Application system evaluation method and system based on big data

Abstract

The present invention discloses an application system evaluation method and system based on big data, comprising: firstly obtaining the current task indication of the target application system, calling the system to execute and obtain the feedback result, and then using the pre-trained application system evaluation model for evaluation. When the evaluation result shows that the system is abnormal, the method will determine the solution strategy based on the semantic features of the feedback result. With such a design, an efficient and accurate evaluation of the performance and stability of the application system is achieved, and the automation and intelligence level of the evaluation is improved.

Description

Technical Field

[0001] The present invention relates to the field of application system testing, and in particular to an application system evaluation method and system based on big data. Background Art

[0002] With the rapid development of big data technology, various application systems are becoming increasingly complex, and the evaluation of their performance, stability and security is becoming more and more important. Traditional application system evaluation methods are often based on fixed indicators and rules, which makes it difficult to fully and dynamically reflect the real performance and potential problems of the system. Summary of the invention

[0003] The purpose of the present invention is to provide an application system evaluation method and system based on big data.

[0004] In a first aspect, an embodiment of the present invention provides an application system evaluation method based on big data, the method comprising:

[0005] Acquire a current task instruction for a target application system, wherein the current task instruction is used to test the target application system;

[0006] Calling the target application system to execute the current task instruction and obtain the current feedback result;

[0007] Inputting the current feedback result into a pre-selected and trained application system evaluation model to obtain a current evaluation result;

[0008] When the evaluation result indicates that an abnormality exists in the target application system, a solution strategy search result of the current feedback result is determined based on the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0010] Compared with the prior art, the beneficial effects provided by the present invention include: adopting the big data-based application system evaluation method and system disclosed in the present invention, by obtaining the current task instructions of the target application system, calling the system to execute and obtain the feedback results, and then using the pre-trained application system evaluation model for evaluation. When the evaluation results show that the system is abnormal, the method will determine the solution strategy based on the semantic features of the feedback results. Such a design realizes efficient and accurate evaluation of the performance and stability of the application system, and improves the automation and intelligence level of the evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0012] Figure 1 A schematic diagram of the steps of the big data-based application system evaluation method provided by an embodiment of the present invention;

[0013] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0014] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0016] In order to solve the technical problems in the aforementioned background technology, Figure 1 The present invention provides a flowchart of an application system evaluation method based on big data in an embodiment of the present invention. The application system evaluation method based on big data is introduced in detail below.

[0017] Step S201, obtaining a current task instruction for a target application system, wherein the current task instruction is used to test the target application system;

[0018] Step S202, calling the target application system to execute the current task instruction and obtain the current feedback result;

[0019] Step S203, inputting the current feedback result into the pre-selected trained application system evaluation model to obtain the current evaluation result;

[0020] Step S204, when the evaluation result indicates that the target application system has an abnormality, based on the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result, a solution strategy search result of the current feedback result is determined.

[0021] In an embodiment of the present invention, for example, an e-commerce platform wants to evaluate the performance of its search recommendation system. The server of the e-commerce platform receives a task instruction, which simulates a user entering the keyword "smartphone" in the search box and requests the search recommendation system to return a list of related product recommendations. The server of the e-commerce platform calls the search recommendation system according to the received task instruction. The system quickly retrieves and returns a list of smartphone recommendations as a feedback result based on the keyword "smartphone". This list contains recommended information of multiple smartphones, such as brand, model, price, sales, etc. The server inputs the smartphone recommendation list returned by the search recommendation system into a pre-trained evaluation model. This evaluation model is trained through historical data and user feedback, and can automatically analyze the quality of the recommendation list, such as the accuracy and diversity of the recommendation. After the model analysis, the server obtains an evaluation result indicating the accuracy and user satisfaction of this search recommendation. Assume that the evaluation model finds that the recommended position of a certain mobile phone in the smartphone recommendation list returned by the search recommendation system does not match its historical sales and user evaluation, which is identified as an abnormality by the evaluation model. At this point, the server will further analyze the two main feature domains of this phone: one is product features (such as brand, price, functions, etc.), and the other is user feedback features (such as user reviews, sales trends, etc.). Based on the semantic analysis of these two feature domains, the server determines a solution strategy, which is to adjust the recommendation weight of the phone to more accurately reflect its market positioning and user preferences. Subsequently, the server will search for relevant optimization algorithms or strategies to automatically or semi-automatically adjust the recommendation system to improve the quality of future recommendations. Through the above four steps, e-commerce platforms can conduct a comprehensive performance evaluation of their search recommendation systems, and promptly identify and resolve potential problems, thereby improving user experience and sales performance.

[0022] In the embodiment of the present invention, the application system evaluation model is obtained in the following manner.

[0023] Obtaining task feedback instances, wherein the task feedback instances include target feedback instances and non-target feedback instances; the target feedback instance is a feedback instance in which the test application system generates a correct feedback result for a historical task indication; the non-target feedback instance is a feedback instance in which the test application system does not generate the correct feedback result for the historical task indication;

[0024] When the task feedback instance meets the instance tuning trigger condition, a compressed sensing parameter of the task feedback instance is obtained based on a compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameter, and a tuned task feedback instance is obtained based on the non-target feedback instance and the target feedback instance completed by compressed sensing;

[0025] An application system evaluation model is obtained by training based on the task feedback examples completed by the tuning;

[0026] The evaluation result of the application system evaluation model is optimized to obtain an optimized application system evaluation model; the optimized application system evaluation model is used to evaluate the feedback result generated by the application system in response to the task instruction.

[0027] In an embodiment of the present invention, illustratively, an online video platform wants to optimize its video recommendation system. The server first collects a large number of task feedback instances. These instances include target feedback instances, i.e., cases where the video recommendation system successfully recommends similar types of videos based on the user's historical viewing records, and non-target feedback instances, i.e., cases where the system fails to accurately recommend videos that the user may be interested in. The server analyzes the collected task feedback instances, and when it is found that the proportion of non-target feedback instances exceeds a certain set threshold (such as 10%), it triggers the instance tuning process. This means that the recommendation accuracy of the system needs to be improved. The server uses compressed sensing technology to extract and compress features of non-target feedback instances. In this way, the server can more effectively identify key information in these instances, such as the user's viewing history, video content features, etc., and compress this information into a small amount of perception parameters. Using the compressed sensing parameters obtained in the previous step, the server performs tuning processing on the non-target feedback instances. This includes re-adjusting the parameters in the recommendation algorithm according to the user's personalized features and video content features to improve the recommendation results. After tuning, the original non-target feedback instances are converted into recommendation results that are more in line with the user's preferences, thereby forming a task feedback instance that has been tuned. The server uses the tuned task feedback instances together with the target feedback instances to train a new application system evaluation model. This model can more accurately evaluate the quality of the recommendation results generated by the video recommendation system based on the user's viewing history. In actual applications, the server will continuously collect new feedback data and optimize the application system evaluation model based on this data. By adjusting the parameters and structure of the model, the server can further improve the evaluation accuracy of the model. The optimized application system evaluation model will more accurately reflect the actual performance of the video recommendation system, thereby helping the platform to continuously improve its recommendation algorithm and enhance user experience.

[0028] In an embodiment of the present invention, the compressed sensing parameters of the task feedback instance are obtained based on the compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameters. The task feedback instance completed by tuning is obtained based on the non-target feedback instance and the target feedback instance completed by compressed sensing. The implementation can be performed through the following examples.

[0029] Obtaining a first task feedback instance and a second task feedback instance from the task feedback instance;

[0030] Obtaining a first compressed sensing parameter corresponding to the first task feedback instance based on a compressed sensing strategy, and obtaining a second compressed sensing parameter corresponding to the second task feedback instance based on a compressed sensing strategy;

[0031] Tuning the first task feedback instance based on the first compressed sensing parameter to obtain a tuned first task feedback instance, and tuning the second task feedback instance based on the second compressed sensing parameter to obtain a tuned second task feedback instance;

[0032] Based on the first task feedback instance after the tuning is completed and the second task feedback instance after the tuning is completed, a task feedback instance after the tuning is completed is obtained.

[0033] In an embodiment of the present invention, illustratively, an online retail platform wants to optimize its product recommendation system. The server selects two representative examples from the collected task feedback examples: the first task feedback example, in which the recommendation system successfully recommends a book that the user may be interested in; the second task feedback example, in which the recommendation system fails to accurately recommend the electronic product that the user needs. The server uses compressed sensing technology to extract and compress the first task feedback example and the second task feedback example respectively. For the first task feedback example, the server extracts the user's historical purchase record, browsing behavior, and key features such as the category and author of the book, and compresses them into a first compressed sensing parameter. For the second task feedback example, the server also extracts the user's electronic product purchase history, search record, and features such as the brand and function of the electronic product, and compresses them into a second compressed sensing parameter. Using the first compressed sensing parameter, the server tunes the first task feedback example. By deeply analyzing the user's reading preferences and book features, the server fine-tunes the parameters in the recommendation algorithm so that the recommendation system can more accurately recommend similar types of books to the user. Similarly, using the second compressed sensing parameter, the server tunes the second task feedback example. By adjusting the weights and filtering conditions in the electronic product recommendation algorithm, the server improved the recommendation accuracy of the recommendation system for electronic products. After tuning, both the first task feedback instance and the second task feedback instance were optimized, forming the first task feedback instance that was tuned and the second task feedback instance that was tuned. The server summarizes and analyzes the first task feedback instance that was tuned and the second task feedback instance that was tuned. By comparing the recommendation results and user feedback data before and after the tuning, the server verified the effectiveness of the tuning effect. Finally, the server integrated these tuned task feedback instances into a more complete and accurate recommendation strategy library to guide future product recommendation work. In this way, the product recommendation system of the online retail platform can be continuously optimized and the user experience can be improved.

[0034] In an embodiment of the present invention, the first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; the compressed sensing strategy includes a category-differentiated compressed sensing method; and the first compressed sensing parameter corresponding to the first task feedback instance is obtained based on the compressed sensing strategy, which can be implemented through the following examples.

