A strategy generation method and device, electronic equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, in particular to a strategy generation method and device, electronic equipment and a storage medium, which are used to improve the generation efficiency and accuracy of a risk control strategy. The method comprises the following steps: obtaining risk assessment data to be mined; the risk assessment data comprises at least one of the following: object feature data related to risk assessment of at least one object and environment influence data related to an external environment; a data mining model is constructed based on the risk assessment data, and a model prediction result is obtained; the data mining model is used to identify potential laws related to a risk control strategy from the risk assessment data to perform risk assessment; the model prediction result is subjected to rule processing to generate an executable rule set; the rule set comprises at least one rule for risk assessment; and a risk control strategy for the at least one object is generated according to the rule set.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a strategy generation method, apparatus, electronic device, and storage medium. Background Technology

[0002] In recent years, with the rapid development of internet finance and the explosive growth of behavioral data accumulated online, the volume of online lending business has increased year by year, relying on massive amounts of data. However, many financial institutions currently focus on pre-loan risk analysis and control. After completing the pre-loan assessment process and disbursement, they tend to relax their oversight of the loan process, neglecting the importance of dynamic analysis, leading to a continuous increase in default rates during repayment. To reduce the risk of default during the loan process, real-time monitoring and dynamic adjustments can be implemented through loan strategies to ensure continuous optimization of risk management measures throughout the entire loan cycle.

[0003] The methods for generating loan strategies in related technologies mainly rely on human experience and traditional data analysis methods. Specifically, risk managers in credit institutions typically formulate strategies based on historical experience and simple statistical analysis. For example, they may set a fixed threshold for the number of days of overdue repayment (usually determined based on the mean, standard deviation, etc. of historical data) to decide whether to initiate a reminder program, or classify risk levels based on the customer's credit score and take corresponding measures.

[0004] However, this approach has significant limitations. First, human experience is subjective and cannot adapt to complex and ever-changing market environments and customer behaviors. Second, traditional data analysis methods struggle to handle large-scale data and fail to uncover deep-seated relationships and patterns within the data. Therefore, loan strategies generated based on these methods are often inefficient, have limited accuracy, and cannot effectively mitigate risk.

[0005] In summary, improving the efficiency and accuracy of risk control strategy (such as loan mid-term strategy) generation is an urgent issue that needs to be addressed. Summary of the Invention

[0006] This application provides a strategy generation method, apparatus, electronic device, and storage medium to improve the efficiency and accuracy of risk control strategy generation.

[0007] This application provides a strategy generation method, including:

[0008] Obtain risk assessment data to be mined; the risk assessment data includes at least one of the following categories: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment;

[0009] A data mining model is constructed based on the risk assessment data, and the model prediction results are obtained; the data mining model is used to identify potential patterns related to risk control strategies from the risk assessment data in order to conduct risk assessment.

[0010] The model prediction results are processed into rules to generate an executable rule set; the rule set contains at least one rule for risk assessment.

[0011] Based on the set of rules, a risk control strategy is generated for the at least one object.

[0012] This application provides a strategy generation apparatus, comprising:

[0013] The data acquisition unit is used to acquire risk assessment data to be mined; the risk assessment data includes at least one of the following: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment;

[0014] The data processing unit is used to construct a data mining model based on the risk assessment data and obtain the model prediction results; the data mining model is used to identify potential patterns related to risk control strategies from the risk assessment data in order to conduct risk assessment.

[0015] The rule generation unit is used to process the model prediction results into rules to generate an executable rule set; the rule set includes at least one rule for risk assessment.

[0016] The strategy generation unit is used to generate a risk control strategy for the at least one object based on the set of rules.

[0017] Optionally, the data acquisition unit is specifically used for:

[0018] Historical risk assessment data is obtained from the data warehouse and used as the initial risk assessment data to be mined.

[0019] The data processing unit is specifically used for:

[0020] An initial data mining model is constructed based on the initial risk assessment data to be mined;

[0021] Furthermore, the data acquisition unit is also used for:

[0022] Real-time monitoring of data changes in the data warehouse; when a data update is detected, the new risk assessment data in the data warehouse is used as the risk assessment data to be mined this time;

[0023] The data processing unit is also used for:

[0024] The data mining model is updated based on the risk assessment data to be mined.

[0025] Optionally, the rule generation unit is specifically used for:

[0026] Based on the relationship between the input and output features of the data mining model, at least one key attribute related to risk assessment is extracted, along with a risk threshold corresponding to each key attribute; the input feature is the risk assessment data, and the output feature is the model prediction result.

[0027] The at least one key attribute and its corresponding risk threshold are processed into rules to generate an executable set of rules.

[0028] Optionally, the rule generation unit is specifically used for:

[0029] Based on the relationship between the input and output features of the data mining model, key attributes whose influence on the model's prediction results is higher than a preset influence threshold are identified.

[0030] For each of the identified key attributes, a risk threshold corresponding to the key attribute is obtained by quantitative analysis of the risk assessment data.

[0031] Optionally, before the strategy generation unit generates a risk control strategy for the at least one object based on the rule set, the rule generation unit is further configured to:

[0032] Perform at least one of the following optimization operations on the set of rules:

[0033] The data mining model is evaluated, and the model parameters of the data mining model are adjusted based on the obtained first evaluation result, so as to indirectly adjust the rule set;

[0034] The rule set is evaluated, and rule pruning is performed on the rule set based on the obtained second evaluation result;

[0035] If there are multiple rule sets, then the multiple rule sets are integrated.

[0036] Optionally, the rule generation unit is specifically used for:

[0037] The data mining model is evaluated to obtain a first evaluation result;

[0038] Adjust the model parameters of the data mining model based on the first evaluation result;

[0039] The risk assessment data is input into the adjusted data mining model to obtain new model prediction results;

[0040] The prediction results of the new model are then processed into rules to generate a new set of executable rules.

[0041] Optionally, the rule generation unit is specifically used for:

[0042] Each rule in the rule set is evaluated to obtain a second evaluation result;

[0043] Based on the second evaluation results, identify the negative key attributes in each rule that reduce the accuracy of the model's prediction results;

[0044] By deleting relevant content of the negative key attributes in each rule, the rule set is pruned.

[0045] Optionally, the rule generation unit is specifically used to integrate multiple rule sets in at least one of the following ways:

[0046] For multiple rules that predict the direction in multiple rule sets, the multiple rules are merged;

[0047] For each rule set, a weight is assigned to each rule set based on the third evaluation result corresponding to the rule set, so as to summarize the model prediction results corresponding to the determined object according to the weight of the rule set to which each rule belongs.

[0048] Optionally, the device further includes:

[0049] The optimization unit is configured to perform at least one of the following optimization operations on the risk strategy:

[0050] The risk control strategy is evaluated based on at least one evaluation indicator to obtain a fourth evaluation result; the risk control strategy is then adjusted based on the fourth evaluation result.

[0051] After applying the risk strategy to actual business operations, collect feedback information from the target entities; and adjust the risk strategy based on the feedback information.

[0052] Optionally, the optimization unit is specifically used for:

[0053] The set of rules for generating the risk strategy is adjusted using a preset rule adjustment method;

[0054] Based on the adjusted set of rules, a new risk strategy is generated.

[0055] The rule adjustment method includes at least one of the following:

[0056] Adjust rule parameters, add new rules, and delete invalid rules.

[0057] Optionally, the optimization unit is specifically used for:

[0058] The risk strategy is adjusted using a preset strategy adjustment method;

[0059] The strategy adjustment method includes at least one of the following:

[0060] Adjust strategy parameters, add new strategies, and delete invalid strategies.

[0061] An electronic device provided in this application includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of any of the above-described strategy generation methods.

[0062] This application provides a computer-readable storage medium including a computer program. When the computer program is run on an electronic device, the computer program is used to cause the electronic device to perform the steps of any of the above-described strategy generation methods.

[0063] This application provides a computer program product, which includes a computer program stored in a computer-readable storage medium. When a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any of the above-described strategy generation methods.

[0064] The beneficial effects of this application are as follows:

[0065] This application provides a strategy generation method, apparatus, electronic device, and storage medium. Through systematic data mining and rule-based processing, this application significantly improves multiple aspects of risk assessment and risk control strategies. Specifically:

[0066] First, this application does not require consideration of human experience. After acquiring the risk assessment data to be mined, a data mining model is constructed based on the risk assessment data. The data mining model is then used to automatically generate risk control strategies, reducing the interference of human factors and ensuring the objectivity and scientific nature of the strategies. At the same time, by processing the model's prediction results into rules, an executable set of rules is generated, making the strategies easy to understand and implement, further improving their operability and effectiveness in practical applications. Second, the ability of this application to automatically generate strategies in real time enables the system to quickly respond to changes in the market and customer behavior, greatly improving decision-making efficiency and adapting to the needs of a dynamic environment.

[0067] Furthermore, this application utilizes advanced data mining technology to fully uncover potential patterns in the data, which not only enhances the accuracy of the strategy but also effectively identifies and controls risks, ensuring the security and stability of financial institutions and other business platforms. Moreover, by considering object characteristic data and external environmental impact data, this method ensures that the generated risk control strategy is more accurate and reliable, while also automatically adjusting the strategy according to different customer groups and market conditions, improving the flexibility and adaptability of the strategy.

[0068] In summary, this method not only overcomes the shortcomings of related technologies, but also brings greater intelligence and flexibility to risk management, making it suitable for various industries and application scenarios.

[0069] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0070] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0071] Figure 1 This is a schematic diagram of an application scenario in an embodiment of this application;

[0072] Figure 2 A flowchart illustrating the implementation of a strategy generation method provided in this application embodiment;

[0073] Figure 3 This is a schematic diagram of a decision tree in an embodiment of this application;

[0074] Figure 4 This is a schematic diagram of another decision tree in an embodiment of this application;

[0075] Figure 5 This is a schematic diagram illustrating the construction of a random forest in an embodiment of this application;

[0076] Figure 6 This is a schematic diagram illustrating the calculation logic of the model prediction results of a random forest in an embodiment of this application;

[0077] Figure 7 This is a schematic diagram illustrating an optimization method for a planning set in an embodiment of this application;

[0078] Figure 8 This is a logical diagram illustrating one type of cross-validation in an embodiment of this application;

[0079] Figure 9 This is a schematic diagram of a confusion matrix in an embodiment of this application;

[0080] Figure 10 This is a schematic diagram of a strategy generation system according to an embodiment of this application;

[0081] Figure 11 This is a schematic diagram illustrating an optimization method of a strategy in an embodiment of this application;

[0082] Figure 12 This is a schematic diagram illustrating the specific process of generating a loan strategy in an embodiment of this application;

[0083] Figure 13 This is an example of an interaction logic diagram between a terminal device and a server in an embodiment of this application;

[0084] Figure 14 This is a schematic diagram of the composition structure of a strategy generation device according to an embodiment of this application;

[0085] Figure 15 This is a schematic diagram of the hardware structure of an electronic device using an embodiment of this application. Detailed Implementation

[0086] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0087] The following describes some of the concepts involved in the embodiments of this application.

[0088] 1. Risk Control Strategy: This refers to a set of rules, processes, and measures established by an organization to identify, assess, and manage potential risks. The aim is to minimize the probability of adverse events and their impact, ensuring the safety, stability, and efficiency of business activities. These strategies, through a systematic risk management approach, cover risk prevention, monitoring, response, and recovery, and are applicable to multiple industries and sectors, such as finance, manufacturing, information technology (IT), transportation, energy, construction and engineering, and retail. In each industry and sector, risk control strategies are customized according to specific risk types and business needs; this document does not impose specific limitations on this.

[0089] 2. Loan Management Strategy: This refers to a series of decision-making rules and action plans formulated to effectively manage loan risk and improve loan returns from the time the loan is issued until its repayment. It is a risk control strategy exemplified in this application. These strategies encompass dynamically monitoring customer repayment behavior, credit changes, and market conditions, adjusting credit terms or taking preventative measures in a timely manner to ensure the safety and efficient operation of the loan, thereby reducing default risk and optimizing loan lifecycle management.

[0090] 3. Data Mining: This refers to the process of automatically searching for hidden information in large amounts of data using algorithms. The aim is to discover patterns, trends, and correlations to support decision-making. It utilizes statistical, machine learning, and artificial intelligence techniques to extract valuable knowledge from complex datasets, helping businesses and other entities make more informed decisions.

[0091] 4. Data Mining Model: This refers to the automatic discovery of patterns, trends, and correlations from large amounts of data using algorithms and / or statistical methods, and the transformation of these findings into a structured expression that can be used for prediction or decision-making. It is built using machine learning, statistics, and artificial intelligence techniques based on specific objectives such as classification, regression, clustering, or association rule learning. Data mining models can not only reveal valuable information hidden in data but also predict future events, supporting business decisions. For example, the data mining model in this application embodiment is used to mine potential patterns and relationships related to risk control strategies in input data.

[0092] 5. Object Feature Data: This refers to data that directly reflects the specific activities and attributes of an object within a business platform. This data records in detail the object's behavioral patterns, attribute information, and interactions on the platform, providing a foundation for analysis and decision-making. A business platform refers to a system or environment that provides specific services or functions and collects, processes, and stores user data, such as credit institutions and e-commerce platforms. These platforms generate rich user activity and attribute data through their business processes and service interactions, supporting various application scenarios such as risk assessment, personalized recommendations, and customer service.

[0093] Taking the loan strategy generation scenario as an example, the corresponding object feature data provides a deep understanding of individual customer borrowing and repayment behavior, which is the foundation for assessing credit risk and formulating personalized loan strategies. Accordingly, object feature data can be divided into two main categories: static attribute information and dynamic behavioral records, specifically including but not limited to the following aspects: basic customer information, loan information, repayment records, credit scores, consumption behavior, financial statements and asset information, transaction records, and customer service interaction records.

[0094] 6. Environmental Impact Data: This refers to information from external data sources. This data provides broader contextual information, helps to understand and predict the impact of the external environment on customer behavior and environmental conditions (such as market, weather, etc.), enhances the comprehensiveness and accuracy of data analysis, helps to identify systemic risks and external uncertainties, and supports more robust risk management decisions.

[0095] Taking the loan strategy generation scenario as an example, the relevant environmental impact data includes, but is not limited to, the following aspects: macroeconomic data, industry data, socioeconomic indicators, third-party credit reports, social media and public data.

[0096] 7. Rule-based approach: This refers to transforming the model results obtained from data mining into explicit, actionable rules, enabling the abstract model output to directly guide actual business practices. Through rule-based approach, complex analytical results can be simplified into concrete decision-making logic, enhancing the transparency and consistency of strategy execution, facilitating understanding and application by non-technical personnel, ensuring efficient and accurate implementation in business practices, and simultaneously increasing the application value and impact of the model.

[0097] 8. Automation: The system can automatically execute the entire strategy generation process without human intervention. In this embodiment, by constructing an automated system that requires no human intervention, the entire process from data analysis to strategy generation and implementation is automated, which can improve efficiency, reduce human error, and ensure rapid response to market changes.

[0098] 9. Data Warehouse: A data warehouse is a subject-oriented, integrated, relatively stable collection of data that reflects data changes and is used to support management decisions. The main purpose of a data warehouse is to provide a centralized platform for data analysis and reporting, integrating data from one or more different sources and preprocessing and optimizing it to meet complex query and analysis needs.

[0099] The design concept of the embodiments of this application is briefly introduced below:

[0100] Credit business, also known as credit asset or loan business, is an important asset business of financial institutions. It involves lending money to recover principal and interest, and then making a profit after deducting costs.

[0101] With the development of internet and big data technologies, online lending has flourished. Leveraging the advantages of the internet, customers can complete all steps of the loan application process from the comfort of their homes, including understanding the application requirements for various loans, preparing application materials, and submitting the application—all efficiently done online. While online lending brings great convenience to customers, this contactless lending model also presents financial institutions with significant risks and challenges.

[0102] Risk primarily refers to the negative uncertainty of future outcomes. For financial institutions, risk refers to the possibility of economic losses or negative fluctuations in returns due to various uncertainties in their operations. It can be divided into credit risk and fraud risk. Credit risk, also known as default risk, refers to the possibility that a borrower, for various reasons, is unwilling or unable to fulfill the contractual terms, thus defaulting and causing losses to the financial institution. Fraud risk refers to intentional acts aimed at deliberately distorting or concealing facts to induce a mistaken understanding in the other party; typically, the purpose of such acts is to profit the fraudster.

[0103] From the time a loan is issued until it is fully repaid, financial institutions need to formulate a series of decision-making rules and action plans to effectively manage risks and improve returns; these are known as loan strategies.

[0104] The methods for generating loan mid-term strategies in related technologies primarily rely on human experience and traditional data analysis methods. Typically, risk managers at credit institutions formulate loan mid-term strategies based on historical experience and some simple statistical analysis. For example, they might set a fixed threshold for the number of days overdue payments to determine whether to issue a reminder, or classify customers into different risk levels based on their credit scores and take corresponding risk control measures. In terms of data analysis, basic statistical tools such as the mean and standard deviation might be used to analyze customer repayment behavior and credit status.

[0105] While this approach can meet basic risk control needs to some extent, it has significant limitations. First, relying on human experience to formulate lending strategies is easily influenced by subjective factors, lacking objectivity and scientific rigor. Furthermore, the process is cumbersome and time-consuming, failing to respond promptly to dynamic changes in market conditions and customer behavior. Second, traditional data analysis methods struggle to fully extract information from data, resulting in low strategy accuracy and potentially hindering effective risk identification and control. Finally, fixed strategies are ill-suited to adapting to changes in different customer groups and market conditions, lacking flexibility and adaptability.

[0106] Of course, the same problem may exist in other scenarios besides those mentioned above, such as traffic scenarios.

[0107] In view of this, this application proposes a strategy generation method, apparatus, electronic device, and storage medium. Through systematic data mining and rule-based processing, this application significantly improves multiple aspects of risk assessment and risk control strategies. Specifically, as follows:

[0108] First, this application does not require consideration of human experience. After acquiring the risk assessment data to be mined, a data mining model is constructed based on the risk assessment data. The data mining model is then used to automatically generate risk control strategies, reducing the interference of human factors and ensuring the objectivity and scientific nature of the strategies. At the same time, by processing the model's prediction results into rules, an executable set of rules is generated, making the strategies easy to understand and implement, further improving their operability and effectiveness in practical applications. Second, the ability of this application to automatically generate strategies in real time enables the system to quickly respond to changes in the market and customer behavior, greatly improving decision-making efficiency and adapting to the needs of a dynamic environment.

[0109] Furthermore, this application utilizes advanced data mining technology to fully uncover potential patterns in the data, which not only enhances the accuracy of the strategy but also effectively identifies and controls risks, ensuring the security and stability of financial institutions and other business platforms. Moreover, by considering object characteristic data and external environmental impact data, this method ensures that the generated risk control strategy is more accurate and reliable, while also automatically adjusting the strategy according to different customer groups and market conditions, improving the flexibility and adaptability of the strategy.

[0110] In summary, this method not only overcomes the shortcomings of related technologies, but also brings greater intelligence and flexibility to risk management, making it suitable for various industries and application scenarios.

[0111] The preferred embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit this application. Furthermore, the embodiments and features in the embodiments of this application can be combined with each other without conflict.

[0112] like Figure 1 The diagram shown illustrates an application scenario in an embodiment of this application. The application scenario diagram includes a terminal device 110 and a server 120.

