A ticket risk control feature automatic mining method, device and equipment and storage medium

Abstract

The embodiment of the application provides a ticket risk control feature automatic mining method, device and equipment and a storage medium, which comprises the following steps: obtaining current state information of a target interface; generating a plurality of candidate feature extraction functions based on the current state information of the target interface; the candidate feature extraction function is used for extracting a feature vector; based on a preset data set, the plurality of candidate feature extraction functions are evaluated, and a reward value corresponding to the plurality of candidate feature extraction functions is obtained; when the reward value corresponding to the plurality of candidate feature extraction functions meets a preset condition, the plurality of candidate feature extraction functions are determined as target feature extraction functions or are determined as target feature extraction functions according to the current state information; and the request message received by the target interface is extracted by the target feature extraction function, and a feature vector corresponding to the request message is obtained. The accuracy of risk feature determination is improved, and the labor cost can be reduced.

Description

Technical Field

[0001] This application relates to the field of computer technology, specifically to a method, apparatus, equipment, and storage medium for automated mining of ticketing risk control features. Background Technology

[0002] In recent years, with the rapid development of high-speed rail, more and more people are choosing to travel by high-speed rail. Malicious behaviors surrounding the ticketing transaction process have become increasingly complex, including but not limited to automated ticket scalping, ticket hoarding by scalpers, fake account orders, and arbitrage fraud. These behaviors are typically implemented through high-frequency access, abnormal parameter combinations, and cross-interface collaborative actions, seriously affecting system fairness, business revenue, and user experience. Ticketing risk control data is core data used to identify, warn against, and block fraudulent behavior, abnormal operations, and illegal arbitrage in ticketing transactions. It covers the entire ticketing lifecycle (purchase, rescheduling, refund, verification) and supports the training and real-time decision-making of risk control models. Its core objective is to protect the rights and interests of ticketing platforms and users, and prevent risks such as ticket hoarding by scalpers, malicious ticket scalping, and payment fraud. To more quickly identify normal transactions and risky behaviors, risk identification models are usually built. When building a risk identification model, it is necessary to determine the characteristics of the risk data. In existing ticketing risk control systems, data feature engineering is the core foundation for the performance of risk identification models and is implemented manually by risk control experts. For example, feature extraction functions used to extract and filter risk features need to be completed manually by risk experts.

[0003] The above methods rely on the human experience of risk control experts to determine risk characteristics, which presents two main problems. First, manual completion is inefficient. When adding or changing a business interface, risk control experts need to re-understand the interface semantics and design feature logic to implement the feature extraction function, resulting in low development efficiency. Second, manual confirmation is highly subjective. Different risk experts may have different understandings and judgments of risk characteristics, potentially introducing redundant or invalid features, reducing the robustness and accuracy of risk identification, and hindering the improvement of overall risk control capabilities. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, equipment and storage medium for automated mining of ticketing risk control features, which facilitates the automatic identification of risk features, reduces manual intervention, thereby improving the accuracy of risk feature identification and reducing costs.

[0005] In a first aspect, embodiments of this application provide an automated method for mining ticketing risk control features, comprising: obtaining current status information of a target interface; the target interface is an interface that needs to be subject to risk control; the current status information includes description information of the target interface, instruction information for describing the risk control features of the target interface, and context data information corresponding to the description information; Based on the current state information of the target interface, multiple candidate feature extraction functions are generated; the candidate feature extraction functions are used to extract feature vectors; wherein, the feature vectors are used to describe the risk characteristics of the request messages received by the target interface; Based on a preset dataset, the multiple candidate feature extraction functions are evaluated, and the reward values ​​corresponding to the multiple candidate feature extraction functions are obtained; When the reward values ​​corresponding to the plurality of candidate feature extraction functions meet the preset conditions, the plurality of candidate feature extraction functions are determined as the target feature extraction function; The target feature extraction function is used to extract features from the request message received by the target interface to obtain the feature vector corresponding to the request message.

[0006] In one possible implementation of the first aspect, the method further includes: When the reward values ​​corresponding to the multiple candidate feature extraction functions do not meet the preset conditions, the example context data information and improvement instruction information are determined based on the multiple candidate feature extraction functions and their corresponding reward values. Based on the example context data information, update the context data information in the current state information, and based on the improved instruction information, update the instruction information in the current state information to update the current state information of the target interface; Based on the updated current state information of the target interface, the steps are re-executed: generating multiple candidate feature extraction functions and steps based on the current state information of the target interface; evaluating the multiple candidate feature extraction functions based on a preset dataset; obtaining the reward values ​​corresponding to the multiple candidate feature extraction functions; until the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions.

[0007] In one possible implementation of the first aspect, determining the example context data information based on the plurality of candidate feature extraction functions and their corresponding reward values ​​includes: Based on the multiple candidate feature extraction functions and their corresponding reward values, the candidate feature extraction functions and their corresponding reward values ​​that have a reward value higher than the reward value in the context data information of the current state information are determined as example context data information. The step of updating the context data information in the current state information based on the example context data information includes: Based on the example context data information, lower-level context data information is determined from the current state information; wherein, the reward value of the lower-level context data information is lower than the reward value of the example context data information; The example context data information is used to replace the lower-level context data information in the current state information to update the context data information in the current state information.

