experience in the data field, problems such as high cost of feature data production and maintenance, high degree of duplication and redundancy, timeout of feature production task readiness, and unclear feature data caliber and difficulty in reuse have become increasingly prominent. These problems not only affect the efficiency of feature data production and use, but also restrict the further improvement of model decision-making effects. Therefore, a new method is urgently needed to optimize the feature derivation process in the model life cycle to solve the problems of existing technologies and improve the overall performance of the model. Summary of the invention

[0005] In view of this, it is necessary to provide a digital feature derivation method in the model life cycle to solve the above-mentioned defects of the prior art.

[0006] In order to solve the above problems, in a first aspect, an embodiment of the present invention provides a method for deriving digital features in a model life cycle, including: Obtain time series data, continuously add incremental data to the head of the sequence, and remove data at the tail of the time series; perform feature statistics on the time series data using the preset feature statistics calculator; The preset real-time computing engine receives data input from the data supply side, processes the input data in real time based on metadata configuration and scenario rule configuration, and outputs real-time feature data.

[0007] Preferably, the acquiring of time series data and performing feature statistics on the time series data using a preset feature statistics calculator include: The repeated aggregation logic between multiple tasks is sunk to the middle layer to aggregate the detailed data according to the user granularity to generate user time series data; Based on the time series data, filter the subscripts of the statistical detail data in the time series through expressions; Substitute the statistical subscript into each field of the time series data to obtain statistical details, and complete the data statistics through the feature statistics calculator.

[0008] Preferably, the operation process of the instant computing engine includes: Select a data source, and obtain real-time data from the selected data source; wherein the data source includes a message queue, a real-time database, and a streaming file system; Select detail tables and dimension tables under different subject domains to perform data cleaning, data conversion, data aggregation and data filtering.

[0009] According to the characteristics of data streams, time semantics is selected to process real-time data; Define the output of the data stream to output the processing results; Configure parallelism and fault tolerance mechanisms based on data volume and computing requirements; Deploy the configured application to the instant computing engine cluster and start running after approval; Monitor the performance and status of computing applications in real time and optimize computing applications based on data latency, throughput, and resource utilization.

[0010] Preferably, the method further comprises: Iteratively optimize the initial model through incremental iteration or non-incremental iteration; The performance of different versions of the model before and after iterative optimization was evaluated using the A / B testing method.

[0011] Preferably, the incremental iterative process includes: Optimize the initial model by adjusting hyperparameters, improving feature engineering, and optimizing model structure; The non-incremental iterative process includes: The initial model is reconstructed based on model selection, feature selection, and model architecture design.

[0012] Preferably, the A / B testing method is used to evaluate the performance of different versions of the model before and after iterative optimization, including: Configure the model traffic replay function and / or the model real-time tracking function according to the optimization goal of incremental iteration or non-incremental iteration; Set up Group A and Group B. Group A uses the current online model, and Group B uses the new iterative model. For the traffic replay function, pre-configure the model call address after retraining, associate the online model and preview the corresponding model parameters; configure the replay start time and end time nodes, and determine the corresponding features and model evaluation indicators; For the real-time tracking function, select the model to be called online in real time and click the Track button to configure the calling address and traffic distribution rules of the tracking model; During the A / B test, collect metrics data for both Group A and Group B; Evaluate the collected data and decide whether to publish the new iterative model online based on the evaluation results.

[0013] In a second aspect, an embodiment of the present invention provides a digital feature derivation system in a model life cycle, including: The time series feature production module is used to continuously add incremental data to the head of the sequence and remove data at the tail of the time series; it performs feature statistics on the time series data through the preset feature statistics calculator; The real-time computing engine module is used to receive data input from the data supply side through a preset real-time computing engine, process the input data in real time based on metadata configuration and scenario rule configuration, and output real-time feature data.

[0014] Preferably, the system further comprises: An incremental iteration module is used to iteratively optimize the initial model through incremental iteration or non-incremental iteration; The A / B testing module is used to evaluate the performance of different versions of the model before and after iterative optimization through the A / B testing method.

