Feature data processing method and device, electronic equipment and storage medium

Abstract

The invention provides a feature data processing method and device, electronic equipment and a storage medium, and is applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic and aided driving, the feature data processing method comprises the following steps: inputting at least two pieces of service feature data to each service model to obtain a first service prediction result; performing feature adjustment on each piece of business feature data in sequence, and inputting at least two pieces of business feature data after feature adjustment to each business model to obtain a second business prediction result after adjustment for each piece of business feature data; generating feature contribution data according to the first service prediction result, the second service prediction result and the service label; obtaining resource adjustment contribution data of each piece of business feature data according to the model resource contribution data and the feature contribution data; and according to the resource adjustment contribution data and the feature storage information, obtaining an importance evaluation result of each piece of business feature data. The method can improve the evaluation precision of the importance degree of the business feature data.

Description

Technical Field

[0001] This application belongs to the field of computer technology, and particularly relates to a method, apparatus, electronic device, and storage medium for processing feature data. Background Art

[0002] The evaluation of feature importance is a relatively important task in machine learning, which is of great help for understanding problems. At the same time, through appropriate feature screening, the computational and storage overhead of the object recommendation system can be reduced.

[0003] Related technologies usually evaluate the importance of features by successively masking features and observing the impact on model metrics after each feature is masked. The greater the impact, the more important the feature is, or by training a tree model and using the information gain metric of each feature in the tree model as the standard for feature importance. However, related technologies mainly evaluate the importance of features for a single model. In an object recommendation system, often a feature is used by multiple models, and the feature value played by the same feature in models that adjust data for different resources is also different. Therefore, considering the resource contribution data and resource consumption data (such as cost) of the model is particularly important for improving the evaluation accuracy of feature importance and reducing the resource consumption of the object recommendation system. Summary of the Invention

[0004] To solve the above technical problems, this application provides a method, apparatus, electronic device, and storage medium for processing feature data.

[0005] On the one hand, this application proposes a method for processing feature data, the method includes:

[0006] Obtain at least two business models used by the object recommendation system at a first time, at least two business feature data used by the object recommendation system within a second time after the first time, and model resource contribution data obtained by each business model in processing business within the second time;

[0007] Input the at least two business feature data into each business model for business prediction processing to obtain a first business prediction result predicted by each business model; and sequentially perform feature adjustment on each business feature data, and input the at least two business feature data after feature adjustment into each business model for business prediction processing to obtain a second business prediction result predicted by each business model for each business feature data after adjustment;

[0008] Evaluate the model performance of each business model according to the first business prediction result, the second business prediction result, and the business labels of the at least two business feature data, and generate feature contribution data for each business feature data after input into each business model according to the model performance evaluation result;

[0009] According to the model resource contribution data and the feature contribution data, perform resource contribution data analysis and processing on each piece of service feature data to obtain the resource adjustment contribution data of each piece of service feature data;

[0010] According to the resource adjustment contribution data and the storage information of each piece of service feature data, perform feature importance evaluation processing on each piece of service feature data to obtain the importance evaluation result of each piece of service feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding service feature data when adjusting the resource consumption data of each service model.

[0011] On the other hand, the present application proposes a feature data processing device, and the device includes:

[0012] A model resource acquisition module, configured to acquire at least two service models used by the object recommendation system at a first time, at least two pieces of service feature data used by the object recommendation system within a second time after the first time, and model resource contribution data obtained by each service model when processing services within the second time;

[0013] A prediction module, configured to input the at least two pieces of service feature data into each service model for service prediction processing to obtain a first service prediction result predicted by each service model; and sequentially perform feature adjustment on each piece of service feature data, input the at least two pieces of service feature data after feature adjustment into each service model for service prediction processing, and obtain a second service prediction result predicted by each service model for each piece of service feature data after adjustment;

[0014] A performance evaluation module, configured to perform model performance evaluation on each service model according to the first service prediction result, the second service prediction result, and the service labels of the at least two pieces of service feature data, and generate feature contribution data of each piece of service feature data after being input into each service model according to the model performance evaluation result;

[0015] A resource contribution analysis module, configured to perform resource contribution data analysis and processing on each piece of service feature data according to the model resource contribution data and the feature contribution data to obtain the resource adjustment contribution data of each piece of service feature data;

[0016] An importance evaluation module, configured to perform feature importance evaluation processing on each piece of service feature data according to the resource adjustment contribution data and the storage information of each piece of service feature data to obtain the importance evaluation result of each piece of service feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding service feature data when adjusting the resource consumption data of each service model.

[0017] On the other hand, the present application proposes an electronic device for processing feature data. The electronic device includes a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the feature data processing method as described above.

[0018] On the other hand, the present application proposes a computer-readable storage medium. At least one instruction or at least one program segment is stored in the computer-readable storage medium, and the at least one instruction or the at least one program segment is loaded and executed by the processor to implement the feature data processing method as described above.

[0019] On the other hand, the present application proposes a computer program product, including a computer program which, when executed by a processor, implements the feature data processing method as described above.

[0020] The feature data processing method, apparatus, electronic device and storage medium proposed in the embodiments of the present application perform model performance evaluation on each business model according to the first business prediction result obtained by predicting at least two business feature data used at a second time after a first time by each business model, the second business prediction result obtained by predicting the business feature data after feature adjustment by each business model, and the business label. Generate feature contribution data after each business feature data is input into each business model according to the model performance evaluation result. Perform resource contribution data analysis processing on each business feature data according to the model resource contribution data obtained by each business model when processing business within the second time and the feature contribution data after each business feature data is input into each business model, to obtain the resource adjustment contribution data of each business feature data. Perform feature importance evaluation processing on each business feature data according to the resource adjustment contribution data and the storage information of each business feature data, to obtain the importance evaluation result of each business feature data. Since the model resource contribution data can reflect the resources contributed by the model to the object recommendation system, and the storage information can reflect the storage cost of the feature, it is possible to quantify the importance of the business feature data used by the object recommendation system from the perspective of model resource contribution data and cost, from the overall perspective of the object recommendation system, improve the evaluation accuracy of the importance of the business feature data, and reduce the resource consumption of the object recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0022] Figure 1 It is a schematic diagram of the implementation environment of a feature data processing method shown according to an exemplary embodiment.

[0023] Figure 2 It is a flowchart of a feature data processing method shown according to an exemplary embodiment Figure 1 .

[0024] Figure 3 It is a flowchart of a feature data processing method shown according to an exemplary embodiment Figure 2 .

[0025] Figure 4 It is a flowchart of a feature data processing method shown according to an exemplary embodiment Figure 3 .

[0026] Figure 5 It is a flowchart of a feature data processing method shown according to an exemplary embodiment Figure 4 .

[0027] Figure 6 It is a flowchart of a process for generating model performance difference information shown according to an exemplary embodiment.

[0028] Figure 7 It is a block diagram of a feature data processing device shown according to an exemplary embodiment.

[0029] Figure 8 It is a hardware structure block diagram of a server provided according to an exemplary embodiment. Detailed implementation

[0030] Artificial Intelligence (AI) uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, a theory, method, technology, and application system that perceives the environment, acquires knowledge, and uses knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable the machines to have the functions of perception, reasoning, and decision-making.

[0031] Artificial intelligence technology is an interdisciplinary subject with a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, large-feature data processing technology, operation / interaction systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0032] Among them, Machine Learning (ML) is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning. The pre-trained model is the latest development result of deep learning, integrating the above technologies.

[0033] Specifically, in the embodiments of the present application, the process of predicting business feature data according to the business model to obtain the business prediction result involves deep learning technology in machine learning.

[0034] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0035] It should be noted that the terms "first", "second", etc. in the specification, claims, and the above-mentioned drawings of the embodiments of the present application are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units does not necessarily have to be limited to those clearly listed steps or units, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products, or devices.

[0036] It should be noted that in the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be fully or partially implemented by using software, hardware (such as processing circuits or memories), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of the overall module or unit that includes the function of the module or unit.

[0037] In an object recommendation system, a piece of business feature data may often be used by multiple business models, such as attribute features like a user's age and gender. For the same piece of business feature data, if it ranks second in the business model that contributes data to the first model and first in the traffic model that contributes data to the second model, and the data contributed by the first model is much larger than that contributed by the second model, it is likely that the characteristic value of the former is relatively more effectively utilized. Therefore, when the production and storage resources of business feature data remain unchanged, the more benefits that business feature data can bring, the higher its importance to the object recommendation system. The embodiments of the present application propose a method for estimating the global importance of business feature data from the perspectives of resource contribution data and cost.

[0038] Figure 1 It is a schematic diagram of an implementation environment of a feature data processing method shown according to an exemplary embodiment. As Figure 1 shown, this implementation environment may at least include a terminal 01 and a server 02, and the terminal 01 and the server 02 can be directly or indirectly connected through wired or wireless communication means, and the embodiments of the present application do not limit this here.