[0008] In one possible implementation of the first aspect, evaluating the plurality of candidate feature extraction functions based on a preset dataset and obtaining the reward values ​​corresponding to the plurality of candidate feature extraction functions includes: Based on a preset dataset, multiple sets of candidate risk control feature vectors are obtained using the multiple candidate feature extraction functions. Based on the multiple sets of candidate risk control feature vectors, reward analysis is performed on the multiple candidate feature extraction functions to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0009] In one possible implementation of the first aspect, the preset dataset includes a first dataset and a second dataset; The step of obtaining multiple sets of candidate risk control feature vectors based on a preset dataset and using the multiple candidate feature extraction functions includes: Based on the first dataset, multiple candidate feature extraction functions are used to obtain multiple sets of candidate risk control feature vectors; The step of performing reward analysis on the multiple candidate feature extraction functions based on the multiple sets of candidate risk control feature vectors, and obtaining the reward values ​​corresponding to the multiple candidate feature extraction functions, includes: Based on the second dataset, multiple sets of risk control feature vectors for training are obtained using the multiple candidate feature extraction functions. Based on the multiple sets of training risk control feature vectors, multiple risk control models are trained and obtained respectively. Based on the multiple sets of candidate risk control feature vectors and the multiple risk control models, obtain the risk control result information corresponding to each of the risk control models; Based on the risk control results, a reward analysis is performed to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0010] In one possible implementation of the first aspect, obtaining the current state information of the target interface includes: Obtain the description information of the target interface; Based on the description information of the target interface, if there is first context data information in the stored context data information, then the first context data information is determined as the context data information corresponding to the description information; wherein, the first context data information is the context data information in the stored context data information that has the highest similarity to the description information of the target interface and the similarity value is greater than a preset similarity threshold. The preset instruction information is determined as the instruction information corresponding to the target interface.

[0011] One possible implementation of the first aspect also includes: If the first context data information is not present in the stored context data information, then the preset second context data information is determined as the context data information corresponding to the description information.

[0012] In one possible implementation of the first aspect, the preset conditions include: the highest reward value corresponding to the plurality of candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface, and the number of times the highest reward value corresponding to the plurality of candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface reaches a preset number threshold.

[0013] In one possible implementation of the first aspect, the method further includes: When the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions, the target feature extraction function and its corresponding reward value are stored as context data information.

[0014] Secondly, embodiments of this application provide an automated ticketing risk control feature mining device, comprising: an acquisition module, used to acquire current status information of a target interface; the target interface is an interface that needs to be subject to risk control; the current status information includes description information of the target interface, instruction information for describing risk control features, and context data information corresponding to the description information; The generation module is used to generate multiple candidate feature extraction functions based on the current state information of the target interface; the candidate feature extraction functions are used to generate feature vectors; the feature vectors are used to describe the risk control features of the request messages received by the target interface. The processing module is also used to evaluate the plurality of candidate feature extraction functions based on a preset dataset and obtain the reward values ​​corresponding to the plurality of candidate feature extraction functions; The processing module is further configured to determine the multiple candidate feature extraction functions as target feature extraction functions when the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions, and to extract features from the request message received by the target interface through the target feature extraction function to obtain the feature vector corresponding to the request message.

[0015] Thirdly, embodiments of this application provide an electronic device, including a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the electronic device is triggered to execute the method described in any of the first aspects above.

[0016] Fourthly, embodiments of this application provide a computer-readable storage medium including a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the method described in any of the first aspects above.

[0017] The solution provided in this application embodiment obtains the current state information of the target interface, which includes: description information of the target interface, instruction information describing the risk control characteristics of the target interface, and context data information corresponding to the description information. Based on the current state information of the target interface, multiple candidate feature extraction functions are generated; based on a preset dataset, the multiple candidate feature extraction functions are evaluated to obtain reward values ​​corresponding to the multiple candidate feature extraction functions; when the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions, the multiple candidate feature extraction functions are determined as target feature extraction functions, so that feature extraction can be performed on the request message received by the target interface through the target feature extraction functions to obtain the feature vector corresponding to the request message. Thus, in this application embodiment, when risk control processing of the target interface is required, the target feature extraction function is first obtained so that feature vectors can be extracted from the request message received by the target interface through the target feature extraction function for risk assessment. Based on this, the current state information of the target interface can be obtained, and multiple candidate feature extraction functions can be generated according to the current state information of the target interface. Based on a preset dataset, the multiple candidate feature extraction functions are evaluated, and the reward value of each candidate feature extraction function is obtained. When the reward values ​​corresponding to multiple candidate feature extraction functions meet preset conditions, it indicates that the feature vector extracted by the current candidate feature extraction function is relatively effective, and the multiple candidate feature extraction functions can be determined as the target feature extraction function. In other words, in this embodiment, multiple candidate feature extraction functions can be automatically generated according to the current state information of the target interface, and the multiple candidate feature extraction functions can be automatically evaluated. When the reward values ​​corresponding to multiple candidate feature extraction functions meet preset conditions, the multiple candidate feature extraction functions can be determined as the target feature extraction function to extract feature vectors from the request messages received by the target interface for risk assessment. In this embodiment, the target feature extraction function can be automatically determined without manual intervention, reducing the time and cost of risk control processing, and improving the accuracy and efficiency of risk feature determination. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating an automated method for mining ticketing risk control features, provided as an embodiment of this application; Figure 2 A flowchart illustrating another automated method for mining ticketing risk control features provided in this application embodiment; Figure 3 A flowchart illustrating another automated method for mining ticketing risk control features provided in this application embodiment; Figure 4 This application provides a schematic diagram of the structure of an automated feature mining device for ticketing risk control. Figure 5 This is a schematic diagram of the structure of an electronic device for automated mining of ticketing risk control features, provided in an embodiment of this application. Detailed Implementation