[0113] In this embodiment, the terminal device 110 includes, but is not limited to, mobile phones, tablets, laptops, desktop computers, e-book readers, smart voice interaction devices, smart home appliances, and in-vehicle terminals. The terminal device may have a policy generation-related client installed. This client can be software (e.g., browsers, instant messaging software, e-commerce software), or a webpage, mini-program, etc. The server 120 is a backend server corresponding to the software, webpage, or mini-program, or a server specifically used for policy generation; this application does not impose specific limitations. The server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0114] It should be noted that the strategy generation method in each embodiment of this application can be executed by an electronic device, which can be a terminal device 110 or a server 120. That is, the method can be executed by the terminal device 110 or the server 120 alone, or by the terminal device 110 and the server 120 together.

[0115] In one alternative implementation, the terminal device 110 and the server 120 can communicate via a communication network.

[0116] In one alternative implementation, the communication network is a wired network or a wireless network.

[0117] It should be noted that, Figure 1 The examples shown are merely illustrative; in reality, the number of terminal devices and servers is unlimited and is not specifically limited in the embodiments of this application.

[0118] In this embodiment of the application, when there are multiple servers, the multiple servers can form a blockchain, and the servers are nodes on the blockchain; as disclosed in the strategy generation method of this embodiment, the data involved can be stored on the blockchain, such as risk assessment data, model prediction results, rule sets, risk control strategies, etc.

[0119] Furthermore, the embodiments of this application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, smart transportation, and assisted driving. Specifically, they can be applied to fields related to policy generation within these scenarios. Below are a few specific examples of application extensions:

[0120] For example, in marketing, businesses can utilize the strategy generation methods described in this application to optimize advertising, promotional activities, and customer relationship management, and combine them with risk assessment to ensure the effectiveness and cost-effectiveness of marketing activities. Financial institutions can use the strategy generation methods described in this application for credit scoring, loan approval, portfolio management, and fraud detection, and improve the accuracy and security of decision-making through risk assessment. Manufacturing companies and retailers can use the strategy generation methods described in this application to optimize inventory management, logistics, and supplier selection, and reduce supply chain risks through risk assessment. Energy companies can use the strategy generation methods described in this application for power dispatching, demand-side management, and renewable energy integration, and combine them with risk assessment to ensure the stability and reliability of energy supply.

[0121] For example, cloud computing platforms can use the strategy generation method described in this application to dynamically adjust virtual machine allocation, load balancing, and storage resources to optimize cloud service performance and cost-effectiveness, and combine this with risk assessment to ensure high system availability and data security. Urban traffic management departments can use the strategy generation method described in this application to automatically generate optimal traffic light switching strategies based on real-time traffic flow data, reducing congestion, improving road traffic efficiency, and combining this with risk assessment to prevent traffic accidents and emergencies. Autonomous vehicles can use the strategy generation method described in this application to generate driving decision strategies in real time, such as lane keeping, automatic parking, and emergency braking, ensuring driving safety, and combining this with risk assessment to improve the reliability and response speed of driving decisions. In these application scenarios, by incorporating risk assessment, the effectiveness and reliability of strategies can be further enhanced, ensuring optimal risk control and decision optimization in different business environments.

[0122] By introducing the strategy generation method described in this application, it can be widely applied to multiple fields such as cloud technology, artificial intelligence, intelligent transportation, and assisted driving. Furthermore, it enables dynamic adjustment and optimization of strategies in each specific scenario. This method significantly improves the scientific rigor and accuracy of decision-making, enhances the ability to cope with complex and ever-changing environments, and provides more flexible and efficient solutions, bringing significant progress to risk management, profit improvement, and the optimization of other key business indicators across various industries.

[0123] It should be emphasized that the relevant data involved in the specific embodiments of this application, such as the risk assessment data listed above, are subject to permission or consent from the target user when applied to specific products or technologies. Furthermore, the collection, use, and processing of the relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0124] The strategy generation method provided by the exemplary embodiments of this application will be described below with reference to the accompanying drawings and the application scenarios described above. It should be noted that the application scenarios described above are only shown to facilitate understanding of the spirit and principles of this application, and the embodiments of this application are not limited in any way in this respect.

[0125] See Figure 2 The diagram shown is an implementation flowchart of a strategy generation method provided in this application embodiment. Taking the server as the execution subject as an example, the specific implementation flow of the method is as follows:

[0126] S21: Obtain the risk assessment data to be mined; the risk assessment data includes at least one of the following: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment.

[0127] The generation of risk control strategies (such as loan-trading strategies) requires a large amount of high-quality data as support. The risk assessment data in this application may include object characteristic data related to risk control management (such as loan-trading association) collected from the internal database of the business platform (such as the credit institution) (also known as internal data); in addition, it may also include environmental impact data obtained from external data sources (also known as external data).

[0128] An alternative implementation is that the risk assessment data includes at least object characteristic data and may further include environmental impact data. This approach can provide a basic risk assessment, is highly flexible, and is more suitable for scenarios with limited resources or less dependence on the external environment.

[0129] Another alternative implementation method is to include object characteristic data and environmental impact data in the risk assessment data. This method provides more comprehensive and accurate data, effectively identifies systemic risks, and is more suitable for complex environments that require highly accurate assessment.

[0130] Of course, in another alternative implementation, the risk assessment data includes at least environmental impact data and may further include object characteristic data. This approach prioritizes the impact of the external environment on the market and customer behavior, and is particularly suitable for industries that are highly sensitive to changes in the external environment, ensuring more robust risk management decisions in a complex and volatile environment.

[0131] Among them, object characteristic data refers to data that directly reflects the specific activities and attributes of objects within the business platform. This data records in detail the object's behavioral patterns, attribute information, and interactions on the platform, providing a foundation for analysis and decision-making. Environmental impact data refers to information from external data sources, used to understand and predict the impact of the external environment on customer behavior and the environment itself, supporting more comprehensive risk management decisions.

[0132] For example, in the field of intelligent transportation, object characteristic data includes, but is not limited to, some or all of the following:

[0133] Driver information, such as driver's license number and validity period, driving experience, traffic violation records, accident history, training status, etc.; information about the vehicle driven by the driver, such as license plate number, vehicle type (e.g., sedan, truck, bus), vehicle age, maintenance records, insurance status, etc.; information about the goods transported by the driver, such as type of goods, value of goods, load capacity, dangerous goods label, etc.

[0134] Environmental impact data includes, but is not limited to, some or all of the following:

[0135] Weather conditions: such as temperature, precipitation (rain, snow, etc.), wind speed and direction, visibility, etc.; Road conditions: such as road type (highway, urban road, rural road, etc.), road maintenance status, construction areas, traffic signal settings, etc.; Traffic flow: such as real-time traffic flow, average driving speed, congestion index, etc.; Infrastructure: such as the distribution of public transportation stops, the location of gas stations / charging stations, emergency service facilities (such as hospitals, fire stations, etc.).

[0136] This data can help risk control systems in the transportation sector to more comprehensively understand potential risk factors, thereby taking appropriate measures to prevent or mitigate adverse effects. For example, by analyzing weather conditions and road conditions, potential traffic accident risks can be warned in advance; and by understanding driver behavior, high-risk drivers can be identified and targeted safety training can be provided, and so on.

[0137] Of course, the embodiments of this application are also applicable to other fields. The following mainly takes the loan strategy generation scenario in the financial field as an example, assuming that the business platform is a credit institution:

[0138] Optionally, target characteristic data related to loan management collected from the credit institution's database may include, but is not limited to, some or all of the following:

[0139] Customer basic information, loan information, repayment history, credit score, consumption behavior, financial statements and asset information, transaction records, and customer service interaction records.

[0140] The customer's basic information includes, but is not limited to, demographic information such as age, gender, education background, and marital status. This information helps to depict a more comprehensive picture of the customer, understand their life stage, and potential consumption habits. Loan information includes, but is not limited to, loan amount, term, interest rate, and loan type (e.g., mortgage, auto loan, personal loan), which helps assess the customer's debt level and repayment ability. Repayment records document each repayment, including whether it was made on time, whether there were any overdue payments, and the specific number of overdue days. Credit score is a comprehensive score calculated based on multiple factors (such as payment history, debt burden, and credit history length), which can be used to quantify the likelihood of customer default. Consumption behavior analysis is used to analyze the customer's daily spending patterns, such as monthly fixed expenses, frequency of non-essential purchases, and large-ticket items. This provides insights into the customer's financial management and lifestyle preferences, providing a basis for personalized services.

[0141] Financial statements are essential documents reflecting a company's or individual's financial position, operating results, and cash flows during a specific period. For corporate loans, financial statements include, but are not limited to, the balance sheet, income statement, and cash flow statement. These statements provide information about the company's financial health, profitability, and cash flow situation. Asset information refers to data describing the resources owned by a company or individual and their value. These resources are expected to bring economic benefits to the company or individual and include, but are not limited to, current assets, non-current assets, fixed assets, intangible assets, and investment assets. This information helps assess their repayment ability and default risk.

[0142] Transaction records refer to a detailed historical record of all financial transactions a customer makes at a credit institution or related platform. These records include, but are not limited to: (1) Credit card transactions: the specific time, merchant name, amount, location, and transaction type (e.g., online payment, offline card swiping) of credit card spending. (2) Account transfer records: fund transfers between customer accounts, including the time, amount, and counterparty account information of the transfer in and out. (3) Deposit and withdrawal records: the time, amount, and channel (e.g., bank counter, online banking) of customer depositing or withdrawing funds. (4) Fee and penalty records: fees incurred for various reasons, such as handling fees, late fees, over-limit fees, etc., and the time and amount incurred. (5) Refund and adjustment records: records related to refunds, bill adjustments, dispute resolution, etc. after a transaction. These transaction records help to gain a more comprehensive understanding of the customer's financial dynamics.

[0143] Customer service interaction records refer to the historical records of communication and interaction between customers and credit institutions through various channels. These records may include, but are not limited to: (1) telephone customer service conversations, which may involve loan consultations, repayment arrangements, problem solving, etc.; (2) online chat messages, which may cover loan application progress inquiries, technical support, etc.; (3) email correspondence, which may record loan term confirmations, document submissions, etc.; (4) social media interactions, reflecting customer opinions and feedback; and self-service platform logs. These customer service interaction records help to more comprehensively understand customer behavior.

[0144] It should be emphasized that the acquisition of the aforementioned object characteristic data is done with the permission or consent of the client, and the collection, use, and processing of the relevant data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0145] Optionally, environmental impact data obtained from external data sources may include, but is not limited to, some or all of the following:

[0146] Macroeconomic data, industry data, social data, third-party credit reports, social media and publicly available data.

[0147] Macroeconomic data includes, but is not limited to, trends in macroeconomic indicators such as the Gross Domestic Product (GDP) growth rate, unemployment rate, and inflation rate. These data affect the stability of the overall market and consumers' income expectations, thereby influencing their repayment ability and willingness. Industry data monitors the prosperity of specific industries, such as the real estate market trend and automobile sales volume. If a customer's industry is in a downturn, their employment stability and income level may be affected, increasing the risk of loan default.

[0148] Social data includes, but is not limited to, policy adjustments (such as tax reform and monetary policy) and changes in laws and regulations (such as updates to consumer protection laws); changes in policies and regulations can directly or indirectly affect the operational norms and service scope of credit business.

[0149] Third-party credit reports are detailed credit reports provided by professional credit reporting agencies to supplement the deficiencies of internal credit assessment systems and provide more accurate credit evaluations.

[0150] Social media and public data refer to the analysis of a customer's social media activities or public statements obtained through legal means, used to understand their lifestyle and social network, and to reflect the customer's integrity and attitude towards life.

[0151] In this embodiment, object characteristic data can directly reflect a customer's borrowing and repayment situation, accurately depicting customer behavior and providing a foundation for risk assessment; environmental impact data can enhance the comprehensiveness and accuracy of data analysis. The generation of loan strategies is a complex and meticulous process. It not only relies on data within the credit institution but can also enrich and deepen the understanding of customer behavior by integrating external data sources. Through data mining of this multi-source data, it is possible to more accurately grasp the customer's actual situation and development trends, thereby providing a solid and accurate risk assessment foundation for loan strategy generation, enriching and deepening the understanding of customer behavior, and effectively improving the accuracy and flexibility of risk management.

[0152] Furthermore, it should be noted that the above-mentioned object characteristic data and environmental impact data are just simple examples. In addition, the data information related to the generation of loan strategies is also applicable to the embodiments of this application, and will not be described in detail here.

[0153] In practical applications, the collected risk assessment data often has problems such as noise, missing values ​​and inconsistencies. Therefore, the collected raw data can be preprocessed to improve data quality and model performance.

[0154] Optionally, data preprocessing includes, but is not limited to, the following aspects: data cleaning, data integration, data transformation, and data reduction.

[0155] The following section will first explain data cleaning:

[0156] Data cleaning refers to identifying, correcting, or removing errors, noise, and inconsistencies in data, aiming to improve data quality and remove noisy data and outliers.

[0157] In this application embodiment, the data cleaning method includes, but is not limited to, some or all of the following:

[0158] (1) Statistical methods: Identify and correct outliers and fill in missing values ​​in the data by applying mathematical models and algorithms. (2) Rule filtering: Set rules based on business logic or domain knowledge to exclude data that does not meet the conditions. (3) Deduplication: Identify and remove identical or highly similar data records; missing value handling: fill in or delete records containing missing values. (4) Format standardization: Standardize the data format to ensure consistency.

[0159] Taking the 3σ principle in statistical methods as an example, the 3σ principle is a statistical method based on normal distribution, used to identify and remove outliers. Its core idea is that in a normal distribution, almost all data points (approximately 99.73%) fall within the range of the mean plus or minus three standard deviations; that is, any data point outside this range is considered an outlier. Therefore, the data cleaning method based on the 3σ principle is as follows:

[0160] Suppose the data set is X = {x1, x2, ..., x...} n The mean of the data is Standard deviation is According to the 3σ principle, the data value falls within... Values ​​outside the specified range are considered outliers and can be removed.

[0161] It should be noted that the data cleaning methods listed above are merely simple examples. Other data cleaning methods are also applicable to the embodiments of this application, and will not be elaborated upon here. Furthermore, in practical applications, these data cleaning methods can be used individually or in combination. Combining them can significantly improve data quality, laying a solid foundation for subsequent analysis and modeling.

[0162] The following section explains data integration and data transformation:

[0163] Data integration refers to combining data from different data sources to ensure data consistency and integrity. Data transformation refers to a series of processes and adjustments performed on the original data to ensure that the data format, scale, and distribution meet the requirements of data mining algorithms, thereby improving the performance and accuracy of the model.

[0164] In the embodiments of this application, the data transformation method includes, but is not limited to, some or all of the following:

[0165] (1) Data standardization: By subtracting the mean and dividing by the standard deviation, different features are made to have zero mean and unit variance. (2) Data normalization: Numerical features are scaled to a specific range (e.g., 0 to 1) to ensure that features of different magnitudes have equal importance in the analysis. (3) Discretization: Continuous variables are converted into discrete variables to facilitate certain types of analysis or simplify model complexity. (4) Encoding: Categorical variables are converted into numerical forms, such as one-hot encoding or multi-class label encoding, so that the algorithm can handle non-numerical data. (5) Feature construction: New features are created based on existing features to enhance the expressiveness of the model, such as interaction terms, multinomial features, etc. (6) Missing value handling: Missing values ​​are filled or deleted using appropriate methods to ensure data integrity.

[0166] It should be noted that the data transformation methods listed above are merely simple examples. Other data transformation methods are also applicable to the embodiments of this application, and will not be elaborated upon here. Furthermore, in practical applications, these data transformation methods can be used individually or in combination. When used in combination, they can significantly improve data quality, laying a solid foundation for subsequent analysis and modeling.

[0167] Finally, the data specification is explained:

[0168] Data reduction refers to reducing the size and dimensions of data without compromising data integrity, thereby improving data processing efficiency. This not only helps to speed up data analysis but also reduces storage costs and makes data easier to manage and understand.

[0169] In this application embodiment, the data reduction method includes, but is not limited to, some or all of the following:

[0170] (1) Dimensionality Reduction: Reduce data dimensionality using techniques such as Principal Component Analysis (PCA), feature selection, and feature extraction while retaining key information. (2) Numerical Reduction: Simplify datasets and reduce data volume using parametric methods (such as regression and discretization) and non-parametric methods (such as histograms, clustering, and sampling). (3) Data Compression: Reduce data storage space using lossy or lossless methods. Lossy compression sacrifices some accuracy to improve the compression ratio, while lossless compression ensures that the decompressed data is exactly the same as the original data. (4) Attribute Construction: Generate new synthetic attributes based on existing attributes to more effectively represent the essential characteristics of the data and improve the model's expressiveness.

[0171] In this embodiment, data reduction can help solve the curse of dimensionality problem caused by high-dimensional data, enabling algorithms to converge faster and improving model performance. Furthermore, when dealing with large-scale data, data reduction helps simplify computational complexity, making the processing more efficient. By appropriately applying the above data reduction methods, the size and dimensionality of data can be significantly reduced without affecting data integrity, thereby optimizing the data processing flow and improving analysis efficiency.

[0172] It should be noted that the data preprocessing methods listed above are just simple examples. In addition, other data preprocessing methods are also applicable to the embodiments of this application. For example, data labeling refers to assigning a label or category (such as risk level) to each sample in the dataset; another example is data balancing, which is used to deal with the class imbalance problem. Based on oversampling of the minority class, undersampling of the majority class, or generating synthetic samples, it ensures that the sample size between different classes is relatively balanced, thereby improving the generalization ability of the model; and so on. These will not be elaborated on in this article.

[0173] In this embodiment of the application, the preprocessed data can be stored in a dedicated data warehouse for subsequent data mining and analysis.

[0174] S22: Build a data mining model based on risk assessment data and obtain the model prediction results; the data mining model is used to identify potential patterns related to risk control strategies from risk assessment data in order to conduct risk assessment.

[0175] In this application embodiment, building a data mining model based on risk assessment data specifically refers to using data-driven modeling or using data to update the model.

[0176] In building the data mining model, risk assessment data is used as input, forming the foundation for model training. During model construction, key features (such as credit scores and debt ratios) are extracted and selected from the risk assessment data, followed by necessary data preprocessing and feature engineering. The model is then trained using the risk assessment data. Furthermore, methods such as cross-validation can be used to ensure the model's effectiveness and robustness. The final data mining model can identify potential patterns in the risk assessment data, providing support for risk control strategies.

[0177] In this application embodiment, whether data-driven modeling or data-updated modeling is used depends on whether a relevant data mining model existed previously. These situations are explained below:

[0178] In one alternative implementation:

[0179] When obtaining risk assessment data to be mined in S21, the process includes: obtaining historical risk assessment data from the data warehouse and using the historical risk assessment data as the initial risk assessment data to be mined.

[0180] In this embodiment, the data warehouse not only stores current data but also retains historical data, recording how the data changes over time. For example, in a loan strategy generation scenario, the data warehouse can contain data extracted and integrated from different business platforms (such as loan management systems, credit scoring systems, customer relationship management systems, etc.). Furthermore, the raw data can be preprocessed to ensure data quality and consistency.

[0181] In practical applications, historical data for a certain period of time can be obtained from the data warehouse as historical risk assessment data, such as the past year, the past quarter, the past month, the past week, etc. The specific data can be determined according to actual needs, and this article does not make any specific restrictions.