[0015] In a third aspect, the present invention further provides an electronic device, including a memory and a processor, wherein: The memory is used to store programs; The processor is coupled to the memory and is used to execute the program stored in the memory to implement the steps in the method for deriving digital features in the model life cycle as described in the embodiment of the first aspect of the present invention.

[0016] In a fourth aspect, the present invention further provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps in the method for deriving digital features in the model life cycle as described in the embodiment of the first aspect of the present invention.

[0017] Compared with the prior art, the digital feature derivation method and system in the model life cycle provided by the present invention have the following beneficial effects: 1) The present invention significantly reduces repeated data processing and calculation, reduces resource consumption, and improves the production speed of feature data through the incremental update of time series.

[0018] 2) Through good feature data production specifications, reasonable feature production mode design and visual configuration, the problems of high storage resource consumption and high computing resource consumption faced by feature production can be solved, and the big data cost of algorithm feature production can be reduced.

[0019] 3) By optimizing the quality of feature data, the effect of online models has been improved. Through a sound model evaluation system, the model effect can be tracked and accurately evaluated in real time, further improving model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a method for deriving digital features in a model life cycle provided by the present invention; Figure 2 A comparison diagram of the old feature production logic of the prior art and the feature variable production method based on time series of the present invention; Figure 3 This is a characteristic statistical principle diagram provided by the present invention.

[0021] Figure 4 A schematic diagram of the instant computing engine provided by the present invention; Figure 5 An architectural diagram of the instant computing engine provided by the present invention; Figure 6 A structural block diagram of a digital feature derivation system in a model life cycle provided by the present invention; Figure 7 This is a structural block diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0022] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0023] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0024] With the continuous accumulation of upper-level algorithm model applications and underlying feature data, the existing feature data production solutions have gradually exposed a series of problems. Since model engineers usually lack development and management experience in the data field, the high cost of feature data production and maintenance, high degree of duplication and redundancy, feature production task readiness timeout, and unclear feature data caliber and difficulty in reuse have become increasingly prominent. These problems not only affect the production and use efficiency of feature data, but also restrict the further improvement of model decision-making effects.

[0025] In view of this, the present invention provides a digital feature derivation method in the model life cycle, which solves the high storage resource consumption and high computing resource consumption problems faced by feature production through good feature data production specifications, reasonable feature production mode design and visual configuration, and reduces the big data cost of algorithm feature production. The following will be explained and introduced through multiple embodiments.

[0026] Figure 1 The flowchart of the digital feature derivation method in the model life cycle provided by the present invention. Figure 1 As shown in the figure, the digital feature derivation method in the model life cycle includes the following: S1, obtain time series data, continuously add incremental data to the head of the sequence, and remove data at the tail of the time series; perform feature statistics on the time series data through the preset feature statistics calculator; Figure 2 This is a comparison chart of the old feature production logic of the prior art and the feature variable production method based on time series of the present invention. Figure 2In the old feature production logic, the production of daily feature values ​​requires reading all detailed data within the statistical period. According to the calculation of 40 million details of daily IOUs, the amount of data read by a single task production is as high as 14.6 billion, and the shuffle overhead caused by the aggregation operation is very large. What's more, many count_distinct operations are added to individual tasks. During the calculation process, 14.6 billion of these operations are expanded into multiple re-aggregations, which has a catastrophic impact on the entire production queue. During the production of such tasks, it is difficult for other tasks to be allocated resources, resulting in large-scale task delays in the entire production queue. The most fundamental reason for the high resource consumption of the old production model is that a large part of the calculations in the daily feature production link are repeated. For example, the feature production of day T-1 will shuffle all detailed data from day 1 to day T-1. This data will continue to be shuffled during the feature production process of day T. The repeated shuffles waste a lot of computing resources. In addition to the repeated shuffles between daily production, this problem also exists between tasks. In order to solve the problem of repeated shuffles, in terms of implementation scheme, the present invention designs a production method of feature variables based on time series.