[0020] To better understand the technical solution of this application, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0021] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0022] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0023] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0024] In existing ticketing risk control systems, data feature engineering is the core foundation for the performance of risk identification models. This is typically done manually by risk control experts, which presents two main problems. First, manual completion is inefficient. When adding or changing a business interface, risk control experts need to re-understand the interface semantics and design feature logic to implement the feature extraction function, leading to low development efficiency. Second, manual confirmation is highly subjective. Different risk experts may have different understandings and judgments of risk features, potentially introducing redundant or invalid features. This reduces the robustness and accuracy of risk identification, hindering the improvement of overall risk control capabilities.

[0025] To address the aforementioned issues, this application provides a method, apparatus, device, and storage medium for automated mining of ticketing risk control features. Specifically, it involves: acquiring the current state information of a target interface, which includes: description information of the target interface, instruction information describing the risk control features of the target interface, and context data information corresponding to the description information; generating multiple candidate feature extraction functions based on the current state information of the target interface; evaluating the multiple candidate feature extraction functions based on a preset dataset to obtain reward values ​​corresponding to the multiple candidate feature extraction functions; and determining the multiple candidate feature extraction functions as target feature extraction functions when the reward values ​​of the multiple candidate feature extraction functions meet preset conditions, so that feature extraction can be performed on the request messages received by the target interface to obtain feature vectors corresponding to the request messages. Thus, in this embodiment of the application, when risk control processing of a target interface is required, the target feature extraction function is first acquired so that feature vectors can be extracted from the request messages received by the target interface for risk assessment. Based on this, the current state information of the target interface can be obtained, and multiple candidate feature extraction functions can be generated according to the current state information of the target interface. Based on a preset dataset, the multiple candidate feature extraction functions are evaluated, and the reward value of each candidate feature extraction function is obtained. When the reward values ​​corresponding to multiple candidate feature extraction functions meet preset conditions, it indicates that the feature vector extracted by the current candidate feature extraction function is relatively effective, and the multiple candidate feature extraction functions can be determined as the target feature extraction function. In other words, in this embodiment, multiple candidate feature extraction functions can be automatically generated according to the current state information of the target interface, and the multiple candidate feature extraction functions can be automatically evaluated. When the reward values ​​corresponding to multiple candidate feature extraction functions meet preset conditions, the multiple candidate feature extraction functions can be determined as the target feature extraction function to extract feature vectors from the request messages received by the target interface for risk assessment. In this embodiment, the target feature extraction function can be automatically determined without manual intervention, reducing the time and cost of risk control processing, and improving the accuracy and efficiency of risk feature determination. A detailed explanation follows.

[0026] See Figure 1 This is a flowchart illustrating a ticketing risk control method provided in an embodiment of this application. Figure 1 As shown, the method includes: Step S101: Obtain the current status information of the target interface.

[0027] The target interface is the interface that requires risk control. The current status information includes the target interface's description, the instruction information describing the target interface's risk control characteristics, and the context data information corresponding to the description.

[0028] In this embodiment, when adding or modifying a business interface, risk control monitoring can be performed on that interface to identify it as a target interface and obtain its current status information. The current status information of the target interface describes its current descriptive information, risk control feature instruction information, and related context data. This current status information can be pre-stored in a storage device and retrieved from the storage device.

[0029] In other words, the description information of the target interface can be obtained from a storage device or can be input by the user. Typically, the description information of the target interface is used to describe relevant information about that target interface, such as its function, inbound information, outbound information, etc.

[0030] When initially acquiring instruction information to describe the risk control features of the target interface, pre-set general instruction information can be used as the instruction information for describing the risk control features of the target interface. In subsequent iterations, this information can be updated based on the evaluation results of the candidate feature extraction function, as detailed in the following steps. The context data information corresponding to the description information can also be pre-set context data information.

[0031] In some embodiments, obtaining the current state information of the target interface includes: obtaining the description information of the target interface; based on the description information of the target interface, if first context data information exists in the stored context data information, then the first context data information is determined as the context data information corresponding to the description information. Preset instruction information is determined as the instruction information corresponding to the target interface.

[0032] Among them, the first context data information is the context data information that has the highest similarity to the description information among the stored context data information, and the similarity value is greater than the preset similarity threshold.

[0033] In this embodiment, when obtaining the current state information of the target interface, the description information, instruction information, and context data information of the target interface can be obtained separately. Since the description information of the target interface can be set simultaneously when setting the target interface, it can be directly obtained from the storage device. After obtaining the description information of the target interface, a search can be performed using similarity in the stored context data. If there is context data in the stored context data that has a similarity greater than a preset similarity threshold with the description information, then the context data with the highest similarity among the context data with a similarity greater than the preset similarity threshold is taken as the first context data, and the first context data is determined as the context data corresponding to the description information. When obtaining the current state information of the target interface for the first time, preset instruction information can be determined as the instruction information corresponding to the target interface.