[0039] Specifically, the terminal 01 can be used to collect business feature data. Optionally, the terminal 01 can include, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc.

[0040] Specifically, the server 02 can be used to evaluate the importance of business feature data to obtain an importance evaluation result of the business feature data. Optionally, the server 02 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, Content Delivery Network (CDN), and big data and artificial intelligence platforms.

[0041] It should be noted that Figure 1This is just an example. In other scenarios, other implementation environments may also be included. For example, the implementation environment may include a terminal that acquires business feature data and evaluates the importance of the business feature data to obtain the importance evaluation result of the business feature data.

[0042] Embodiments of the present invention can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc.

[0043] It can be understood that in the specific implementation of this application, user information is involved. For example, data related to age, gender, etc. When the above embodiments of this application are applied to specific products or technologies, user permission or consent needs to be obtained, and the collection, use, and processing of relevant data need to comply with relevant laws, regulations, and standards in relevant countries and regions.

[0044] Figure 2 It is a flowchart showing a method for processing feature data according to an exemplary embodiment. Figure 1 This method can be used in Figure 1 the implementation environment in. This specification provides the method operation steps as described in the embodiments or flowcharts, but based on routine or non-creative labor, more or fewer operation steps may be included. The step order listed in the embodiments is only one way among the execution orders of numerous steps and does not represent the only execution order. When the actual system or server product executes, it can be executed in the order of the embodiments or the method shown in the drawings, or executed in parallel (for example, in an environment of parallel processors or multi-threaded processing). Specifically, as Figure 2 shown, this method may include:

[0045] S101. Obtain at least two business models used by the object recommendation system at a first time, at least two business feature data used by the object recommendation system within a second time after the first time, and the model resource contribution data obtained by each business model in processing business within the second time.

[0046] Optionally, the type of the object recommendation system is determined according to the type of the object. The type of the object may include but not be limited to advertisements, commodities, etc. When the object is an advertisement, the object recommendation system may be an advertisement recommendation system, and when the object is a commodity, the object recommendation system may be a commodity recommendation system.

[0047] Optionally, the first time can be any time during the operation of the object recommendation system, and no specific limitation is imposed thereon. The at least two business models can be the business models used by the object recommendation system at the first time. Taking the object recommendation system as an advertisement recommendation system as an example, for an advertisement slot, at least two business models may be required. The at least two business models may include, but are not limited to, a click-through rate prediction model, a conversion rate prediction model, etc.

[0048] Optionally, the second time may refer to a period of time after the first time, and the duration of the second time is not limited in the embodiments of the present application. The at least two business feature data can be the data used by the object recommendation system within the second time after the first time. Taking the object recommendation system as an advertisement recommendation system as an example, assume that the first time is time T. At time T, at least two business models used by the object recommendation system are obtained, and then the business feature data used by the object recommendation system within a period of time t after time T can be obtained from the object recommendation system. Among them, the second time refers to the time period from T to T + t.

[0049] It should be noted that the business feature data can be a series of feature data used to characterize the terminal account and the object, and no specific limitation is imposed thereon. Taking the object recommendation system as an advertisement recommendation system as an example, the business feature data can be a series of feature data used to characterize the terminal account and the advertisement. For example, the attributes of the terminal account (age, gender, etc.), the attributes of the advertisement (advertisement type, advertisement duration, advertisement size, etc.).

[0050] Optionally, each business model receives its respective traffic data within the second time, and serves its respective traffic data to obtain the model resource contribution data of each business model. The model resource contribution data is used to characterize the resources contributed by the business model. For example, the resources contributed by the model to the object recommendation system. Taking the object recommendation system as an advertisement recommendation system as an example, assume that the business models include a click-through rate prediction model and a conversion rate prediction model. The click-through rate prediction model accounts for 40% of the traffic data, and the conversion rate prediction model accounts for 60% of the traffic data. The total model resource contribution data of the advertisement recommendation system is A. Then, the model resource contribution data of the click-through rate prediction model is A×40%, and the model resource contribution data of the conversion rate prediction model is A×60%. Among them, the model resource contribution data can specifically refer to the resource contribution data generated when the advertisement recommendation system recommends interesting advertisements to users within the time window from T to T + t, the users click on the interesting advertisements, and the advertisers are willing to pay.

[0051] S103. Input at least two pieces of service feature data into each service model for service prediction processing to obtain a first service prediction result predicted by each service model; and sequentially perform feature adjustment on each piece of service feature data, input the at least two pieces of service feature data after feature adjustment into each service model for service prediction processing to obtain a second service prediction result predicted by each service model for each piece of service feature data after adjustment.

[0052] In the embodiments of the present application, each service model has its respective corresponding service prediction ability. After obtaining at least two pieces of service feature data, the at least two pieces of service feature data can be respectively input into each service model, and each service model respectively performs service prediction processing on the at least two pieces of service feature data to obtain a first service prediction result predicted by each service model. Taking the object recommendation system as an advertisement recommendation system as an example, assume that the service models include a click-through rate prediction model and a conversion rate prediction model. The click-through rate prediction model is pre-trained and has the function of predicting whether a user clicks on the displayed advertisement, and the conversion rate prediction model is pre-trained and has the function of predicting whether a user converts the clicked advertisement (registering an account, downloading an APP, following a public account, form submission, etc.). The at least two pieces of service feature data can be respectively input into the click-through rate prediction model and the conversion rate prediction model to obtain a click-through rate prediction result predicted by the click-through rate prediction model and a conversion rate prediction result predicted by the conversion rate prediction model. The click-through rate prediction result and the conversion rate prediction result can be regarded as the first service prediction result.

[0053] In addition, to determine the second business prediction result after adjusting each piece of business feature data, each piece of business feature data among at least two pieces of business feature data can be adjusted in sequence to obtain at least two pieces of business feature data after feature adjustment, and the at least two pieces of business feature data after feature adjustment are respectively input into each business model for business prediction processing, so as to obtain the second business prediction result after adjustment for each piece of business feature data predicted by each business model. Taking the object recommendation system as an advertisement recommendation system as an example, assuming that the business models include a click-through rate prediction model and a conversion rate prediction model, and the business feature data are business feature data 1, business feature data 2, ……, business feature data N, business feature data 1 can be adjusted first to obtain at least two pieces of business feature data after adjustment (only business feature data 1 among the at least two pieces of business feature data is adjusted, and other business feature data remain unchanged), and the at least two pieces of business feature data after adjustment are respectively input into the click-through rate prediction model and the conversion rate prediction model to obtain the click-through rate prediction result predicted by the click-through rate prediction model and the conversion rate prediction result predicted by the conversion rate prediction model. The click-through rate prediction result and the conversion rate prediction result can be regarded as the second business prediction result after adjustment for business feature data 1. Then business feature data 2 can be adjusted, and the above operations can be repeated to obtain the second business prediction result after adjustment for business feature data 2, and so on, to obtain the second business prediction result after adjustment for each piece of business feature data.

[0054] Optionally, the adjustment of the business feature data may include, but is not limited to, masking processing, random perturbation processing, etc.

[0055] S105. Evaluate the model performance of each business model according to the first business prediction result, the second business prediction result, and the business labels of at least two pieces of business feature data, and generate feature contribution data for each piece of business feature data after being input into each business model according to the model performance evaluation result.

[0056] In the embodiment of the present application, after obtaining the first business prediction result and the second business prediction result, the model performance of each business model can be evaluated according to the first business prediction result, the second business prediction result, and the business labels of at least two pieces of business feature data to obtain the model performance evaluation results of each business model, and the feature contribution data for each piece of business feature data after being input into each business model are generated according to the model performance evaluation results of each business model.

[0057] Among them, the business label refers to the business result pre-labeled for at least two pieces of business feature data. Taking the object recommendation system as an advertisement recommendation system as an example, assuming that the business models include a click-through rate prediction model and a conversion rate prediction model, then the business label can be a click-through rate label, a conversion rate label, etc.

[0058] Among them, the feature contribution data after each business feature data is input into each business model can refer to: after each business feature is input into each business model, the data contributed to the resources contributed by the model of each business model to the object recommendation system.

[0059] Optionally, the model performance evaluation results may include, but are not limited to: AUC data, recall rate, F1 score, precision rate. Among them, the AUC data refers to the area enclosed by the ROC curve and the coordinate axes. The recall rate refers to the proportion of samples that are actually positive examples and are predicted as positive examples by the model. The precision rate refers to the proportion of samples predicted as positive examples by the model that are actually positive examples. The F1 score can be regarded as a harmonic mean of the precision rate and the recall rate.

[0060] S107. According to the model resource contribution data and the feature contribution data, perform resource contribution data analysis and processing on each business feature data to obtain the resource adjustment contribution data of each business feature data.

[0061] In the embodiment of the present application, after obtaining the model resource contribution data and the feature contribution data, from the perspective of resource contribution, resource contribution data analysis and processing can be performed on each business feature data to obtain the resource adjustment contribution data of each business feature data.