[0027] Reference Figure 2 , the production method of the characteristic variables based on time series provided in this embodiment ( Figure 2 In the new feature production logic, first, aggregate the detailed data of a given period as time series data. The given period is preferably set to one year. At initialization, aggregate one year's detailed data to form a time series. On subsequent days, add the newly generated incremental data to the head of the time series. As the statistical time window slides, the old data at the tail of the time series is removed to maintain the timeliness and accuracy of the time series.

[0028] The present invention significantly reduces repeated data processing and calculation, reduces resource consumption, and improves the production speed of feature data by means of time series incremental update.

[0029] Among them, detailed data refers to the original, unaggregated or processed data records, each record contains a specific behavior or state information of the user, as well as the timestamp of the behavior or state. In the present invention, detailed data is the basis for generating time series data. By aggregating and processing detailed data at user granularity, time series data for feature statistics and calculations can be generated. Time series data refers to a series of data points arranged in chronological order, which reflect the changes of a variable or multiple variables over time. In the present invention, time series data specifically refers to data aggregated at user granularity.

[0030] Then, feature statistics are performed on the time series data through the pre-developed feature statistics operator. Since the aggregated time series data ensures that the data to be processed by a single user is on the same node, the feature statistics task can complete data production through the map-only mode, which improves processing efficiency. In the map stage, the system traverses the time series data of each user and performs calculations according to the predefined feature statistics operators. These operators process the data based on the detailed data in the time series data according to specific statistical logic, and finally generate the required feature variables. Since the map-only mode avoids the transmission and merging of data between different nodes, it can significantly improve the speed and efficiency of data processing. At the same time, this mode also reduces the complexity and resource consumption of the system, making the production of feature variables more efficient and reliable.

[0031] Figure 3 The characteristic statistical principle diagram provided by the present invention is shown in FIG. Figure 1 and Figure 3 In step S1, time series data is obtained, and feature statistics are performed on the time series data using a preset feature statistics calculator, which may specifically include: The aggregation logic repeated between multiple tasks is sunk to the middle layer to achieve, and the detailed data is aggregated according to the user granularity to generate user time series data. In this embodiment, the aggregation logic is sunk to the middle layer, making the feature production process more flexible and able to adapt to different data structures and statistical requirements.

[0032] Based on the time series data, filter the subscripts of the statistical detail data in the time series through expressions; Substitute the statistical subscript into each field of the time series data to obtain statistical details, and complete the data statistics through the feature statistical calculator. In this embodiment, the feature statistical calculator performs statistics based on the detailed data after the time series is filtered, ensuring the accuracy and reliability of the statistical results.

[0033] The time series-based feature variable production method provided in the embodiment of the present invention greatly reduces the computing resource consumption and running time of feature production through time series incremental update and feature map-only calculation. This method reduces repeated calculations and improves production efficiency by optimizing the process of data aggregation and feature statistics. At the same time, it also improves the quality of feature data and provides a better data foundation for subsequent model construction and training.

[0034] S2 receives data input from the data supply side through a preset real-time computing engine, processes the input data in real time based on metadata configuration and scenario rule configuration, and outputs real-time feature data.

[0035] It is understandable that precomputed features have the inherent characteristics of high cost and slow delivery. As feature data continues to increase, the precomputed production cost invested is not proportional to the business benefits, and many features are used very rarely. Precomputed features refer to feature data that is precomputed and stored before model training or inference.

[0036] In the context of the large exhaustive space of pre-calculation features, high storage and computing costs, and low immediacy of business delivery, the embodiment of the present invention provides a new instant computing engine. Instant computing based on details is a comprehensive supplement to the pre-calculation features. The instant computing engine provides interface configuration rules, which can be directly configured by most business personnel, facilitating quick trial and error. A small number of requirements can be developed quickly and put online at low cost.