[0034] In some embodiments, empty instruction information can be used as preset instruction information, and the preset instruction information can be determined as the instruction information corresponding to the target interface.

[0035] In some embodiments, obtaining the current status information of the target interface further includes: If the first context data information is not present in the stored context data information, then the preset second context data information is determined as the context data information corresponding to the description information.

[0036] In this embodiment, if the first context data information is not present in the stored context data information, it indicates that the stored context data information does not contain context data information similar to the description information of the target interface. In this case, the preset second context data information can be determined as the context data information corresponding to the description information.

[0037] In other words, if no similar stored context data is found by searching for text similarity in the stored context data information based on the description information of the target interface (e.g., based on the business type of the target interface, such as query, payment, etc.), the preset second context data information can be determined as the context data information corresponding to the description information.

[0038] Step S102: Based on the current state information of the target interface, generate multiple candidate feature extraction functions.

[0039] The candidate feature extraction function is used to extract feature vectors. These feature vectors describe the risk characteristics of the request messages received by the target interface.

[0040] In this embodiment, after obtaining the current state information of the target interface, it becomes clear what information the generated feature extraction function should contain. At this point, multiple candidate feature extraction functions can be generated based on the current state information of the target interface.

[0041] In some embodiments, a feature extraction function generation model can be pre-trained. This allows the current state information of the target interface to be used as input to the feature extraction function generation model, generating multiple candidate feature extraction functions. For example, the pre-trained feature extraction function generation model may have a preset generation strategy as follows: The current status information of the target interface can be obtained using This is represented by the expression t, where t represents the number of iterations. The current state information of the target interface is used as the input value to the feature extraction function generation model, which then generates multiple candidate feature extraction functions. .in, Let i represent the candidate feature extraction function in the t-th iteration. For random sampling, such as sampling temperature, random seed, etc.

[0042] It should be understood that the number of candidate feature extraction functions generated can be preset by the user according to actual needs, such as 5, 10, or 3, etc., and this application embodiment does not limit this.

[0043] It should be noted that, in order to determine the final target feature extraction function in this embodiment, after generating multiple candidate feature extraction functions, it is necessary to determine whether a candidate feature extraction function is the target feature extraction function by obtaining its reward value. If not, the current state information of the target interface needs to be adjusted and a new candidate feature extraction function needs to be generated. Thus, through a continuous iterative process, the final target feature extraction function is determined. Therefore, in the above embodiment, t represents the iteration number, i.e., the t-th iteration. The t-th iteration can be taken as the current iteration process.

[0044] Step S103: Based on the preset dataset, evaluate multiple candidate feature extraction functions and obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0045] In this embodiment, after generating multiple candidate feature extraction functions, the feature vectors extracted by these functions may not be accurate. It is necessary to determine whether the generated candidate function is the final target function so that the target function can be used to extract the feature vectors of the request message received by the target interface. If it is not the final target function, the target interface's status information needs to be updated to regenerate candidate functions. Therefore, after generating multiple candidate function, a preset dataset can be used to evaluate the functions and determine their effectiveness in extracting feature vectors, thereby obtaining reward values ​​for each function. For example, data from the dataset can be used as input to multiple candidate function sets. Each function extracts features from the input data, resulting in multiple sets of candidate risk control feature vectors. These sets are then analyzed to obtain the reward values ​​for each function.

[0046] In some embodiments, the above-mentioned evaluation of multiple candidate feature extraction functions based on a preset dataset to obtain reward values ​​corresponding to the multiple candidate feature extraction functions includes: Based on a pre-defined dataset, multiple candidate feature extraction functions are used to obtain multiple sets of candidate risk control feature vectors. Based on these sets, reward analysis is performed on the candidate feature extraction functions to obtain their corresponding reward values.

[0047] In this embodiment, to evaluate candidate feature extraction functions, a dataset can be pre-set to evaluate the candidate feature extraction functions using data from the pre-set dataset. Specifically, a pre-set dataset can be obtained in a storage device, and the data in the pre-set dataset can be used as input to multiple candidate feature extraction functions. For each of the multiple candidate feature extraction functions, a set of candidate risk control feature vectors corresponding to the data in the pre-set dataset is calculated using that candidate feature extraction function. Thus, a set of candidate risk control feature vectors can be obtained for each candidate feature extraction function, resulting in multiple sets of candidate risk control feature vectors. After obtaining multiple sets of candidate risk control feature vectors, each set of candidate feature vectors can be analyzed to obtain the reward value corresponding to each of the multiple candidate feature extraction functions.

[0048] For ease of implementation, in some embodiments, the preset dataset includes a first dataset and a second dataset. Based on the preset dataset, multiple candidate feature extraction functions are used to obtain multiple sets of candidate risk control feature vectors, including: based on the first dataset, multiple candidate feature extraction functions are used to obtain multiple sets of candidate risk control feature vectors.

[0049] Based on the aforementioned multiple sets of candidate risk control feature vectors, reward analysis is performed on multiple candidate feature extraction functions to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions, including: Based on the second dataset, multiple candidate feature extraction functions are used to obtain multiple sets of risk control feature vectors for training. Based on these multiple sets of risk control feature vectors, multiple risk control models are trained and obtained.