[0182] If the historical risk assessment data (hereinafter referred to as historical data) obtained from the data warehouse has undergone data preprocessing, the subsequent processes can be executed directly; if the historical risk assessment data obtained from the data warehouse has not undergone data preprocessing, it can be preprocessed according to the methods listed above before executing the subsequent processes to improve efficiency and accuracy.

[0183] Accordingly, when constructing a data mining model based on risk assessment data in S22, it includes: constructing an initial data mining model based on the initial risk assessment data to be mined.

[0184] Specifically, appropriate data mining models can be selected based on business needs, such as decision trees, random forests, support vector machines (SVMs), and neural networks. Then, historical data is used to train the model to ensure it can capture long-term trends and patterns. It should be noted that this is only a brief overview; the specific implementation methods for building data mining models based on risk assessment data in this step will be explained in detail below and will not be repeated here.

[0185] In another alternative implementation:

[0186] When acquiring risk assessment data to be mined in S21, it also includes: real-time detection of data changes in the data warehouse; when data updates are detected, the new risk assessment data in the data warehouse is used as the risk assessment data to be mined this time.

[0187] In practical applications, as business environments, market demands, and external conditions change, data in data warehouses are updated periodically or in real-time to ensure accuracy and timeliness. Below are some examples of data warehouse update mechanisms and their application scenarios:

[0188] (1) Updated at a fixed period.

[0189] Specifically, this refers to batch updating of data in a data warehouse at predetermined time intervals (such as daily, weekly, or monthly). For example, financial institutions can extract the latest loan applications, repayment records, and other data from multiple business platforms daily or monthly to update the historical risk assessment data in the data warehouse.

[0190] Taking daily updates as an example, many financial institutions update their data every day in the early morning, such as between 0:00 and 2:00. This is because after the day's business concludes, various transaction data and repayment records have stabilized. For example, in consumer loan business, customers continuously make purchases and repayments during the day, and these records are temporarily stored in the business platform's temporary database. Then, at night, the system batch-updates every new purchase detail, repayment amount, and overdue status to the data warehouse, ensuring that the data warehouse contains complete and up-to-date customer spending and repayment dynamics.

[0191] Taking monthly updates as an example, the end of each month is a common time for data updates. For instance, a customer's monthly earnings data may be related to their payroll cycle, as many companies concentrate salary payments at the end of the month. Similarly, lending institutions will obtain the latest customer earnings data from partner companies on the last day of each month or the first day of the following month and update it in their data warehouse. Furthermore, for some long-term loan products, monthly financial statements and asset-liability information are also updated at this time to more accurately assess a customer's repayment ability.

[0192] The aforementioned fixed-period update method ensures the stable operation of the data warehouse and avoids performance impacts from frequent data changes. Furthermore, setting fixed update times simplifies data management and maintenance.

[0193] (2) Event-triggered update.

[0194] Specifically, this means that when a specific event occurs, the data warehouse is immediately updated. For example, when a borrower experiences a major credit event (such as overdue payments or default), the relevant records in the data warehouse are immediately updated so that risk control strategies can be adjusted in a timely manner.

[0195] Taking major transaction events as an example, when a customer makes a significant financial transaction, the data is updated immediately. For instance, if a customer applies for a large mortgage, this application instantly triggers multiple data update processes. On one hand, the customer's debt information is updated immediately, adding detailed data such as the mortgage amount, term, and repayment plan; on the other hand, the associated credit assessment data also changes in real time, because large debts may affect their overall credit risk. This updated data is immediately transmitted to the data warehouse for subsequent strategy adjustments.

[0196] Taking late payment triggering as an example, once a customer defaults on a payment, the relevant data is updated immediately. The system detects the default event and, while updating the default record, also updates related derivative data concerning the customer's credit score, feeding this data back to the data warehouse in real time. Subsequently, the automated loan strategy generation system, based on this latest data, quickly decides whether to initiate a reminder process (such as a repayment reminder) or adjust the credit limit to ensure the safety of loan funds.

[0197] The event-triggered update method described above enables rapid response to critical events, ensuring the timeliness and accuracy of data. Furthermore, this method only updates data relevant to the event, reducing unnecessary processing overhead.

[0198] (3) External data update synchronization.

[0199] Specifically, this refers to maintaining synchronization with external data sources, acquiring the latest external data in real time or periodically, and updating the data warehouse accordingly. For example, synchronizing with external data sources such as credit rating agencies and macroeconomic databases to obtain the latest information on credit scores, interest rate changes, and market dynamics in real time, and updating the environmental impact data in the data warehouse.

[0200] Taking macroeconomic data as an example, macroeconomic data providers regularly release new data, such as when banks adjust interest rates or release inflation data. Financial institutions' data warehouses will update this external information simultaneously. Because the macroeconomic environment has a potential impact on customers' repayment ability and default risk, credit institutions need to incorporate this new data to reassess the risk status of existing loan customers and adjust their lending strategies.

[0201] Taking third-party credit data as an example, every so often, lending institutions obtain the latest credit report data of customers from third-party credit reporting platforms and update it to their own data warehouse. If the third-party credit reporting agency discovers that a customer has added new credit card information, after synchronizing the data, financial institutions can use this information to improve the characteristics of the target customer and optimize their loan management strategies.

[0202] The aforementioned methods of synchronizing external data ensure that the data warehouse contains the latest internal and external information, providing a more comprehensive analytical foundation. Furthermore, it can promptly reflect changes in the external environment, enhancing the scientific rigor and foresight of decision-making.

[0203] It should be noted that the above-listed update methods are just simple examples. Other update methods are also applicable to the embodiments of this application, and will not be described in detail here.

[0204] In this embodiment, detection tools can be used to monitor data changes in the data warehouse in real time, ensuring timely responses to the latest business and market dynamics. For example, triggers can be configured in the source system of the data warehouse to automatically trigger event notifications when new data is inserted, updated, or deleted. Another example is capturing data changes by monitoring changes in log files, suitable for systems that do not support direct triggers. Loading data and setting short time intervals (e.g., every minute) allows for periodic queries of key tables in the data warehouse to check for new records or updates.

[0205] For newly captured risk assessment data, the data preprocessing methods listed above can be used to preprocess the data before proceeding with subsequent processes to improve efficiency and accuracy.

[0206] Correspondingly, when building a data mining model based on risk assessment data in S22, it also includes updating the data mining model according to the risk assessment data to be mined this time.

[0207] In this embodiment of the application, when retraining the data mining model based on new risk assessment data, an incremental learning method can be used to learn only from the new risk assessment data, rather than retraining the entire data mining model, in order to improve efficiency.

[0208] Let the initial data mining model be M0, and the new risk assessment data be D. new Then the data mining model after incremental learning can be represented as: M = update(M0, D new ).

[0209] Here, update(.,.) is the incremental learning function, which updates the initial data mining model based on the new risk assessment data.

[0210] The first implementation method described above relies on a large amount of historical data, ensuring high model stability and reliability. Furthermore, based on detailed historical records, it accurately captures long-term trends and patterns. The real-time update mechanism in the second implementation method ensures the model always reflects the latest market and target object (e.g., borrowers, merchants) conditions, improving its ability to respond to sudden changes. It also enables rapid response to macroeconomic changes, policy adjustments, and market dynamics, maintaining the model's flexibility and adaptability. Moreover, combining the latest internal and external data improves the accuracy and refinement of risk assessment, helping to identify systemic risks and external uncertainties. The combined application of these two methods allows for a more comprehensive and accurate risk assessment, optimizing risk control strategies (such as loan processing strategies).

[0211] The following explains how to build a data mining model based on risk assessment data:

[0212] Specifically, based on data preprocessing, a data mining model is constructed, and the model's prediction results (or simply prediction results) are used to uncover potential patterns and relationships within the data. The data mining model is explained below:

[0213] In this application embodiment, the data mining model includes, but is not limited to, any one or more of the following:

[0214] Decision trees, random forests, SVMs, neural networks, regression models, logistic regression, gradient boosting decision trees (GBDT), association rule learning, and Naive Bayes.

[0215] Taking the loan strategy generation scenario as an example, by building a data mining model, we can uncover potential patterns and relationships related to loan strategies, such as customer repayment behavior and credit risk, providing a basis for subsequent strategy generation.

[0216] The following is a brief explanation of how to construct a decision tree:

[0217] A decision tree is a classification and regression method based on a tree structure. It constructs a decision tree by recursively partitioning the data, where each internal node represents a test of an attribute, each branch represents a test result, and each leaf node represents a category or predicted value.

[0218] Attributes are data items used to describe the characteristics or variables of an object / environment in a specific business scenario. These attributes can be numerical, categorical, or textual, and can help decision-makers understand data more accurately and formulate corresponding strategies.

[0219] For example, in the scenario of loan strategy generation, the attributes include, but are not limited to, some or all of the following:

[0220] (1) Credit Score: An important indicator for measuring a borrower's credit risk. A high credit score usually means a lower risk of default. (2) Return Level: A borrower's monthly or annual return reflects their repayment ability. A higher return level usually means a stronger repayment ability. (3) Historical Repayment Behavior: A borrower's past repayment records (such as whether they made timely payments, number of overdue payments, etc.) directly reflect their credit status and willingness to repay. (4) Debt Ratio: The ratio of a borrower's total debt to their income, used to assess their financial health. (5) Asset Holdings: Various assets owned by the borrower, such as real estate, vehicles, investments, etc., used to assess their wealth level and repayment ability. (6) Loan Purpose: The specific purpose for which the borrower applies for the loan, such as purchasing a house, a car, education, consumption, etc., used to assess the purpose and rationality of the loan.

[0221] For example, in the context of intelligent transportation, attributes include, but are not limited to, some or all of the following:

[0222] (1) Weather conditions: Affect road safety and traffic flow, and are important factors in assessing accident risk and planning travel. (2) Traffic flow: Helps predict potential traffic bottlenecks and accident-prone areas, and guides dynamic traffic restriction rules and road condition warnings. (3) Driver behavior characteristics: Helps identify high-risk driving behaviors and provides a basis for personalized safety reminders and driver training. (4) Vehicle status: Can be used to assess vehicle safety and reliability, and assist in developing targeted risk control measures, such as strengthening the inspection of specific vehicles or restricting their driving areas.

[0223] These attributes can help build a more intelligent and responsive decision tree model, thereby effectively supporting risk management and optimized scheduling in intelligent transportation.

[0224] It should be noted that the scenarios and attributes listed above are just simple examples. Other scenarios and attributes are also applicable to the embodiments of this application, and will not be described in detail here.

[0225] Correspondingly, the test results are judgments or classifications of attribute values, while categories or predicted values ​​are conclusions or predictions drawn based on the test results.

[0226] The following example illustrates the loan strategy generation scenario:

[0227] In the scenario of loan strategy generation, the attribute corresponding to a certain node may be a credit score. This attribute is a numerical value that comprehensively reflects the customer's credit status. It can be given by professional credit rating agencies or internal algorithms of financial institutions, and the value range may be 300-850 points.

[0228] Correspondingly, the test results may be "above 600 points" or "below 600 points," which can initially screen out two customer groups with better and worse credit.

[0229] Correspondingly, the categories can include, but are not limited to: "low-risk customers," "medium-risk customers," and "high-risk customers," which can be denoted as the category set C = {low risk, medium risk, high risk}. This classification helps financial institutions decide whether to lend and how much to lend. For example, customers with high credit scores, stable income, and low debt are likely to be classified as low-risk customers.

[0230] Correspondingly, the predicted value may represent the probability of a customer defaulting on future payments, with a value between 0 and 1. For example, a leaf node outputting a predicted value of 0.1 means that the customer has a 10% chance of defaulting. Financial institutions can use this value to fine-tune their subsequent loan management strategies, such as whether to remind customers to repay in advance.

[0231] For example, the attribute corresponding to a certain node might be the borrower's monthly or annual income, reflecting their repayment ability.

[0232] Correspondingly, the test results can be "monthly income greater than 10,000 yuan" and "monthly income less than 10,000 yuan" to distinguish between high-income and low-income customers.

[0233] Accordingly, the categories can include, but are not limited to: "high-yield customers", "medium-yield customers", and "low-yield customers", which can be denoted as the category set C = {high-yield, medium-yield, low-yield}. For example, customers with high monthly returns are generally considered to be of lower risk.

[0234] Correspondingly, the predicted value may represent the probability of a customer's default, with a value between 0 and 1. For example, a leaf node outputting a predicted value of 0.05 means that the customer is predicted to have a 5% chance of defaulting, and financial institutions can adjust credit limits or approval standards accordingly.

[0235] For example, the attribute corresponding to a certain node may be the borrower's past repayment records (such as whether repayments were made on time, number of overdue payments, etc.), which directly reflects their credit status and willingness to repay.

[0236] Correspondingly, the test results can be "no overdue payments in the past 12 months" or "one or more overdue payments in the past 12 months".

[0237] Accordingly, the categories can include, but are not limited to: "customers with good repayment records", "customers with average repayment records", and "customers with poor repayment records", which can be denoted as the category set C = {good, average}. For example, customers who have repeatedly defaulted on payments in the past 12 months may be classified as high-risk customers.

[0238] Correspondingly, the predicted value may represent the probability of a customer's future default, with a value between 0 and 1. For example, a leaf node outputting a predicted value of 0.2 means that the customer is predicted to have a 20% chance of defaulting, and the financial institution may take more stringent monitoring measures.

[0239] It should be noted that the decision tree-related attributes, test results, categories / predicted values, etc. listed above are just simple examples, and this article does not make specific limitations on them.

[0240] Optionally, during the construction of a decision tree, an optimal splitting attribute can be selected to maximize information gain or minimize impurity. This involves finding the attribute that most effectively divides the dataset into multiple purer subsets. By selecting the optimal splitting attribute, the decision tree can achieve the best data split at each node, thereby improving the model's predictive performance and interpretability.

[0241] In simple terms, after obtaining the risk assessment data, the decision tree construction process is as follows:

[0242] Step 1: Calculate the impurity of the root node.

[0243] Optional methods for measuring impurity include, but are not limited to: entropy (corresponding to information gain), Gini coefficient, classification error rate, chi-square test, gain ratio, etc.

[0244] The original dataset consists of risk assessment data to be mined.

[0245] The following example demonstrates how to calculate the impurity of the original dataset using entropy or the Gini index:

[0246] Step 2: Calculate the partitioning effect of each attribute.

[0247] Specifically, for each possible partitioning attribute, calculate its information gain (if entropy is used as a measure of impurity) or Gini index reduction (if Gini index is used as a measure of impurity).

[0248] Step 3: Select the best splitting attribute.

[0249] Specifically, the information gain or Gini index reduction of each attribute is compared, and the attribute with the largest information gain or the smallest Gini index is selected as the best splitting attribute for the current node.

[0250] Step 4: Recursively perform the same operation on the leaf nodes.

[0251] Specifically, for each newly created leaf node, repeat the above process until the stopping conditions are met, such as reaching the preset maximum depth, the number of samples in the leaf node being less than a certain threshold, all samples belonging to the same category, or no more attributes being available.

[0252] The following is a brief explanation of the construction of decision trees in the scenario of loan strategy generation, using information gain as an example:

[0253] Information gain is a concept based on entropy, used to measure the amount of information lost when a dataset is partitioned by a certain attribute. Entropy measures the degree of uncertainty or disorder in a system; in a perfectly deterministic system, entropy is 0. For classification problems, information gain can be used to evaluate the effectiveness of a given attribute in reducing uncertainty about class labels.

[0254] Assumptions: The dataset D, consisting of risk assessment data to be mined, contains several samples (the exact number depends on the actual situation), and the corresponding category set of this dataset is C = {c1, c2, ..., c...}. j The sample belongs to category c. i The probability is in This indicates that data in dataset D belongs to category c. i The sample set, express The proportion of the sample size in dataset D to the total number of samples in dataset D, therefore the probability p i Specifically, category c i The proportion of data in dataset D.

[0255] Therefore, the information entropy H(D) of dataset D is calculated using the following formula 1:

[0256]

[0257] Assumption: For attribute A, the dataset D is divided into n data subsets D1, D2, ..., Dn. n .

[0258] Therefore, the information gain of attribute A, Gain(A), is calculated using the following formula 2:

[0259]

[0260] in, Represents each data subset D i The proportion of the number of samples in dataset D to the total number of samples in dataset D.

[0261] Referring to Formulas 1 and 2 above, select the attribute with the largest information gain as the splitting attribute, repeat this process until the stopping condition is met, and a decision tree can be constructed.

[0262] By constructing the decision tree in the above manner, it is ensured that the impurity of the decision tree can be reduced as much as possible at each layer, thereby improving classification performance and generating a more accurate data mining model.

[0263] like Figure 3 As shown, this is a schematic diagram of a decision tree in an embodiment of this application. The structure of the decision tree is described as follows:

[0264] Figure 3 The decision tree shown contains three internal nodes, with internal node 1 serving as the root node, used for initial screening of customer creditworthiness. The root node's question is whether the customer's credit score is greater than 700. This root node has two branches: root node branch 1, corresponding to the answer "yes"; and root node branch 2, corresponding to the answer "no".

[0265] For customers with a credit score above 700 (i.e., following the "Yes" branch), further assessment is conducted through internal node 2 (beyond root node branch 1) to determine if they have any overdue payment records in the past 6 months. This internal node 2 also has two branches: Internal node 2 branch 1: If there are no overdue payment records, the customer is classified as low risk; Internal node 2 branch 2: If there are overdue payment records, the customer is classified as medium risk.

[0266] For customers with a credit score of no more than 700 (i.e., those on the "No" branch), internal node 3 (related to root node branch 2) checks whether their current outstanding loan balance exceeds 50,000 yuan. This internal node 3 also has two branches: Internal node 3 branch 1: If the loan balance exceeds 50,000 yuan, the customer is classified as high-risk; Internal node 3 branch 2: If the loan balance does not exceed 50,000 yuan, the customer is also classified as medium-risk.

[0267] In summary, this decision tree starts from the root node, initially classifying customers based on their credit scores. Then, through two further internal nodes, it assesses the customer's repayment history and loan balance, ultimately categorizing customers as low-risk, medium-risk, or high-risk. This structure systematically assesses each customer's credit risk and makes corresponding loan management decisions.

[0268] In practical applications, different decision trees can be constructed by utilizing information from different dimensions.

[0269] like Figure 4 As shown, this is a schematic diagram of another decision tree in an embodiment of this application. The structure of this decision tree is described as follows:

[0270] Figure 4 The decision tree shown contains nine internal nodes, with internal node 1 serving as the root node, used for initial screening of customer social activity levels. The root node's question asks whether the customer's social activity level is 10 times per day. This root node has two branches: root node branch 1, corresponding to the answer "yes"; and root node branch 2, corresponding to the answer "no".

[0271] For customers with more than 10 social interactions per day (i.e., following the "Yes" branch), their online spending frequency is further assessed via internal node 2 (related to root node branch 1) to determine if it exceeds 5 times per month. Internal node 2 also has two branches: Internal node 2 branch 1: If it exceeds, their credit score is further assessed via internal node 4 to determine if it exceeds 650. If yes, they are classified as low-risk. Otherwise, their income level is assessed via internal node 8 to determine if it exceeds 8000 yuan per month. If yes, they are classified as medium-risk; otherwise, they are classified as high-risk. Internal node 2 branch 2: If it does not exceed, their credit score is further assessed via internal node 5 to determine if it exceeds 700. If yes, they are classified as low-risk; otherwise, they are classified as high-risk.