[0037] Figure 4 The schematic diagram of the instant computing engine provided by the present invention is shown in FIG. Figure 3 The real-time computing engine receives data input in the form of Kafka topic streams, processes the input data in real time based on metadata configuration and scenario rule configuration, and generates real-time feature data. These feature data are continuously output in the form of calculation result streams and delivered to downstream delivery platforms or business systems through dedicated message channels.

[0038] Figure 5 The architecture diagram of the instant computing engine provided by the present invention is shown in FIG. Figure 5 ,The instant computing engine can be composed of the expression layer, the translation layer, the execution layer and the output layer.

[0039] The expression layer is used to process the source data through a series of scenario rule expressions. The translation layer is used to convert Flink SQL (including DDL data definition language and DML data manipulation language) into Flink jobs, realize the conversion from SQL to execution plan, and ensure the correct execution of data processing logic. Flink SQL is a SQL interface provided by Apache Flink for processing stream data and batch data. It includes DDL (data definition language) and DML (data manipulation language). Flink job refers to the data processing task submitted and executed on the Apache Flink platform. The execution layer is responsible for the scheduling and execution of Flink jobs. This layer performs syntax translation, job generation and scheduling generation to ensure that the job can be executed efficiently and accurately. The output layer is used to process the output results of Flink jobs. Including the selection of output channels, adaptation of output formats and flow control, to ensure that the processing results can be accurately and efficiently transmitted to the target system or storage medium.

[0040] In this embodiment, the operation process of the real-time computing engine specifically includes: Select a data source and obtain real-time data from the selected data source; wherein the data source includes a message queue (Apache Kafka, RabbitMQ), a real-time database (Redis, Apache Cassandra) and a streaming file system (Apache Hadoop HDFS); Select detail tables and dimension tables under different subject domains to perform data cleaning, data conversion, data aggregation and data filtering to meet business needs.

[0041] According to the characteristics of the data stream, time semantics are selected to process real-time data. In this embodiment, the characteristics of the data stream refer to that events in the data stream usually have different time attributes, such as event time, processing time, and ingestion time. Appropriate time semantics are selected to process real-time data.

[0042] Define the output of the data stream to output the processing results; Configure parallelism and fault tolerance mechanisms based on data volume and computing requirements; fault tolerance mechanisms are used to handle failures and data loss.

[0043] Deploy the configured application to the instant computing engine cluster and start running after approval; Monitor the performance and status of computing applications in real time, and optimize computing applications based on data latency, throughput, and resource utilization. Monitor the quality and volatility of data output and issue timely alerts.

[0044] The instant computing engine provided by the present invention can solve the problems of high cost, slow delivery, low frequency of use, etc. of pre-calculated features. It provides instant computing capabilities based on details to meet the urgent needs of businesses for real-time data processing. The engine is committed to reducing data processing delays, improving the immediacy of data delivery, and optimizing the cost-effectiveness ratio, so that businesses can quickly respond to market changes and achieve data-driven intelligent decision-making.

[0045] As a comprehensive supplement to pre-calculated features, the real-time computing engine further improves the immediacy of feature data and the efficiency of business delivery. The real-time computing engine configures rules through the interface, so that most business needs can be directly configured and implemented by business personnel, reducing development costs and speeding up the launch.

[0046] In a preferred embodiment of the present invention, the method for deriving digital features in the model life cycle provided by the present invention further includes: S3, iteratively optimizing the initial model through incremental iteration or non-incremental iteration; Specifically, model iteration is divided into two categories: incremental iteration refers to improving and optimizing the existing model by modifying certain aspects of the model to improve performance, which generally corresponds to retraining the model. Non-incremental iteration means redesigning and building the entire model each iteration instead of modifying the existing model. Based on the platform architecture, the platform provides the ability to quickly produce, iterate, and standardize features, improving the efficiency of building the entire model life cycle by more than 80%. The complete A / B testing solution includes solutions such as model traffic replay and real-time model tracking.

[0047] In this embodiment, the incremental iterative process includes: optimizing the initial model by adjusting hyperparameters, improving feature engineering and optimizing model structure. The non-incremental iterative process includes: reconstructing the initial model based on model selection, feature selection and model architecture design.