[0062] S109. According to the resource adjustment contribution data and the storage information of each business feature data, perform feature importance evaluation processing on each business feature data to obtain the importance evaluation result of each business feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding business feature data when adjusting the resource consumption data of each business model.

[0063] In the embodiment of the present application, after obtaining the resource adjustment contribution data of each business feature data, according to the resource adjustment contribution data of each business feature data and the storage information of each business feature data, feature importance evaluation processing can be performed on each business feature data to obtain the importance evaluation result of each business feature data. Since the model resource contribution data can reflect the resources contributed by the model to the object recommendation system, and the storage information can reflect the storage cost of the feature, the importance of the business feature data used by the object recommendation system can be quantified from the perspective of resource contribution data and cost, from the overall perspective of the object recommendation system, improving the evaluation accuracy of the importance of the business feature data and reducing the resource consumption of the object recommendation system.

[0064] It should be noted that the step of sequentially performing feature adjustment on each piece of business feature data, inputting at least two pieces of business feature data after feature adjustment into each business model for business prediction processing, and obtaining the second business prediction results adjusted for each piece of business feature data predicted by each business model can be implemented in various ways, and no specific limitation is made thereto.

[0065] In one implementation, the feature adjustment may refer to feature masking. Then, the step of sequentially performing feature adjustment on each piece of business feature data, inputting at least two pieces of business feature data after feature adjustment into each business model for business prediction processing, and obtaining the second business prediction results adjusted for each piece of business feature data predicted by each business model may include:

[0066] Sequentially performing masking processing on each piece of business feature data to obtain at least two pieces of business feature data after masking, inputting the at least two pieces of business feature data after masking into each business model for business prediction processing, and obtaining the second business prediction results adjusted for each piece of business feature data predicted by each business model.

[0067] In this embodiment, any piece of business feature data that has not been masked among the at least two pieces of business feature data can be used as the current business feature data, masking processing is performed on the current business feature data to obtain at least two pieces of business feature data after masking, the at least two pieces of business feature data after masking are input into each business model for business prediction processing, and the second business prediction results adjusted for the current business feature data predicted by each business model are obtained; repeat the operation of using any piece of business feature data that has not been masked among the at least two pieces of business feature data as the current business feature data until the above-mentioned second business prediction results adjusted for the current business feature data predicted by each business model are obtained until all of the at least two pieces of business feature data have been subjected to masking processing.

[0068] Taking an object recommendation system as an example of an advertisement recommendation system, assuming that the business model includes a click-through rate prediction model and a conversion rate prediction model, and the business feature data is business feature data 1, business feature data 2,..., business feature data N, it is possible to first mask business feature data 1 to obtain at least two masked business feature data, and input the at least two masked business feature data into the click-through rate prediction model and the conversion rate prediction model respectively to obtain the click-through rate prediction result predicted by the click-through rate prediction model and the conversion rate prediction result predicted by the conversion rate prediction model. The click-through rate prediction result and the conversion rate prediction result can be regarded as the second business prediction results for the masked business feature data 1. Then, business feature data 2 can be adjusted, and the above operations can be repeated to obtain the second business prediction results for the masked business feature data 2, and so on, to obtain the second business prediction results for each masked business feature data. Thus, by masking each business feature data in turn, it is possible to accurately obtain the business prediction results for each masked business feature data, that is, accurately obtain the impact of each masked business feature data on the performance of the business model, thereby improving the accuracy of the importance evaluation of business feature data.

[0069] In another embodiment, feature adjustment may refer to feature random perturbation. Then, the above-mentioned sequentially performing feature adjustment on each business feature data, inputting the at least two business feature data after feature adjustment into each business model for business prediction processing, and obtaining the second business prediction results predicted by each business model for each business feature data after adjustment may include:

[0070] Performing random perturbation processing on each business feature data in turn to obtain at least two business feature data after random perturbation, and inputting the at least two business feature data after random perturbation into each business model for business prediction processing to obtain the second business prediction results predicted by each business model for each business feature data after adjustment.

[0071] In this embodiment, any one of the at least two business feature data that has not been randomly perturbed can be used as the current business feature data, perform random perturbation processing on the current business feature data to obtain at least two business feature data after random perturbation, input the at least two business feature data after random perturbation into each business model for business prediction processing to obtain the second business prediction results predicted by each business model for the current business feature data after random perturbation; repeat the above operation of using any one of the at least two business feature data that has not been randomly perturbed as the current business feature data until the above-mentioned second business prediction results predicted by each business model for the current business feature data after random perturbation are obtained, until all of the at least two business feature data have been randomly perturbed.

[0072] It should be noted that the random perturbation may include, but is not limited to: randomly assigning a value to the service feature data, or adding a noise to the service feature data, etc.

[0073] Taking the object recommendation system as an example of the advertisement recommendation system, assume that the service model includes a click-through rate prediction model and a conversion rate prediction model, and the service feature data includes service feature data 1, service feature data 2,..., service feature data N. First, random perturbation can be performed on the service feature data 1 to obtain at least two pieces of service feature data after random perturbation, and the at least two pieces of service feature data after random perturbation are respectively input into the click-through rate prediction model and the conversion rate prediction model to obtain the click-through rate prediction result predicted by the click-through rate prediction model and the conversion rate prediction result predicted by the conversion rate prediction model. The click-through rate prediction result and the conversion rate prediction result can be regarded as the second service prediction result after random perturbation of the service feature data 1. Then, random perturbation can be performed on the service feature data 2, and the above operations can be repeated to obtain the second service prediction result after random perturbation of the service feature data 2, and so on, to obtain the second service prediction result after random perturbation of each piece of service feature data. Thus, by randomly perturbing each piece of service feature data in turn, the service prediction result after random perturbation of each piece of service feature data can be accurately obtained, that is, the influence of each piece of service feature data on the performance of the service model after being masked can be accurately obtained, thereby improving the accuracy of the importance evaluation of the service feature data.

[0074] It should be noted that in the above step S105, the model performance evaluation of each service model according to the first service prediction result, the second service prediction result, and the service labels of at least two pieces of service feature data, and generating the feature contribution data of each piece of service feature data after being input into each service model according to the model performance evaluation result can be implemented in various ways, and no specific limitation is made here.

[0075] Figure 3 It is a schematic flowchart of a feature data processing method shown according to an exemplary embodiment Figure 2 , such as Figure 3 shown, in an implementation manner, the model performance evaluation of each service model according to the first service prediction result, the second service prediction result, and the service labels of at least two pieces of service feature data, and generating the feature contribution data of each piece of service feature data after being input into each service model according to the model performance evaluation result may include:

[0076] S1051. Evaluate the model performance of each business model based on the first business prediction result and business labels to obtain the first model performance evaluation result, and evaluate the model performance of each business model based on the second business prediction result and business labels to obtain the second model performance evaluation result adjusted for each business feature data; the first model performance evaluation result and the second model performance evaluation result are used to represent the probability that the positive examples predicted by the business model are ranked in front of the negative examples.

[0077] S1053. Generate model performance difference information for each business feature data based on the first model performance evaluation result and the second model performance evaluation result.

[0078] S1055. Normalize the model performance difference information for each business feature data to obtain the feature contribution data after each business feature data is input into each business model.

[0079] Optionally, in the above step S1051, the first model performance evaluation result and the second model performance evaluation result can be used to represent the probability that the positive examples predicted by the business model are ranked in front of the negative examples, that is, the first model performance evaluation result and the second model performance evaluation result can be AUC data, and the model performance of each business model can be evaluated through the AUC data.

[0080] Exemplarily, taking the AUC data as an example, the above-mentioned evaluation of the model performance of each business model based on the first business prediction result and business labels to obtain the first model performance evaluation result is described as follows: after each business feature data is input into each business model, the first business prediction result for each business feature data predicted by each business model will be obtained, and each business feature data also has its corresponding business label. The business feature data in at least two business feature data can be sorted from large to small according to the corresponding first business prediction result. Then, starting from the business feature data with the largest first business prediction result, each business feature data is marked as a positive sample one by one, and the true positive rate (TPR) and false positive rate (FPR) are calculated at this time. Among them, TPR is the proportion of samples predicted as positive samples among all samples that are actually positive samples, and FPR is the proportion of samples predicted as positive samples among all samples that are actually negative samples. Then, starting from the business feature data with the largest first business prediction result, each business feature data is marked as a negative sample one by one, and the true positive rate and false positive rate are calculated at this time. The rectangular areas under the calculated true positive rates and false positive rates are finally added up, and the area under the ROC curve is the AUC data, that is, the first model performance evaluation result.