[0035] Based on the category-differentiated compressed sensing mode, a response variation coefficient of a first target feedback instance is obtained according to the first target feedback instance and the first non-target feedback instance; a first compressed sensing parameter corresponding to the first task feedback instance is obtained based on the response variation coefficient of the first target feedback instance and the feedback distribution feature instance of the first target feedback instance;

[0036] The obtaining, based on the compressed sensing strategy, a second compressed sensing parameter corresponding to the second task feedback instance includes:

[0037] Based on the category-differentiated compressed sensing method, a response variation coefficient of a second target feedback instance is obtained according to the second target feedback instance and the second non-target feedback instance; and a second compressed sensing parameter corresponding to the second task feedback instance is obtained based on the response variation coefficient of the second target feedback instance and the feedback distribution feature instance of the second target feedback instance.

[0038] In an embodiment of the present invention, exemplarily, in a recommendation system of an e-commerce platform, a server collects a large amount of user feedback data on recommended products as a first task feedback instance. Among them, the first target feedback instance is a record of a user expressing satisfaction or purchasing a recommended product, while the first non-target feedback instance is a record of a user not being interested in or not generating a purchase behavior for the recommended product. The server uses a category-differentiated compressed sensing method to analyze the first target feedback instance and the first non-target feedback instance. This method can help the server more accurately identify the key differences between the two types of feedback instances. By comparing the first target feedback instance and the first non-target feedback instance, the server calculates a response variation coefficient, namely the first target feedback instance response variation coefficient, which reflects the changing trend of the user's satisfaction with the recommended product. The server combines the first target feedback instance response variation coefficient and the feedback distribution characteristics of the first target feedback instance (such as purchase quantity, browsing time, etc.), and calculates the first compressed sensing parameter corresponding to the first task feedback instance through a specific algorithm. This parameter combines the user's feedback information and behavior characteristics, and provides an important basis for the subsequent optimization of the recommendation algorithm. In the same e-commerce platform, the server also collects another set of user feedback data as a second task feedback instance. Similar to the first task feedback instance, the second target feedback instance is a record of user satisfaction with the recommended product, while the second non-target feedback instance is a record of user dissatisfaction. The server also uses category-differentiated compressed sensing to process the second task feedback instance. By comparing the second target feedback instance with the second non-target feedback instance, the server calculates the response variation coefficient of the second target feedback instance, reflecting another set of changing trends in user satisfaction. Finally, the server combines the response variation coefficient of the second target feedback instance and the feedback distribution characteristics of the second target feedback instance, and calculates the second compressed sensing parameter corresponding to the second task feedback instance through an algorithm. This parameter also provides valuable data support for further optimization of the recommendation system. Through these steps, the server can more accurately evaluate and optimize the recommendation system of the e-commerce platform, thereby improving user experience and sales performance.

[0039] In an embodiment of the present invention, the first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; the compressed sensing strategy includes a step-by-step compressed sensing method; the embodiment of the present invention also provides the following implementation methods.

[0040] Obtaining a target feedback instance response variation coefficient according to the target feedback instance and the non-target feedback instance;

[0041] The obtaining, based on the compressed sensing strategy, a first compressed sensing parameter corresponding to the first task feedback instance includes:

[0042] According to the step-by-step compressed sensing method, a first compressed sensing parameter corresponding to the first task feedback instance is obtained based on the response variation coefficient of the target feedback instance and the feedback distribution feature instance of the first target feedback instance;

[0043] The obtaining, based on the compressed sensing strategy, a second compressed sensing parameter corresponding to the second task feedback instance includes:

[0044] According to the step-by-step compressed sensing method, a second compressed sensing parameter corresponding to the second task feedback instance is obtained based on the response variation coefficient of the target feedback instance and the feedback distribution feature instance of the second target feedback instance.

[0045] In an embodiment of the present invention, exemplarily, in an online learning platform, the server collects user feedback data on course recommendations. These data are divided into a first task feedback instance and a second task feedback instance. In each task feedback instance, the server further distinguishes between target feedback instances (users are satisfied with the recommended course or have completed the course learning) and non-target feedback instances (users are not interested in the recommended course or have not completed the learning). The server compares and analyzes all target feedback instances and non-target feedback instances, and calculates a response variation coefficient through a specific algorithm. This coefficient reflects the dynamic changes in user satisfaction with the recommended course, and provides an important reference for subsequent recommendation optimization. When processing the first task feedback instance, the server adopts a step-by-step compressed sensing method. First, the server performs preliminary perception based on the response variation coefficient of the target feedback instance and the feedback distribution characteristics of the first target feedback instance (such as course completion rate, learning time, etc.). Then, through a step-by-step in-depth manner, the server continuously refines and refines the compressed sensing parameters of the first task feedback instance, and finally obtains the first compressed sensing parameters. For the second task feedback instance, the server also uses a step-by-step compressed sensing method for processing. Similar to the first task feedback instance, the server perceives the response variation coefficient of the target feedback instance and the feedback distribution characteristics of the second target feedback instance. Through step-by-step in-depth analysis and calculation, the server obtains the second compressed sensing parameter corresponding to the second task feedback instance. After these steps are completed, the server has more accurate user feedback data and evaluation results of the course recommendation effect. This helps the online learning platform to further optimize its course recommendation algorithm and improve user experience and learning effects. For example, based on these compressed sensing parameters, the platform can adjust the recommendation strategy to recommend courses that are more in line with users' learning needs and interests.

[0046] In an embodiment of the present invention, the compressed sensing strategy includes a basic compressed sensing method; the compressed sensing parameters of the task feedback instance are obtained based on the compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameters, and the task feedback instance is optimized based on the non-target feedback instance and the target feedback instance completed by compressed sensing. The implementation can be performed through the following examples.

[0047] Based on the basic compressed sensing method, obtaining a target feedback instance response variation coefficient according to the target feedback instance and the non-target feedback instance;

[0048] Obtaining a compressed sensing parameter based on the target feedback instance response variation coefficient and the target feedback instance feedback distribution feature instance;

[0049] removing selected non-target feedback instances from the non-target feedback instances according to the compressed sensing parameters to obtain non-target feedback instances completed by compressed sensing;

[0050] The non-target feedback instance completed by the compressed sensing and the target feedback instance are combined to obtain a task feedback instance completed by tuning.

[0051] In an embodiment of the present invention, exemplarily, in an intelligent advertising delivery system, the server collects data on user clicks and conversions on advertisements as task feedback instances. These data are divided into target feedback instances (users clicked and converted advertisements) and non-target feedback instances (users did not click or did not convert after clicking). The server uses a basic compressed sensing method to calculate a target feedback instance response variation coefficient by comparing the difference between the target feedback instance and the non-target feedback instance, which reflects the dynamic change of the user's response to the advertisement. After obtaining the target feedback instance response variation coefficient, the server further combines the feedback distribution characteristics of the target feedback instance (such as click time, conversion rate, etc.) to calculate a compressed sensing parameter through a specific algorithm. This parameter comprehensively considers the user's response changes and specific feedback behaviors, and provides an important basis for subsequent advertising optimization. The server screens non-target feedback instances based on the calculated compressed sensing parameters. By comparing the matching degree of each non-target feedback instance with the compressed sensing parameter, the server removes those non-target feedback instances that are significantly different from the parameter or obviously do not meet the advertising optimization target. The remaining non-target feedback instances constitute a set of non-target feedback instances completed by compressed sensing. Finally, the server merges the non-target feedback instances processed by compressed sensing with the target feedback instances. This merged task feedback instance set contains both the advertising data that users actively respond to and the non-response data after optimization and screening. This optimized task feedback instance provides more comprehensive and accurate data support for further optimization of the advertising delivery system. Through the above steps, the intelligent advertising delivery system can more accurately evaluate the advertising effect and optimize the delivery strategy, thereby increasing the click-through rate and conversion rate of the advertisement, and improving the advertising effect and user experience.

[0052] In an embodiment of the present invention, the compressed sensing strategy includes a segmented compressed sensing method; the compressed sensing parameters of the task feedback instance are obtained based on the compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameters, and the task feedback instance is optimized based on the non-target feedback instance and the target feedback instance completed by compressed sensing. The implementation can be performed through the following examples.

[0053] Obtaining a first task feedback instance and a second task feedback instance from the task feedback instance;

[0054] Obtaining a compressed sensing parameter combination based on the segmented compressed sensing method; the compressed sensing parameter combination includes a first compressed sensing parameter and a second compressed sensing parameter;

[0055] Tuning the first task feedback instance based on the first compressed sensing parameter to obtain a tuned first task feedback instance, and tuning the second task feedback instance based on the second compressed sensing parameter to obtain a tuned second task feedback instance;

[0056] Based on the first task feedback instance after the tuning is completed and the second task feedback instance after the tuning is completed, a task feedback instance after the tuning is completed is obtained.

[0057] In an embodiment of the present invention, exemplarily, in an e-commerce platform, the server collects user feedback data on product recommendations as task feedback instances. These data can be divided into a first task feedback instance and a second task feedback instance according to different product categories or user groups. For example, the first task feedback instance can be user feedback on electronic product recommendations, while the second task feedback instance can be user feedback on clothing product recommendations. The server uses a segmented compressed sensing method to process the first task feedback instance and the second task feedback instance respectively. In this way, the server can more accurately capture the key information in different task feedback instances. During the processing, the server will calculate the first compressed sensing parameter and the second compressed sensing parameter respectively, which reflect the feedback characteristics and behavior patterns of users in different task feedback instances. Finally, the server merges the two parameters into a compressed sensing parameter merger. After obtaining the compressed sensing parameter merger, the server begins to tune the first task feedback instance and the second task feedback instance respectively. For the first task feedback instance (such as electronic product recommendation feedback), the server adjusts the recommendation algorithm according to the first compressed sensing parameter to more accurately meet the user's demand for electronic products. Similarly, for the second task feedback instance (such as clothing product recommendation feedback), the server performs tuning based on the second compressed sensing parameter to improve the recommendation effect of clothing products. After the above tuning process, the server obtains the first task feedback instance that has been tuned and the second task feedback instance that has been tuned. These instances contain optimized recommendation results and user feedback data. Finally, the server merges the two tuned instances into a set of tuned task feedback instances. This set not only contains the user's optimized recommendation feedback data for electronic products and clothing products, but also reflects the improvement of the server's recommendation performance in different product categories. Through the above steps and scenario examples, it can be seen that the segmented compressed sensing method can help e-commerce platforms more accurately evaluate and optimize the performance of their recommendation systems in different product categories, thereby improving user experience and sales performance.