[0272] For customers whose social activity level is less than 10 times / day (i.e., those on the "No" branch), their online spending frequency is further assessed through internal node 3 (related to root node branch 2) to see if it exceeds 3 times / month. This internal node 3 also has two branches: Internal node 3 branch 1: If it exceeds, then internal node 6 is used to assess whether their credit score exceeds 700. If so, they are classified as low-risk. Otherwise, internal node 9 is used to assess whether their income level exceeds 10,000 yuan / month. If so, they are classified as medium-risk; otherwise, they are classified as high-risk. Internal node 3 branch 2: If it does not exceed, then internal node 7 is used to assess whether their credit score exceeds 600. If so, they are classified as medium-risk; otherwise, they are classified as high-risk.

[0273] It should be noted that the above Figure 3 or Figure 4 The decision trees listed are just simple examples; this article does not impose any specific limitations on them.

[0274] In this way, decision trees can progressively refine customer classification based on multiple attributes in loan strategy generation scenarios, ultimately determining the customer's risk level at the leaf nodes. Furthermore, this method can intuitively display the decision logic, helping to identify high-risk customers, dynamically adjust loan strategies, and improve the accuracy and transparency of risk management. Of course, other scenarios follow the same principle, which will not be elaborated upon here.

[0275] The following is a brief explanation of how to construct a random forest:

[0276] Random forests are an ensemble learning method consisting of multiple decision trees. By training multiple decision trees on different subsamples and using voting or averaging to make the final prediction, the accuracy and robustness of the model are improved.

[0277] In short, after obtaining the risk assessment data, the construction process of the random forest is as follows:

[0278] Step 1: Randomly select multiple data subsets (i.e., sample subsets) with replacement from the original dataset. The original dataset consists of risk assessment data to be mined.

[0279] Step 2: For each subset of data, randomly select a portion of attributes and construct a decision tree.

[0280] Step 3: Repeat the above steps to construct multiple decision trees.

[0281] Step 4: For a new sample, let each decision tree make a prediction, and then use a voting method or an averaging method to determine the final model prediction result.

[0282] like Figure 5 As shown, this is a schematic diagram illustrating the construction of a random forest in an embodiment of this application. Figure 5 As shown, following step 1 above, m data subsets can be randomly selected from the original dataset, denoted as: data subset 1, data subset 2, data subset 3, ..., data subset m. Following steps 2 and 3 above, a decision tree can be constructed for each data subset, as shown... Figure 5 The decision trees shown are 1, 2, 3, ..., m, which together form a random forest.

[0283] Based on this, for each new sample, the model prediction result corresponding to that sample can be obtained by following step 4 above.

[0284] like Figure 6The diagram shown illustrates the calculation logic of a random forest model prediction result in an embodiment of this application. For a new sample x, the prediction result of decision tree 1 is y1, the prediction result of decision tree 2 is y2, the prediction result of decision tree 3 is y3, ..., the prediction result of decision tree m is ym. These m prediction results are calculated using either a voting method (as shown in Formula 3 below) or an averaging method (as shown in Formula 5 below) to obtain the model prediction result output by the random forest.

[0285] The following is a brief explanation of the voting method:

[0286] Assumptions: There are m decision trees in the random forest. For a sample x, the prediction result of the i-th decision tree is y. i (representing a category), the prediction result of the random forest can be obtained by voting, calculated as shown in Formula 3 below:

[0287]

[0288] Here, I(·) is an indicator function; I(·) is 1 when the condition within the parentheses is true, and 0 otherwise. argmax is used to synthesize the prediction results of multiple decision trees to obtain a final model prediction result corresponding to this random forest.

[0289] In the scenario of loan strategy generation, if it is necessary to determine whether a customer will default on repayment, then the category set formed by the categories corresponding to the decision tree is C = {0, 1}, where c is an element in the category set C, which is all possible categories, i.e., 0 or 1.

[0290] y i This indicates the prediction result of the i-th decision tree for sample x, i.e. whether it is overdue. This sample x corresponds to the relevant loan data of a specific customer.

[0291] It is important to note that in this random forest, the prediction result y of the m decision trees has the same meaning, all of which give a predicted value about the target variable (whether it is overdue) for the same sample x. However, the training data subset and the selected feature subset of each decision tree are different, so the specific predicted value may be different.

[0292] When checking y i Whether c equals 1 means checking if the prediction given by the i-th decision tree for sample x matches a specific category in the category set C. For example, if c = 1, it means checking if the decision tree predicts that the customer will default. If y i =1 means that this decision tree predicts that the customer will default; if y i =0, then it is predicted that the customer will not default.

[0293] Suppose a random forest contains 5 decision trees (i.e., m=5). When classifying a sample x, the prediction results of these 5 trees are as follows:

[0294] Tree 1: Prediction result y1 = 1; Tree 2: Prediction result y2 = 0; Tree 3: Prediction result y3 = 1; Tree 4: Prediction result y4 = 0; Tree 5: Prediction result y5 = 1.

[0295] The number of votes for each category can then be calculated using the following formula 4:

[0296]

[0297] The number of votes for category 1: count(1) = I(y1=1) + I(y2=1) + I(y3=1) + I(y4=1) + I(y5=1) = 3;

[0298] The number of votes for category 0: count(0) = I(y1=0) + I(y2=0) + I(y3=0) + I(y4=0) + I(y5=0) = 2;

[0299] Based on the above calculations, Category 1 received the most votes, therefore the final prediction result is Category 1, which means that the repayment will be overdue.

[0300] In this way, random forests utilize the collective wisdom of multiple decision trees to select the most likely category as the final prediction result through majority voting, thereby improving the accuracy and robustness of classification.

[0301] In lending, the final forecast result is used to quickly assess a customer's repayment risk. If a customer is predicted to default (as in the example above where the final forecast result is 1), the lending institution may decide to temporarily suspend loan disbursements, or tighten loan limits for customers who have already received loans, strengthen loan monitoring, and prepare reminder measures in advance, thereby reducing the risk of bad debts and ensuring the safety of funds.

[0302] The following is a brief explanation of the averaging method:

[0303] Assumptions: There are m decision trees in the random forest. For a sample x, the prediction result of the i-th decision tree is y. i (representing a probability value), the prediction result of the random forest can be obtained by averaging, calculated as shown in Formula 5 below:

[0304]

[0305] In other words, in the scenario of loan strategy generation, random forests train multiple decision trees and integrate their predictions, using voting or averaging to make the final decision. This method not only improves the model's accuracy and robustness but also handles high-dimensional data and complex relationships, effectively identifying potential high-risk customers. Random forests help dynamically adjust loan strategies, optimize risk management, and improve loan returns and security. Of course, the same principle applies to other scenarios, which will not be elaborated upon here.

[0306] The following is a brief explanation of SVM:

[0307] The core objective of SVM is to find an optimal hyperplane that can distinguish different categories of data as clearly as possible in the feature space.

[0308] In loan strategy generation scenarios, SVM can accurately distinguish between high- and low-risk customers by finding the optimal hyperplane to maximize the margin between different risk categories. For example, loan data consists of features such as customer income, credit score, and past repayment history. SVM uses kernel functions (such as linear kernels and Gaussian kernels) to map the low-dimensional original feature space to a high-dimensional space, making the originally linearly inseparable data linearly separable.

[0309] When conducting data mining, SVM focuses on the boundary conditions of the data, finding the points closest to the boundary between the two classes of samples—the support vectors—to determine the hyperplane that best separates different repayment behaviors and credit risk categories. For example, separating the data of customers who repay on time and those who repay late allows us to discover which feature combinations are most critical in distinguishing between the two types of customers. For instance, customers with high credit scores and stable incomes are more likely to repay on time, revealing a potential pattern.

[0310] Based on the above, and according to the different categories identified, a correlation can be established between the characteristics of the input data and the loan strategy. For example, for customers whose data falls on the overdue payment side, the subsequent loan strategy might be to increase the frequency of reminders and reduce the credit limit; while for customers on the other side, regular service can be maintained.

[0311] The following is a brief explanation of neural networks:

[0312] Neural networks mimic the workings of neurons in the human brain. They consist of multiple layers of neurons (such as input layers, hidden layers, and output layers) and are trained by adjusting connection weights. They can automatically learn complex nonlinear relationships in the input data.

[0313] In this application embodiment, neural networks can be used to uncover complex nonlinear relationships, predict customer behavior, and provide support for risk management.

[0314] In loan strategy generation scenarios, neural networks can capture complex nonlinear relationships in customer data through a multi-layered structure (input layer, hidden layer, and output layer). For example, the input layer receives feature data such as customer age and loan amount, and performs nonlinear transformations and weight adjustments through the activation function of the hidden layer to continuously fit the patterns in the data. The activation function can be a Rectified Linear Unit (ReLU), Sigmoid, etc., and this paper does not specify a particular one.

[0315] In data mining, neural networks, with their massive parameters and multi-layered architecture, can capture extremely subtle and high-order feature interactions. For example, Recurrent Neural Networks (RNNs) or Long Short-Term Memory Networks (LSTMs) excel at processing sequential data, enabling them to monitor a customer's repayment sequence over a period of time and discover patterns that are difficult to summarize using simple rules, such as periodic delinquencies and gradual changes in repayment ability.

[0316] The trained neural network outputs a model prediction, which is correlated with the customer's credit risk rating, repayment tendency, and other factors. Correspondingly, the output is linked to different lending strategies. For example, if the customer's credit risk is predicted to rise, the corresponding lending strategy is to tighten credit conditions; if the customer's repayment performance is predicted to be stable and positive, appropriate interest rate discounts may be offered.

[0317] The following are some data mining neural network models suitable for this scenario:

[0318] (1) Multilayer Perceptron (MLP). An MLP is a typical feedforward neural network consisting of an input layer, one or more hidden layers, and an output layer. It can handle classification and regression tasks.

[0319] In the context of loan strategy generation, MLP can be used to predict the likelihood of future defaults by analyzing a customer's loan history, repayment records, and other relevant characteristics. It can also be combined with multiple data sources (such as macroeconomic indicators and industry data) to assess the customer's overall risk level.

[0320] (2) LSTM. LSTM is a special type of recurrent neural network specifically designed to solve the long-term dependency problem in traditional RNNs. It excels at processing time series data.

[0321] In the context of loan strategy generation, LSTM can be used to predict future repayment behavior based on a customer's historical repayment records, thereby identifying potential default risks in advance. It can also flexibly adjust credit limits based on changes in a customer's real-time consumption patterns and financial situation.

[0322] (3) Graph Neural Networks (GNNs). Graph neural networks are neural network models specifically designed for processing graph-structured data. They can capture the relationships and interactions between nodes and are suitable for datasets involving complex relationships.

[0323] In the context of loan strategy generation, GNNs can be used to analyze social relationships among customers, identify potential influencers or groups, and help understand the impact of group behavior on individual decisions. They can also consider the connections between customers (such as family members, business partners, etc.) to assess the risk of joint accounts or secured loans.

[0324] It should be noted that each of the aforementioned neural network models has its own characteristics and performs well in different application scenarios. Appropriate models can be selected or combined according to specific needs to improve the accuracy and effectiveness of lending strategies. Furthermore, these neural networks are merely simple examples; other neural networks that can be used for data mining are also applicable to the embodiments of this application, such as autoencoders and convolutional neural networks (CNNs), which will not be elaborated upon here.

[0325] The following is a brief explanation of other data mining models:

[0326] Regression models are a class of statistical methods used to predict continuous-valued outputs. Common types include linear regression and multinomial regression. These models make predictions by fitting the relationship between input features and a target variable. Regression models are generally simple, easy to interpret, and can provide good performance in some cases. Linear regression assumes a linear relationship between input features and the target variable. Multinomial regression, on the other hand, allows for more complex non-linear relationships by introducing higher-order terms of the features.

[0327] Despite its name containing "regression," logistic regression is actually a classification algorithm, primarily used for binary classification problems. It estimates the probability of an event occurring (usually 0 or 1) and classifies it based on a set threshold. Logistic regression uses the sigmoid function (or logistic function) to map the results of linear combinations to the interval [0,1], representing probabilities.

[0328] It's important to note that regression models and logistic regression are traditional statistical machine learning methods, primarily used for modeling simple linear or nonlinear relationships. Neural networks, on the other hand, are a more complex and powerful class of models, capable of handling highly nonlinear and complex pattern recognition tasks. Therefore, while both models can be used for data mining tasks, their working principles, applicable scenarios, and complexity differ significantly. The choice of model depends on the specific application requirements and data characteristics.

[0329] GBDT is a reinforcement learning method that corrects errors in an existing model by progressively adding new weak learners (usually decision trees). Each new tree attempts to minimize the residual. In the context of loan strategy generation, GBDT improves prediction accuracy by progressively adding new decision trees, each minimizing the residual. It can handle complex nonlinear relationships and high-dimensional data, such as features like customer credit scores and repayment behavior.

[0330] Association rule learning aims to discover frequent co-occurrence relationships between itemsets in large amounts of transaction data.

[0331] Naive Bayes uses a probabilistic classifier based on Bayes' theorem, assuming that features (such as credit scores, repayment history, etc.) are independent of each other, and quickly calculates posterior probabilities for classification. It can learn the relationship between customer characteristics and default risk based on historical data, and assess the likelihood of new customers defaulting in real time.

[0332] It should be noted that each of the above data mining models has its unique advantages and applicable scenarios. The selection of a suitable algorithm depends on the specific application requirements, data characteristics, and business objectives. Furthermore, these data mining models are merely simple examples; other data mining models or algorithms are also applicable to the embodiments of this application, such as K-means clustering and hierarchical clustering, which will not be elaborated upon here.

[0333] S23: Perform rule-based processing on the model prediction results to generate an executable rule set; the rule set contains at least one rule for risk assessment.

[0334] In practical applications, the prediction results obtained from data mining models are usually complex mathematical expressions or probability values, such as expressions derived from decision trees, expressions derived from regression models, probability values ​​predicted by logistic regression, probability groups predicted by neural networks, etc. These results are difficult to apply directly to actual business.

[0335] To ensure that these model prediction results can be effectively transformed into risk control strategies in actual operation, in S23 of this application embodiment, the model prediction results need to be processed into rules and transformed into executable rule forms to generate an operable rule set. This not only improves the applicability of the model results, but also ensures the scientific nature and executability of the risk control strategy.

[0336] Taking a decision tree model as an example, during its growth process, internal nodes are divided multiple times based on attributes. The conditions (inputs) on each path of the decision tree determine the final category or predicted value (output), reflecting the logical relationship between input and output. When the decision tree is transformed into rules, the combination of conditions on the path forms an expression. For example, when assessing the credit risk of a loan customer, a decision tree might be divided by three attributes: "credit score, income level, and debt situation." The rule condition part corresponding to the last branch could be written as a mathematical expression like: "(credit score < 600) and (monthly income < 5000) and (debt / income > 0.8)." The calculations and logical relationships here comprehensively reflect a certain risk characteristic of the customer and are typical data expressions. However, they are too complex and difficult for business personnel to interpret and implement intuitively.

[0337] Therefore, in S23, the structure of the decision tree can be transformed into rules. Specifically, each leaf node of the decision tree corresponds to a rule. The condition part of the rule consists of attribute tests along the path from the root node to the leaf node, and the conclusion part of the rule is the category or predicted value of the leaf node. Therefore, the process of summarizing business rules is as follows:

[0338] Starting from the root node of the decision tree, each path leads to a leaf node, representing a specific rule. Upon reaching an internal node, the test conditions of that node become part of the rule. All conditions on each path are connected using a logical AND operator to form a complete IF-THEN rule. Upon reaching a leaf node, the classification or prediction result of that node serves as the conclusion of the rule.

[0339] For example, in a decision tree used to predict customer default risk, one rule might be: if a customer's credit score is less than a1 and the number of late payments in the last three months is greater than a2, then the customer is predicted to have a high default risk.

[0340] This type of business rule based on decision tree path induction is highly interpretable, easy to understand, and easy to communicate. Of course, the random forest model works on a similar principle, which will not be repeated here.

[0341] For example, linear regression is often used to predict loan amounts. It predicts the dependent variable (loan amount) based on multiple independent variables (such as customer age, years of employment, asset size, etc.), resulting in an expression like "Loan amount = 3000 + 500 × age + 800 × years of employment - 1000 × debt size". In actual business, loan officers find it difficult to make quick decisions based on such formulas. They need to translate these into easily understandable rules, such as "Customers aged 30 and above, with more than 5 years of employment and debt below a certain value, can be approved for higher loan amounts."

[0342] For example, in logistic regression, a common approach to binary classification problems like predicting customer default, the model outputs the probability that a sample belongs to a particular category. For instance, to predict whether a customer will default on a loan, the model outputs a probability value between 0 and 1, such as 0.3, meaning the customer has a 30% probability of defaulting. Credit institutions cannot directly use this probability to implement strategies; business rules need to clearly define the boundaries, such as "customers with a probability value greater than 0.5 are marked as high-risk and require early intervention and alerts."

[0343] For example, when deep neural networks perform credit risk rating, the output layer, after being processed by the softmax function, provides a probability distribution of a customer belonging to each risk level. For instance, the output might be [0.1, 0.2, 0.7], corresponding to low, medium, and high risk probabilities, respectively. To implement this in business, this result needs to be transformed into rules, such as "when the high-risk probability exceeds 0.5, suspend the issuance of new credit lines," clarifying the ambiguous probabilities and turning them into actionable actions.

[0344] For other types of models, such as random forests and support vector machines, approximate methods can be used to transform the model results into rules. For example, by analyzing the input-output relationship of the model, some key features and thresholds can be extracted to construct rules.

[0345] In summary, this process involves not only simple rule generation, but also the relationship between input and output features based on data mining models, the extraction of key attributes and their corresponding risk thresholds, and the formalization of these into executable strategies.

[0346] Therefore, one optional implementation is that S23 can be implemented according to the following process, including the following S231 to S232 ( Figure 2 (Not shown):

[0347] S231: Based on the relationship between the input and output features of the data mining model, extract at least one key attribute related to risk assessment, and the risk threshold corresponding to each key attribute; wherein the input feature is risk assessment data, and the output feature is the model prediction result.

[0348] In this embodiment, step S231 involves identifying the features that have the greatest impact on the prediction results from the model's input and output features, so that these key features can be subsequently transformed into explicit business rules or logical conditions for better understanding and application. Specifically, this process aims to identify key attributes that have a significant impact on the model's prediction results based on the relationship between the model's input and output features, and to determine the risk thresholds corresponding to these key attributes through various methods. This not only ensures that the key attributes and their risk thresholds extracted from the prediction results of the data mining model accurately reflect risk assessment needs, but also generates a set of rules with practical operability, improving the scientific nature and applicability of risk control strategies.

[0349] Key attributes refer to features or variables that have a significant impact on the results of risk assessment in a specific business scenario. These attributes can be identified through data analysis, feature importance assessment, or other methods, and can play an important role in model prediction and actual operation, helping decision-makers to more accurately understand the data and formulate corresponding risk control strategies.