[0081] Exemplarily, taking the AUC data as an example, the above-mentioned model performance evaluation of each business model based on the second business prediction result and business label, and obtaining the adjusted second model performance evaluation result for each business feature data is described as follows: Sort each business feature data in at least two business feature data from largest to smallest according to the corresponding second business prediction result. Then, starting from the business feature data with the largest second business prediction result, mark each business feature data as a positive sample one by one, and calculate the true positive rate (TPR) and false positive rate (FPR) at this time. Among them, TPR is the proportion of samples predicted as positive samples among all samples that are actually positive samples, and FPR is the proportion of samples predicted as positive samples among all samples that are actually negative samples. Then, starting from the business feature data with the largest second business prediction result, mark each business feature data as a negative sample one by one, and calculate the true positive rate and false positive rate at this time. Then calculate the rectangular area under each true positive rate and false positive rate, and finally add up these rectangular areas. The area under the ROC curve is the adjusted AUC data for each business feature data, that is, the adjusted second model performance evaluation result for each business feature data.

[0082] Optionally, in the above step S1053, the difference between the first model performance evaluation result and the second model performance evaluation result can be directly calculated to generate model performance difference information for each business feature data. It is also possible to estimate the weights of the first model performance evaluation result and the second model performance evaluation result for the estimation of feature importance respectively, so as to assign a first weight to the first model performance evaluation result, assign a second weight to the second model performance evaluation result, calculate the first product between the first model performance evaluation result and the first weight, calculate the second product between the second model performance evaluation result and the second weight, and calculate the difference between the first product and the second product to obtain the model performance difference information.

[0083] Taking the first model performance evaluation result and the second model performance evaluation result as AUC data as an example, the calculation formula of the model performance difference information can be as follows:

[0084] AUC gap =AUC real -AUC drop ;

[0085] Wherein, AUC gap refers to the model performance difference information, AUC real refers to the first model performance evaluation result, and AUC real refers to the second model performance evaluation result.

[0086] In other embodiments, a first weight may be assigned to the first model performance evaluation result, and a second weight may be assigned to the second model performance evaluation result, and the first product of the first weight and AUC real and the second product of the second weight and AUC real are calculated, and the difference between the first product and the second product is calculated to obtain model performance difference information.

[0087] It should be noted that in addition to evaluating the model performance through AUC data, the model performance can also be evaluated by recall rate, F1 score, precision, etc.

[0088] Optionally, in the above step S1055, the model performance difference information for each business feature data can be normalized to obtain the feature contribution data after each business feature data is input into each business model.

[0089] In one way, continuing as Figure 3 shown, the above step S1055 may include:

[0090] S10551. Obtain the model performance difference information greater than the first performance difference threshold from the model performance difference information for each business feature data, and obtain the model performance difference information less than the second performance difference threshold from the model performance difference information for each business feature data; the second model performance threshold is less than the first model performance threshold.

[0091] S10553. Determine the first difference between the model performance difference information for each business feature data and the model performance difference information less than the second performance difference threshold, and determine the second difference between the model performance difference information greater than the first performance difference threshold and the model performance difference information less than the second performance difference threshold.

[0092] S10555. Normalize the model performance difference information for each business feature data according to the ratio between the first difference and the second difference to obtain the feature contribution data after each business feature data is input into each business model.

[0093] In this embodiment, the above first performance difference threshold and second performance difference threshold can be set according to actual business requirements, and no specific limitation is made thereto. For example, the model performance difference information greater than the first performance difference threshold can be the largest model performance difference information among all model performance difference information, and the model performance difference information less than the second performance difference threshold can be the smallest model performance difference information among all model performance difference information. The first difference between the model performance difference information for each business feature data and the smallest model performance difference information can be calculated, as well as the second difference between the largest model performance difference information and the smallest model performance difference information. Then, the ratio between the first difference and the second difference is calculated to normalize the model performance difference information for each business feature data, and the feature contribution data after each business feature data is input into each business model is obtained. Alternatively, the weights of the first difference and the second difference for feature importance evaluation are estimated respectively. The first difference is multiplied by the corresponding weight to obtain a third product, and the second difference is multiplied by the corresponding weight to obtain a fourth product. The ratio of the third product to the fourth product is calculated to obtain the feature contribution data after each business feature data is input into each business model.

[0094] Taking the first model performance evaluation result and the second model performance evaluation result as AUC data as an example, the calculation formula for the feature contribution data after each business feature data is input into each business model can be as follows:

[0095]

[0096] Among them, score i refers to the feature contribution data after each business feature data is input into each business model, AUC gapi refers to the model performance difference information for each business feature data, min(AUC gap ) refers to the smallest model performance difference information, max(AUC gap ) refers to the largest model performance difference information, (AUC gapi - min(AUC gap ) refers to the first difference, (max(AUC gap ) - min(AUC gap )) refers to the second difference.

[0097] In other embodiments, a third weight can also be assigned to the first difference and a fourth weight can be assigned to the second difference. The product of the first difference and the third weight is calculated, the product of the second difference and the fourth weight is calculated, and the ratio of the two products is calculated to obtain the feature contribution data after each business feature data is input into each business model.

[0098] Thus, the model performance difference information for each business feature data is normalized by the model performance difference information greater than the first performance difference threshold and the model performance difference information less than the second performance difference threshold, so as to scale the model performance difference information with different scales and ranges to the same range, thereby eliminating the influence caused by the dimension and unit differences, enabling different business feature data to be compared and analyzed on the same scale, facilitating improving the determination accuracy of the feature contribution data after each business feature data is input into each business model, and thus improving the accuracy of the importance evaluation of the business feature data.

[0099] Since the first model performance evaluation result is the evaluation result before the business feature data is adjusted, and the second model performance evaluation result is the evaluation result after the feature is adjusted, the model performance difference information for each business feature data can be accurately determined according to the difference between the two. Normalizing the model performance difference information for each business feature data enables different business feature data to be compared and analyzed on the same scale, facilitating improving the determination accuracy of the feature contribution data after each business feature data is input into each business model, and thus improving the accuracy of the importance evaluation of the business feature data.

[0100] It should be noted that in the above step S107, there are various ways to perform resource contribution data analysis processing on each business feature data according to the model resource contribution data and the feature contribution data to obtain the resource adjusted contribution data for each business feature data, and no specific limitation is made here.

[0101] Figure 4 It is a flowchart showing a method for processing feature data according to an exemplary embodiment Figure 3 , such as Figure 4 shown. In one implementation, in the above step S107, performing resource contribution data analysis processing on each business feature data according to the model resource contribution data and the feature contribution data to obtain the resource adjusted contribution data for each business feature data may include:

[0102] S1071. Determine the sum of the model resource contribution data obtained by each business model in processing the business within the second time to obtain the total model resource contribution data.

[0103] In this embodiment, it is possible to calculate the model resource contribution data obtained by each business model receiving its respective traffic data within the second time period and serving its respective traffic data, and calculate the sum of the model resource contribution data of each business model to obtain the total model resource contribution data. Taking the object recommendation system as an example of an advertising recommendation system, assume that the business models include a click-through rate prediction model and a conversion rate prediction model. The model resource contribution data obtained by the click-through rate prediction model serving the traffic data within the second time period is model resource contribution data 1, and the model resource contribution data obtained by the conversion rate prediction model serving the traffic data within the second time period is model resource contribution data 2. Calculate the sum of model resource contribution data 1 and model resource contribution data 2 to obtain the total model resource contribution data.

[0104] S1073. Normalize the model resource contribution data obtained by each business model processing the business within the second time period according to the model resource contribution data obtained by each business model processing the business within the second time period and the total model resource contribution data, to obtain the resource adjustment proportion information of each business model.

[0105] In one way, it is possible to directly calculate the ratio of the model resource contribution data obtained by each business model processing the business within the second time period to the total model resource contribution data, to obtain the resource adjustment proportion information of each business model, thereby improving the calculation convenience of the resource adjustment proportion information of each business model and reducing the resource consumption of the object recommendation system. The calculation formula can be as follows:

[0106]

[0107] where income j refers to the model resource contribution data obtained by each business model processing the business within the second time period, refers to the total model resource contribution data, refers to the resource adjustment proportion information of each business model.

[0108] In another way, it is also possible to estimate the weight of each business model's evaluation of the importance of business feature data, calculate the product of the model resource contribution data obtained by each business model processing the business within the second time period and its respective weight, calculate the ratio of this product to the total model resource contribution data, to obtain the resource adjustment proportion information of each business model.

[0109] S1075. Perform resource contribution data analysis processing on each business feature data according to the resource adjustment proportion information of each business model and the feature contribution data, to obtain the resource adjustment contribution data of each business feature data.

[0110] In this embodiment, resource contribution degree analysis processing can be performed on each business feature data according to the resource adjustment proportion information of each business model and the feature contribution data after each business feature data is input into each business model, so as to obtain the resource adjustment contribution data of each business feature data.