[0048] S4, the performance of different versions of the model before and after iterative optimization is evaluated through the A / B testing method.

[0049] Specifically, according to the optimization goal of incremental iteration or non-incremental iteration, the model traffic replay function and / or the model real-time following function are configured; Set up Group A and Group B. Group A uses the current online model, and Group B uses the new iterative model. For the traffic replay function, pre-configure the model call address after retraining, associate the online model and preview the corresponding model parameters; configure the replay start time and end time nodes, and determine the corresponding features and model evaluation indicators; For the real-time tracking function, select the model to be called online in real time and click the Track button to configure the calling address and traffic distribution rules of the tracking model; During the A / B test, collect metrics data for both Group A and Group B; Evaluate the collected data and decide whether to publish the new iterative model online based on the evaluation results.

[0050] As a preferred implementation, during the A / B test process, according to the optimization goal of incremental iteration or non-incremental iteration, the model traffic replay function and / or the model real-time following function are configured. Specifically, the configuration of the model traffic replay function and the model real-time following function can be described through the following embodiments: 1. Incremental iteration goal: Model improvement and optimization. Based on the existing model, some aspects are modified and adjusted to improve performance. This may include adjusting hyperparameters, improving feature engineering, optimizing model structure, etc. Retraining is a common way of incremental iteration.

[0051] 1.1 Model traffic replay function: Model developers need to configure the model calling address after retraining. Multiple addresses are supported. Associate the online model and preview the corresponding online model input parameter data; Configure the corresponding replay start time and end time nodes; Configure and check the corresponding features and model evaluation indicators, such as Accuracy, Precision, Recall, F1Score, ROC, AUC, etc. This embodiment provides the corresponding interface to add custom indicators; Generate the corresponding experimental report and draw the corresponding conclusion. If the effect after incremental iteration is better, it can be directly provided to the decision flow for calling.

[0052] 1.2 Real-time model tracking function: When it is impossible to accurately determine the model effect, the model can be tracked in real time; The model developer selects the model to be called online in real time and clicks the Follow button; Configure the corresponding call address of the following model. Here, the input parameters of the following model are allowed to be less than or equal to the online real-time call model. Configure the following traffic distribution rules, such as 100% traffic distribution, 50% traffic distribution, etc. After the configuration is completed, click Go Online, and the online traffic will be copied to the following model according to the percentage; Provide corresponding real-time indicators to observe the effect of the following model.

[0053] 2. Non-incremental iteration: Each iteration redesigns and builds the entire model instead of modifying the existing model. This may include model selection, feature selection, model architecture design, etc. Non-incremental iteration is usually used for fundamental model improvement and innovation. The model input here is inconsistent with the model input in real-time online calls. 2.1 Model traffic replay function: The input parameters of multiple models are different, which makes it impossible to directly use online model data for traffic replay; Model developers need to configure the model calling address after retraining. Multiple addresses are supported. The associated online model is used for model effect comparison; Configure the corresponding model input data, which involves the processing of different offline and real-time indicators as well as the use and processing of online model data; Configure and select the corresponding features and model evaluation indicators, such as Accuracy, Precision, Recall, F1Score, ROC, AUC, etc. We also provide corresponding interfaces to add custom indicators; Generate the corresponding experimental report and draw the corresponding conclusion. If the effect after non-incremental iteration is better, it can be directly provided to the decision flow for calling.

[0054] 2.2. Real-time model tracking function: When it is impossible to accurately determine the model effect, the model can be tracked in real time; The model developer selects the model to be called online in real time and clicks the Follow button; Configure the corresponding follow-up model calling address; After obtaining the primary key data of the online model, it obtains the input parameters of the corresponding non-incremental iteration model. There will be a delay of 5 to 10 seconds. After configuring the following traffic allocation rules, such as 100% traffic allocation, 50% traffic allocation, etc., click Go Online, and the online traffic will be copied to the following model according to the percentage.