[0350] The following is a brief explanation of the key attributes in different scenarios:

[0351] For example, in the scenario of loan strategy generation, key attributes may include, but are not limited to, some or all of the following:

[0352] (1) Credit score: A borrower's credit score is an important indicator for measuring their credit risk; a high credit score usually means a lower risk of default. (2) Return level: A borrower's monthly or annual return reflects their repayment ability; a higher return level usually means a stronger repayment ability. (3) Historical repayment behavior: A borrower's past repayment records (such as whether they made timely payments, number of overdue payments, etc.) directly reflect their credit status and willingness to repay. (4) Debt ratio: The ratio of a borrower's total debt to their income, used to assess their financial health.

[0353] For example, in a strategy generation system, if the model predicts a borrower has a high probability of default, it might be due to a low credit score and multiple late payments in their repayment history. These two key attributes together determine the borrower's high-risk status, and the system may automatically adjust their credit limit or increase monitoring frequency.

[0354] For example, in the context of intelligent transportation, key attributes may include, but are not limited to, some or all of the following: weather conditions, traffic flow, driver behavior characteristics, etc.

[0355] It should be noted that the key attributes listed above are merely examples, and other key attributes also apply to the embodiments of this application, without specific limitations. In the embodiments of this application, by identifying and utilizing key attributes, enterprises and institutions can more accurately assess risks, optimize strategies, and make informed decisions.

[0356] Specifically, data mining models such as decision trees and neural networks learn from massive amounts of input data. Taking a decision tree as an example, it recursively divides the data space, continuously subdividing the data based on the value ranges of different attributes. This process uncovers hidden patterns in the data. For instance, a decision tree might learn that customers with low credit scores, high debt, and unstable income are more likely to default on loans. Once the decision tree is fully grown, each path from the root node to a leaf node naturally forms the prototype of a set of rules. The condition part of each rule is the attribute judgment along the path, and the conclusion part is the classification result corresponding to the leaf node. See the above examples for details; repeated points will not be elaborated upon.

[0357] Neural networks work similarly. Hidden layer neurons perform complex transformations on the input data, learning the non-linear relationships between features, and finally, the output layer provides a predicted classification or probability value. The output results can be further refined into rules. For example, if the output layer determines that a certain type of customer has high credit risk, the corresponding rule is to identify the characteristics of this type of customer, which serves as the basis for subsequent risk control measures.

[0358] In practical applications, to accurately extract key attributes related to risk assessment and their corresponding risk thresholds, the relationship between the input and output features of the data mining model is first analyzed to identify key attributes that significantly impact the model's prediction results. Then, for each identified key attribute, quantitative analysis methods are used to determine its corresponding risk threshold. This process ensures that the selection of key attributes and the setting of risk thresholds are both scientific and have practical application value.

[0359] Accordingly, an optional implementation method is to carry out S231 according to the following process, including the following S2311 to S2312 ( Figure 2 (Not shown):

[0360] S2311: Based on the relationship between the input and output features of a data mining model, identify key attributes whose influence on the model's prediction results exceeds a preset influence threshold.

[0361] Taking the loan strategy generation scenario as an example, the input features come from risk assessment data, such as the borrower's credit score, income level, and historical repayment behavior. The output features refer to the model's prediction results, such as the probability of default and credit rating.

[0362] In the embodiments of this application, by analyzing the feature importance of the model, such as feature importance score and SHAP value, key attributes that have a significant impact on the model's prediction results can be identified. For example, if a certain feature (such as credit score) has a significant impact on the prediction of default probability, it is identified as a key attribute.

[0363] Feature importance score is a method for measuring the contribution of each feature (corresponding to the attributes in this paper) to the model's prediction results. The calculation method for feature importance score varies depending on the model type. For example, in decision trees, the importance of a feature can be measured by the purity gain (such as the Gini coefficient or information gain) it brings when used to split nodes; for ensemble models (such as random forests and gradient boosting trees), the feature importance score is the average of the feature's importance across all trees; in linear regression or logistic regression, the importance of a feature can be measured by the absolute value of its coefficient, with a larger coefficient indicating a greater impact on the model's prediction results; and so on. This paper does not impose specific limitations on these methods.

[0364] The SHAP value is used to explain the specific contribution of each feature (i.e., attribute) to a single prediction result. It assesses the importance of a feature by considering its average marginal contribution across all possible combinations of features, which is not specifically limited in this paper. The SHAP value not only provides global feature importance (i.e., which feature is most important overall) but also local explanatory power (i.e., which features have the greatest impact on the prediction result for a specific sample), taking into account the interaction effects between features and providing a more comprehensive explanation.

[0365] Specifically, by using one or more of the above methods, the influence of each attribute in the input features on the model prediction results can be determined, and then key attributes can be selected by comparing them with a preset influence threshold.

[0366] In this embodiment of the application, in order to identify key attributes that have a significant impact on the model prediction results, a preset impact threshold needs to be set. Optionally, the preset impact threshold can be determined according to any one or more of the following methods:

[0367] (1) Rule of thumb: Set a reasonable threshold based on the experience of domain experts; for example, in financial risk assessment, features with an importance score exceeding a certain percentage (e.g., 5%) are generally considered key attributes. (2) Statistical methods: Use statistical tests (e.g., t-test, chi-square test) to determine whether the feature importance score is significantly higher than the level of random noise. (3) Cross-validation: Evaluate the model performance under different thresholds through cross-validation and select the threshold that optimizes the model performance; for example, select the threshold that maximizes the accuracy on the validation set. (4) Visualization tools: Use visualization tools (e.g., feature importance charts, SHAP value summary charts) to visually observe the importance and distribution of features, assisting in setting thresholds; etc.

[0368] Suppose that the random forest model finds that the feature importance scores of "credit score", "income level" and "historical repayment record" are 0.35, 0.25 and 0.20 respectively, which are significantly higher than other features. Furthermore, based on the rule of thumb and cross-validation results, features with a feature importance score greater than 0.15 are set as key attributes. Therefore, "credit score", "income level" and "historical repayment record" can be identified as key attributes.

[0369] Of course, you can also choose not to set a preset influence threshold, that is, set the preset influence threshold to 0, so that all attributes can be regarded as key attributes.

[0370] In this embodiment, these key attributes will be used for subsequent risk threshold setting and other optimization operations to ensure the scientific nature and effectiveness of the risk control strategy. The specific process is as follows:

[0371] S2312: For each identified key attribute, the risk threshold corresponding to the key attribute is obtained by quantitative analysis of the risk assessment data.

[0372] This step involves determining the corresponding risk threshold for each identified key attribute through quantitative analysis of the risk assessment data. Quantitative analysis is a technique that uses mathematical and statistical methods to systematically study data, aiming to extract meaningful information from large amounts of data.

[0373] In the embodiments of this application, risk assessment data can be quantitatively analyzed and risk thresholds corresponding to key attributes can be determined through one or more means such as numerical models, statistical tests, and machine learning algorithms to ensure the accuracy and reliability of decision-making.

[0374] The following is a simple example illustrating how to determine risk thresholds through quantitative analysis:

[0375] Method 1: Obtained through statistical analysis of risk assessment data.

[0376] Specifically, by conducting statistical analysis on a large amount of historical risk assessment data, reasonable risk thresholds for key attributes can be determined. Below is a simple example of Method One:

[0377] (1) Frequency distribution and quantiles.

[0378] For example, collect a large amount of data on past customers' credit scores, number of late payments, and corresponding default situations. Create a frequency distribution histogram for the key attribute of credit score to observe data clustering. For instance, if it is found that customers with credit scores below 550 have a significantly higher frequency of subsequent defaults than those in other score ranges, then 550 may become a risk threshold corresponding to the credit score.

[0379] Alternatively, quantiles can be used for calculation. For example, if the 25th percentile is calculated, and the probability of default for customers at or below this quantile reaches 30%, which is higher than the overall average probability of default, then the credit score corresponding to this quantile is worth considering as a credit score threshold, i.e., the risk threshold corresponding to the credit score.

[0380] Regarding the key attribute of the number of overdue payments, we also statistically analyzed the proportion of defaulting customers corresponding to different numbers of overdue payments. If the proportion of defaulting customers who are overdue for 3 times or more suddenly jumps, then 3 times may be the appropriate threshold for the number of overdue payments, that is, the risk threshold corresponding to the number of overdue payments.

[0381] (2) Correlation analysis.

[0382] Specifically, the correlation coefficients between credit score, number of delinquencies, and default status can be calculated, such as the Pearson correlation coefficient. A high correlation means that the variable has a strong explanatory power for whether or not a default will occur. For example, a correlation coefficient of 0.7 between the number of delinquencies and default status indicates that the number of delinquencies is a key factor (i.e., a key attribute). Then, by observing the changing trend of the number of delinquencies and the probability of default through a scatter plot, the node at which the probability of default rises rapidly can be found, thereby determining the corresponding risk threshold.

[0383] It should be noted that the methods listed above are just simple examples. In addition, other methods based on statistical analysis are also applicable to the embodiments of this application, and will not be described in detail here.

[0384] Method 2: Obtained by evaluating risk assessment data based on machine learning models.

[0385] Risk assessment data is evaluated using machine learning models, and risk thresholds are dynamically adjusted. For example, reinforcement learning or online learning algorithms are used to continuously optimize risk thresholds based on the latest data. A simple example of method two is given below:

[0386] (1) Grid search.

[0387] For example, historical risk assessment data can be divided into training and validation sets according to a certain ratio. A series of possible threshold combinations can be set, such as credit score thresholds ranging from 500 to 700 in 50-point increments; and thresholds for the number of overdue payments ranging from 1 to 5. For each combination, a decision tree model can be built on the training set, and then the model's prediction accuracy, recall, F1 score, and other metrics can be evaluated using the validation set.

[0388] Based on the above, record the performance of evaluation indicators under each threshold combination, and select the combination that yields the optimal overall indicator. If a certain combination achieves an accuracy of 85%, a recall of 60%, and an F1 score of 70%, which are higher than other combinations, then the threshold in that combination can be used as the risk threshold corresponding to the relevant key attribute. For example, the risk threshold corresponding to the information score is 600, and the risk threshold corresponding to the number of overdue payments is 1.

[0389] For example, in a strategy generation system, if a data mining model predicts a borrower has a high probability of default, it might be due to a low credit score and multiple late payments in their repayment history. Through quantitative analysis, borrowers with a credit score below 600 and one or more late payments in the past 12 months can be identified as high-risk customers. The system can then automatically adjust credit limits or increase monitoring frequency based on these key attributes and their risk thresholds, thereby managing risk more accurately.

[0390] (2) Cross-validation.

[0391] Using k-fold cross-validation (e.g., k=5), the historical risk assessment data is divided into 5 equal parts, and 4 parts are used to train the decision tree and 1 part for validation in turn. For different thresholds, the average precision, recall, F1 score, and other metrics are calculated for each validation fold. This allows for a more robust evaluation of the threshold setting, avoids bias from a single validation set, and selects the most stable threshold.

[0392] It should be noted that the methods listed above are just simple examples. In addition, other methods based on machine learning model evaluation are also applicable to the embodiments of this application, and will not be described in detail here.

[0393] Furthermore, it should be noted that the quantitative analysis methods listed in Method 1 or Method 2 above are merely simple examples. In addition, other quantitative analysis methods are also applicable to the embodiments of this application, and will not be described in detail here.

[0394] Of course, in addition to quantitative analysis, risk thresholds can also be determined by referring to domain expertise, as in method three below:

[0395] Method 3: Determined based on domain expertise.

[0396] Specifically, by combining business experience and expert knowledge, the risk thresholds corresponding to key attributes can be determined. Below is a simple example of method three:

[0397] (1) Industry practice.

[0398] Long-term practice in the financial industry has led to some common understandings. In the field of personal lending, credit scoring systems are relatively mature, and many institutions generally consider 600 points as the approximate boundary between customers with good and bad credit. This can serve as a reference for setting initial risk thresholds, which can then be fine-tuned based on their own business data.

[0399] (2) Risk control team experience.

[0400] The organization's in-house risk control experts, with years of experience handling default cases, can also offer reasonable suggestions. If they find that among the high-risk clients they have handled in the past, it is extremely common for them to have more than two overdue payments in the past three months, they will suggest setting the threshold for the number of overdue payments to 2, and then verify its feasibility through data analysis.

[0401] It should be noted that the methods listed above are just simple examples. In addition, other methods that combine domain expertise are also applicable to the embodiments of this application, and will not be described in detail here.

[0402] In addition, the methods listed above can be used in combination, such as combining the risk threshold of the same key attribute determined by multiple methods with the average or median, etc., which will not be elaborated here.

[0403] In the above implementation, extracting key attributes and their corresponding risk thresholds based on the relationship between the input and output of the data mining model can accurately reflect risk assessment needs and help generate a set of rules with practical operability. Furthermore, by identifying key attributes that have a significant impact on the model's prediction results and determining their risk thresholds through multiple methods, risks can be assessed more accurately and corresponding strategies can be formulated, thereby improving overall operational efficiency and risk management levels.

[0404] S232: Perform rule-based processing on at least one key attribute and its corresponding risk threshold to generate an executable set of rules.

[0405] In this embodiment of the application, the rule set includes at least one rule used for risk assessment.

[0406] In this step, the determined risk thresholds are transformed into specific business rules or logical conditions, ensuring that these rules can be directly applied to actual operations and are easy to understand and interpret. Through this process, complex model predictions can be converted into clear operational guidelines, improving the transparency and operability of decision-making. The specific implementation method is as follows:

[0407] First, the key attributes and their corresponding risk thresholds identified in the previous steps (such as S2311 and S2312) serve as the basis for rule-making; and the rule structure is defined, that is, the basic structure of the rules is designed. For example, the "if...then..." format can be used to ensure that each rule is clear, concise, and easy to understand. Then, specific triggering conditions are set according to the key attributes and risk thresholds. For example, "If the credit score is below 600, then mark it as a high-risk customer."

[0408] Based on the above, a rule set can be generated. Specifically, all rules corresponding to key attributes and their risk thresholds can be integrated into a single rule base to form a complete rule set. Alternatively, some rules corresponding to key attributes and their risk thresholds can be integrated into one rule base, while other rules corresponding to key attributes and their risk thresholds can be integrated into yet another rule base, and so on, forming multiple rule sets.

[0409] Furthermore, it's crucial to check the logical consistency between rules during or after rule generation to avoid conflicts or duplication, ensuring the completeness and accuracy of the rule set. Rule expression can also be simplified and optimized to ensure not only accuracy but also ease of execution. For example, natural language processing techniques can be used to make rules more closely resemble business language.

[0410] For example, in a strategy generation system, suppose the following key attributes and their risk thresholds are determined through quantitative analysis: credit score: 700 points; historical repayment behavior: 1 overdue payment; current loan balance: 50,000 yuan. (Continuing with the above...) Figure 3 As shown in the example, the following four rules can be derived from this decision tree, and these four rules form a rule set:

[0411] If your credit score is >700 and you have no overdue payment records in the past 6 months, the risk level is low.

[0412] If your credit score is >700 and you have a history of delinquency within the past 6 months, the risk level is medium.

[0413] If your credit score is ≤700 and your current loan balance is >50,000 yuan, then the risk level is high.

[0414] If your credit score is ≤700 and your current loan balance is ≤50,000 yuan, then your risk level is medium.

[0415] Through this rule-based processing, the system can automatically apply corresponding risk control strategies based on the borrower's circumstances, which improves efficiency and enhances the scientific nature and reliability of decision-making.

[0416] In the above implementation, by regularizing the model prediction results to generate an executable set of rules, the scientific nature and transparency of risk assessment are ensured. Simultaneously, this approach extracts key attributes and their risk thresholds based on data mining models, making risk control strategies both accurate and easy to implement, improving decision-making efficiency and the effectiveness of risk management, while also enhancing the system's adaptability and robustness.

[0417] In this embodiment of the application, to further improve the accuracy and reliability of the rules, some optimization methods can be used to optimize the rule set after it is generated. Specifically, the rules can be optimized indirectly through an optimization model, or optimization work can be carried out directly on the rules.

[0418] An alternative implementation is that, before S24, at least one of the following optimization operations—optimization operation one, optimization operation two, and optimization operation three—can be performed on the rule set.

[0419] like Figure 7 The diagram illustrates an optimization method for a planning set in an embodiment of this application. Optimization of the planning set can be categorized into two main approaches: one is to directly optimize the model, thereby indirectly optimizing the rules, as shown in Rule Optimization Operation 1 below; the other is to directly optimize the rules, as shown in Rule Optimization Operations 2 and 3 below. Rule Optimization Operation 2 primarily employs rule pruning, while Rule Optimization Operation 3 primarily employs rule integration.

[0420] In the embodiments of this application, these optimization operations can ensure the quality and effectiveness of the rule set. These measures improve the accuracy and practicality of the rule set and ensure the effectiveness of the risk control strategy.

[0421] The following sections provide a detailed explanation of each of these rule optimization operations:

[0422] Rule optimization operation one: Evaluate the data mining model and adjust the model parameters of the data mining model based on the first evaluation result, so as to indirectly adjust the rule set.

[0423] Rule optimization refers to indirectly optimizing rules by optimizing the model. Specifically, one or more methods, such as validation sets, cross-validation, and confusion matrices, can be used to evaluate the performance of the data mining model and obtain initial evaluation results (e.g., precision, recall, F1 score). Then, based on these initial evaluation results, the model's hyperparameters (e.g., learning rate, regularization coefficient, tree depth) are adjusted to optimize model performance. Thus, by improving the model, more accurate key attributes and risk thresholds are generated, indirectly optimizing the rule set and ensuring its greater accuracy and reliability.

[0424] In other words, when indirectly optimizing the rule set by optimizing the data mining model, one possible implementation method is as follows:

[0425] The data mining model is evaluated to obtain a first evaluation result; the model parameters of the data mining model are adjusted based on the first evaluation result; the risk assessment data is input into the adjusted data mining model to obtain a new model prediction result; the new model prediction result is processed into rules to generate a new set of executable rules.

[0426] The processes of inputting risk assessment data into the adjusted data mining model to obtain new model prediction results, and of performing rule-based processing on the new model prediction results to generate a new set of executable rules, are the same as those described in S22 and S23 above, and will not be repeated here.

[0427] The following examples illustrate the process of optimizing data mining models:

[0428] (1) Optimize the data mining model through cross-validation.

[0429] Cross-validation is a technique for evaluating the performance of machine learning models. It involves dividing a dataset (which can be the original dataset consisting of risk assessment data to be mined, or a validation set or other risk assessment data) into multiple subsets (i.e., folding), and then using these subsets in turn for training and validation, thereby obtaining a more reliable and stable performance estimate. Each subset typically has a similar size and is a random sample of the dataset.

[0430] Specifically, when using cross-validation to optimize a data mining model, the dataset is divided into multiple parts, and the model is trained and validated using different subsets of data in turn. During this process, the model parameters are continuously adjusted, making the data mining model more generalizable. For example, the depth of the decision tree and the node splitting criteria become more reasonable. Since the rules originate from the data mining model, as the model becomes more accurate, the rules derived from it better reflect the true distribution of the data. Rules that originally coarsely categorized customer risk can now, after cross-validation optimization of the data mining model, more finely differentiate customers at different risk levels, improving the accuracy of the rules.

[0431] Suppose the dataset is divided into k folds. For each fold i, a model is trained using the data excluding fold i. Then, the trained model is used to predict fold i to obtain the prediction result. The cross-validation error of the model can then be defined as follows: Formula 6:

[0432]

[0433] Among them, y iLet be the true label of fold i, and L(·,·) be the loss function used to measure the difference between the predicted result and the true label.