[0058] In an embodiment of the present invention, the first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; the compressed sensing parameter merging obtained based on the segmented compressed sensing method can be implemented through the following examples.

[0059] Based on the segmented compressed sensing method, obtaining a plurality of first undetermined response variation coefficients according to the first target feedback instance and the first non-target feedback instance;

[0060] Based on the segmented compressed sensing method, obtaining a plurality of second undetermined response variation coefficients according to the second target feedback instance and the second non-target feedback instance;

[0061] Construct a response variation coefficient merger according to the plurality of first pending response variation coefficients and the plurality of second pending response variation coefficients; the response variation coefficient merger includes a first target feedback instance response variation coefficient and a second target feedback instance response variation coefficient, the first target feedback instance response variation coefficient belongs to the plurality of first pending response variation coefficients, and the second target feedback instance response variation coefficient belongs to the plurality of second pending response variation coefficients;

[0062] The first compressed sensing parameter is obtained based on the response variation coefficient of the first target feedback instance and the feedback distribution feature instance of the first target feedback instance, and the second compressed sensing parameter is obtained based on the response variation coefficient of the second target feedback instance and the feedback distribution feature instance of the second target feedback instance;

[0063] A compressed sensing parameter combination is obtained based on the first compressed sensing parameter and the second compressed sensing parameter.

[0064] In an embodiment of the present invention, exemplarily, in an online video platform, the server collects user feedback data on recommended videos, wherein the first task feedback instance is feedback on science fiction videos. The server first distinguishes between the first target feedback instance (the user watched the recommended video and gave a positive evaluation) and the first non-target feedback instance (the user did not watch or gave a negative evaluation). Then, based on the segmented compressed sensing method, the server compares and analyzes these two types of instances and calculates multiple first pending response variation coefficients, which reflect the different degrees of user reactions to science fiction video recommendations. Also in the online video platform, the second task feedback instance is feedback on comedy videos. The server similarly distinguishes between the second target feedback instance and the second non-target feedback instance, and calculates multiple second pending response variation coefficients based on the segmented compressed sensing method, which reveal the differences in user reactions to comedy video recommendations. The server conducts a comprehensive analysis of the pending response variation coefficients calculated in the first task feedback instance and the second task feedback instance, and constructs a response variation coefficient merger. This merged coefficient set not only includes the user's reaction coefficient to science fiction videos (the response variation coefficient of the first target feedback instance), but also includes the user's reaction coefficient to comedy videos (the response variation coefficient of the second target feedback instance). The server further uses the response variation coefficient and feedback distribution characteristics (such as viewing time, number of likes, etc.) of the first target feedback instance to calculate a first compressed sensing parameter for science fiction video recommendation. Similarly, the server also uses the response variation coefficient and feedback distribution characteristics of the second target feedback instance to calculate a second compressed sensing parameter for comedy video recommendation. Finally, the server merges the first compressed sensing parameter and the second compressed sensing parameter to obtain a comprehensive compressed sensing parameter merger. This merged parameter not only takes into account the user's reaction to science fiction videos, but also the user's reaction to comedy videos, thereby providing strong data support for further optimization of the video recommendation system. Through the above steps and scenario examples, it can be seen that the segmented compressed sensing method can help online video platforms more accurately evaluate and optimize the performance of their recommendation systems on different video categories, thereby improving user experience and viewing satisfaction.

[0065] In an embodiment of the present invention, the first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; the first task feedback instance is tuned based on the first compressed sensing parameter to obtain the tuned first task feedback instance, which can be implemented through the following example.

[0066] Performing compressed sensing on a first non-target feedback instance in the first task feedback instance according to the first compressed sensing parameter, and obtaining a first task feedback instance that has been tuned based on the first non-target feedback instance and the first target feedback instance completed by compressed sensing;

[0067] The step of tuning the second task feedback instance based on the second compressed sensing parameter to obtain the tuned second task feedback instance includes:

[0068] Compressed sensing is performed on the second non-target feedback instance in the second task feedback instance according to the second compressed sensing parameter, and a tuned second task feedback instance is obtained based on the second non-target feedback instance and the second target feedback instance completed by compressed sensing.

[0069] In an embodiment of the present invention, exemplarily, in an online shopping platform, the server first collects first task feedback instances about the sale of electronic products, which include successful cases of users purchasing electronic products (first target feedback instances) and cases of not purchasing or returning products after purchase (first non-target feedback instances). The server uses the first compressed sensing parameters calculated previously to perform compressed sensing analysis on these first non-target feedback instances. This process mainly identifies and optimizes the key factors that cause users not to purchase or return products, such as excessive prices, unclear product descriptions, etc. After performing compressed sensing analysis, the server adjusts and optimizes the problems in the first non-target feedback instances, such as adjusting the price strategy of electronic products and improving the detail and accuracy of product descriptions. Then, the server merges these optimized first non-target feedback instances with the original first target feedback instances to form a set of first task feedback instances that have been tuned. This set not only includes cases of successful purchases by users, but also includes cases of non-purchase or returns after optimization. In the same online shopping platform, the server also collected second task feedback instances about clothing sales, which also included successful cases of users purchasing clothing (second target feedback instances) and cases of not purchasing or returning after purchase (second non-target feedback instances). Using the second compressed sensing parameters calculated previously, the server performed similar compressed sensing analysis on these second non-target feedback instances to identify and optimize the key factors affecting clothing sales. After compressed sensing analysis, the server made corresponding adjustments and optimizations to the problems in the second non-target feedback instances, such as improving the style design of clothing, optimizing size selection, etc. Then, the server merged these optimized second non-target feedback instances with the original second target feedback instances to form a set of tuned second task feedback instances. This set reflects successful cases in clothing sales and problem cases that have been optimized. It can be seen from the above steps and scenario examples that the process of tuning task feedback instances using compressed sensing parameters is a process of continuously improving and optimizing the performance of business application systems. This method can help online shopping platforms more accurately identify and solve problems in the sales process, thereby improving user shopping experience and platform sales performance.

[0070] In the embodiment of the present invention, the optimization of the evaluation result of the application system evaluation model to obtain the optimized application system evaluation model can be implemented through the following examples.

[0071] Obtaining a sample evaluation result of the application system evaluation model for the task feedback instance;

[0072] Obtaining a first optimization mapping relationship for the application system evaluation model according to the sample evaluation result and the actual response label of the task feedback instance;

[0073] The evaluation result of the application system evaluation model is adjusted according to the first optimization mapping relationship to obtain an optimized application system evaluation model.

[0074] In an embodiment of the present invention, exemplarily, in an intelligent customer service system, a server deploys an application system evaluation model for evaluating the quality of the system's response to customer service requests. In order to optimize this evaluation model, the server first uses the model to evaluate a batch of task feedback instances of customer service requests, and obtains sample evaluation results of the model for these instances. These results may include satisfaction scores, resolution speed scores, etc. for each service request. The server then compares the sample evaluation results given by the model with the actual response labels of the task feedback instances. The actual response labels may be feedback directly provided by the customer, such as the results of a satisfaction survey, or measured by other objective indicators (such as repeat request rate, problem resolution rate, etc.). Through comparison, the server finds that there is a large difference between the evaluation results of the model and the actual response labels on certain specific types of service requests. For example, for processing requests for complex problems, the satisfaction scores given by the model are generally high. Based on these findings, the server constructs a first optimization mapping relationship, which indicates the deviation between the model evaluation results and the actual response and its adjustment direction. According to the first optimization mapping relationship, the server adjusts the evaluation algorithm of the application system evaluation model. For example, for processing requests for complex problems, the model now considers more factors, such as processing time, subsequent customer feedback, etc., to give a more accurate satisfaction score. After such adjustments, the consistency between the model's evaluation results for different types of service requests and the actual response labels has been significantly improved. In the end, the server obtained an optimized application system evaluation model, which can more accurately evaluate the response quality of the intelligent customer service system to customer service requests. It can be seen from the above steps and scenario examples that by optimizing the evaluation results of the application system evaluation model, the accuracy of the model evaluation can be improved, thereby better guiding the improvement and optimization of the business application system.

[0075] In the embodiment of the present invention, the optimization of the evaluation result of the application system evaluation model to obtain the optimized application system evaluation model can be implemented through the following examples.

[0076] Obtaining a second optimization mapping relationship for the application system evaluation model according to the compressed sensing parameter;

[0077] The evaluation result of the application system evaluation model is adjusted according to the second optimization mapping relationship to obtain an optimized application system evaluation model.

[0078] In an embodiment of the present invention, exemplarily, in an intelligent logistics system, a server uses big data technology to analyze and optimize the logistics transportation process. An application system evaluation model is deployed in the system to evaluate the efficiency and cost-effectiveness of logistics transportation. In order to further improve the accuracy of the evaluation model, the server considers using the compressed sensing parameters previously obtained by compressed sensing technology. Specifically, the server first analyzes various parameters in the logistics transportation process obtained by compressed sensing technology, such as transportation time, cost, cargo loss rate, etc. These parameters reflect the key performance and efficiency indicators in the actual transportation process. By deeply analyzing the relationship between these parameters and the output results of the evaluation model, the server discovered an important mapping relationship, that is, there is a significant deviation between certain specific compressed sensing parameter values ​​and the system evaluation results. Based on these findings, the server constructs a second optimization mapping relationship. This mapping relationship clearly points out how the output results of the evaluation model should be adjusted under different compressed sensing parameter values ​​to more accurately reflect the actual logistics transportation efficiency and cost-effectiveness. After determining the second optimization mapping relationship, the server begins to adjust the application system evaluation model. Specifically, the server integrates the second optimization mapping relationship into the evaluation model, so that the model can automatically make corresponding adjustments according to the current compressed sensing parameter value when outputting the evaluation results. For example, when the compressed sensing parameter shows that the cost of a certain logistics transportation is high and the transportation time is long, according to the second optimization mapping relationship, the evaluation model will automatically lower the efficiency score of the transportation and increase the negative evaluation of cost-effectiveness accordingly. The adjusted evaluation model can more accurately reflect the actual logistics transportation situation and provide more reliable decision support for management. It can be seen from the above steps and scenario examples that using compressed sensing parameters to optimize the application system evaluation model can significantly improve the evaluation accuracy of the model and make it better adapt to complex actual application scenarios. This optimization method is of great significance for improving the performance and efficiency of big data application systems.