[0050] Based on multiple sets of candidate risk control feature vectors and multiple risk control models, the risk control result information corresponding to each risk control model is obtained. Reward analysis is then performed based on the risk control result information corresponding to each risk control model to obtain the reward values ​​corresponding to multiple candidate feature extraction functions. When determining the reward score, a pre-set correspondence between different accuracy rates and reward scores can be established, and the reward value corresponding to different candidate feature extraction functions is determined based on this correspondence.

[0051] In this embodiment, the preset dataset includes a first dataset and a second dataset. The second dataset is used to train the risk control model, and the first dataset is used to obtain the reward values ​​of the candidate feature extraction functions. Based on this, after obtaining multiple candidate feature extraction functions, the data in the first dataset can be used as input values ​​for these functions. The data in the first dataset can then be processed using these multiple candidate feature extraction functions to extract feature vectors, resulting in multiple sets of candidate risk control feature vectors. It should be understood that one data point can be processed by one candidate feature extraction function to extract one feature vector. Thus, for each of the multiple candidate feature extraction functions, feature extraction is performed on all data in the first dataset using that function, resulting in multiple candidate risk control feature vectors. These multiple candidate risk control feature vectors are then used as a set of candidate risk control feature vectors. Therefore, for each candidate feature extraction function, a set of candidate risk control feature vectors can be obtained, thereby forming a set of multiple sets of candidate risk control feature vectors.

[0052] In other words, the first dataset contains multiple data points. For each data point, features can be extracted using multiple candidate feature extraction functions to obtain multiple sets of candidate risk control feature vectors. For example, if the first dataset contains m data points, features can be extracted for each of the m data points using n candidate feature extraction functions to obtain n sets of candidate risk control feature vectors, with each set containing m candidate risk control feature vectors.

[0053] To reduce manual intervention, automate the design process, and improve the accuracy of evaluating candidate feature extraction functions, a pre-defined network model can be trained on a second dataset for each feature extraction function, resulting in a risk control model corresponding to each function. Different feature extraction functions correspond to different risk control models. The risk control model analyzes the input feature vectors and outputs whether the input feature vectors represent risk features. Based on this, the second dataset can be used to train the pre-defined network model to obtain the risk control model. At this point, features can be extracted from the data in the second dataset using multiple candidate feature extraction functions, resulting in multiple sets of training risk control feature vectors. For each set of training risk control feature vectors, this set is used as training data to train a pre-defined network model, resulting in a risk control model. For example, all the training risk control feature vectors in a certain set can be input into a pre-defined network model, which analyzes the set and outputs whether the set represents a risk feature. The second dataset contains not only the data input to the candidate feature extraction function, but also information on whether this data is risky. Therefore, after determining whether the risk control feature vector output by the pre-defined network model is a risky feature, it's possible to check if the difference between the output of the pre-defined network model and the risk data information recorded in the second dataset is within a preset range. If it's not within the preset range, it indicates that the output of the pre-defined network model is inaccurate. Based on the risk data information recorded in the second dataset and whether the risk control feature vector output by the pre-defined network model is a risky feature, the parameters of the pre-defined network model can be adjusted, and the above training process can be repeated until the difference between the output of the pre-defined network model and the risk data information recorded in the second dataset is within the preset range. If the difference between the output of the pre-defined network model and the risk data information recorded in the second dataset is within the preset range, it indicates that the output of the pre-defined network model is relatively accurate, and the training of the pre-defined network model can be considered complete. This yields the risk control model corresponding to the set of training risk control feature vectors, which, for ease of description, will be referred to as the first risk control model below. After obtaining the first risk control model, the candidate risk control feature vectors for the corresponding groups obtained above can be used as input values ​​to the first risk control model. The first risk control model can analyze the input candidate risk control feature vectors for the corresponding groups and output the result of whether the candidate risk control feature vectors for the corresponding groups are risk features, which is the corresponding risk control result information. This risk control result information is the information output by the first risk control model that characterizes whether the candidate risk control feature vectors for the corresponding groups are risk features. The first dataset not only contains the data input to the candidate feature extraction function, but also contains information on whether this data is risky data.Therefore, after the first risk control model outputs the risk control result information corresponding to the candidate risk control feature vectors for that group, a reward analysis can be performed on the risk control result information corresponding to the candidate risk control feature vectors for that group based on the risk data information recorded in the first dataset and the risk control result information corresponding to the candidate risk control feature vectors output by the first risk control model, to obtain the reward value corresponding to the candidate feature extraction function. The reward value can characterize the accuracy of the candidate feature vectors extracted by the candidate feature extraction function.

[0054] It should be noted that the candidate feature extraction function corresponding to the training risk control feature vector used for training the first risk control model is the same as the candidate feature extraction function corresponding to a set of candidate feature vectors input into the first risk control model. That is, different risk control models are trained for different candidate feature extraction functions; one candidate feature extraction function corresponds to one risk control model.