[0111] Since the model resource contribution data and the total sum of the model resource contribution data obtained by each business model processing the business within the second time can accurately determine the resource adjustment proportion information of each business model, and based on the resource adjustment proportion information, the resource adjustment contribution data of each business feature data is analyzed, which realizes calculating the resource adjustment contribution data of each business feature data from the perspective of the model resource contribution data (for example, the perspective of the resource contribution data), thereby improving the evaluation accuracy of the importance of the business feature data and reducing the resource consumption of the object recommendation system.

[0112] In a specific embodiment, continuing as Figure 4 shown, in the above step S1075, the resource contribution data analysis processing of each business feature data according to the resource adjustment proportion information and the feature contribution data of each business model to obtain the resource adjustment contribution data of each business feature data may include:

[0113] S10751. Determine the product of the resource adjustment proportion information of each business model and the feature contribution data after each business feature data is input into each business model, so as to obtain the resource contribution score after each business feature data is input into each business model.

[0114] S10753. Generate the resource adjustment contribution data of each business feature data according to the sum of the resource contribution scores after each business feature data is input into each business model.

[0115] In this embodiment, for each business feature data, a resource contribution score will be obtained after it is input into each business model. The sum of the resource contribution scores obtained after each business feature data is input into each business model can be directly calculated to obtain the resource adjustment contribution data of each business feature data. The calculation formula can be as follows:

[0116]

[0117] Wherein, refers to the resource adjustment contribution data of each business feature data, refers to the feature contribution data after each business feature data is input into each business model, It refers to the resource adjustment proportion information of each business model. Since in an object recommendation system, each business feature data may be used by multiple business models, and by summing the resource contribution scores of each business feature data input into each business model, the resource adjustment contribution data of each business feature data is generated, which can fully consider the resource contribution of each business feature data used by multiple business models. Thus, the importance of each business feature data can be evaluated from the overall perspective of the object recommendation system, rather than evaluating the importance of business feature data for a single business model, and then the global importance of the features used by the object recommendation system can be accurately quantified.

[0118] In other embodiments, among the resource contribution scores of each business feature data input into each business model, the resource contribution scores with significant differences from other resource contribution scores are filtered to obtain the resource contribution scores of each business feature data input into each business model after filtering. According to the sum of the resource contribution scores of each business feature data input into each business model after filtering, the resource adjustment contribution data of each business feature data is generated. For example, among the resource contribution scores of each business feature data input into each business model, there are resource contribution score 1, resource contribution score 2, and resource contribution score 3. If resource contribution score 3 is significantly different from resource contribution score 1 and resource contribution score 2, then resource contribution score 3 can be filtered out first, and the sum of resource contribution score 1 and resource contribution score 2 can be directly calculated to obtain the resource adjustment contribution data of each business feature data.

[0119] It should be noted that in the above step S109, the above-mentioned feature importance evaluation process for each business feature data based on the resource adjustment contribution data and the storage information of each business feature data to obtain the importance evaluation result of each business feature data can be carried out in various ways, and no specific limitation is made here.

[0120] Figure 5 is a flowchart illustration of a feature data processing method shown according to an exemplary embodiment Figure 4 , such as Figure 5 shown, in an alternative embodiment, in the above step S109, the above-mentioned feature importance evaluation process for each business feature data based on the resource adjustment contribution data and the storage information of each business feature data to obtain the importance evaluation result of each business feature data may include:

[0121] S1091. Obtain the storage capacity of each business feature data.

[0122] S1093. Normalize the storage capacity of each business feature data to obtain the normalized result of the storage capacity of each business feature data.

[0123] S1095. According to the resource adjustment contribution data of each business feature data and the normalized result of the storage capacity of each business feature data, perform feature importance evaluation processing on each business feature data to obtain the importance evaluation result of each business feature data.

[0124] In this embodiment, the storage capacity of the business feature data can be used to indicate the size of the business feature data. The larger the storage capacity, the larger the size of the business feature data and the higher the storage cost. Conversely, the smaller the size of the business feature data, the lower the storage cost. After obtaining the storage capacity of each business feature data, the storage capacity of each business feature data can be normalized to a standard range to obtain the normalized result of the storage capacity of each business feature data. Finally, according to the resource adjustment contribution data of each business feature data and the normalized result of the storage capacity of each business feature data, perform feature importance evaluation processing on each business feature data to obtain the importance evaluation result of each business feature data. Since normalization can scale storage capacities with different scales and ranges to the same range, thereby eliminating the influence caused by differences in dimensions and units, enabling different storage capacities to be compared and analyzed on the same scale, which is beneficial to improving the evaluation accuracy of feature importance evaluation for each business feature data; and since the storage capacity can reflect the storage cost of the business feature data, and the resource adjustment contribution data can reflect the resources contributed by the model to the object recommendation system, therefore, it is possible to quantify the global importance of the business feature data used by the object recommendation system model from the dimensions of resource contribution data and cost, and evaluate the importance of the business feature data from the overall perspective of the object recommendation system, improving the evaluation accuracy of the importance of the business feature data.

[0125] In other embodiments, the storage cost of the business feature data can also be used to replace the storage capacity of the business feature data. For example, if the storage capacity of the business feature data is 1GB, the 1GB storage capacity can be converted into a storage cost of 100 yuan, and this storage cost is normalized to obtain the normalized result of the storage cost of each business feature data. According to the resource adjustment contribution data of each business feature data and the normalized result of the storage cost of each business feature data, perform feature importance evaluation processing on each business feature data to obtain the importance evaluation result of each business feature data.

[0126] In an alternative embodiment, continuing as Figure 5 shown, in the above step S1093, the normalization of the storage capacity of each business feature data to obtain the normalized result of the storage capacity of each business feature data may include:

[0127] S10931. Obtain the storage capacity greater than the first capacity threshold from the storage capacity of each business feature data, and obtain the storage capacity less than the second capacity threshold from the storage capacity of each business feature data; the first capacity threshold is greater than the second capacity threshold.

[0128] S10933. Determine the third difference between the storage capacity of each business feature data and the storage capacity less than the second capacity threshold, and the fourth difference between the storage capacity greater than the first capacity threshold and the storage capacity less than the first capacity threshold.

[0129] S10935. Normalize the storage capacity of each business feature data according to the third difference and the fourth difference to obtain the normalized result of the storage capacity of each business feature data.

[0130] In this embodiment, the above first capacity threshold and second capacity threshold can be set according to actual business requirements, and no specific limitation is made here. For example, the storage capacity greater than the first capacity threshold can be the largest storage capacity among all storage capacities, and the storage capacity less than the second capacity threshold can be the smallest storage capacity among all storage capacities. The third difference between the storage capacity of each business feature data and the smallest storage capacity can be calculated, as well as the fourth difference between the largest storage capacity and the smallest storage capacity. Then calculate the ratio between the third difference and the fourth difference to normalize the storage capacity of each business feature data to obtain the normalized result of the storage capacity of each business feature data. Or, estimate the weights of the third difference and the fourth difference for the evaluation of feature importance respectively. Multiply the third difference by the corresponding weight to get a product, multiply the fourth difference by the corresponding weight to get a product, and calculate the ratio of the two products to obtain the normalized result of the storage capacity of each business feature data. The calculation formula can be as follows:

[0131]

[0132] Among them, refers to the normalized result of the storage capacity of each business feature data, S i refers to the storage capacity of each business feature data, min(S) refers to the smallest storage capacity, max(S) refers to the largest storage capacity, S i -min(S) refers to the third difference, and max(S)-min(S) refers to the fourth difference.

[0133] Thus, the storage capacity for each piece of service feature data is normalized by a storage capacity greater than the first capacity threshold and less than the second capacity threshold, so as to scale storage capacities with different scales and ranges to the same range, thereby eliminating the influence brought about by differences in dimension and unit, enabling different storage capacities to be compared and analyzed on the same scale, and facilitating the improvement of the accuracy of the importance evaluation of service feature data.

[0134] In an alternative embodiment, continuing as Figure 5 shown, in step S1095 above, the importance evaluation process for each piece of service feature data is performed based on the resource adjustment contribution data of each piece of service feature data and the normalized result of the storage capacity of each piece of service feature data, and the importance evaluation result of each piece of service feature data can include:

[0135] S10951. Obtain the preset resource adjustment coefficient of the object recommendation system.

[0136] Optionally, the preset resource adjustment coefficient can be a coefficient used to characterize the relationship between resource contribution data and cost, and it is a preset hyperparameter. In the case where the object recommendation system is an advertisement recommendation system, the preset resource adjustment coefficient can be the return on investment (ROI) coefficient.

[0137] S10953. Determine the product of the preset resource adjustment coefficient and the normalized result of the storage capacity of each piece of service feature data.

[0138] S10955. Based on the resource adjustment contribution data of each piece of service feature data and the product, perform an importance evaluation process for each piece of service feature data to obtain the importance evaluation result of each piece of service feature data.