[0079] In the embodiment of the present invention, the optimization of the evaluation result of the application system evaluation model to obtain the optimized application system evaluation model can be implemented through the following examples.

[0080] The evaluation result of the application system evaluation model is adjusted according to the target feedback instance response variation coefficient to obtain an optimized application system evaluation model.

[0081] In an embodiment of the present invention, exemplarily, in an online learning platform, a server runs an application system evaluation model, which is used to evaluate the teaching effect and user satisfaction of the platform. In order to continuously improve the teaching quality and user experience, the server will regularly collect user feedback on the course and optimize the evaluation model based on these feedbacks. In this scenario, the server first analyzes the user feedback data on a specific course over a period of time, which includes user satisfaction with the course content, evaluation of the lecturer's teaching method and other information. Through in-depth mining of these data, the server discovered some rules, such as certain specific types of content or teaching methods are more likely to cause positive feedback from users. Based on these findings, the server calculates a "target feedback instance response variation coefficient", which reflects the degree of influence of different teaching content or teaching methods on user feedback. For example, if the content of a certain interactive link is particularly popular with users, then the response variation coefficient of this link will be relatively high. Next, the server uses this response variation coefficient to adjust the evaluation results of the application system evaluation model. Specifically, for those teaching contents or teaching methods with high response variation coefficients, the model will give higher evaluation weights; conversely, for those contents with low response variation coefficients or those that cause negative feedback, the model will reduce their evaluation weights. Through such adjustments, the application system evaluation model can more accurately reflect the user's real feelings and needs when evaluating teaching effects and user satisfaction, thereby providing more targeted improvement suggestions for the online learning platform. After the above steps, the server has obtained an optimized application system evaluation model. This model can not only more accurately evaluate the current teaching effect and user satisfaction, but also provide strong data support for the platform's future teaching content design and teaching method improvement. From this example, it can be seen that optimizing the application system evaluation model using the target feedback instance response variation coefficient can significantly improve the evaluation accuracy and practicality of the model, thereby helping the business application system better meet user needs and improve user experience.

[0082] In the embodiment of the present invention, the application system evaluation model is obtained by training according to the task feedback examples completed by the tuning, which can be implemented through the following examples.

[0083] A first evaluation model is trained based on the first task feedback instance of the tuning completed included in the task feedback instance of the tuning completed; the first evaluation model is used to evaluate the feedback result generated by the application system when the task instruction is not correctly executed; the first task feedback instance of the tuning completed is obtained by performing compressed sensing on the first task feedback instance, and the first task feedback instance is a feedback instance for generating feedback results when the test application system generates feedback results when the historical task instruction is not correctly executed;

[0084] A second evaluation model is trained based on the second task feedback instance of the tuning completed included in the task feedback instance of the tuning completed; the second evaluation model is used to evaluate the feedback result generated by the application system when the task instruction is correctly executed; the second task feedback instance of the tuning completed is obtained by performing compressed sensing on the second task feedback instance, and the second task feedback instance is a feedback instance that generates feedback results when the test application system is correctly executed for the historical task instruction;

[0085] An application system evaluation model is obtained by merging the first evaluation model and the second evaluation model.

[0086] In an embodiment of the present invention, exemplarily, in an intelligent order processing system, the server records a large number of feedback instances of historical order processing tasks. Among them, some feedback instances are generated when the system fails to correctly perform the order processing task, that is, the first task feedback instance. The server first uses compressed sensing technology to tune these first task feedback instances, removes noise and redundant information, and highlights key features. Then, the server uses these tuned first task feedback instances to train the first evaluation model. This model is specifically used to evaluate the feedback results generated when the system fails to correctly perform the order processing task. For example, it can predict the possibility that the system fails to correctly process an order under a certain specific situation, and the impact of this situation on user experience and business processes. In the same intelligent order processing system, another part of the feedback instances is generated when the system correctly performs the order processing task, that is, the second task feedback instance. The server also uses compressed sensing technology to tune these instances. Then, the server uses these tuned second task feedback instances to train the second evaluation model. This model is used to evaluate the feedback results generated when the system correctly performs the order processing task. For example, it can analyze the probability of the system successfully processing an order in different scenarios, and the impact of these successful processing on user satisfaction and business efficiency. After training the first and second evaluation models, the server merges the two models to create a comprehensive application system evaluation model. This comprehensive model can simultaneously consider the feedback results of the system when it performs order processing tasks correctly and incorrectly. By merging the two models, the server obtains an evaluation model that can comprehensively evaluate the performance of the intelligent order processing system. This model can not only predict the performance of the system in different situations, but also provide valuable insights and suggestions for optimizing system performance and improving user experience. In general, through the above steps, the server can build an accurate and comprehensive application system evaluation model based on the feedback examples of the tasks completed by tuning, so as to better understand and improve the performance of the intelligent order processing system.

[0087] In the embodiment of the present invention, the aforementioned step S204 may be implemented through the following example execution.

[0088] Acquire semantic features of a first feature domain and a second feature domain of a current feedback result, and semantic features of a first feature domain and a second feature domain of each standard solution strategy in a first solution strategy set, wherein the first feature domain is smaller than the second feature domain;

[0089] Pairing the semantic features of the first feature domain of the current feedback result with the semantic features of the first feature domain of each standard solution strategy in the first solution strategy set to obtain a corresponding first pairing adaptation degree, and selecting a preset number of standard solution strategies ranked top from the first solution strategy set based on the first pairing adaptation degree to obtain a second solution strategy set;

[0090] The semantic features of the second feature domain of the current feedback result are paired with the semantic features of the second feature domain of each standard solution strategy in the second solution strategy set to obtain a corresponding second pairing adaptation degree, and the solution strategy search result of the current feedback result is determined based on the second pairing adaptation degree.

[0091] In an embodiment of the present invention, exemplarily, in an intelligent customer service system, a server receives a printer failure problem reported by a user. In order to quickly locate and solve the problem, the server needs to automatically find the most suitable solution strategy from a large number of standard solution strategies. The server first analyzes the current fault description reported by the user, and extracts the semantic features of the first feature domain (such as key information such as fault type and device model) and the second feature domain (such as more detailed information such as specific error code and fault phenomenon). At the same time, the server also extracts the semantic features of the first feature domain and the second feature domain corresponding to each standard solution strategy from the pre-established first solution strategy set. In this example, the first feature domain may include keywords such as "printer" and "paper jam", while the second feature domain may include more specific detailed information such as "error code E02" and "paper cannot pass normally". Next, the server pairs the semantic features of the first feature domain of the current feedback with the semantic features of the first feature domain of each standard solution strategy in the first solution strategy set. By calculating the similarity or matching degree, the corresponding first pairing adaptation degree is obtained. Based on this first pairing adaptation degree, the server selects the top preset number (such as 5) of standard solution strategies from the first solution strategy set to form a second solution strategy set. In this way, the server quickly narrows down the range of possible solution strategies. Then, the server pairs the second feature domain semantic features of the current feedback with the second feature domain semantic features of each standard solution strategy in the second solution strategy set. Through more sophisticated matching calculations, the corresponding second pairing adaptation degree is obtained. Based on this second pairing adaptation degree, the server finally determines the solution strategy search result for the current feedback result. In this example, the server may have found a standard solution strategy that is highly matched with the current fault description, which provides detailed instructions on how to clear the printer paper jam. In this way, the server can efficiently provide users with accurate fault solutions, greatly improving the response speed and quality of customer service.

[0092] In the embodiment of the present invention, the acquisition of the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result may be implemented through the following examples.

[0093] Deconstructing the current feedback result to obtain a first semantic description and a second semantic description corresponding to the current feedback result;

[0094] A feature extraction operation is performed on the first semantic description and the second semantic description corresponding to the current feedback result respectively to obtain semantic features of the first feature domain and semantic features of the second feature domain of the current feedback result.

[0095] In an embodiment of the present invention, exemplarily, in an online shopping platform, the server is responsible for processing user feedback on shopping experience, product quality, logistics and distribution, etc. In order to more effectively analyze and respond to these feedbacks, the server adopts a feature extraction method to extract key information in the feedback. When a user submits a feedback, such as: "The size of the shoes I bought is not suitable, and the logistics speed is too slow." The server first deconstructs the feedback. The purpose of the deconstruction process is to decompose the feedback content into different semantic description parts. In this example, the server will deconstruct the feedback into two parts: the first part is about the problem of the product ("The size of the shoes I bought is not suitable"), and the second part is about the problem of logistics ("And the logistics speed is too slow"). These two parts correspond to the first semantic description and the second semantic description respectively. Next, the server performs feature extraction operations on the first semantic description and the second semantic description respectively. The purpose of feature extraction is to extract the key features in each semantic description, which will be used for subsequent matching and analysis. For the first semantic description "The size of the shoes I bought is not suitable", the key features that the server can extract include "shoes", "size", and "unsuitable". These features constitute the semantic features of the first feature domain. For the second semantic description "and the logistics speed is too slow", the key features that the server can extract include "logistics", "speed", and "too slow". These features constitute the semantic features of the second feature domain. Through deconstruction processing and feature extraction operations, the server successfully obtains the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result. These features will be used in the subsequent solution strategy search and matching process to help the server locate problems more quickly and provide corresponding solutions.

[0096] In an embodiment of the present invention, the feature extraction operation is performed on the first semantic description and the second semantic description corresponding to the current feedback result respectively to obtain the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result, which can be implemented through the following examples.

[0097] Performing preliminary feature extraction operations on the first semantic description and the second semantic description of the current feedback result respectively to obtain a first preliminary semantic feature and a second preliminary semantic feature of the current feedback result;

[0098] Performing feature aggregation operations on the first preliminary semantic feature and the second preliminary semantic feature of the current feedback result respectively to obtain a first aggregated semantic feature and a second aggregated semantic feature corresponding to the current feedback result;

[0099] A feature selection operation is performed on the first summary semantic feature and the second summary semantic feature respectively to obtain a semantic feature of a first feature domain and a semantic feature of a second feature domain corresponding to the current feedback result.