[0055] The embodiment of the present invention enables real-time tracking and accurate evaluation of model effects through a sound model evaluation system, thereby further improving model performance.

[0056] Figure 6 The structural block diagram of the digital feature derivation system in the model life cycle provided by the present invention is as follows: Figure 6 As shown, an embodiment of the present invention provides a digital feature derivation system 600 in a model life cycle, including: The time series feature production module 601 is used to continuously add incremental data to the head of the sequence and remove the data at the tail of the time series; perform feature statistics on the time series data through a preset feature statistics calculator; The real-time computing engine module 602 is used to receive data input from the data supply side through a preset real-time computing engine, process the input data in real time based on metadata configuration and scenario rule configuration, and output real-time feature data.

[0057] Preferably, the system further comprises: The incremental iteration module 603 is used to iteratively optimize the initial model through incremental iteration or non-incremental iteration; The A / B testing module 604 is used to evaluate the performance of different versions of the model before and after iterative optimization through the A / B testing method.

[0058] The digital feature derivation system in the model life cycle provided by the present invention is used to execute the digital feature derivation method in the model life cycle provided by the above-mentioned embodiments through the above-mentioned modules. The digital feature derivation method in the model life cycle has been described in detail in the above-mentioned embodiments, and will not be repeated in this embodiment.

[0059] Figure 7The structural block diagram of the electronic device provided by the present invention is as follows: Figure 7 As shown, the present invention also provides an electronic device, and the electronic device 700 can be a computing device such as a mobile terminal, a desktop computer, a notebook, a palm computer, and a server. The electronic device 700 includes a processor 701 and a memory 702, wherein the memory 702 stores a digital feature derivation program 703 in the model life cycle.

[0060] In some embodiments, the memory 702 may be an internal storage unit of a computer device, such as a hard disk or memory of the computer device. In other embodiments, the memory 702 may also be an external storage device of the computer device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (SecureDigital, SD) card, a flash card (Flash Card), etc. equipped on the computer device. Furthermore, the memory 702 may also include both an internal storage unit of the computer device and an external storage device. The memory 702 is used to store application software and various types of data installed on the computer device, such as program codes installed on the computer device. The memory 702 can also be used to temporarily store data that has been output or is to be output. In one embodiment, when the digital feature derivation program 703 in the model life cycle is executed by the processor 701, the following steps are implemented: Obtain time series data, continuously add incremental data to the head of the sequence, and remove data at the tail of the time series; perform feature statistics on the time series data using the preset feature statistics calculator; The preset real-time computing engine receives data input from the data supply side, processes the input data in real time based on metadata configuration and scenario rule configuration, and outputs real-time feature data.

[0061] In some embodiments, the processor 701 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run program codes or process data stored in the memory 702, such as executing a digital feature derivation program in the model life cycle.

[0062] This embodiment further provides a computer-readable storage medium on which a digital feature derivation program in a model life cycle is stored. When the digital feature derivation program in the model life cycle is executed by a processor, the following steps are implemented: Obtain time series data, continuously add incremental data to the head of the sequence, and remove data at the tail of the time series; perform feature statistics on the time series data using the preset feature statistics calculator; The preset real-time computing engine receives data input from the data supply side, processes the input data in real time based on metadata configuration and scenario rule configuration, and outputs real-time feature data.

[0063] In summary, the digital feature derivation method and system in the model life cycle provided by the present invention have the following beneficial effects compared with the prior art: 1) The present invention significantly reduces repeated data processing and calculation, reduces resource consumption, and improves the production speed of feature data through the incremental update of time series.

[0064] 2) Through good feature data production specifications, reasonable feature production mode design and visual configuration, the problems of high storage resource consumption and high computing resource consumption faced by feature production can be solved, and the big data cost of algorithm feature production can be reduced.

[0065] 3) By optimizing the quality of feature data, the effect of online models has been improved. Through a sound model evaluation system, the model effect can be tracked and accurately evaluated in real time, further improving model performance.