[0139] In one approach, the difference between the resource adjustment contribution data of each piece of service feature data and the above product can be calculated to obtain the importance evaluation result of each piece of service feature data, and the calculation formula can be as follows:

[0140]

[0141] where score i refers to the importance evaluation result of each piece of service feature data, refers to the resource adjustment contribution data of each piece of service feature data, λ refers to the preset resource adjustment coefficient, refers to the storage capacity of each piece of service feature data, refers to the product of the preset resource adjustment coefficient and the normalized result of the storage capacity of each piece of service feature data.

[0142] Since the storage capacity can reflect the storage cost of business feature data, the preset resource adjustment coefficient can reflect the relationship between the preset cost and the resource contribution data, and the resource adjustment contribution data can reflect the resources contributed by the real model to the object recommendation system. Therefore, by evaluating the product of the storage capacity and the preset resource adjustment coefficient, as well as the importance of the model resource contribution data to the business feature data, it is possible to further quantify the global importance of the business feature data used by the object recommendation system model from the dimensions of resource contribution data and cost, and evaluate the importance of the business feature data from the overall perspective of the object recommendation system.

[0143] In other embodiments, it is also possible to directly calculate the ratio of the resource adjustment contribution data of each business feature data to the normalization result of the storage capacity of each business feature data to obtain the importance evaluation result of each business feature data.

[0144] In an alternative embodiment, after obtaining the importance evaluation result of each business feature data, the above method further includes:

[0145] Sort each business feature data according to the importance evaluation result of each business feature data to obtain a business feature data sequence.

[0146] Adjust the resource consumption data of each business model. When it is determined that the resource consumption data of each business model is greater than the preset consumption threshold, offline the business feature data that meets the preset conditions in the business feature data sequence.

[0147] In this embodiment, each business feature data can be sorted in ascending order or descending order according to the importance evaluation result of each business feature data to obtain a business feature data sequence. When adjusting the resource consumption data of each business model, for example, optimizing and adjusting the cost of each business model, if it is found that the cost of a certain business model is relatively high, then the business feature data that meets the preset conditions in the business feature data sequence corresponding to this business model can be taken offline. Specifically, if the business feature data sequence is a sequence sorted in ascending order, then the top preset number of business feature data can be taken offline. If the business feature data sequence is a sequence sorted in descending order, then the bottom preset number of business feature data can be taken offline, so as to realize the optimization and adjustment of the cost of this business model and reduce the calculation and storage overhead of the object recommendation system.

[0148] It should be noted that the offline business feature data may be able to bring more resource contribution data to other business models. Then, the importance of this offline business feature data is relatively high for this business model, and this offline business feature data can be used in other business models.

[0149] Taking the object recommendation system as an example of an advertising recommendation system, assume that the business recommendation model includes a click-through rate prediction model and a conversion rate prediction model. The cost of the click-through rate prediction model is the first cost, and the resource contribution data is the first resource contribution data. The cost of the conversion rate prediction model is the second cost, and the resource contribution data is the second resource contribution data. Among them, the first cost is greater than the second cost, and the first resource contribution data is much smaller than the second resource contribution data. Then the cost of the click-through rate prediction model is relatively high, that is, the resource contribution data brought after the business feature data is input into the click-through rate prediction model is not ideal, indicating that some business feature data is not important business feature data for the click-through rate prediction model. Therefore, the business feature data that meets the preset conditions among at least two business feature data for the click-through rate preset model can be taken offline. The taken-offline business feature data can be provided to the click-through rate prediction model for use under appropriate circumstances.

[0150] Next, taking the object recommendation system as an advertising recommendation system and the model performance evaluation result as AUC data as an example, the above feature data processing method will be described as a whole:

[0151] Step 1: Figure 6 It is a flowchart showing a process of generating model performance difference information according to an exemplary embodiment. As Figure 6 shown, at least two business models used by the object recommendation system at the first time (for example, at time T) can be obtained.

[0152] Step 2: Obtain at least two business feature data used by the object recommendation system within the second time after the first time (for example, in the time period from T to T + t).

[0153] Step 3: Input at least two business feature data into each business model for business prediction processing to obtain the first business prediction results predicted by each business model. Perform model performance evaluation on each business model according to the first business prediction results and business labels to obtain the first model performance evaluation results, that is, obtain the true AUC data.

[0154] Step 4: Adjust each business feature data in turn. Input the at least two business feature data after feature adjustment into each business model for business prediction processing to obtain the second business prediction results predicted by each business model for each business feature data after adjustment. Perform model performance evaluation on each business model according to the second business prediction results and business labels to obtain the second model performance evaluation results for each business feature data after adjustment, that is, obtain the AUC data for each business feature data after adjustment.

[0155] Step 5: Calculate the difference between the true AUC data and the AUC data for each business feature data after adjustment to obtain the model performance difference information for each business feature data.

[0156] Step 6: Normalize the model performance difference information for each business feature data to obtain the feature contribution data after each business feature data is input into each business model.

[0157] Step 7: Obtain the model resource contribution data obtained by each business model in processing business within the second time period, calculate the ratio of the model resource contribution data obtained by each business model in processing business within the second time period to the total sum of the model resource contribution data, and obtain the resource adjustment proportion information of each business model.

[0158] Step 8: Determine the product of the resource adjustment proportion information of each business model and the feature contribution data after each business feature data is input into each business model to obtain the resource contribution score after each business feature data is input into each business model; generate the resource adjustment contribution data of each business feature data according to the sum of the resource contribution scores after each business feature data is input into each business model.

[0159] Step 9: Obtain the storage capacity of each business feature data. Normalize the storage capacity of each business feature data to obtain the normalized result of the storage capacity of each business feature data.

[0160] Step 10: Obtain the ROI coefficient of the object recommendation system, calculate the product of the preset resource adjustment coefficient and the normalized result of the storage capacity of each business feature data, calculate the difference between the resource adjustment contribution data of each business feature data and this product, and obtain the importance evaluation result of each business feature data.

[0161] Taking the object recommendation system as an example of the advertisement recommendation system, the business feature data processing method in the embodiments of the present application can be applied to news applications (Application, APP), video APPs, instant messaging APPs, and other APPs. The operation chain of users on advertisements in these APPs can generally be summarized as: display (exposure) -> click -> conversion (register an account, download the APP, follow the official account, form submission, etc.). Among them, there is generally a business model at the arrow in the above chain to predict the probability of generating the behavior on the right side of the arrow given the left side, such as display -> click, which is the click-through rate prediction model, and click -> conversion, which is the conversion rate prediction model. By improving the accuracy of these business models, the advertisement content that users are interested in and that advertisers are willing to pay for can be recommended. And the input of these business models during training is a series of business feature data depicting users and advertisements. Obviously, the importance of different business feature data for business models is different. Therefore, through the business feature data processing method provided by the embodiments of the present application, the global importance of features used by the advertisement recommendation system model can be quantified, and the cost performance of features can be evaluated from the overall perspective of the advertisement recommendation system, which is beneficial to comprehensively considering the resource consumption and resource contribution of business models and reducing the calculation and storage overhead of the advertisement recommendation system.

[0162] Figure 7 is a block diagram of a feature data processing device shown according to an exemplary embodiment, as Figure 7 shown, the feature data processing device includes:

[0163] A model resource acquisition module 201, configured to acquire at least two business models used by the object recommendation system at a first time, at least two business feature data used by the object recommendation system within a second time after the first time, and model resource contribution data obtained by each business model processing business within the second time.

[0164] A prediction module 203, configured to input the at least two business feature data into each business model for business prediction processing to obtain a first business prediction result predicted by each business model; and sequentially perform feature adjustment on each business feature data, and input the at least two business feature data after feature adjustment into each business model for business prediction processing to obtain a second business prediction result predicted by each business model for each business feature data after adjustment.

[0165] A performance evaluation module 205, configured to perform model performance evaluation on each business model according to the first business prediction result, the second business prediction result, and the business labels of the at least two business feature data, and generate feature contribution data after each business feature data is input into each business model according to the model performance evaluation result.

[0166] A resource contribution analysis module 207, configured to perform resource contribution data analysis and processing on each piece of service feature data according to the model resource contribution data and the feature contribution data, so as to obtain resource adjustment contribution data for each piece of service feature data.

[0167] An importance evaluation module 209, configured to perform feature importance evaluation processing on each piece of service feature data according to the resource adjustment contribution data and the storage information of each piece of service feature data, so as to obtain an importance evaluation result for each piece of service feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding service feature data when adjusting the resource consumption data of each service model.

[0168] In an optional embodiment, the prediction module includes:

[0169] A masking unit, configured to sequentially perform masking processing on each piece of service feature data to obtain the at least two pieces of service feature data after masking, and input the at least two pieces of service feature data after masking into each service model for service prediction processing, so as to obtain a second service prediction result adjusted for each piece of service feature data predicted by each service model; or,

[0170] A random perturbation unit, configured to sequentially perform random perturbation processing on each piece of service feature data to obtain the at least two pieces of service feature data after random perturbation, and input the at least two pieces of service feature data after random perturbation into each service model for service prediction processing, so as to obtain a second service prediction result adjusted for each piece of service feature data predicted by each service model.