[0100] In an embodiment of the present invention, for example, on a social networking platform, users often provide feedback on platform functions, interface design, advertising push, etc. In order to improve platform services, the server needs to conduct in-depth analysis of these feedbacks. The server receives a user feedback, for example: "The interface design of the new version is very novel, but the advertising push is too frequent." First, the server deconstructs this feedback to obtain a first semantic description: "The interface design of the new version is very novel", and a second semantic description: "But the advertising push is too frequent." Then, the server performs preliminary feature extraction operations on the first semantic description and the second semantic description respectively. For the first semantic description, the preliminary features that may be extracted include "new version", "interface design", and "novel"; for the second semantic description, the preliminary features extracted may include "advertising push" and "frequent". After the preliminary feature extraction, the server summarizes the first preliminary semantic features and the second preliminary semantic features respectively. The purpose of this step is to organize and merge the preliminary features for further analysis. For the preliminary features of the first semantic description, "new version", "interface design" and "novel" can be summarized as "new version interface design is novel". Similarly, the preliminary features of the second semantic description "advertising push" and "frequent" can be summarized as "advertising push is frequent". Finally, the server performs feature selection operations on the first summary semantic feature and the second summary semantic feature. The purpose of this step is to select the most representative features from the summary features for subsequent feedback analysis and processing. For the first summary semantic feature "the new version interface design is novel", the server may select "new version" and "interface design" as the semantic features of the first feature domain because they are directly related to the theme of user feedback. For the second summary semantic feature "frequent advertising push", the server may select "advertising push" as the semantic feature of the second feature domain because it directly reflects the user's problem. Through the above steps, the server successfully extracts the key semantic features of the first feature domain and the semantic features of the second feature domain from the user feedback. These features will be used for further user feedback analysis to help the server understand the user's needs and problems, thereby improving the service quality of the social networking platform. This method not only improves the efficiency of feedback processing, but also enables the server to respond to user needs and problems more accurately.

[0101] In an embodiment of the present invention, the performing of feature extraction operations on the first semantic description and the second semantic description corresponding to the current feedback result is implemented by calling a strategy feature analysis model, wherein the strategy feature analysis model includes a preliminary semantic feature encoder, a feature aggregation component, and a second feature selection component;

[0102] The feature extraction operation is performed on the first semantic description and the second semantic description corresponding to the current feedback result to obtain the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result, which can be implemented through the following example.

[0103] By means of the preliminary semantic feature encoder, performing preliminary feature extraction operations on the first semantic description and the second semantic description respectively, to obtain a first preliminary semantic feature and a second preliminary semantic feature of the current feedback result;

[0104] By means of the feature summarization component, a feature summarization operation is performed on the first preliminary semantic feature and the second preliminary semantic feature respectively, so as to obtain a first summarized semantic feature and a second summarized semantic feature corresponding to the current feedback result;

[0105] Performing a feature selection operation on the first summary semantic feature through the second feature selection component to obtain a semantic feature of the first feature domain corresponding to the current feedback result;

[0106] A feature compression operation is performed on the second summarized semantic feature to obtain a semantic feature of a second feature domain corresponding to the current feedback result.

[0107] In the embodiment of the present invention, for example, in an e-commerce platform, user feedback is an important basis for improving services and products. The server performs in-depth analysis of user feedback by calling a strategic feature analysis model to extract key information and respond. When a user submits feedback, such as "the purchase process is too complicated and the coupon is not convenient to use", the server first processes the feedback through a preliminary semantic feature encoder. The encoder deconstructs the feedback into two semantic descriptions: "the purchase process is too complicated" and "the coupon is not convenient to use". The encoder performs preliminary feature extraction on these two semantic descriptions respectively. For "the purchase process is too complicated", the features that may be extracted include "purchase process" and "complex"; for "the coupon is not convenient to use", the extracted features may include "coupon" and "inconvenient to use". Next, the feature aggregation component aggregates the above-mentioned preliminary extracted features. This step merges similar features, removes redundancy, and enhances the expressive power of the features. For "the purchase process is too complicated", the aggregated feature may be "complex purchase process"; for "the coupon is not convenient to use", the aggregated feature may be "barrier to coupon use". Afterwards, the second feature selection component selects the aggregated features. For "complex purchase process", "purchase process" may be selected as the semantic feature of the first feature domain because it directly points to the core of the problem. For "coupon usage barriers", in addition to feature selection, feature compression operations are also performed to further refine the features. The compressed features may focus more on the inconvenience of "coupon usage", so "coupon usage" is selected as the semantic feature of the second feature domain. Through the processing of the strategic feature analysis model, the server successfully extracted the key semantic features in the user feedback. These features will be used for subsequent problem location and solution strategy search to help e-commerce platforms quickly respond to user feedback and improve service processes. This approach not only improves the intelligence level of user feedback processing, but also enables e-commerce platforms to grasp user needs more accurately and improve user experience.

[0108] In an embodiment of the present invention, the strategy feature analysis model further includes a first feature selection component. The embodiment of the present invention also provides the following implementation methods.

[0109] Obtain the original strategy feature analysis model and solution strategy instance;

[0110] Based on the solution strategy example, a training process is performed on the preliminary semantic feature encoder in the original strategy feature analysis model to obtain a first strategy feature analysis model corresponding to the original strategy feature analysis model;

[0111] Fixing the parameters of the preliminary semantic feature encoder in the first strategic feature analysis model, executing a training process on the first feature selection component in the first strategic feature analysis model, and obtaining a second strategic feature analysis model corresponding to the first strategic feature analysis model;

[0112] Fix the parameters of the preliminary semantic feature encoder in the first strategy feature analysis model and the parameters of the first feature selection component in the first strategy feature analysis model, execute a training process for the second feature selection component in the second strategy feature analysis model, obtain a third strategy feature analysis model corresponding to the second strategy feature analysis model, and use the third strategy feature analysis model as the strategy feature analysis model.

[0113] In an embodiment of the present invention, for example, in a financial service platform, credit risk assessment is one of the core businesses. In order to more accurately assess the credit risk of borrowers, the server needs to use a strategy feature analysis model to process and analyze a large amount of credit application data. The server first obtains an original strategy feature analysis model, which includes a preliminary semantic feature encoder, a feature aggregation component, a first feature selection component, and a second feature selection component. At the same time, the server also collects a batch of processed credit application data as a solution strategy instance, which contains various information about the borrower and the final credit decision result. The server uses the solution strategy instance to train the preliminary semantic feature encoder in the original strategy feature analysis model. In this process, the server continuously adjusts the parameters of the encoder so that it can more accurately extract preliminary semantic features from the credit application data. After multiple rounds of training, the server obtains an optimized first strategy feature analysis model. Next, the server fixes the parameters of the preliminary semantic feature encoder in the first strategy feature analysis model, and then trains the first feature selection component. The task of this component is to select the most useful features for credit risk assessment from the preliminary semantic features. By continuously adjusting the parameters of the first feature selection component, the server enables it to more accurately screen out key features. After the training is completed, the server obtains the second policy feature analysis model. Finally, the server fixes the parameters of the preliminary semantic feature encoder and the parameters of the first feature selection component in the first policy feature analysis model, and then trains the second feature selection component. This component is responsible for further refining and compressing features to improve the accuracy of credit risk assessment. After multiple rounds of training and optimization, the server obtains the final third policy feature analysis model and applies it as a policy feature analysis model to the actual credit risk assessment task. By gradually training and optimizing the various components of the policy feature analysis model, the server successfully improves the accuracy of credit risk assessment. This method based on big data and machine learning not only improves the risk control capabilities of the financial service platform, but also provides borrowers with more fair and scientific credit assessment services.

[0114] In an embodiment of the present invention, based on the solution strategy instance, a training process is performed on the preliminary semantic feature encoder in the original strategy feature analysis model to obtain a first strategy feature analysis model corresponding to the original strategy feature analysis model, which can be implemented through the following example.

[0115] Deconstruct the solution strategy instance to obtain a first semantic description and a second semantic description corresponding to the solution strategy instance, and use the first semantic description corresponding to the solution strategy instance as a reference semantic description, the second semantic description corresponding to the solution strategy instance as a target semantic description, and other solution strategy instances as non-target semantic descriptions;

[0116] Calling the preliminary semantic feature encoder in the original strategy feature analysis model to perform preliminary feature extraction operations on the reference semantic description, the target semantic description and the non-target semantic description, respectively, to obtain corresponding reference semantic description features, target semantic description features and non-target semantic description features;

[0117] Acquire a first feature distance between the reference semantic description feature and the target semantic description feature, and a second feature distance between the reference semantic description feature and the non-target semantic description feature, and construct a first cost parameter of the original strategy feature analysis model based on the first feature distance and the second feature distance;

[0118] The preliminary semantic feature encoder in the original strategy feature analysis model is selected based on the first cost parameter to obtain a first strategy feature analysis model corresponding to the original strategy feature analysis model.

[0119] In an embodiment of the present invention, for example, in an intelligent customer service system of a large enterprise, the server needs to accurately classify the questions raised by users in order to quickly provide solutions. To achieve this goal, the server uses a strategic feature analysis model and improves the accuracy of problem classification through training. The server first collects a batch of solved problem instances, which contain the original text of the user's question and the corresponding classification label. The server deconstructs these instances and decomposes each problem instance into two parts: the first semantic description (i.e., the core content of the problem) and the second semantic description (i.e., the specific details or background information of the problem). For example, for the question "What if my mobile phone cannot connect to WiFi?", the first semantic description can be "The mobile phone cannot connect to WiFi", and the second semantic description can contain specific information such as the mobile phone model and the operating system version. Next, the server calls the preliminary semantic feature encoder in the original strategic feature analysis model to perform preliminary feature extraction operations on the reference semantic description (the core content of the correctly classified problem), the target semantic description (the core content of the problem to be classified), and the non-target semantic description (the core content of other classified problems). The purpose of this step is to extract key feature information from the text to provide a basis for subsequent classification tasks. The server calculates the first feature distance between the reference semantic description feature and the target semantic description feature, and the second feature distance between the reference semantic description feature and the non-target semantic description feature. These distances measure the similarity between the core contents of different questions. Based on these distances, the server constructs a cost parameter to evaluate the performance of the preliminary semantic feature encoder. This cost parameter reflects the accuracy of the encoder in distinguishing different question categories. Finally, the server uses the above cost parameter to select features for the preliminary semantic feature encoder in the original strategy feature analysis model. By optimizing this cost parameter, the server gradually adjusts the parameters of the encoder so that it can more accurately extract key features from the question text. After multiple rounds of training, the server obtains the optimized first strategy feature analysis model, which shows higher accuracy in the question classification task. Through the above training process, the question classifier of the intelligent customer service system has been significantly improved. In practical applications, this classifier can more accurately identify the types of questions raised by users and quickly provide corresponding solutions. This not only improves customer service efficiency, but also improves user satisfaction.