[0434] like Figure 8 As shown, it is a logical schematic diagram of cross-validation in an embodiment of this application. Wherein, Figure 8 Taking k=10 as an example, suppose there is a dataset D containing 1000 samples, and it is decided to use 10-fold cross-validation. The dataset is divided into 10 folds, each with approximately 100 samples, denoted as D1, D2, D3, ..., D10.

[0435] In the first round, the first nine folds D1-D9 form the training set for training the data mining model, and the tenth fold D10 serves as the validation set to validate the previously trained data mining model. Similarly, in the second round, the first eight folds D1-D8 and the tenth fold D10 are used for training, and the ninth fold D9 is used for validation. In the third round, the first seven folds D1-D7 and the last two folds D9-D10 are used for training, and the eighth fold D8 is used for validation. This process continues until each fold has been used as a validation set at least once.

[0436] In this way, k-fold cross-validation provides a robust method to evaluate the performance of data mining models while making the most of limited data resources.

[0437] For a given fold, which is essentially a subset of data, it may contain multiple samples. For example, in the above embodiment, each fold contains 100 samples. Therefore, in Equation 6, Specifically, it represents the prediction result of a sample in fold i, and correspondingly, y i For the true label of this sample in fold i, each sample participates in the calculation when calculating the loss function L(·,·), which will not be elaborated here.

[0438] Correspondingly, Figure 7 The prediction results shown are 1, 2, 3, ..., 10, which are the outputs obtained by inputting samples from the validation set into the model trained on the training set. For each fold, there are 100 samples per fold, so there are 100 prediction results, totaling 200. In each round, the cross-validation error can be calculated by referring to Formula 6.

[0439] Building upon the above, model parameters can be adjusted based on cross-validation error, and different parameter values ​​will significantly affect model performance. Specifically, hyperparameters manually set before model training can be adjusted, such as learning rate, regularization coefficient, tree depth, and number of nodes; internal parameters in the machine learning model can also be adjusted, such as weights, biases, neurons, and number of layers, but this paper does not impose specific limitations on these.

[0440] Taking random forest as an example, hyperparameters include the number of decision trees and the maximum depth of each tree. By trying different combinations of hyperparameters multiple times, the corresponding cross-validation error is calculated to find the hyperparameter settings that minimize this error. For example, by gradually increasing the number of decision trees and observing the trend of cross-validation error, when the error stops decreasing or even starts to increase, the current number is determined to be the optimal value.

[0441] Optionally, this application can also optimize the feature selection and engineering of the data mining model. For example, if a high cross-validation error is found, it may be due to certain features in the data causing the model to overfit or underfit. Attempts can be made to remove highly correlated redundant features to reduce noise interference. For instance, in a loan risk assessment model, if two features, "credit card debt" and "total debt," are highly correlated, removing one of them and retraining the model to calculate the cross-validation error can help improve the model's performance. Features can also be transformed, such as through logarithmic transformation or standardization, to make the model learn more smoothly and reduce errors.

[0442] Optionally, this application can also compare the cross-validation errors of different types of data mining models to determine the more suitable base model. If the linear regression model has a large error, a non-linear support vector machine or neural network can be tried. Additionally, using ensemble learning methods to combine multiple different data mining models (such as multiple decision trees forming a random forest) can usually reduce the error. The ensembled data mining models are then cross-validated again, continuously optimizing the combination method until the cross-validation error reaches an acceptable range.

[0443] In addition, it should be noted that, besides the standard k-fold cross-validation methods listed above, other variations can also be used, such as Leave-One-Out Cross-Validation (LOOCV) or hierarchical k-fold cross-validation, etc. This article does not make specific restrictions on these methods.

[0444] (2) Optimize the data mining model using the confusion matrix.

[0445] A confusion matrix is ​​a specific table layout used in machine learning and statistics to describe the performance of classification models. It helps to understand the performance of a model (specifically a classification model) by comparing predicted results with actual results.

[0446] like Figure 9The diagram shown illustrates a confusion matrix in one embodiment of this application. Positive examples refer to the type of instances the model attempts to identify. Negative examples refer to instances that are not positive examples. The confusion matrix displays the correspondence between the model's predictions and the true values, such as the number of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Specifically, TP represents the number of instances correctly predicted as positive by the model; FP represents the number of instances that are actually negative but incorrectly predicted as positive by the model; TN represents the number of instances correctly predicted as negative by the model; and FN represents the number of instances that are actually positive but incorrectly predicted as negative by the model.

[0447] Through analysis Figure 9 The confusion matrix shown can identify misjudgments by the data mining model in specific categories (e.g., a medium-risk customer being misclassified as high-risk), thereby understanding the model's performance under different circumstances and identifying which sample types it is prone to errors on. Based on these understandings, model parameters or structure can be adjusted to reduce misjudgments, optimize the model's accuracy and reliability, and ensure more precise and effective risk assessment.

[0448] For example, if a decision tree model consistently misclassifies medium-risk customers as high-risk, the model structure or parameters can be adjusted to reduce these errors. Correspondingly, rules generated based on the corrected model will be more accurate in identifying medium-risk customers, indirectly optimizing the rules themselves.

[0449] The above implementation focuses on the evaluation and adjustment of data mining models. By evaluating the model and adjusting its parameters based on the initial evaluation results, more accurate prediction results can be generated. The new model prediction results are then processed into a more precise and executable set of rules. This method ensures that the rule set is always based on the latest and optimized model, thereby improving the adaptability and accuracy of risk control strategies.

[0450] Rule optimization operation two: Evaluate the rule set and prune the rules based on the obtained second evaluation results.

[0451] Unlike rule optimization operation one, rule optimization operation two directly optimizes the rules. Specifically, rule pruning is an optimization technique primarily used to reduce the complexity and redundancy of the generated rule set, such as removing unnecessary or ineffective rules and retaining the most effective subset of rules. Rule pruning starts directly from the rules themselves, removing redundant or harmful parts, thereby effectively improving the interpretability and generalization ability of data mining models and increasing the credibility of the rules.

[0452] In this application embodiment, when performing rule-based pruning, an optional implementation method is as follows:

[0453] Each rule in the rule set is evaluated to obtain a second evaluation result; based on the second evaluation result, the negative key attributes in each rule that reduce the accuracy of the model prediction results are identified; and the rule set is pruned by deleting the relevant content of the negative key attributes in each rule.

[0454] Negative key attributes refer to features or conditions in the rule set that negatively impact the accuracy of the model's predictions. The presence of these attributes may cause the model to make incorrect classifications or predictions. This could be because they introduce noise, are irrelevant to the target variable, or mislead the model, preventing it from correctly learning patterns in the data.

[0455] In the embodiments of this application, rule evaluation can be carried out through actual application or simulation test. For example, the quality of each rule can be evaluated using various indicators (such as accuracy, coverage, confidence, support, etc.) to obtain a second evaluation result; or statistical significance tests (such as chi-square test) can be used to evaluate the effectiveness of the rule to obtain a second evaluation result. This document does not make specific limitations on this.

[0456] Based on the second evaluation results, feature selection algorithms (such as recursive feature elimination, minimum absolute shrinkage and selection operator (LASSO) regression) or model-based feature importance scores (such as decision trees and random forests) can be used to determine which attributes have a significant impact on model predictions. Then, the correlations between attributes are examined to identify redundant attributes that may lead to multicollinearity. Highly correlated attributes may cancel each other out, or one attribute may contribute very little to the model. Finally, experimental verification or other methods can be used to confirm whether an attribute is a negative key attribute.

[0457] For example, if adding a certain attribute to a rule reduces accuracy instead of increasing it, then that redundant attribute should be removed. Take the rule "Credit score < 600, Age > 30, Monthly income < 5000 -> High risk" as an example. If removing "Age > 30" increases the accuracy of predicting high-risk customers, then this more concise rule can be used as the optimized result and replace the original rule.

[0458] In the above implementation, after evaluating the rule set, rule pruning by identifying and removing negative key attributes that reduce prediction accuracy can effectively reduce noise and redundant information in the model. This improves the model's simplicity and interpretability, while enhancing its generalization ability and reducing the risk of overfitting. The optimized model typically performs more stably and accurately on new data. For example, after the decision tree is built, pruning can effectively remove overly specific leaf nodes, preventing overfitting and enabling the model to more accurately identify truly high-risk customers, thereby improving the quality and efficiency of business decisions.

[0459] In another alternative implementation, after obtaining the second evaluation result, unnecessary, redundant, or conflicting rules can be identified based on the second evaluation result. These invalid or redundant rules can then be removed or merged to simplify the rule set and improve its efficiency and interpretability. For example, rules with extremely low coverage can be deleted or rules with similar logic can be merged.

[0460] In practice, other criteria for pruning can be determined through one or more of the following methods, such as: (1) Minimum support / confidence: Rules below a certain threshold are removed. (2) Coverage: Only rules that cover a sufficient number of data points are retained. (3) Redundancy detection: Rules that are highly similar to other rules or are completely contained by other rules are removed. (4) Complexity penalty: A complexity penalty term is introduced to prioritize the retention of simple rules; etc.

[0461] In practical applications, rules that do not meet the requirements can be removed step by step according to the selected criteria. Alternatively, a recursive approach can be used to repeatedly evaluate and remove rules until all pruning conditions are met.

[0462] By pruning rules, the model becomes simpler and easier to understand while maintaining good predictive performance. Furthermore, after pruning, the pruned rule set can be tested on an independent validation set to ensure that its performance has not significantly decreased and that the model remains simpler and easier to interpret.

[0463] Rule optimization operation 3: If there are multiple rule sets, integrate the multiple rule sets.

[0464] In real-world applications, it's possible to generate multiple different rule sets. Rules within the same rule set may share the same, partially share, or differ in their key attributes; similarly, different rule sets may also share the same, partially share, or differ in their key attributes.

[0465] For example, in loan risk assessment, different sets of rules can be obtained based on different subsets of data, combinations of features, or model algorithms. Here's a simple example:

[0466] (1) Different data subsets: For example, two different data subsets are randomly selected from all loan customer data, and decision tree models are constructed for each subset, resulting in two sets of rules. Assume that the first set of rules is based on data from young customers and focuses on key attributes such as income growth trends and consumption habits; the second set of rules comes from elderly customers and focuses on key attributes such as asset stability and medical insurance status. This results in two different sets of rules, each set of rules corresponding to a decision tree.

[0467] (2) Combination of different attributes: For example, when constructing one set of rules, only financially relevant key attributes such as credit score and debt level are selected; another set of rules adds non-traditional key attributes such as social activity and online consumption frequency, and uses information from different dimensions to generate different rule systems.

[0468] (3) Different model algorithms: For example, a batch of rules can be generated using decision trees, whose rule form and logic depend on the partitioning of the tree structure; at the same time, an association rule mining algorithm can be used to process the data and produce another set of rules with a completely different style. In this way, two sets of rules can be generated based on completely different underlying logic.

[0469] It should be noted that the above-listed methods for generating different rule sets are just simple examples. Other feasible methods are also applicable to the embodiments of this application, and will not be described in detail here.

[0470] Once you have multiple sets of rules from different sources, you can use rule integration methods to combine these sets of rules, which can improve overall performance.

[0471] In the embodiments of this application, by selecting appropriate integration strategies, such as voting mechanisms, weighted averages, stacking, etc., multiple rule sets are integrated into a comprehensive rule set. This can take advantage of the advantages of multiple rule sets, improve overall performance and robustness, and reduce the possible bias or limitations of a single rule set.

[0472] In one optional implementation, when integrating multiple rule sets, at least one of the following integration methods one and integration method two is included, but is not limited to:

[0473] In summary, such as Figure 7 The rule integration shown is further divided into rule-level integration and weight allocation-level integration. Among them, integration method one belongs to rule-level integration, and integration method two belongs to weight allocation-level integration.

[0474] The following sections will explain each of these integration methods in detail:

[0475] Integration Method 1: For multiple rules that are related to the prediction direction in multiple rule sets, merge the multiple rules.

[0476] This method is a rule-level integration, which involves merging rules. When the prediction directions of multiple rules are related, these rules can be merged.

[0477] Prediction direction association refers to multiple rules logically having similar prediction results or behavioral tendencies. This means that when processing a specific input, these rules tend to arrive at the same or very similar conclusions. Specifically, prediction direction association can manifest in the following ways:

[0478] (1) Similar preconditions and identical conclusions: If multiple rules have similar preconditions (i.e., triggering conditions) and identical conclusions (i.e., output results or action recommendations), then these rules can be considered to be related in predictive direction. For example, in loan management, two rules are "Credit score < 550, debt > 80% of return -> high risk" and "Credit score < 500, debt > 90% of return -> high risk".

[0479] (2) Different premises but complementary, same conclusion: Even if the premises are different, if they all point to the same conclusion, these rules may be related in predictive direction. For example, one rule is based on "credit score below 600 points" and another rule is based on "income level below 5,000 yuan", but both conclude that the customer is "marked as a high-risk customer". Although these rules have different premises, they are complementary and enhance the reliability of the conclusion.

[0480] (3) Similar premises, different but related conclusions: The premises of the rules are similar, but the conclusions are slightly different but related. For example, one rule may recommend "reducing credit limits," while another rule recommends "increasing monitoring frequency." Although these two conclusions are not exactly the same, they are related measures in risk management, aiming to address the same type of risk.

[0481] (4) Different Premises but Consistent Conclusions: The premises are completely different, but the final conclusions are consistent. For example, one rule is based on "good historical repayment record," and another rule is based on "stable employment background," but both lead to the conclusion of "low-risk customer." This situation shows that different combinations of characteristics can lead to the same conclusion.

[0482] It should be noted that the above-listed cases are just simple examples. Other cases also apply to the embodiments of this application, and will not be described in detail here.

[0483] In one alternative implementation, when integrating rules, multiple rules can be merged into a single rule with broader coverage and greater adaptability to improve the overall reliability and usability of the rules.

[0484] As listed above, for example, two rules, "Credit score < 550, debt > 80% of return -> High risk" and "Credit score < 500, debt > 90% of return -> High risk," can be combined into "Credit score < 500, debt > 80% of return -> High risk." Another example is that original rule 1, "If the credit score is below 600, reduce the credit limit," and original rule 2, "If the return is below 5,000 yuan, increase the monitoring frequency," can be combined into "If the credit score is below 600 or the return is below 5,000 yuan, reduce the credit limit and increase the monitoring frequency"; and so on.

[0485] In another alternative implementation, a voting method can be used when integrating rules, but here it applies to the rules rather than the prediction results. The rule with the highest number of votes is selected as the final merged result.

[0486] Suppose there are three sets of rules. The first set contains the rule "Credit score < 600 and debt / income > 0.8 -> High risk"; the second set contains the rule "Number of overdue payments in the past three months > 2 -> High risk"; and the third set contains the rule "Number of credit cards > 5 and monthly spending fluctuation > 30% -> High risk". When assessing a customer, if the first rule is met, one vote is recorded as high risk; if the second rule is met, another vote is recorded; and if the third rule is met, another vote is recorded. Finally, the votes are tallied, and the risk level corresponding to the majority of votes is the integrated rule-based judgment result. This is operation directly at the rule condition level.

[0487] In the above implementation, by identifying and merging rules related to prediction direction, integration can be achieved at the rule level, simplifying the rule set and improving its efficiency and interpretability. Rule merging not only reduces redundancy but also enhances the transparency and consistency of the system, making the decision-making process clearer and easier to understand. Furthermore, the merged rules are easier to maintain and optimize, ensuring the continuous improvement and adaptability of risk control strategies.

[0488] Integration Method 2: For each rule set, a weight is assigned to each rule set based on the third evaluation result corresponding to the rule set, so as to summarize the model prediction results corresponding to the determined object according to the weight of the rule set to which each rule belongs.

[0489] In this way, each rule is individually scored based on the weight of its set, ensuring that the final risk category of each object is based on a multi-dimensional assessment. This allows for the generation of risk control strategies for that object based on the model's predictions. For example, if the rules determine that an object is ultimately high-risk, the alert frequency for that object can be increased from the original frequency. Conversely, if the rules determine that an object is ultimately low-risk, the alert frequency for that object can be decreased from the original frequency, and so on.

[0490] Specifically, this method belongs to weighted allocation integration, which assigns weights to different rule sets. These weights can be allocated based on metrics such as accuracy and recall in the past validation of these rule sets, and can also be dynamically adjusted based on these metrics to ensure that high-performing rule sets have a greater weight in the final decision.

[0491] For example, for each rule set, rule evaluation can be carried out through actual application or simulation testing, such as using various indicators (such as accuracy, coverage, confidence, support, etc.) to evaluate the quality of each rule; or statistical significance tests (such as chi-square test) can be used to evaluate the effectiveness of the rules; based on this, the third evaluation result corresponding to the rule set is comprehensively determined according to the quality or effectiveness of each rule in a rule set. The specific implementation method is not specifically limited in this article.

[0492] Taking accuracy as the third evaluation result as an example, the higher the accuracy of the rule set, the higher its corresponding weight. The two are positively correlated. For example, the weight of the rule set with high accuracy is set to 0.6, the weight of the rule set with slightly lower accuracy is set to 0.3, and the weight of the rule set with even lower accuracy is set to 0.1.

[0493] In this way, for customer data, each rule is scored according to the weight of its respective set. The final risk category is determined by summarizing the scores. This integration method is based on the attributes of the rules themselves, rather than simply summarizing prediction results. This integration method based on the attributes of the rules themselves not only considers the content and conditions of the rules, but also reflects the historical performance of each rule set through weights, thereby more comprehensively assessing risks and ensuring the accuracy and reliability of the model's prediction results.

[0494] Of course, if there is only one set of rules, different weights can be assigned to different rules within this set. Similarly, for customer data, each rule is scored according to the weight of its set, and the final risk category is determined after the scores are aggregated. The specific implementation method is as follows:

[0495] Let the rule set R = {r1, r2, ..., r} n For sample x, each rule r i The prediction result is y iThe prediction results of rule integration can then be obtained through weighted voting:

[0496]

[0497] Among them, w i For rule r i The weights can be learned from the training data, and this article does not make any specific restrictions on them.

[0498] In summary, the rule integration in the embodiments of this application can be carried out at the level of rule elements such as rule content, conditions, and weights, and is not limited to the level of prediction results.

[0499] The above implementation improves the accuracy and robustness of risk control strategies by integrating multiple rule sets. Rules related to predictive direction are merged to reduce redundancy and simplify the decision-making process. Furthermore, weights are assigned based on the third evaluation results of each rule set, ensuring that more effective rule sets have a greater weight in the final model prediction. This approach not only enhances the model's predictive accuracy but also improves its ability to handle complex situations, guaranteeing the consistency and reliability of decisions.

[0500] In this embodiment, rule-based processing transforms the results of data mining models into clear, actionable rules, facilitating their application and understanding in real-world business scenarios. Furthermore, rule optimization further ensures the reliability of decision-making.

[0501] S24: Generate a risk control strategy for at least one object based on the set of rules.

[0502] In this application embodiment, the risk control strategy aims to ensure the robust operation of the business and protect asset security, traffic safety, etc. by identifying, assessing and controlling risks.