[0171] In an optional embodiment, the performance evaluation module includes:

[0172] A model performance evaluation generation unit, configured to perform model performance evaluation on each service model according to the first service prediction result and the service label to obtain a first model performance evaluation result, and perform model performance evaluation on each service model according to the second service prediction result and the service label to obtain a second model performance evaluation result adjusted for each piece of service feature data; the first model performance evaluation result and the second model performance evaluation result are used to represent the probability that the positive examples predicted by the service model are ranked in front of the negative examples.

[0173] A difference information generation subunit, configured to generate model performance difference information for each piece of service feature data according to the first model performance evaluation result and the second model performance evaluation result.

[0174] A difference information normalization unit, configured to perform normalization processing on the model performance difference information for each piece of service feature data to obtain feature contribution data for each piece of service feature data input into each service model.

[0175] In an optional embodiment, the difference information normalization unit includes:

[0176] A difference information acquisition subunit, configured to acquire model performance difference information greater than a first performance difference threshold from the model performance difference information for each business feature data, and acquire model performance difference information less than a second performance difference threshold from the model performance difference information for each business feature data; the second model performance threshold is less than the first model performance threshold.

[0177] A difference information difference determination subunit, configured to determine a first difference between the model performance difference information for each business feature data and the model performance difference information less than the second performance difference threshold, and determine a second difference between the model performance difference information greater than the first performance difference threshold and the model performance difference information less than the second performance difference threshold.

[0178] A feature contribution data generation subunit, configured to normalize the model performance difference information for each business feature data according to the ratio between the first difference and the second difference, to obtain feature contribution data after each business feature data is input into each business model.

[0179] In an optional embodiment, the resource contribution analysis module includes:

[0180] A model resource contribution data total determination unit, configured to determine the sum of the model resource contribution data obtained by each business model in processing services within the second time period, to obtain the total model resource contribution data.

[0181] A resource normalization unit, configured to normalize the model resource contribution data obtained by each business model in processing services within the second time period according to the model resource contribution data obtained by each business model in processing services within the second time period and the total model resource contribution data, to obtain the resource adjustment proportion information of each business model.

[0182] A resource contribution data analysis unit, configured to perform resource contribution data analysis processing on each business feature data according to the resource adjustment proportion information of each business model and the feature contribution data, to obtain the resource adjustment contribution data of each business feature data.

[0183] In an optional embodiment, the resource normalization unit includes:

[0184] The proportion information determination subunit is configured to determine the ratio between the model resource contribution data obtained by each business model in processing services within the second time period and the total sum of the model resource contribution data, so as to obtain the resource adjustment proportion information of each business model.

[0185] In an alternative embodiment, the resource contribution data analysis unit includes:

[0186] The resource contribution score generation subunit is configured to determine the product of the resource adjustment proportion information of each business model and the feature contribution data after each business feature data is input into each business model, so as to obtain the resource contribution score after each business feature data is input into each business model.

[0187] The resource adjustment contribution data generation subunit is configured to generate the resource adjustment contribution data of each business feature data according to the sum of the resource contribution scores after each business feature data is input into each business model.

[0188] In an alternative embodiment, the storage information is storage capacity, and the importance evaluation module includes:

[0189] The storage capacity acquisition unit is configured to acquire the storage capacity of each business feature data.

[0190] The storage capacity normalization result generation unit is configured to normalize the storage capacity of each business feature data to obtain the storage capacity normalization result of each business feature data.

[0191] The importance evaluation result generation unit is configured to perform feature importance evaluation processing on each business feature data according to the resource adjustment contribution data of each business feature data and the storage capacity normalization result of each business feature data, so as to obtain the importance evaluation result of each business feature data.

[0192] In an alternative embodiment, the storage capacity normalization result generation unit includes:

[0193] The storage capacity acquisition subunit is configured to acquire the storage capacity greater than the first capacity threshold from the storage capacity of each business feature data, and acquire the storage capacity less than the second capacity threshold from the storage capacity of each business feature data; the first capacity threshold is greater than the second capacity threshold.

[0194] The capacity difference determination subunit is configured to determine the third difference between the storage capacity of each business feature data and the storage capacity less than the second capacity threshold, and the fourth difference between the storage capacity greater than the first capacity threshold and the storage capacity less than the first capacity threshold.

[0195] A storage capacity normalization subunit, configured to normalize the storage capacity of each service feature data according to the third difference and the fourth difference, so as to obtain a storage capacity normalization result of each service feature data.

[0196] In an alternative embodiment, the importance evaluation result generation unit includes:

[0197] An adjustment coefficient acquisition subunit, configured to acquire a preset resource adjustment coefficient of the object recommendation system.

[0198] A capacity product generation subunit, configured to determine the product of the preset resource adjustment coefficient and the storage capacity normalization result of each service feature data.

[0199] A feature importance evaluation subunit, configured to perform feature importance evaluation processing on each service feature data according to the resource adjustment contribution data of each service feature data and the product, so as to obtain an importance evaluation result of each service feature data.

[0200] In an alternative embodiment, the apparatus further includes:

[0201] A sorting module, configured to sort each service feature data according to the importance evaluation result of each service feature data, so as to obtain a service feature data sequence.

[0202] An offline module, configured to adjust the resource consumption data of each service model, and offline the service feature data that meets the preset conditions in the service feature data sequence when it is determined that the resource consumption data of each service model is greater than a preset consumption threshold.

[0203] It should be noted that the apparatus embodiment provided in the embodiment of the present application is based on the same inventive concept as the above method embodiment.

[0204] The embodiment of the present application further provides an electronic device for processing feature data. The electronic device includes a processor and a memory. At least one instruction or at least one program segment is stored in the memory, and at least one instruction or at least one program segment is loaded and executed by the processor to implement the feature data processing method provided in any of the above embodiments.

[0205] The embodiment of the present application further provides a computer-readable storage medium. The computer-readable storage medium can be set in a terminal to store at least one instruction or at least one program segment for implementing a feature data processing method in the method embodiment. At least one instruction or at least one program segment is loaded and executed by the processor to implement the feature data processing method provided in the above method embodiment.

[0206] Optionally, in the embodiments of this specification, the storage medium may be located in at least one of the multiple network servers of a computer network. Optionally, in this embodiment, the above storage medium may include, but is not limited to, various media that can store program codes, such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs.

[0207] The memory in the embodiments of this specification can be used to store software programs and modules. The processor executes various functional application programs and characteristic data processing by running the software programs and modules stored in the memory. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for functions, etc.; the data storage area can store data created according to the use of the device, etc. In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid-state storage devices. Correspondingly, the memory may further include a memory controller to provide the processor with access to the memory.

[0208] The embodiments of this application also provide a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the characteristic data processing method provided in the above method embodiments.

[0209] The embodiments of the characteristic data processing method provided in the embodiments of this application can be executed on a terminal, a computer terminal, a server, or a similar computing device. Taking running on a server as an example, Figure 8 is a hardware structure block diagram of a server provided according to an exemplary embodiment. As Figure 8As shown, the server 300 can vary significantly due to differences in configuration or performance. It may include one or more central processing units (CPUs) 310 (the central processing unit 310 may include, but is not limited to, processing devices such as a microprocessor MCU or a field-programmable gate array FPGA), a memory 330 for storing data, and one or more storage media 320 for storing application programs 323 or data 322 (such as one or more mass storage devices). Among them, the memory 330 and the storage media 320 can be transient storage or persistent storage. The program stored in the storage media 320 may include one or more modules, and each module may include a series of instruction operations on the server. Further, the central processing unit 310 can be configured to communicate with the storage media 320 and execute a series of instruction operations in the storage media 320 on the server 300. The server 300 may also include one or more power supplies 360, one or more wired or wireless network interfaces 350, one or more input / output interfaces 340, and / or one or more operating systems 321, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, and so on.

[0210] The input / output interface 340 can be used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by the communication provider of the server 300. In one example, the input / output interface 340 includes a network interface controller (NIC), which can be connected to other network devices through a base station and thus communicate with the Internet. In one example, the input / output interface 340 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0211] Those of ordinary skill in the art can understand that Figure 8 the structure shown is only schematic and does not limit the structure of the above-mentioned electronic device. For example, the server 300 may also include more or fewer components than Figure 8 shown, or have a different configuration from Figure 8 shown.

[0212] It should be noted that the above order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. And the above specific embodiments of this specification have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0213] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device and server embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments.

[0214] Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by a program instructing relevant hardware. The program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk, an optical disk, etc.