[0120] In the embodiment of the present invention, the training process is executed on the first feature selection component in the first strategy feature analysis model to obtain the second strategy feature analysis model corresponding to the first strategy feature analysis model, which can be implemented through the following example.

[0121] Calling the first feature selection component in the first strategy feature analysis model to perform feature selection operations on the reference semantic description feature, the target semantic description feature, and the non-target semantic description feature, respectively, to obtain corresponding reference semantic description selection features, target semantic description selection features, and non-target semantic description selection features;

[0122] Acquire a third feature distance between the reference semantic description selection feature and the target semantic description selection feature, and a fourth feature distance between the reference semantic description selection feature and the non-target semantic description selection feature, and construct a second cost parameter of the first strategy feature analysis model based on the third feature distance and the fourth feature distance;

[0123] The first feature selection component in the first strategy feature analysis model selects features based on the second cost parameter to obtain a second strategy feature analysis model corresponding to the first strategy feature analysis model.

[0124] In an embodiment of the present invention, exemplarily, in a social media platform, the server needs to recommend relevant content to the user based on the user's interests and historical behaviors. In order to improve the accuracy of the recommended content and user satisfaction, the server uses a strategic feature analysis model and optimizes the feature selection component in the model through training. The server first calls the first feature selection component in the first strategic feature analysis model. The function of this component is to select the most useful features for the recommendation task from the initially extracted features. The server performs feature selection operations on the reference semantic description features (content features of known user interests), the target semantic description features (features of the content to be recommended), and the non-target semantic description features (features of other content). This step is intended to screen out the features most relevant to the user's interests to improve the accuracy of the recommendation. Next, the server calculates the third feature distance between the reference semantic description selection feature and the target semantic description selection feature, and the fourth feature distance between the reference semantic description selection feature and the non-target semantic description selection feature. These distances reflect the similarity between the selected features and the user's interests. Based on these distances, the server constructs a second cost parameter for evaluating the performance of the first feature selection component. The smaller this cost parameter is, the more the selected features match the user's interests. Finally, the server uses the above-mentioned second cost parameter to further optimize the first feature selection component in the first strategy feature analysis model. By adjusting the parameters of the component, the server gradually reduces the second cost parameter, thereby selecting features that are closer to the user's interests. After multiple rounds of training, the server obtains the optimized second strategy feature analysis model. Through the above training process, the content recommendation system of the social media platform has been significantly improved. In practical applications, this optimized system can more accurately recommend relevant content based on the user's interests and historical behaviors. This not only improves user satisfaction, but also increases user stickiness, bringing more active users to the social media platform. At the same time, because the recommended content is more accurate, it also improves the advertising conversion rate and commercial value of the platform.

[0125] In the embodiment of the present invention, the training process is executed on the second feature selection component in the second strategy feature analysis model to obtain the third strategy feature analysis model corresponding to the second strategy feature analysis model, which can be implemented through the following example.

[0126] Acquire a strategy feature instance set, the strategy feature instance set comprising feature instances corresponding to a plurality of solution strategy instances and a feature instance target value corresponding to each of the feature instances, the feature instance target value being used to characterize a selection feature of a feature selection operation performed on the feature instance by the second feature selection component;

[0127] Calling the second feature selection component in the second strategy feature analysis model to perform feature selection operations on each of the feature instances to obtain inferred selected features corresponding to each of the feature instances;

[0128] Determine the feature distance between each of the inferred selected features and the corresponding feature instance target value, and calculate the mean of each of the feature distances to obtain a third cost parameter of the strategy feature analysis model;

[0129] Based on the third cost parameter, the second feature selection component in the second strategy feature analysis model selects features to obtain a third strategy feature analysis model corresponding to the second strategy feature analysis model.

[0130] In an embodiment of the present invention, illustratively, on a large e-commerce platform, the server needs to accurately recommend products that users may be interested in. In order to achieve this goal, the server uses a strategic feature analysis model and continuously optimizes the feature selection components in the model through multiple stages of training, thereby improving the accuracy of product recommendations. The server first collects a batch of user purchase records as a set of strategic feature instances. These instance sets contain feature instances (such as product type, price range, user historical purchase behavior, etc.) corresponding to multiple solution strategy instances (i.e., user purchase records) and feature instance target values ​​corresponding to each feature instance. These target values ​​represent the selected features expected to be obtained after the feature instance is subjected to feature selection operation by the second feature selection component, i.e., the product features that the user finally purchased. Next, the server calls the second feature selection component in the second strategic feature analysis model to perform feature selection operation on each feature instance respectively. The purpose of this step is to filter out the features that have the greatest impact on the user's purchase decision from the rich feature set. Through this operation, the server obtains the inferred selection features corresponding to each feature instance. The server then determines the feature distance between each inferred selection feature and the corresponding feature instance target value. This distance reflects the difference between the user preference inferred by the model and the actual user purchase behavior. In order to obtain a global evaluation index, the server averages these feature distances and obtains the third cost parameter of the strategy feature analysis model. Finally, the server further trains and optimizes the second feature selection component in the second strategy feature analysis model based on the third cost parameter. By adjusting the parameters of the component, the server aims to reduce the third cost parameter, that is, to reduce the difference between the model inference and the actual behavior of the user. After multiple rounds of training, the server obtains the optimized third strategy feature analysis model. Through the above training process, the product recommendation system of the e-commerce platform has been significantly improved. In actual applications, this optimized system can more accurately recommend relevant products based on the user's personal preferences and historical purchasing behavior. This not only improves the user's shopping experience, but also increases the platform's sales and user satisfaction. At the same time, because the recommendation is more accurate, it also reduces the time cost of users browsing and searching, and improves the overall user experience.

[0131] In the embodiment of the present invention, obtaining the semantic features of the first feature domain and the semantic features of the second feature domain of each standard solution strategy in the first solution strategy set may be implemented through the following examples.

[0132] The following processing is performed on each standard solution strategy in the first solution strategy set:

[0133] Deconstructing the standard solution strategy to obtain a third semantic description corresponding to the standard solution strategy;

[0134] Performing a preliminary feature extraction operation on the third semantic description corresponding to the standard solution strategy to obtain a third preliminary semantic feature corresponding to the standard solution strategy;

[0135] Performing a feature aggregation operation on the third preliminary semantic feature corresponding to the standard solution strategy to obtain a third aggregated semantic feature corresponding to the standard solution strategy;

[0136] A preset multi-level feature selection operation is performed on the third summary semantic feature to obtain the semantic features of the first feature domain and the semantic features of the second feature domain corresponding to the standard solution strategy.

[0137] In an embodiment of the present invention, illustratively, in a financial risk control system, a server needs to automatically approve a credit application. In order to improve the accuracy and efficiency of approval, the server uses big data and machine learning technology to analyze and evaluate the credit risk of the applicant. In this scenario, how to obtain the semantic features of the first feature domain and the second feature domain in the credit approval strategy (i.e., the standard solution strategy) will be described in detail. The server first obtains a credit approval strategy set (a first solution strategy set), which contains multiple standard solution strategies. For each standard solution strategy, the server performs a deconstruction process. For example, a standard solution strategy may be "if the applicant is older than 25 years old and younger than 60 years old, and the monthly income exceeds 5,000 yuan, then approve the credit application." The server deconstructs this strategy into a third semantic description, such as "age older than 25 years old", "age younger than 60 years old", and "monthly income exceeds 5,000 yuan". Next, the server performs a preliminary feature extraction operation on the deconstructed third semantic description. For the third semantic description in the above example, the preliminary features that the server may extract include "age", "greater than 25 years old", "less than 60 years old", "monthly income" and "more than 5,000 yuan". These preliminary features reflect the key information points in the credit approval strategy. The server then aggregates the initially extracted features. In this step, the server may aggregate the two features "age greater than 25 years old" and "age less than 60 years old" into a higher-level feature of "age range". Similarly, "monthly income" and "more than 5,000 yuan" can be aggregated into the feature of "monthly income level". Through feature aggregation, the server can reduce the redundancy of features and improve the efficiency of subsequent processing. Finally, the server performs a preset multi-level feature selection operation on the aggregated semantic features. The purpose of this step is to filter out the features that have the greatest impact on the credit approval decision. For example, in the first level of selection, the server can retain features that are directly related to credit risk assessment, such as "age range" and "monthly income level"; while in the second level of selection, the server may further filter out features that are more important for the approval of specific credit products. Through multi-level feature selection, the server can build a more accurate and efficient credit approval model. Through the above processing flow, the server successfully obtains the semantic features of the first feature domain and the second feature domain in the credit approval strategy. These features not only reflect the key information points of the credit approval strategy, but also provide strong data support for subsequent credit approval decisions. In actual applications, this optimized system can more accurately assess the credit risk of applicants and make reasonable credit approval decisions, thereby improving the risk control capabilities and business efficiency of financial institutions.

[0138] In the embodiment of the present invention, the performing of a preliminary feature extraction operation on the third semantic description corresponding to the standard solution strategy to obtain the third preliminary semantic feature corresponding to the standard solution strategy may be implemented through the following example.

[0139] Performing a preliminary feature extraction operation on the third semantic description corresponding to the standard solution strategy to obtain a plurality of description feature representations corresponding to the standard solution strategy, and using the plurality of description feature representations as third preliminary semantic features;

[0140] The performing a feature aggregation operation on the third preliminary semantic feature corresponding to the standard solution strategy to obtain the third aggregated semantic feature corresponding to the standard solution strategy includes:

[0141] A feature integration operation is performed on the multiple descriptive feature representations to obtain a third summary semantic feature corresponding to the standard solution strategy.