[0503] Taking the loan intermediary strategy as an example, feasible loan intermediary strategies include, but are not limited to, some or all of the following:

[0504] (1) Increase the frequency of reminders to high-risk customers, such as increasing the number of phone reminders or sending more reminder text messages. (2) Adjust credit limits, and reduce the credit limit for customers with consecutive overdue payments. (3) Optimize repayment plans, and provide customers with excessively high debt ratios with options to extend the repayment period or make installment payments. (4) Strengthen customer interaction and support, and provide high-quality customers with first-time overdue payments with services such as dedicated customer service and temporary grace periods.

[0505] These strategies work together to improve the effectiveness of risk management and maintain good customer relationships.

[0506] For example, in the context of intelligent transportation, feasible risk control strategies include, but are not limited to, some or all of the following:

[0507] (1) Real-time traffic monitoring and early warning: such as issuing early warning information to drivers through various channels (such as mobile applications, in-vehicle navigation systems, and electronic road signs) to help drivers avoid congested road sections or dangerous areas in advance and reduce the probability of traffic accidents. (2) High-risk driving behavior identification and intervention: such as automatically sending reminders or suggestions to drivers when abnormal behavior is detected, correcting bad driving habits in a timely manner, and reducing the risk of accidents caused by improper operation. (3) Dynamically adjusting traffic restriction rules to optimize traffic flow distribution, alleviate peak-hour congestion, and improve air quality.

[0508] These strategies work together to form a comprehensive and efficient risk control framework within the intelligent transportation system, which not only improves traffic safety but also enhances urban management and service capabilities.

[0509] Of course, risk control strategies in other scenarios follow the same principle, which will not be elaborated on here.

[0510] Specifically, this step involves applying the rules previously extracted through data mining to specific business scenarios to achieve risk control. The following is a detailed explanation of this process:

[0511] First, it's necessary to clarify which entities(s) will be the targets for risk control strategies. These entities can be individual customers, corporate accounts, transaction activities, or any other entity that may be involved in risk assessment and management. Specifically, they are entities related to the risk assessment data to be mined. For example, if the risk assessment data includes characteristic data for Entity 1, Entity 2, and Entity 3, then specific risk control strategies can be developed for these entities. If the risk assessment data only contains environmental impact data, then specific risk control strategies can be developed for some pre-defined entities.

[0512] Next, we will conduct an in-depth analysis of the existing set of rules, including but not limited to understanding the meaning, applicable conditions, and expected effects of each rule.

[0513] Based on this, specific loan-making strategies can be formulated according to the generated rule set and applied to actual business operations. Alternatively, specific loan-making strategies can be formulated based on the generated rule set and the model prediction results of the target, and then applied to actual business operations.

[0514] For example, if the rule set contains a rule that “increases the frequency of reminders for customers with credit scores below xx,” then this rule will be transformed into a specific reminder action plan, such as increasing the number of phone reminders, sending reminder text messages, email notifications, etc.

[0515] For example, if a customer's last three bills are overdue, their credit limit will be reduced. If the system automatically detects three consecutive overdue payments, it will immediately lower the customer's credit limit to 70% of the original limit. The system can also inform the customer of the credit limit adjustment via SMS or email and provide consultation services to help the customer understand the reasons and how to restore the normal credit limit. Furthermore, it can issue early warnings for transactions approaching the new credit limit to prevent over-limitation.

[0516] The following example illustrates the loan strategy generation scenario. To achieve automated, real-time generation of loan strategies, this application also establishes an automated strategy generation system. This system can monitor data changes in real time, and when data is updated, it automatically triggers data mining and rule-making processes to generate the latest loan strategy.

[0517] like Figure 10 The diagram shown is a schematic of a strategy generation system according to an embodiment of this application. The main components of the strategy generation system 1000 include a data monitoring module 1001, a model update module 1002, a rule generation module 1003, and a strategy execution module 1004.

[0518] The data monitoring module 1001 is responsible for monitoring data changes in the data warehouse in real time. When a data update is detected, it notifies the model update module 1002. For example, if the data warehouse contains risk assessment data for objects 1, 2, 3, ..., N, and the risk assessment data for objects 1 and 2 changes, then this new data is used as risk assessment data to be mined, and the model update module 1002 is notified.

[0519] The model update module 1002 retrains the data mining model based on the new data. For example, an incremental learning method can be used to learn only from the new data instead of retraining the entire model to improve efficiency. For specific update methods, please refer to the above embodiments, and repeated details will not be elaborated here.

[0520] The rule generation module 1003 performs rule-based processing on the updated model results to generate a new set of rules.

[0521] The strategy execution module 1004 formulates specific loan strategies based on the generated rule set and applies these strategies to actual business operations. For example, it executes the corresponding loan strategies for object 1 and object 2.

[0522] An automated real-time risk control strategy generation system can quickly respond to data changes, adjust risk control strategies in a timely manner, and improve decision-making efficiency and accuracy.

[0523] Furthermore, in order to ensure the stable operation and continuous improvement of the system, system maintenance and updates can also be performed.

[0524] Optionally, system maintenance includes, but is not limited to, data backup, troubleshooting, and performance optimization. Regularly backing up system data prevents data loss; timely troubleshooting ensures normal system operation; and optimizing system performance improves response speed and processing capacity.

[0525] Optionally, system updates include model updates, rule updates, and strategy updates. Over time, data distribution and customer behavior may change, necessitating regular updates to data mining models and rule sets to ensure strategy effectiveness. Simultaneously, lending strategies should be adjusted promptly based on business needs and market changes to improve their adaptability.

[0526] In this embodiment of the application, after generating a risk control strategy for at least one object based on the rule set, the generated loan strategy needs to be evaluated and optimized to ensure its effectiveness and adaptability in order to ensure the effectiveness and stability of the strategy.

[0527] Therefore, this application also provides a series of optimization operations. These optimization measures aim to improve the performance of risk control strategies through continuous evaluation and adjustment, making them more aligned with actual business needs and able to respond promptly to market changes. One optional implementation involves performing at least one of the following optimization operations: Strategy Optimization Operation One and Strategy Optimization Operation Two.

[0528] like Figure 11 The diagram shown illustrates one optimization method of a strategy in an embodiment of this application. When optimizing the planning set, there are two specific methods, as described below: Strategy Optimization Operation One and Strategy Optimization Operation Two. These two strategy optimization operations are explained in detail below:

[0529] Strategy optimization operation one: Evaluate the risk control strategy based on at least one evaluation indicator to obtain the fourth evaluation result; adjust the risk control strategy based on the fourth evaluation result.

[0530] In this embodiment of the application, a series of preset evaluation indicators are used to quantitatively evaluate the risk control strategy and obtain a fourth evaluation result. These evaluation indicators can objectively reflect the performance of the risk control strategy in different dimensions.

[0531] In practical applications, strategy evaluation can be conducted from multiple perspectives. Specific evaluation indicators can be set to monitor the effectiveness of the strategy. When an evaluation indicator shows abnormalities, the system can automatically adjust the strategy or notify manual intervention, thereby responding to changes in a timely manner and maintaining the optimal performance of the strategy.

[0532] Optionally, the evaluation indicators include, but are not limited to, some or all of the following:

[0533] The strategy's accuracy, recall, risk control effectiveness, and cost-effectiveness, among other factors.

[0534] The calculation methods for each evaluation indicator are explained below:

[0535] (1) Accuracy assessment can be performed by comparing the predicted results of the risk control strategy with the actual results.

[0536] One possible calculation method is as follows: Let p be the probability that the risk control strategy predicts the customer's default, and q be the actual probability of the customer's default. Then, the accuracy of the risk control strategy can be expressed as either the absolute error |pq| or the relative error. To measure.

[0537] Another optional calculation method is as follows: Let the accuracy rate of the risk control strategy be P, then the formula for calculating P is as shown in Formula 8 below:

[0538]

[0539] Wherein, TP stands for True Positive Example, which is the number of samples correctly predicted as positive (i.e., positive examples); FP stands for False Positive Example, which is the number of samples incorrectly predicted as positive (i.e., positive examples); and FN stands for False Negative Example, which is the number of samples incorrectly predicted as negative (i.e., negative examples).

[0540] (2) The calculation method for recall is similar to that for precision. One possible calculation method is:

[0541] Let the recall rate of the risk control strategy be R, then the formula for calculating R is as follows: Formula 9:

[0542]

[0543] Wherein, TP stands for True Positive Example, which is the number of samples correctly predicted as positive; FP stands for False Positive Example, which is the number of samples incorrectly predicted as positive; and FN stands for False Negative Example, which is the number of samples incorrectly predicted as negative. These will not be repeated here.

[0544] (3) The effectiveness of risk control can be evaluated by analyzing the changes in indicators such as the default rate and loss rate of customers before and after the implementation of risk control strategies. For example, the percentage reduction in default rate and loss rate can be used as a measure.

[0545] One possible calculation method is as follows: Let the customer's default rate before the risk control strategy is implemented be r1, and the loss rate be l1; after the risk control strategy is implemented, the customer's default rate is r2, and the loss rate is l2. Then, the risk control effect of the risk control strategy can be expressed as the percentage reduction in the default rate. This can also be measured by the percentage reduction in loss rate. To measure.

[0546] (4) Cost-benefit assessment can be conducted by analyzing the costs and benefits of implementing risk control strategies.

[0547] One possible calculation method is as follows: Let the cost of implementing the risk control strategy be C and the benefit be B. Then the cost-effectiveness of the risk control strategy can be expressed as the benefit-cost ratio. To measure.

[0548] It should be noted that the evaluation metrics and their calculation methods listed above are just simple examples. In addition, other evaluation metrics and corresponding calculation methods are also applicable to the embodiments of this application, such as the F1 score (the harmonic mean of accuracy and recall, which provides a more balanced performance evaluation), etc. This article does not make specific limitations on them.

[0549] The aforementioned evaluation method employs more comprehensive assessment indicators, covering multiple aspects from technical performance to business impact, ensuring a thorough evaluation of the effectiveness and stability of the loan strategy. Furthermore, this method combines data-driven quantitative analysis with practical experience, guaranteeing the effectiveness and adaptability of the risk control strategy.

[0550] After obtaining the fourth assessment result through the above method, the aspects that need improvement can be identified based on the fourth assessment result, and the risk control strategy can be optimized accordingly. One possible implementation method is as follows: when optimizing the risk control strategy, the rules can be optimized directly, indirectly optimizing the strategy; the strategy can be optimized directly; or the rules can be optimized directly while simultaneously optimizing the strategy.

[0551] like Figure 11 As shown, strategy optimization can be divided into two main categories of optimization ideas: one is to directly optimize rules, thereby indirectly optimizing the strategy; the other is to directly optimize the strategy. When directly optimizing rules, optimization methods include adjusting rule parameters, adding new rules, and deleting invalid rules. Similarly, when directly optimizing the strategy, optimization methods include adjusting strategy parameters, adding new strategies, and deleting invalid strategies.

[0552] The following is a detailed explanation of these two optimization ideas:

[0553] (a) Optimize rules directly and optimize strategies indirectly.

[0554] Specifically, the set of rules for generating risk strategies is adjusted using a preset rule adjustment method; then, a new risk strategy is regenerated based on the adjusted rule set. In this embodiment, continuous optimization and simplification of the rule set is key to maintaining the efficiency and relevance of the risk control strategy.

[0555] The rule adjustment methods include at least one of the following: adjusting rule parameters, adding new rules, and deleting invalid rules.

[0556] In this application embodiment, adjusting rule parameters refers to modifying parameters (including conditions) in existing rules, such as risk control thresholds, key attributes, weights, etc., to adapt to new risk situations. Adding new rules refers to adding new rules based on the latest data analysis results or business needs to cover unforeseen situations. Deleting invalid rules refers to removing those rules that are no longer applicable or ineffective, simplifying the rule set and improving efficiency.

[0557] The process of regenerating a risk strategy based on the adjusted rule set is the same as S24 above, and will not be repeated here. A new risk strategy is then generated based on the adjusted rule set to ensure its scientific validity and operability.

[0558] For example, if a rule is found to have low accuracy, its threshold (e.g., lowering the credit score threshold from 600 to 580) or conditional clauses (e.g., adding criteria for detecting abnormal transaction behavior) can be adjusted to improve accuracy. By refining the rule logic, such as by incorporating considerations of recent transaction frequency, amount fluctuations, and changes in geographical location, rules can more accurately capture risk signals, thereby enhancing the overall effectiveness of the risk control system. Such adjustments help to effectively identify high-risk behaviors without impacting user experience.

[0559] If a particular customer group is found to be at higher risk, special rules can be added for that group (such as special terms for specific industries, stricter credit review processes, or higher margin requirements) to more accurately manage the potential risks of these customers.

[0560] If a rule (such as an outdated industry standard) is rarely triggered in practice, consider deleting or updating it to improve the efficiency of risk control strategies. Specifically, infrequently triggered rules may mean that their conditions are too strict or no longer applicable to the current market environment. Removing these rules can not only simplify the system and reduce unnecessary computational burden, but also avoid false alarms, allowing resources to be focused on the truly important risk points.

[0561] The above implementation details how to optimize risk strategies through preset rule adjustments. This rule adjustment mechanism ensures the scientific rigor and operability of the rule set, and can promptly reflect changes in the latest data and business needs. By adjusting rule parameters, existing rules can be fine-tuned to better adapt to new risk situations; adding new rules can cover unforeseen circumstances, ensuring the comprehensiveness of the strategy; deleting invalid rules simplifies the rule set and improves efficiency. Regenerating new risk strategies keeps the entire system up-to-date, providing more accurate decision support. This not only enhances the interpretability and transparency of the model but also strengthens the practical application of risk control strategies, helping financial institutions make more scientific and rational decisions.

[0562] (ii) Direct optimization strategy.

[0563] Specifically, risk strategies are adjusted through preset strategy adjustment methods.

[0564] The strategy adjustment methods include at least one of the following: adjusting strategy parameters, adding new strategies, and deleting invalid strategies.

[0565] In this embodiment, adjusting strategy parameters refers to modifying specific parameters in the risk control strategy, such as adjusting approval standards or increasing the monitoring frequency of high-risk customers, to optimize strategy effectiveness. Adding new strategies refers to introducing new strategies, such as providing value-added services for high-quality customers, to address newly emerging risk factors or business scenarios. Deleting ineffective strategies refers to removing strategies that are no longer effective or applicable, such as eliminating outdated policies, to ensure the overall strategy system is streamlined and efficient.

[0566] For example, if a strategy is found to have low accuracy, its parameters can be adjusted to improve its accuracy. If a customer group is found to be high-risk, specific strategies targeting that group can be added. If a strategy is rarely triggered in practice, it can be considered for removal to improve the efficiency of risk control strategies.

[0567] The above implementation methods ensure the overall strategy system is streamlined and efficient, enabling flexible responses to emerging risk factors or business scenarios. Adjusting strategy parameters can optimize specific operations such as approval standards and monitoring frequency, improving the accuracy and responsiveness of the strategy. Adding new strategies provides more options for handling complex and ever-changing market conditions, such as special terms for specific industries or value-added services for premium clients. Deleting ineffective strategies removes those that are no longer applicable or ineffective, reducing unnecessary complexity and resource waste. In this way, financial institutions can ensure that their risk control strategies are always optimal, maximizing the effectiveness of risk management while improving customer satisfaction and operational efficiency. In summary, these strategy adjustment methods significantly enhance the flexibility and effectiveness of the risk control system.

[0568] Strategy optimization operation two: After applying the risk strategy to actual business, collect feedback information from the target; adjust the risk strategy based on the feedback information.

[0569] In this embodiment, a feedback mechanism can be established to achieve continuous optimization of the strategy. When the strategy is applied in actual business, at least customer feedback information can be collected, and customer business data can also be collected. This information is fed back to the strategy generation system, which uses this information to automatically adjust the strategy to ensure its continuous optimization and adaptability.

[0570] Customer feedback refers to customers' opinions and evaluations of the risk control strategy's implementation results, reflecting their risk tolerance. This includes, but is not limited to, satisfaction surveys, encountered problems and suggestions, and records of abnormal situations, helping to understand the actual impact of the strategy. Business data, on the other hand, consists of objective information such as customer transaction records and behavioral patterns during service usage, reflecting the true state of business activities.

[0571] In one alternative implementation, after collecting feedback information from the object, and even the object's business data, strategy optimization can be performed based on this information. For example... Figure 11 As shown, strategy optimization operation two can also be divided into the two main optimization ideas mentioned above: one is to directly optimize the rules, thereby indirectly optimizing the strategy; the other is to directly optimize the strategy. When directly optimizing the rules, the optimization methods include adjusting rule parameters, adding new rules, and deleting invalid rules. Similarly, when directly optimizing the strategy, the optimization methods include adjusting strategy parameters, adding new strategies, and deleting invalid strategies. Specific implementation methods can be found in the above embodiments; repeated details will not be elaborated further.

[0572] In another alternative implementation, let the feedback information be F and the policy generation system be S, then the feedback mechanism can be expressed as:

[0573] S = update(S, F) (Formula 10)

[0574] Among them, update(.,.) is a function that updates the strategy generation system based on feedback information.

[0575] Taking credit limit satisfaction as an example, some long-term customers with good credit feel that their current loan limit is insufficient to meet their needs for expanding their business, renovating their homes, etc., and hope to increase their limit; conversely, some customers feel that the limit is too high and worry about the difficulty of repayment later, causing them great psychological pressure. Based on this feedback, the strategy generation system will reassess customer risk and adjust the credit limit allocation strategy to better match the limit with the customer's needs and risk tolerance.

[0576] Taking interest rate feedback as an example, high interest rates can place a heavy financial burden on loan customers. When a large number of customers report that the interest rate is too high, and even consider early repayment or switching to other financial institutions for loans, the strategy generation system will weigh the benefits and customer retention, and may appropriately lower the interest rate, especially for high-quality, low-risk customer groups, and enhance competitiveness through differentiated interest rate strategies.

[0577] In this embodiment, analysis of the feedback information can identify unforeseen problems or potential areas for improvement during strategy implementation. Based on this, necessary adjustments can be made to the risk control strategy to make it more aligned with actual conditions, more user-friendly, and more efficient. Through strategy evaluation and optimization, the performance of the risk control strategy can be continuously improved, enabling it to better adapt to market changes and customer needs.

[0578] The above implementation method introduces optimization operations for risk strategies, dynamically adjusting them through evaluation and feedback information. Specifically, the risk control strategy is evaluated based on at least one assessment indicator to obtain a fourth assessment result, and the strategy is adjusted accordingly. Furthermore, feedback information is collected from the target audience after the risk strategy is applied to actual business operations to further optimize the strategy. This method ensures that the risk control strategy can be continuously improved to adapt to the ever-changing market environment and customer needs. Dynamically adjusting the loan strategy allows financial institutions to update management measures in real time based on the latest data analysis results, such as adjusting credit limits and changing repayment plans, thereby improving the flexibility and responsiveness of risk management. Simultaneously, building an intelligent risk early warning system can identify potential problems in advance and take preventative measures to reduce the risk of loss. In summary, these optimization operations improve the accuracy and effectiveness of the risk control strategy and enhance the system's adaptability and robustness.

[0579] The following example mainly focuses on loan strategy generation in e-commerce scenarios:

[0580] In e-commerce scenarios, e-commerce platforms provide loan services to merchants. To manage loan risk and improve loan returns, the automated real-time loan strategy generation method described in this application can be used. (See also...) Figure 12 The diagram shown is a flowchart of an automatic loan generation strategy in an embodiment of this application. The specific implementation process of this method is as follows: S121 to S127:

[0581] S121: Collect risk assessment data related to loan management, including basic information of merchants, store operation data, loan information, repayment records, etc. At the same time, obtain some supplementary information from external data sources, such as industry trends and market competition.