[0055] As illustrated in the example above, for ease of description, the following explanation uses the i-th candidate feature extraction function out of n candidate feature extraction functions as an example. Based on the second dataset, assuming the second dataset contains k data points, the i-th candidate feature extraction function extracts features from the k data points in the second dataset, outputting the i-th set of training risk control feature vectors. This i-th set of training risk control feature vectors contains k training risk control feature vectors. The i-th set of training risk control feature vectors is used to train the i-th preset network model, resulting in the i-th risk control model. Similarly, on the first dataset, the i-th candidate feature extraction function extracts features from the m data points in the first dataset, outputting the i-th set of candidate risk control feature vectors, which contains m candidate risk control feature vectors. The i-th set of candidate risk control feature vectors can be used as input values ​​to the i-th risk control model. The i-th risk control model analyzes the i-th set of candidate risk control feature vectors and outputs the risk control result corresponding to the i-th set of candidate risk control feature vectors. The risk control result is used to characterize whether the i-th group of candidate risk control feature vectors is a risk feature. The first dataset records not only m data points but also information on whether each of the m data points is risky. Therefore, based on the risk control result corresponding to the i-th group of candidate risk control feature vectors output by the i-th risk control model and the information on whether the corresponding m data points recorded in the first dataset are risky, an accuracy analysis can be performed on the risk control result corresponding to the i-th group of candidate risk control feature vectors output by the i-th risk control model, and the reward value corresponding to the i-th candidate feature extraction function can be calculated. This reward value characterizes the accuracy of feature extraction by the i-th candidate feature extraction function. In some embodiments, to facilitate calculation, a pre-set correspondence between reward value and accuracy can be established. Thus, based on the accuracy of the risk control result, the pre-set correspondence between reward value and accuracy can be used to determine the reward value corresponding to the candidate feature extraction function.

[0056] In this way, the reward value corresponding to each of the n candidate feature extraction functions can be obtained.

[0057] In some embodiments, the reward values ​​corresponding to multiple candidate feature extraction functions can also be obtained in other ways. For example, the accuracy of the risk control result and the conversion formula between the reward values ​​can be preset, and the reward values ​​corresponding to multiple candidate feature extraction functions can be calculated based on the formula.

[0058] In some embodiments, it can be done through formula This indicates that the candidate feature extraction function extracts features from the data in the second dataset, resulting in multiple sets of training risk control feature vectors. Let i represent the candidate feature extraction function in the t-th iteration. This represents the second dataset. This is a Python executor, specifically a code sandbox and feature executor, capable of executing candidate feature extraction functions. Let represent the set of risk control feature vectors extracted by the i-th candidate feature extraction function during the t-th iteration.

[0059] After obtaining the set of risk control feature vectors for training, the formula can be used. This indicates that the risk control model is being trained. Here, y represents whether the training data output by the risk control model represents risk characteristics. This represents the trainable parameters in the risk model. A trained risk control model can then be used... express. It can be any binary classification model such as LightGBM, logistic regression, or neural networks.

[0060] It can be done through formula This indicates whether each element in the set of candidate risk control feature vectors output by the risk control model is identified as a risk feature. This indicates whether the set of candidate risk control feature vectors extracted by the i-th candidate feature extraction function during the t-th iteration is a risk feature. Let represent the set of candidate risk control feature vectors extracted by the i-th candidate feature extraction function.

[0061] In some embodiments, when performing reward analysis based on the risk control result information corresponding to the candidate risk control feature vectors and obtaining the reward values ​​corresponding to multiple candidate feature extraction functions, the formula can be used. The reward values ​​corresponding to different candidate feature extraction functions are calculated. This represents the reward value corresponding to the i-th candidate feature extraction function in the t-th iteration. This indicates online computation overhead, feature dimensions, execution time, etc. The preset weighting coefficients, This represents the risk outcome information corresponding to the m data points in the first dataset. It's important to note that the first dataset includes not only the historical data input to the candidate feature extraction function, but also the corresponding risk outcome information. This risk outcome information can be determined by risk control experts through analysis of historical data. Therefore, based on whether the set of candidate risk control feature vectors extracted by the i-th candidate feature extraction function output by the risk control model represents a risk feature, and the risk outcome information corresponding to the input data recorded in the first dataset, the reward value corresponding to the i-th candidate feature extraction function in the t-th iteration can be calculated using the reward function. For example, the more instances where the candidate risk control feature vector extracted by the i-th candidate feature extraction function in the risk control model is identical to the risk result information corresponding to the input data recorded in the first dataset, the closer the prediction result output by the risk control model based on the feature vector extracted by a certain candidate feature extraction function is to the risk result information corresponding to the input data recorded in the first dataset. This indicates that the more accurate the candidate risk control feature vector extracted by the candidate feature extraction function, the higher the reward value calculated by the reward function.

[0062] The fewer the number of times the candidate risk control feature vector extracted by the i-th candidate feature extraction function output by the risk control model is the same as the risk result information corresponding to the input data recorded in the first dataset, the smaller the deviation between the prediction result output by the risk control model based on the feature vector extracted by the candidate feature extraction function and the risk result information corresponding to the input data recorded in the first dataset. This indicates that the candidate risk control feature vector extracted by the candidate feature extraction function is inaccurate, and the lower the reward value calculated by the reward function.

[0063] Step S104: When the reward values ​​corresponding to multiple candidate feature extraction functions meet the preset conditions, determine the multiple candidate feature extraction functions as the target feature extraction function or determine them as the target feature extraction function based on the current state information.