[0142] In an embodiment of the present invention, for example, in an intelligent customer service system of a large enterprise, the server needs to automatically classify the questions raised by the user and give corresponding automatic replies. In order to improve the accuracy of classification and the satisfaction of the reply, the server will use the machine learning model to extract and summarize the features of the standard solution strategy (i.e., problem classification and reply rules). When a user asks a question to the intelligent customer service system, the server will first perform semantic analysis on the question to obtain the third semantic description of the question. Then, the server will perform a preliminary feature extraction operation on this third semantic description. For example, for the question "What should I do if my mobile phone cannot connect to Wi-Fi?", the server will extract the following descriptive feature representations: mobile phone; cannot connect; Wi-Fi; these descriptive feature representations are used by the server as the third preliminary semantic features. These features directly reflect the key information in the question and the focus of user concern. After obtaining multiple descriptive feature representations, the server will perform feature integration operations on these features to obtain more comprehensive and advanced semantic features. In this example, the server will integrate the three descriptive feature representations of "mobile phone", "cannot connect" and "Wi-Fi" into a higher-level feature-"mobile phone Wi-Fi connection problem". This integrated feature is the third summary semantic feature, which describes the essence of the user's problem more comprehensively. Through the above steps, the intelligent customer service system can accurately understand the core of the user's problem and quickly match the corresponding solution strategy based on the third summary semantic feature. For example, in this scenario, the system will automatically find the relevant solution strategy based on the feature of "mobile phone Wi-Fi connection problem" and give a targeted reply, such as "Please check your Wi-Fi settings and make sure the network name and password are entered correctly." This processing method not only improves the response speed of the customer service system, but also improves user satisfaction. In general, by performing preliminary feature extraction and feature aggregation operations on standard solution strategies, the server can more effectively understand and handle user problems, thereby providing more accurate and efficient automatic reply services.

[0143] In the embodiment of the present invention, the preset multi-level feature selection operation is performed on the third summary semantic feature to obtain the semantic features of the first feature domain and the semantic features of the second feature domain corresponding to the standard solution strategy, which can be implemented through the following examples.

[0144] Respectively obtaining a first dimensionality reduction feature and a second dimensionality reduction feature for reducing the dimension of the feature domain of the third summary semantic feature;

[0145] Performing a product operation on the first dimensionality reduction feature and the third summary semantic feature to obtain a first product, and performing a product operation on the second dimensionality reduction feature and the third summary semantic feature to obtain a second product;

[0146] The first product is mapped to obtain semantic features of a first feature domain corresponding to the standard solution strategy, and the second product is mapped to obtain semantic features of a second feature domain corresponding to the standard solution strategy.

[0147] In an embodiment of the present invention, exemplarily, on a large e-commerce platform, a server provides users with accurate product recommendations through a personalized recommendation system. To achieve this goal, the system needs to perform in-depth analysis and feature extraction on the user's shopping behavior, preferences and other data. In this scenario, it will be described how to extract the semantic features of the first feature domain and the semantic features of the second feature domain from the third summary semantic feature by presetting a multi-level feature selection operation. The server first obtains the first dimension reduction feature and the second dimension reduction feature for reducing the dimension of the third summary semantic feature. These dimension reduction features are obtained by training a machine learning algorithm (such as principal component analysis PCA or autoencoder Autoencoder) to reduce the dimension of the data while retaining its main features. For example, the first dimension reduction feature can focus on the user's purchase history and browsing behavior, while the second dimension reduction feature may focus more on the user's search keywords and click-through rate and other data. Next, the server multiplies the first dimension reduction feature with the third summary semantic feature to obtain a first product. Similarly, the server also multiplies the second dimension reduction feature with the third summary semantic feature to obtain a second product. These product operations help map the original high-dimensional features to a low-dimensional space while retaining the most important information. Finally, the server maps the first product to obtain the semantic features of the first feature domain corresponding to the standard solution strategy. This first feature domain can represent the user's long-term shopping preferences and habits. At the same time, the server also maps the second product to obtain the semantic features of the second feature domain corresponding to the standard solution strategy. This second feature domain may reflect the user's short-term interests and needs more. Through the above steps, the personalized recommendation system of the e-commerce platform can more accurately capture the user's shopping preferences and needs. The semantic features of the first feature domain help the system understand the user's long-term stable preferences, while the semantic features of the second feature domain reflect the user's immediate interests and needs. Based on these features, the system can provide users with more accurate and personalized product recommendations, thereby improving user satisfaction and shopping experience. For example, if a user has been buying sports shoes for a long time, the semantic features of the first feature domain will reflect this. If the user has recently searched and browsed multiple smart watches, the semantic features of the second feature domain will capture this short-term interest in a timely manner. Combining the information of these two feature domains, the recommendation system can recommend products that meet both the user's long-term stable preferences and his short-term interests.

[0148] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned application system evaluation method based on big data. Figure 2 As shown, Figure 2The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0149] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A method for evaluating an application system based on big data, characterized in that: The method comprises: Acquire a current task instruction for a target application system, wherein the current task instruction is used to test the target application system; Calling the target application system to execute the current task instruction and obtain the current feedback result; Inputting the current feedback result into a pre-selected and trained application system evaluation model to obtain a current evaluation result; When the evaluation result indicates that the target application system is abnormal, based on the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result, determining a solution strategy search result of the current feedback result; The application system evaluation model is obtained by: Obtaining task feedback instances, wherein the task feedback instances include target feedback instances and non-target feedback instances; the target feedback instance is a feedback instance in which the test application system generates a correct feedback result for a historical task indication; the non-target feedback instance is a feedback instance in which the test application system does not generate the correct feedback result for the historical task indication; When the task feedback instance meets the instance tuning trigger condition, a compressed sensing parameter of the task feedback instance is obtained based on a compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameter, and a tuned task feedback instance is obtained based on the non-target feedback instance and the target feedback instance completed by compressed sensing; A first evaluation model is trained based on the first task feedback instance of the tuning completed included in the task feedback instance of the tuning completed; the first evaluation model is used to evaluate the feedback result generated by the application system when the task instruction is not correctly executed; the first task feedback instance of the tuning completed is obtained by performing compressed sensing on the first task feedback instance, and the first task feedback instance is a feedback instance for generating feedback results when the test application system generates feedback results when the historical task instruction is not correctly executed; A second evaluation model is trained based on the second task feedback instance of the tuning completed included in the task feedback instance of the tuning completed; the second evaluation model is used to evaluate the feedback result generated by the application system when the task instruction is correctly executed; the second task feedback instance of the tuning completed is obtained by performing compressed sensing on the second task feedback instance, and the second task feedback instance is a feedback instance that generates feedback results when the test application system is correctly executed for the historical task instruction; Obtaining an application system evaluation model based on combining the first evaluation model and the second evaluation model; The evaluation result of the application system evaluation model is optimized to obtain an optimized application system evaluation model; the optimized application system evaluation model is used to evaluate the feedback result generated by the application system in response to the task instruction.

2. The method according to claim 1, characterized in that: The method of obtaining a compressed sensing parameter of the task feedback instance based on a compressed sensing strategy, performing compressed sensing on the non-target feedback instance according to the compressed sensing parameter, and obtaining a tuned task feedback instance based on the non-target feedback instance and the target feedback instance completed by compressed sensing includes: Obtaining a first task feedback instance and a second task feedback instance from the task feedback instance; Obtaining a first compressed sensing parameter corresponding to the first task feedback instance based on a compressed sensing strategy, and obtaining a second compressed sensing parameter corresponding to the second task feedback instance based on a compressed sensing strategy; Tuning the first task feedback instance based on the first compressed sensing parameter to obtain a tuned first task feedback instance, and tuning the second task feedback instance based on the second compressed sensing parameter to obtain a tuned second task feedback instance; Based on the first task feedback instance after the tuning is completed and the second task feedback instance after the tuning is completed, a task feedback instance after the tuning is completed is obtained.

3. The method according to claim 2, characterized in that The first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; the compressed sensing strategy includes a category-differentiated compressed sensing method; The obtaining, based on the compressed sensing strategy, a first compressed sensing parameter corresponding to the first task feedback instance includes: Based on the category-differentiated compressed sensing mode, a response variation coefficient of a first target feedback instance is obtained according to the first target feedback instance and the first non-target feedback instance; a first compressed sensing parameter corresponding to the first task feedback instance is obtained based on the response variation coefficient of the first target feedback instance and the feedback distribution feature instance of the first target feedback instance; The obtaining, based on the compressed sensing strategy, a second compressed sensing parameter corresponding to the second task feedback instance includes: Based on the category-differentiated compressed sensing mode, a response variation coefficient of a second target feedback instance is obtained according to the second target feedback instance and the second non-target feedback instance; a second compressed sensing parameter corresponding to the second task feedback instance is obtained based on the response variation coefficient of the second target feedback instance and the feedback distribution feature instance of the second target feedback instance; The compressed sensing strategy also includes a step-by-step compressed sensing method; the method also includes: Obtaining a target feedback instance response variation coefficient according to the target feedback instance and the non-target feedback instance; The obtaining of the first compressed sensing parameter corresponding to the first task feedback instance based on the compressed sensing strategy also includes: According to the step-by-step compressed sensing method, a first compressed sensing parameter corresponding to the first task feedback instance is obtained based on the response variation coefficient of the target feedback instance and the feedback distribution feature instance of the first target feedback instance; The obtaining, based on the compressed sensing strategy, a second compressed sensing parameter corresponding to the second task feedback instance includes: According to the step-by-step compressed sensing method, a second compressed sensing parameter corresponding to the second task feedback instance is obtained based on the response variation coefficient of the target feedback instance and the feedback distribution feature instance of the second target feedback instance.