[0066] The above-described embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for deriving digital features in a model life cycle, characterized in that: include: Get time series data, continuously add incremental data to the head of the sequence, and remove data at the tail of the time series; Perform feature statistics on time series data through preset feature statistics calculators; The preset real-time computing engine receives data input from the data supply side, processes the input data in real time based on metadata configuration and scenario rule configuration, and outputs real-time feature data.

2. The method for deriving digital features in the model life cycle according to claim 1, characterized in that: The step of obtaining the time series data and performing feature statistics on the time series data using a preset feature statistics calculator includes: The repeated aggregation logic between multiple tasks is sunk to the middle layer to aggregate the detailed data according to the user granularity to generate user time series data; Based on the time series data, filter the subscripts of the statistical detail data in the time series through expressions; Substitute the statistical subscript into each field of the time series data to obtain statistical details, and complete the data statistics through the feature statistics calculator.

3. The method for deriving digital features in the model life cycle according to claim 1, characterized in that: The operation process of the instant computing engine includes: Select a data source, and obtain real-time data from the selected data source; wherein the data source includes a message queue, a real-time database, and a streaming file system; Select detail tables and dimension tables under different subject domains to perform data cleaning, data conversion, data aggregation and data filtering. According to the characteristics of data streams, time semantics is selected to process real-time data; Define the output of the data stream to output the processing results; Configure parallelism and fault tolerance mechanisms based on data volume and computing requirements; Deploy the configured application to the instant computing engine cluster and start running after approval; Monitor the performance and status of computing applications in real time and optimize computing applications based on data latency, throughput, and resource utilization.

4. The method for deriving digital features in the model life cycle according to claim 1, characterized in that: The method further comprises: Iteratively optimize the initial model through incremental iteration or non-incremental iteration; The performance of different versions of the model before and after iterative optimization was evaluated using the A / B testing method.

5. The method for deriving digital features in the model life cycle according to claim 4, characterized in that: The incremental iterative process includes: Optimize the initial model by adjusting hyperparameters, improving feature engineering, and optimizing model structure; The non-incremental iterative process includes: The initial model is reconstructed based on model selection, feature selection, and model architecture design.

6. The method for deriving digital features in the model life cycle according to claim 1, characterized in that: The A / B testing method is used to evaluate the performance of different versions of the model before and after iterative optimization, including: Configure the model traffic replay function and / or the model real-time tracking function according to the optimization goal of incremental iteration or non-incremental iteration; Set up Group A and Group B. Group A uses the current online model, and Group B uses the new iterative model. For the traffic replay function, pre-configure the model call address after retraining, associate the online model and preview the corresponding model parameters; configure the replay start time and end time nodes, and determine the corresponding features and model evaluation indicators; For the real-time tracking function, select the model to be called online in real time and click the Track button to configure the calling address and traffic distribution rules of the tracking model; During the A / B test, collect metrics data for both Group A and Group B; Evaluate the collected data and decide whether to publish the new iterative model online based on the evaluation results.

7. A digital feature derivation system in a model life cycle for use in the digital feature derivation method in a model life cycle as claimed in any one of claims 1 to 6, comprising: The time series feature production module is used to continuously add incremental data to the head of the sequence and remove data at the tail of the time series; Perform feature statistics on time series data through preset feature statistics calculators; The real-time computing engine module is used to receive data input from the data supply side through a preset real-time computing engine, process the input data in real time based on metadata configuration and scenario rule configuration, and output real-time feature data.

8. The method for deriving digital features in the model life cycle according to claim 7, characterized in that: The system further comprises: An incremental iteration module is used to iteratively optimize the initial model through incremental iteration or non-incremental iteration; The A / B testing module is used to evaluate the performance of different versions of the model before and after iterative optimization through the A / B testing method.

9. An electronic device, It is characterized in that comprising a memory and a processor, wherein: The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the method for deriving digital features in the model life cycle as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps of the method for deriving digital features in the model life cycle as described in any one of claims 1 to 6 above.

Images