[0064] In this embodiment, after obtaining the reward values ​​corresponding to multiple candidate feature extraction functions, it can be checked whether the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions. For example, the preset conditions include that the highest reward value corresponding to the multiple candidate feature extraction functions is not higher than the reward value recorded in the context data information of the current state information of the target interface. At this time, it can be checked whether the highest value among the reward values ​​corresponding to the multiple candidate feature extraction functions is higher than the preset reward value of the target interface. If it is higher, it is determined that the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions.

[0065] When the reward values ​​corresponding to multiple candidate feature extraction functions meet the preset conditions, it indicates that the feature vectors extracted by the multiple candidate feature extraction functions are relatively accurate. Therefore, there is no need to adjust the multiple candidate feature extraction functions, and they can be determined as the target feature extraction function. Alternatively, to improve stability, the multiple candidate feature extraction functions recorded in the context information of the current state can also be determined as the target feature extraction function.

[0066] In some embodiments, the preset conditions include: the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface, and the number of times the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface reaches a preset number threshold.

[0067] In other words, to improve the accuracy of the target feature extraction function, the preset condition can be set as follows: the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the target interface's current state information, and the number of times the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the target interface's current state information reaches a preset threshold. That is, the preset condition is set as follows: the reward score has not increased further, and the number of times it has not increased reaches a preset threshold. This means that when the reward values ​​corresponding to multiple candidate feature extraction functions meet the preset condition—that is, when the highest reward value among the multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the target interface's current state information, and the number of times the highest reward value in the context data information of the target interface's current state information is not exceeded during the iteration process reaches the preset threshold—it indicates that the feature vector extracted by the current candidate feature extraction function is relatively stable, and the current multiple candidate feature extraction functions can be directly determined as the target feature extraction function. Alternatively, to improve stability, the multiple candidate feature extraction functions recorded in the context information of the current state information can also be determined as the target feature extraction function.

[0068] Step S105: Extract features from the request message received by the target interface using the target feature extraction function to obtain the feature vector corresponding to the request message.

[0069] In this embodiment, after determining the target feature extraction function, risk control monitoring of the target interface can be achieved. At this time, the target feature extraction function can be used to extract features from the request messages received by the target interface, obtaining the feature vector corresponding to the request message. Risk control analysis of the request message is then performed using this feature vector to determine whether the request message received by the target interface poses a risk.

[0070] In this embodiment, when risk control processing of the target interface is required, a target feature extraction function is first obtained. This function is then used to extract feature vectors from the request messages received by the target interface for risk assessment. Based on this, the current state information of the target interface can be obtained, and multiple candidate feature extraction functions are generated. Using a preset dataset, these candidate functions are evaluated, and a reward value is obtained for each. When the reward values ​​of multiple candidate function extraction functions meet preset conditions, it indicates that the feature vectors extracted by the current candidate function are relatively effective, and these multiple candidate function extraction functions can be determined as the target feature extraction function. In other words, in this embodiment, multiple candidate feature extraction functions can be automatically generated based on the current state information of the target interface. These functions are then automatically evaluated, and when the reward values ​​of multiple candidate function extraction functions meet preset conditions, they can be determined as the target feature extraction function to extract feature vectors from the request messages received by the target interface for risk assessment. In this embodiment, the target feature extraction function can be automatically determined without manual intervention, reducing the time and cost of risk control processing and improving the accuracy and efficiency of risk feature determination.

[0071] In some embodiments, reference Figure 2 and Figure 3 As shown, the above method also includes: Step S106: When the reward values ​​corresponding to multiple candidate feature extraction functions do not meet the preset conditions, determine the example context data information and improvement instruction information based on the multiple candidate feature extraction functions and their corresponding reward values.

[0072] In this embodiment, when the reward values ​​corresponding to multiple candidate feature extraction functions do not meet preset conditions, it indicates that the current candidate feature extraction functions need further adjustment. At this time, a candidate feature extraction function for updating the context data information in the current state information of the target interface can be selected based on the multiple candidate feature extraction functions and their corresponding reward values, and this candidate feature extraction function is determined as the example context data information. Based on the multiple candidate feature extraction functions and their corresponding reward values, improvement instruction information in natural language form is generated. For example, the generated improvement instruction information indicates which extracted feature vectors are unsuitable and which feature vectors are more suitable.

[0073] In some embodiments, a meta-inference model can be pre-trained to generate improved instruction information based on multiple candidate feature extraction functions and their corresponding reward values.

[0074] In some embodiments, meta-inference model Receives structured input and outputs improved instructions on natural language strategies. , , This represents the (t+1)th iteration. Indicates the highest reward Group candidate feature extraction function.

[0075] In some embodiments, determining example context data information based on multiple candidate feature extraction functions and their corresponding reward values ​​includes: Based on multiple candidate feature extraction functions and their corresponding reward values, the candidate feature extraction functions and their corresponding reward values ​​that have a reward value higher than the reward value in the context data information of the current state information are determined as example context data information.