[0582] S122: Preprocess the collected risk assessment data, remove noisy data and outliers, and then perform data integration, data transformation and data reduction.

[0583] S123: Build data mining models, such as decision trees or random forests, based on preprocessed risk assessment data to uncover potential patterns and relationships related to loan strategies, such as merchants' repayment behavior and operational risks.

[0584] For example, by analyzing indicators such as sales volume, order volume, and customer reviews in risk assessment data, the risk of default by merchants can be predicted.

[0585] S124: The model prediction results are processed into rules and converted into executable rules.

[0586] For example, if a merchant's sales decline for more than three consecutive months and the customer complaint rate exceeds a certain threshold, it is predicted that the merchant has a high risk of default and requires enhanced monitoring during the loan process.

[0587] S125: Establish an automated, real-time strategy generation system to monitor changes in merchant data in real time. When data updates are detected, automatically trigger data mining and rule-making processes to generate the latest lending strategies.

[0588] Specifically, the strategy generation system is as follows: Figure 10 As shown, this will not be repeated here. For example, if a merchant's sales suddenly drop, the strategy generation system will automatically adjust its lending strategy, such as reducing the loan amount or increasing the frequency of reminders.

[0589] S126: Evaluate and optimize the generated loan strategy. Analyze indicators such as the accuracy, risk control effectiveness, and cost-effectiveness of the strategy, and continuously adjust the strategy to improve its performance.

[0590] For example, if a rule is found to have low accuracy, its threshold or condition can be adjusted to improve its accuracy.

[0591] S127: Regularly maintain and update the strategy generation system to ensure stable operation and continuous improvement of the system.

[0592] For example, regularly update data mining models and rule sets to adapt to changes in merchant data and the market environment.

[0593] In this embodiment, S124 is used to lay the foundation and generate initial rules and strategies, while S125 is used for dynamic optimization, allowing the rules and strategies to be dynamically updated based on feedback from actual applications to suit complex and ever-changing business scenarios.

[0594] Specifically, S124 focuses on automated real-time strategy generation, producing initial loan strategies. When the data warehouse detects updates, it triggers model retraining. For example, in e-commerce scenarios, when there are new changes in a merchant's store's operating data, repayment records, etc., the data mining model is trained based on the updated data. The model results are then regularized and transformed into clear "if...then..." rules, such as "If a merchant's sales have decreased by more than 30% month-on-month in the past three months and the negative review rate is higher than 5%, then it is marked as a high-risk merchant." Subsequently, corresponding loan strategies are formulated based on these rules, such as tightening subsequent loan amounts for high-risk merchants and increasing the frequency of repayment reminders.

[0595] Once S125 is activated, multi-dimensional feedback is collected, including both subjective feedback from merchants regarding strategy implementation and business data feedback. For example, if, following the S124 strategy, a loan limit is tightened for an e-commerce merchant, the merchant reports that the recent short-term losses were due to store promotions and not poor management, while business data shows that the merchant's customer retention rate remains high. This feedback is input into the strategy evaluation phase to begin measuring the accuracy of the strategy.

[0596] Furthermore, rules and strategies are dynamically updated. Specifically, based on evaluation results, optimization is initiated when existing rules and strategies deviate significantly from actual conditions. Taking the aforementioned e-commerce merchant as an example, the thresholds for sales decline and negative review rate in the rules may be adjusted to more accurately screen out truly high-risk merchants. Correspondingly, the lending strategy also changes, no longer blindly tightening credit limits, but instead setting a moderate observation period, and then flexibly making decisions based on the recovery of store operations. As merchant data continues to change, repeated evaluations and adjustments are made to maintain the effectiveness and adaptability of rules and strategies.

[0597] In addition, it should be noted that Figure 12 The process shown is just a simple example. For specific implementation methods, please refer to the above embodiments, which will not be repeated here.

[0598] For example Figure 13 The diagram shown illustrates the interaction logic between a terminal device and a server in one embodiment of this application. In practical applications, the strategy generation system can be deployed on the server side. The server can collect risk assessment data through interaction with the terminal device or other electronic devices, preprocess the data, store it in a data warehouse, and monitor data changes in the data warehouse through the strategy generation system to automatically generate loan strategies in real time. The specific execution logic of the strategy generation system can be found in the above embodiment and will not be repeated here.

[0599] Specifically, this application has the following effects: (1) Improve the efficiency and accuracy of loan management: By automatically generating loan strategies in real time, manual intervention is reduced and decision-making efficiency is improved. At the same time, data mining technology is used to fully mine the information in the data, improving the accuracy of the strategy and effectively identifying and controlling risks. (2) Enhance the flexibility and adaptability of loan management: The strategy is automatically adjusted according to different customer groups and market conditions, improving the flexibility and adaptability of the strategy and better meeting business needs. (3) Reduce the cost of loan management: The automated strategy generation process reduces labor costs, and at the same time, effective risk control reduces loss costs. (4) Improve customer experience: Through reasonable loan strategies, customer needs can be better met and customer satisfaction can be improved.

[0600] Based on the same inventive concept, embodiments of this application also provide a strategy generation apparatus. For example... Figure 14 As shown, this is a schematic diagram of the strategy generation device 1400, which may include:

[0601] The data acquisition unit 1401 is used to acquire risk assessment data to be mined; the risk assessment data includes at least one of the following: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment;

[0602] The data processing unit 1402 is used to construct a data mining model based on the risk assessment data and obtain the model prediction results; the data mining model is used to identify potential patterns related to risk control strategies from the risk assessment data in order to conduct risk assessment.

[0603] The rule generation unit 1403 is used to perform rule-based processing on the model prediction results to generate an executable rule set; the rule set includes at least one rule for risk assessment.

[0604] The strategy generation unit 1404 is used to generate a risk control strategy for the at least one object based on the set of rules.

[0605] Optionally, the data acquisition unit 1401 is specifically used for:

[0606] Historical risk assessment data is obtained from the data warehouse and used as the initial risk assessment data to be mined.

[0607] The data processing unit 1402 is specifically used for:

[0608] An initial data mining model is constructed based on the initial risk assessment data to be mined;

[0609] Furthermore, the data acquisition unit 1401 is also used for:

[0610] Real-time monitoring of data changes in the data warehouse; when a data update is detected, the new risk assessment data in the data warehouse is used as the risk assessment data to be mined this time;

[0611] The data processing unit 1402 is further configured to:

[0612] The data mining model is updated based on the risk assessment data to be mined.

[0613] Optionally, the rule generation unit 1403 is specifically used for:

[0614] Based on the relationship between the input and output features of the data mining model, at least one key attribute related to risk assessment is extracted, along with a risk threshold corresponding to each key attribute; the input feature is the risk assessment data, and the output feature is the model prediction result.

[0615] The at least one key attribute and its corresponding risk threshold are processed into rules to generate an executable set of rules.

[0616] Optionally, the rule generation unit 1403 is specifically used for:

[0617] Based on the relationship between the input and output features of the data mining model, key attributes whose influence on the model's prediction results is higher than a preset influence threshold are identified.

[0618] For each of the identified key attributes, a risk threshold corresponding to the key attribute is obtained by quantitative analysis of the risk assessment data.

[0619] Optionally, before the strategy generation unit 1404 generates a risk control strategy for the at least one object based on the rule set, the rule generation unit 1403 is further configured to:

[0620] Perform at least one of the following optimization operations on the set of rules:

[0621] The data mining model is evaluated, and the model parameters of the data mining model are adjusted based on the obtained first evaluation result, so as to indirectly adjust the rule set;

[0622] The rule set is evaluated, and rule pruning is performed on the rule set based on the obtained second evaluation result;

[0623] If there are multiple rule sets, then the multiple rule sets are integrated.

[0624] Optionally, the rule generation unit 1403 is specifically used for:

[0625] The data mining model is evaluated to obtain a first evaluation result;

[0626] Adjust the model parameters of the data mining model based on the first evaluation result;

[0627] The risk assessment data is input into the adjusted data mining model to obtain new model prediction results;

[0628] The prediction results of the new model are then processed into rules to generate a new set of executable rules.

[0629] Optionally, the rule generation unit 1403 is specifically used for:

[0630] Each rule in the rule set is evaluated to obtain a second evaluation result;

[0631] Based on the second evaluation results, identify the negative key attributes in each rule that reduce the accuracy of the model's prediction results;

[0632] By deleting relevant content of the negative key attributes in each rule, the rule set is pruned.

[0633] Optionally, the rule generation unit 1403 is specifically used to integrate multiple rule sets in at least one of the following ways:

[0634] For multiple rules that predict the direction in multiple rule sets, the multiple rules are merged;

[0635] For each rule set, a weight is assigned to each rule set based on the third evaluation result corresponding to the rule set, so as to summarize the model prediction results corresponding to the determined object according to the weight of the rule set to which each rule belongs.

[0636] Optionally, the device further includes:

[0637] Optimization unit 1405 is configured to perform at least one of the following optimization operations on the risk strategy:

[0638] The risk control strategy is evaluated based on at least one evaluation indicator to obtain a fourth evaluation result; the risk control strategy is then adjusted based on the fourth evaluation result.

[0639] After applying the risk strategy to actual business operations, collect feedback information from the target entities; and adjust the risk strategy based on the feedback information.

[0640] Optionally, the optimization unit 1405 is specifically used for:

[0641] The set of rules for generating the risk strategy is adjusted using a preset rule adjustment method;

[0642] Based on the adjusted set of rules, a new risk strategy is generated.

[0643] The rule adjustment method includes at least one of the following:

[0644] Adjust rule parameters, add new rules, and delete invalid rules.

[0645] Optionally, the optimization unit 1405 is specifically used for:

[0646] The risk strategy is adjusted using a preset strategy adjustment method;

[0647] The strategy adjustment method includes at least one of the following:

[0648] Adjust strategy parameters, add new strategies, and delete invalid strategies.

[0649] This application significantly improves multiple aspects of risk assessment and risk control strategies through systematic data mining and rule-based processing. This method not only overcomes the shortcomings of related technologies but also brings greater intelligence and flexibility to risk management, making it applicable to various industries and application scenarios.

[0650] For ease of description, the above sections are divided into modules (or units) according to their functions and described separately. Of course, in implementing this application, the functions of each module (or unit) can be implemented in one or more software or hardware components.

[0651] In this application embodiment, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0652] Having introduced the strategy generation method and apparatus according to exemplary embodiments of this application, we will now introduce an electronic device according to another exemplary embodiment of this application.

[0653] Those skilled in the art will understand that various aspects of this application can be implemented as a system, method, or program product. Therefore, various aspects of this application can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or a combination of hardware and software implementations, collectively referred to herein as a "circuit," "module," or "system."

[0654] Based on the same inventive concept as the above-described method embodiments, this application also provides an electronic device. In one embodiment, the electronic device may be a server, such as... Figure 1 The server 120 is shown. In this embodiment, the structure of the electronic device can be as follows: Figure 15 As shown, it includes a memory 1501, a communication module 1503, and one or more processors 1502.

[0655] The memory 1501 is used to store computer programs executed by the processor 1502. The memory 1501 may mainly include a program storage area and a data storage area. The program storage area may store the operating system and programs required to run instant messaging functions, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0656] Memory 1501 may be volatile memory, such as random-access memory (RAM); memory 1501 may also be non-volatile memory, such as read-only memory, flash memory, hard disk drive (HDD), or solid-state drive (SSD); or memory 1501 may be any other medium capable of carrying or storing a desired computer program having the form of instructions or data structures and accessible by a computer, but is not limited thereto. Memory 1501 may be a combination of the above-described memories.

[0657] Processor 1502 may include one or more central processing units (CPUs) or digital processing units, etc. Processor 1502 is used to implement the above-described strategy generation method when it calls the computer program stored in memory 1501.

[0658] The communication module 1503 is used to communicate with terminal devices and other servers.

[0659] This application embodiment does not limit the specific connection medium between the memory 1501, communication module 1503, and processor 1502. This application embodiment... Figure 15 The memory 1501 and the processor 1502 are connected via a bus 1504, and the bus 1504 is in Figure 15 The diagram uses thick lines to describe the connections between other components; these are for illustrative purposes only and should not be considered limiting. The 1504 bus can be divided into address bus, data bus, control bus, etc. For ease of description, Figure 15 It is described using only a thick line, but does not indicate that there is only one bus or one type of bus.

[0660] The memory 1501 stores a computer storage medium, which stores computer-executable instructions for implementing the strategy generation method of this application embodiment. The processor 1502 is used to execute the above-described strategy generation method, such as... Figure 2 As shown.

[0661] In some possible implementations, various aspects of the strategy generation method provided in this application can also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program causes the electronic device to perform the steps in the strategy generation method according to the various exemplary embodiments of this application described above. For example, the electronic device can perform actions such as... Figure 2 The steps are shown in the figure.

[0662] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0663] The program product of the embodiments of this application may employ a portable compact disc read-only memory (CD-ROM) and include a computer program, and may run on an electronic device. However, the program product of this application is not limited thereto. In this document, the readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with a command execution system, apparatus, or device.

[0664] A readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying a readable computer program. This propagated data signal may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable signal medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with a command execution system, apparatus, or device.

[0665] Computer programs contained on readable media may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0666] Computer programs for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The computer program can execute entirely on the user's electronic device, partially on the user's electronic device, as a standalone software package, partially on the user's electronic device and partially on a remote electronic device, or entirely on a remote electronic device or server. In cases involving remote electronic devices, the remote electronic device can be connected to the user's electronic device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external electronic device (e.g., via the Internet using an Internet service provider).

[0667] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0668] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0669] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing a computer-usable computer program.

[0670] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0671] These computer program commands may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the commands stored in the computer-readable storage medium produce an article of manufacture including command means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0672] These computer program commands can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing the commands executed on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0673] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0674] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A strategy generation method, characterized in that, The method includes: Obtain risk assessment data to be mined; the risk assessment data includes at least one of the following categories: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment; A data mining model is constructed based on the risk assessment data, and the model prediction results are obtained; the data mining model is used to identify potential patterns related to risk control strategies from the risk assessment data in order to conduct risk assessment. The model prediction results are processed into rules to generate an executable rule set; the rule set contains at least one rule for risk assessment. Based on the set of rules, a risk control strategy is generated for the at least one object.

2. The method as described in claim 1, characterized in that, Obtain the risk assessment data to be mined, and construct a data mining model based on the risk assessment data, including: Historical risk assessment data is obtained from the data warehouse and used as initial risk assessment data to be mined; an initial data mining model is constructed based on the initial risk assessment data to be mined. In addition, acquiring risk assessment data to be mined, and constructing a data mining model based on the risk assessment data, further includes: Real-time monitoring of data changes in the data warehouse; when a data update is detected, the new risk assessment data in the data warehouse is used as the risk assessment data to be mined this time; the data mining model is updated based on the risk assessment data to be mined this time.

3. The method as described in claim 1, characterized in that, The step of performing rule-based processing on the model prediction results to generate an executable rule set includes: Based on the relationship between the input and output features of the data mining model, at least one key attribute related to risk assessment is extracted, along with a risk threshold corresponding to each key attribute; the input feature is the risk assessment data, and the output feature is the model prediction result. The at least one key attribute and its corresponding risk threshold are processed into rules to generate an executable set of rules.

4. The method as described in claim 3, characterized in that, The step of extracting at least one key attribute related to risk assessment, and a risk threshold corresponding to each key attribute, based on the relationship between the input and output features of the data mining model, includes: Based on the relationship between the input and output features of the data mining model, key attributes whose influence on the model's prediction results is higher than a preset influence threshold are identified. For each of the identified key attributes, a risk threshold corresponding to the key attribute is obtained by quantitative analysis of the risk assessment data.

5. The method according to any one of claims 1 to 4, characterized in that, Before generating a risk control strategy for the at least one object based on the set of rules, the method further includes: Perform at least one of the following optimization operations on the set of rules: The data mining model is evaluated, and the model parameters of the data mining model are adjusted based on the obtained first evaluation result, so as to indirectly adjust the rule set; The rule set is evaluated, and rule pruning is performed on the rule set based on the obtained second evaluation result; If there are multiple rule sets, then the multiple rule sets are integrated.

6. The method as described in claim 5, characterized in that, After evaluating the data mining model, adjusting the model parameters of the data mining model based on the obtained first evaluation result to indirectly adjust the rule set includes: The data mining model is evaluated to obtain a first evaluation result; Adjust the model parameters of the data mining model based on the first evaluation result; The risk assessment data is input into the adjusted data mining model to obtain new model prediction results; The prediction results of the new model are then processed into rules to generate a new set of executable rules.

7. The method as described in claim 5, characterized in that, The step of evaluating the rule set and pruning the rule set based on the obtained second evaluation result includes: Each rule in the rule set is evaluated to obtain a second evaluation result; Based on the second evaluation results, identify the negative key attributes in each rule that reduce the accuracy of the model's prediction results; By deleting relevant content of the negative key attributes in each rule, the rule set is pruned.

8. The method as described in claim 5, characterized in that, The integration of multiple rule sets includes at least one of the following methods: For multiple rules that predict the direction in multiple rule sets, the multiple rules are merged; For each rule set, a weight is assigned to each rule set based on the third evaluation result corresponding to the rule set, so as to summarize the model prediction results corresponding to the determined object according to the weight of the rule set to which each rule belongs.

9. The method according to any one of claims 1 to 4, characterized in that, The method further includes: Perform at least one of the following optimization operations on the risk strategy: The risk control strategy is evaluated based on at least one evaluation indicator to obtain a fourth evaluation result; the risk control strategy is then adjusted based on the fourth evaluation result. After applying the risk strategy to actual business operations, collect feedback information from the target entities; and adjust the risk strategy based on the feedback information.

10. The method as described in claim 9, characterized in that, The adjustment of the risk strategy includes: The set of rules for generating the risk strategy is adjusted using a preset rule adjustment method; Based on the adjusted set of rules, a new risk strategy is generated. The rule adjustment method includes at least one of the following: Adjust rule parameters, add new rules, and delete invalid rules.

11. The method as described in claim 9, characterized in that, The adjustment of the risk strategy includes: The risk strategy is adjusted using a preset strategy adjustment method; The strategy adjustment method includes at least one of the following: Adjust strategy parameters, add new strategies, and delete invalid strategies.

12. A strategy generation apparatus, characterized in that, include: The data acquisition unit is used to acquire risk assessment data to be mined; the risk assessment data includes at least one of the following: object characteristic data related to the risk assessment of at least one object, and environmental impact data related to the external environment; The data processing unit is used to construct a data mining model based on the risk assessment data and obtain the model prediction results; the data mining model is used to identify potential patterns related to risk control strategies from the risk assessment data in order to conduct risk assessment. The rule generation unit is used to process the model prediction results into rules to generate an executable rule set; the rule set includes at least one rule for risk assessment. The strategy generation unit is used to generate a risk control strategy for the at least one object based on the set of rules.

13. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any of the methods described in claims 1 to 11.

14. A computer-readable storage medium, characterized in that, It includes a computer program that, when run on an electronic device, causes the electronic device to perform the steps of any of the methods described in claims 1 to 11.

15. A computer program product, characterized in that, The method includes a computer program stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, causing the electronic device to perform the steps of any one of claims 1 to 11.

Images