[0215] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for processing feature data, characterized in that, The method includes: obtaining at least two service models used by the object recommendation system at a first time, at least two service feature data used by the object recommendation system within a second time after the first time, and model resource contribution data obtained by each service model in processing services within the second time; inputting the at least two service feature data into each service model for service prediction processing to obtain a first service prediction result predicted by each service model; and sequentially performing feature adjustment on each service feature data, inputting the at least two service feature data after feature adjustment into each service model for service prediction processing to obtain a second service prediction result predicted by each service model for each service feature data after adjustment; performing model performance evaluation on each service model according to the first service prediction result, the second service prediction result, and the service labels of the at least two service feature data, and generating feature contribution data after each service feature data is input into each service model according to the model performance evaluation result; performing resource contribution data analysis processing on each service feature data according to the model resource contribution data and the feature contribution data to obtain resource adjustment contribution data for each service feature data; performing feature importance evaluation processing on each service feature data according to the resource adjustment contribution data and the storage information of each service feature data to obtain an importance evaluation result for each service feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding service feature data when adjusting the resource consumption data of each service model.

2. The characteristic data processing method according to claim 1, wherein The sequentially performing feature adjustment on each service feature data, inputting the at least two service feature data after feature adjustment into each service model for service prediction processing, and obtaining a second service prediction result predicted by each service model for each service feature data after adjustment includes: sequentially performing masking processing on each service feature data to obtain the at least two service feature data after masking, inputting the at least two service feature data after masking into each service model for service prediction processing, and obtaining a second service prediction result predicted by each service model for each service feature data after adjustment; or, sequentially performing random perturbation processing on each service feature data to obtain the at least two service feature data after random perturbation, inputting the at least two service feature data after random perturbation into each service model for service prediction processing, and obtaining a second service prediction result predicted by each service model for each service feature data after adjustment.

3. The feature data processing method according to claim 1, wherein The performing model performance evaluation on each service model according to the first service prediction result, the second service prediction result, and the service labels of the at least two service feature data, and generating feature contribution data after each service feature data is input into each service model according to the model performance evaluation result includes: Performing model performance evaluation on each business model according to the first business prediction result and the business label to obtain a first model performance evaluation result, and performing model performance evaluation on each business model according to the second business prediction result and the business label to obtain a second model performance evaluation result adjusted for each business feature data; the first model performance evaluation result and the second model performance evaluation result are used to characterize the probability that the positive examples predicted by the business model are ranked in front of the negative examples; Generating model performance difference information for each business feature data according to the first model performance evaluation result and the second model performance evaluation result; Performing normalization processing on the model performance difference information for each business feature data to obtain feature contribution data after each business feature data is input into each business model.

4. The feature data processing method according to claim 3, characterized in that, The performing normalization processing on the model performance difference information for each business feature data to obtain feature contribution data after each business feature data is input into each business model includes: Obtaining model performance difference information greater than a first performance difference threshold from the model performance difference information for each business feature data, and obtaining model performance difference information less than a second performance difference threshold from the model performance difference information for each business feature data; the second model performance threshold is less than the first model performance threshold; Determining a first difference between the model performance difference information for each business feature data and the model performance difference information less than the second performance difference threshold, and determining a second difference between the model performance difference information greater than the first performance difference threshold and the model performance difference information less than the second performance difference threshold; Performing normalization processing on the model performance difference information for each business feature data according to the ratio between the first difference and the second difference to obtain feature contribution data after each business feature data is input into each business model.

5. The feature data processing method according to claim 1, characterized in that The performing data analysis processing on the resource contribution of each business feature data according to the model resource contribution data and the feature contribution data to obtain the resource adjustment contribution data of each business feature data includes: Determining the sum of the model resource contribution data obtained by each business model in processing the business during the second time to obtain the total model resource contribution data; Performing normalization processing on the model resource contribution data obtained by each business model in processing the business during the second time according to the model resource contribution data obtained by each business model in processing the business during the second time and the total model resource contribution data to obtain the resource adjustment proportion information of each business model; Performing data analysis processing on the resource contribution of each business feature data according to the resource adjustment proportion information of each business model and the feature contribution data to obtain the resource adjustment contribution data of each business feature data.

6. The feature data processing method according to claim 5, wherein Normalize the model resource contribution data obtained by each business model in processing the service during the second time according to the model resource contribution data obtained by each business model in processing the service during the second time and the total sum of the model resource contribution data, to obtain the resource adjustment proportion information of each business model, including: Determine the ratio between the model resource contribution data obtained by each business model in processing the service during the second time and the total sum of the model resource contribution data, to obtain the resource adjustment proportion information of each business model.

7. The feature data processing method according to claim 5, characterized in that, Perform resource contribution data analysis processing on each service feature data according to the resource adjustment proportion information of each business model and the feature contribution data, to obtain the resource adjustment contribution data of each service feature data, including: Determine the product of the resource adjustment proportion information of each business model and the feature contribution data after each service feature data is input into each business model, to obtain the resource contribution score after each service feature data is input into each business model; Generate the resource adjustment contribution data of each service feature data according to the sum of the resource contribution scores after each service feature data is input into each business model.

8. The characteristic data processing method according to claim 1, wherein The stored information is the storage capacity. Perform feature importance evaluation processing on each service feature data according to the resource adjustment contribution data and the stored information of each service feature data, to obtain the importance evaluation result of each service feature data, including: Obtain the storage capacity of each service feature data; Normalize the storage capacity of each service feature data to obtain the normalized result of the storage capacity of each service feature data; Perform feature importance evaluation processing on each service feature data according to the resource adjustment contribution data of each service feature data and the normalized result of the storage capacity of each service feature data, to obtain the importance evaluation result of each service feature data.

9. The characteristic data processing method according to claim 8, wherein The step of normalizing the storage capacity of each service feature data to obtain the normalized result of the storage capacity of each service feature data includes: Obtain the storage capacity greater than the first capacity threshold from the storage capacity of each service feature data, and obtain the storage capacity less than the second capacity threshold from the storage capacity of each service feature data; the first capacity threshold is greater than the second capacity threshold; Determine the third difference between the storage capacity of each service feature data and the storage capacity less than the second capacity threshold, and the fourth difference between the storage capacity greater than the first capacity threshold and the storage capacity less than the first capacity threshold; Normalize the storage capacity of each service feature data according to the third difference and the fourth difference to obtain the normalized result of the storage capacity of each service feature data.

10. The characteristic data processing method according to claim 8, wherein Perform feature importance evaluation processing on each service feature data according to the resource adjustment contribution data of each service feature data and the normalized result of the storage capacity of each service feature data, to obtain the importance evaluation result of each service feature data, including: Obtain the preset resource adjustment coefficient of the object recommendation system; Determine the product of the preset resource adjustment coefficient and the normalized result of the storage capacity of each service feature data; Based on the resource adjustment contribution data of each service feature data and the product, perform a feature importance evaluation process on each service feature data to obtain the importance evaluation result of each service feature data.

11. The feature data processing method according to any one of claims 1 to 10, characterized in that After obtaining the importance evaluation result of each service feature data, the method further includes: Sort each service feature data according to the importance evaluation result of each service feature data to obtain a service feature data sequence; Adjust the resource consumption data of each service model. When it is determined that the resource consumption data of each service model is greater than the preset consumption threshold, take offline the service feature data that meets the preset conditions in the service feature data sequence.

12. A feature data processing device, characterized in that, The device includes: A model resource acquisition module, configured to acquire at least two service models used by the object recommendation system at a first time, at least two service feature data used by the object recommendation system within a second time after the first time, and the model resource contribution data obtained by each service model in processing services within the second time; A prediction module, configured to input the at least two service feature data into each service model for service prediction processing to obtain a first service prediction result predicted by each service model; and sequentially perform feature adjustment on each service feature data, and input the at least two service feature data after feature adjustment into each service model for service prediction processing to obtain a second service prediction result predicted by each service model for each service feature data after adjustment; A performance evaluation module, configured to perform model performance evaluation on each service model according to the first service prediction result, the second service prediction result, and the service labels of the at least two service feature data, and generate feature contribution data after each service feature data is input into each service model according to the model performance evaluation result; A resource contribution analysis module, configured to perform resource contribution data analysis processing on each service feature data according to the model resource contribution data and the feature contribution data to obtain the resource adjustment contribution data of each service feature data; An importance evaluation module, configured to perform feature importance evaluation processing on each service feature data according to the resource adjustment contribution data and the storage information of each service feature data to obtain the importance evaluation result of each service feature data; the importance evaluation result is used to evaluate whether to take offline the corresponding service feature data when adjusting the resource consumption data of each service model.

13. An electronic device for processing feature data, characterized in that, The electronic device includes a processor and a memory, and at least one instruction or at least one program segment is stored in the memory. The at least one instruction or the at least one program segment is loaded and executed by the processor to perform the feature data processing method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, At least one instruction or at least one program segment is stored in the computer-readable storage medium. The at least one instruction or the at least one program segment is loaded and executed by a processor to implement the feature data processing method according to any one of claims 1 to 11.

Images