4. The method according to claim 2, characterized in that: The compressed sensing strategy includes a basic compressed sensing method; the compressed sensing parameters of the task feedback instance are obtained based on the compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameters, and the optimized task feedback instance is obtained based on the non-target feedback instance and the target feedback instance completed by compressed sensing, including: Based on the basic compressed sensing method, obtaining a target feedback instance response variation coefficient according to the target feedback instance and the non-target feedback instance; Obtaining a compressed sensing parameter based on the target feedback instance response variation coefficient and the target feedback instance feedback distribution feature instance; removing selected non-target feedback instances from the non-target feedback instances according to the compressed sensing parameters to obtain non-target feedback instances completed by compressed sensing; Combining the non-target feedback instance completed by the compressed sensing and the target feedback instance to obtain a task feedback instance completed by tuning; The compressed sensing strategy also includes a segmented compressed sensing method; the compressed sensing parameters of the task feedback instance are obtained based on the compressed sensing strategy, and compressed sensing is performed on the non-target feedback instance according to the compressed sensing parameters, and the task feedback instance completed by tuning is obtained based on the non-target feedback instance and the target feedback instance completed by compressed sensing, and further includes: Obtaining a first task feedback instance and a second task feedback instance from the task feedback instance; the first task feedback instance includes a first target feedback instance and a first non-target feedback instance; the second task feedback instance includes a second target feedback instance and a second non-target feedback instance; Based on the segmented compressed sensing method, obtaining a plurality of first undetermined response variation coefficients according to the first target feedback instance and the first non-target feedback instance; Based on the segmented compressed sensing method, obtaining a plurality of second undetermined response variation coefficients according to the second target feedback instance and the second non-target feedback instance; Construct a response variation coefficient merger according to the plurality of first pending response variation coefficients and the plurality of second pending response variation coefficients; the response variation coefficient merger includes a first target feedback instance response variation coefficient and a second target feedback instance response variation coefficient, the first target feedback instance response variation coefficient belongs to the plurality of first pending response variation coefficients, and the second target feedback instance response variation coefficient belongs to the plurality of second pending response variation coefficients; The first compressed sensing parameter is obtained based on the response variation coefficient of the first target feedback instance and the feedback distribution feature instance of the first target feedback instance, and the second compressed sensing parameter is obtained based on the response variation coefficient of the second target feedback instance and the feedback distribution feature instance of the second target feedback instance; Obtaining a compressed sensing parameter combination based on the first compressed sensing parameter and the second compressed sensing parameter; the compressed sensing parameter combination includes the first compressed sensing parameter and the second compressed sensing parameter; Performing compressed sensing on a first non-target feedback instance in the first task feedback instance according to the first compressed sensing parameter, and obtaining a first task feedback instance that has been tuned based on the first non-target feedback instance and the first target feedback instance completed by compressed sensing; The step of tuning the second task feedback instance based on the second compressed sensing parameter to obtain the tuned second task feedback instance includes: performing compressed sensing on a second non-target feedback instance in the second task feedback instance according to the second compressed sensing parameter, obtaining a tuned second task feedback instance based on the second non-target feedback instance and the second target feedback instance completed by compressed sensing, and tuning the second task feedback instance based on the second compressed sensing parameter to obtain a tuned second task feedback instance; Based on the first task feedback instance after the tuning is completed and the second task feedback instance after the tuning is completed, a task feedback instance after the tuning is completed is obtained.

5. The method according to claim 3, characterized in that: The optimizing the evaluation result of the application system evaluation model to obtain the optimized application system evaluation model includes: Obtaining a sample evaluation result of the application system evaluation model for the task feedback instance; Obtaining a first optimization mapping relationship for the application system evaluation model according to the sample evaluation result and the actual response label of the task feedback instance; Adjusting the evaluation result of the application system evaluation model according to the first optimization mapping relationship to obtain an optimized application system evaluation model; The step of optimizing the evaluation result of the application system evaluation model to obtain an optimized application system evaluation model further includes: Obtaining a second optimization mapping relationship for the application system evaluation model according to the compressed sensing parameter; Adjusting the evaluation result of the application system evaluation model according to the second optimization mapping relationship to obtain an optimized application system evaluation model; The step of optimizing the evaluation result of the application system evaluation model to obtain an optimized application system evaluation model further includes: The evaluation result of the application system evaluation model is adjusted according to the target feedback instance response variation coefficient to obtain an optimized application system evaluation model.

6. The method according to claim 1, characterized in that The step of determining a solution strategy search result for the current feedback result based on the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result includes: Deconstructing the current feedback result to obtain a first semantic description and a second semantic description corresponding to the current feedback result; Performing feature extraction operations on the first semantic description and the second semantic description corresponding to the current feedback result respectively, obtaining semantic features of the first feature domain and the second feature domain of the current feedback result, and semantic features of the first feature domain and the second feature domain of each standard solution strategy in the first solution strategy set, wherein the first feature domain is smaller than the second feature domain; Pairing the semantic features of the first feature domain of the current feedback result with the semantic features of the first feature domain of each standard solution strategy in the first solution strategy set to obtain a corresponding first pairing adaptation degree, and selecting a preset number of standard solution strategies ranked top from the first solution strategy set based on the first pairing adaptation degree to obtain a second solution strategy set; The semantic features of the second feature domain of the current feedback result are paired with the semantic features of the second feature domain of each standard solution strategy in the second solution strategy set to obtain a corresponding second pairing adaptation degree, and the solution strategy search result of the current feedback result is determined based on the second pairing adaptation degree.

7. The method according to claim 6, characterized in that The performing of feature extraction operations on the first semantic description and the second semantic description corresponding to the current feedback result respectively is implemented by calling a strategy feature analysis model, wherein the strategy feature analysis model includes a preliminary semantic feature encoder, a feature aggregation component, and a second feature selection component; The performing feature extraction operations on the first semantic description and the second semantic description corresponding to the current feedback result respectively to obtain the semantic features of the first feature domain and the semantic features of the second feature domain of the current feedback result includes: By means of the preliminary semantic feature encoder, performing preliminary feature extraction operations on the first semantic description and the second semantic description respectively, to obtain a first preliminary semantic feature and a second preliminary semantic feature of the current feedback result; By means of the feature summarization component, a feature summarization operation is performed on the first preliminary semantic feature and the second preliminary semantic feature respectively, so as to obtain a first summarized semantic feature and a second summarized semantic feature corresponding to the current feedback result; Performing a feature selection operation on the first summary semantic feature through the second feature selection component to obtain a semantic feature of the first feature domain corresponding to the current feedback result; Performing a feature compression operation on the second summarized semantic feature to obtain a semantic feature of a second feature domain corresponding to the current feedback result; The strategy feature analysis model further includes a first feature selection component, and the method further includes: Obtain the original strategy feature analysis model and solution strategy instance; Deconstruct the solution strategy instance to obtain a first semantic description and a second semantic description corresponding to the solution strategy instance, and use the first semantic description corresponding to the solution strategy instance as a reference semantic description, the second semantic description corresponding to the solution strategy instance as a target semantic description, and other solution strategy instances as non-target semantic descriptions; Calling the preliminary semantic feature encoder in the original strategy feature analysis model to perform preliminary feature extraction operations on the reference semantic description, the target semantic description and the non-target semantic description, respectively, to obtain corresponding reference semantic description features, target semantic description features and non-target semantic description features; Acquire a first feature distance between the reference semantic description feature and the target semantic description feature, and a second feature distance between the reference semantic description feature and the non-target semantic description feature, and construct a first cost parameter of the original strategy feature analysis model based on the first feature distance and the second feature distance; Selecting features from the preliminary semantic feature encoder in the original strategy feature analysis model based on the first cost parameter to obtain a first strategy feature analysis model corresponding to the original strategy feature analysis model; Fixing the parameters of the preliminary semantic feature encoder in the first strategic feature analysis model, calling the first feature selection component in the first strategic feature analysis model, performing feature selection operations on the reference semantic description feature, the target semantic description feature, and the non-target semantic description feature, respectively, to obtain corresponding reference semantic description selection features, target semantic description selection features, and non-target semantic description selection features; Acquire a third feature distance between the reference semantic description selection feature and the target semantic description selection feature, and a fourth feature distance between the reference semantic description selection feature and the non-target semantic description selection feature, and construct a second cost parameter of the first strategy feature analysis model based on the third feature distance and the fourth feature distance; Selecting features from the first feature selection component in the first strategy feature analysis model based on the second cost parameter to obtain a second strategy feature analysis model corresponding to the first strategy feature analysis model; Fixing the parameters of the preliminary semantic feature encoder in the first strategy feature analysis model and the parameters of the first feature selection component in the first strategy feature analysis model, obtaining a strategy feature instance set, wherein the strategy feature instance set includes feature instances corresponding to a plurality of solution strategy instances and a feature instance target value corresponding to each of the feature instances, wherein the feature instance target value is used to characterize the selected features of the feature selection operation performed on the feature instance by the second feature selection component; Calling the second feature selection component in the second strategy feature analysis model to perform feature selection operations on each of the feature instances to obtain inferred selected features corresponding to each of the feature instances; Determine the feature distance between each of the inferred selected features and the corresponding feature instance target value, and calculate the mean of each of the feature distances to obtain a third cost parameter of the strategy feature analysis model; Based on the third cost parameter, the second feature selection component in the second strategy feature analysis model selects features to obtain a third strategy feature analysis model corresponding to the second strategy feature analysis model, and the third strategy feature analysis model is used as the strategy feature analysis model.

8. The method according to claim 6, characterized in that Acquiring the semantic features of the first feature domain and the semantic features of the second feature domain of each standard solution strategy in the first solution strategy set includes: The following processing is performed on each standard solution strategy in the first solution strategy set: Deconstructing the standard solution strategy to obtain a third semantic description corresponding to the standard solution strategy; Performing a preliminary feature extraction operation on the third semantic description corresponding to the standard solution strategy to obtain a plurality of description feature representations corresponding to the standard solution strategy, and using the plurality of description feature representations as third preliminary semantic features; The performing a feature aggregation operation on the third preliminary semantic feature corresponding to the standard solution strategy to obtain the third aggregated semantic feature corresponding to the standard solution strategy includes: Performing a feature integration operation on the multiple description feature representations to obtain a third summary semantic feature corresponding to the standard solution strategy; Performing a feature aggregation operation on the third preliminary semantic feature corresponding to the standard solution strategy to obtain a third aggregated semantic feature corresponding to the standard solution strategy; Respectively obtaining a first dimensionality reduction feature and a second dimensionality reduction feature for reducing the dimension of the feature domain of the third summary semantic feature; Performing a product operation on the first dimensionality reduction feature and the third summary semantic feature to obtain a first product, and performing a product operation on the second dimensionality reduction feature and the third summary semantic feature to obtain a second product; The first product is mapped to obtain semantic features of a first feature domain corresponding to the standard solution strategy, and the second product is mapped to obtain semantic features of a second feature domain corresponding to the standard solution strategy.

9. A computer device, characterized in that: The computer device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor is used to run the computer program to execute the method according to any one of claims 1 to 8.

Images