[0076] In other words, the context data recorded in the current state information of the target interface can include candidate feature extraction functions and their corresponding reward values ​​generated during completed iterations. Therefore, to improve the accuracy of candidate feature extraction functions generated in subsequent iterations, the context data in the current state information of the target interface can be updated based on the reward values ​​of the candidate feature extraction functions generated in the current iteration. Based on this, after obtaining the reward values ​​corresponding to the candidate feature extraction functions in the current iteration, if it is determined that the reward values ​​of multiple candidate feature extraction functions do not meet the preset conditions, the candidate feature extraction functions need to be regenerated. At this time, the current state information of the target interface can be updated first. Among the multiple candidate feature extraction functions, the candidate feature extraction function whose reward value is higher than the reward value in the context data information of the current state information can be identified as example context data.

[0077] For example, the current state information of the target interface records six candidate feature extraction functions and their corresponding reward values. These are: candidate feature extraction function a1 (7 points), candidate feature extraction function a2 (6 points), candidate feature extraction function a3 (4 points), candidate feature extraction function a4 (4 points), candidate feature extraction function a5 (4 points), and candidate feature extraction function a6 (4 points). Assume that the candidate feature extraction functions obtained in the current iteration are b1, b2, b3, b4, b5, and b6. Specifically, candidate feature extraction function b1 has a reward score of 10 points, candidate feature extraction function b2 has a reward score of 10 points, candidate feature extraction function b3 has a reward score of 9 points, candidate feature extraction function b4 has a reward score of 8 points, candidate feature extraction function b5 has a reward score of 5 points, and candidate feature extraction function b6 has a reward score of 1 point. At this point, in the candidate feature extraction functions b1-b6, the candidate feature extraction functions with reward scores higher than those recorded in the current state information can be used as example context data. For example, if a1-a6 and b1-b6 are sorted according to the reward scores, resulting in b1-b2-b3-b4-a1-a2-b5-a3-a4-a5-a6-b6, then b1, b2, b3, and b4 can be used as example context data.

[0078] Step S107: Update the context data information in the current status information according to the example context data information, and update the instruction information in the current status information according to the improved instruction information, so as to update the current status information of the target interface.

[0079] In this embodiment, after determining the example context data information, the example context data information can be used to replace the context data information in the current state information, thereby updating the context data information in the current state information. Improved instruction information can be used to replace the instruction information in the previous state information to achieve the current state information of the target interface.

[0080] In some embodiments, updating the context data information in the current state information based on example context data information includes: Based on the example context data, determine the next-level context data in the current state information. Replace the next-level context data in the current state information with the example context data to update the context data in the current state information.

[0081] The reward value for the lower-level context data information is lower than the reward value for the example context data information.

[0082] In other words, after obtaining the example context data, the context data in the current state information can be updated using the example context data. At this point, based on the example context data, context data with a reward value lower than the reward value in the example context data can be identified in the current state information and designated as the lower-level context data. The example context data then replaces the first context data in the current state information, thus updating the context data in the current state information.

[0083] As described in the example above, after determining b1, b2, b3, and b4 as example context data, the six candidate feature extraction functions whose scores are lower than those in the example context data (a3, a4, a5, and a6) recorded in the current state information of the target interface are identified as lower-level context data. The lower-level context data in the current state information is then replaced with the new context data. The updated context data of the target interface's current state information includes the candidate feature extraction functions b1, b2, b3, b4, a1, and a2, along with their corresponding reward scores.

[0084] In some embodiments, the J group of candidate feature extraction functions with high reward values ​​and the context data information in the current state information can be fused to obtain the context data information required for the next iteration, i.e., through the formula... , to be integrated. Among them, use Replace the J context data items that are least relevant to the current target interface description. .

[0085] Step S108: Re-execute the steps based on the updated current state information of the target interface. Based on the current state information of the target interface, generate multiple candidate feature extraction functions and steps. Based on the preset dataset, evaluate the multiple candidate feature extraction functions and obtain the reward values ​​corresponding to the multiple candidate feature extraction functions until the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions.

[0086] In this embodiment, after updating the current state information of the target interface, steps S102 and S103 can be re-executed based on the updated current state information of the target interface to achieve the next iteration. That is, candidate feature extraction functions can be regenerated based on the updated current state information of the target interface. Since the current state information has been updated, the regenerated candidate feature extraction function is different from the previously generated candidate feature extraction function. At this time, the newly generated candidate feature extraction function can be re-evaluated to obtain the reward value corresponding to the regenerated candidate feature extraction function. See step S103 for details, which will not be elaborated here. After obtaining the reward values ​​corresponding to multiple candidate feature extraction functions, it can be checked whether the newly obtained reward values ​​of the multiple candidate feature extraction functions meet the preset conditions. If not, the current state information of the target interface needs to be updated again based on the reward values ​​and the multiple candidate feature extraction functions, and steps S102 and S103 need to be re-executed until the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions.

[0087] It should be understood that steps S104-S105 and steps S106-S108 are parallel schemes. When executing steps S104-S105, steps S106-S108 are not executed.

[0088] In this way, when the stored context data is empty, a preset second context data can be used as the context information in the current state information of the target interface. During the iteration process, example context data can be generated based on the comparison results between multiple candidate feature extraction functions and their corresponding reward values ​​and the context information in the current state information. The example context data is then used to update the context data in the current state information of the target interface. With continuous iteration, searchable context data can be gradually formed, achieving bootstrapping evolution. Furthermore, in the above process, during the regeneration of candidate feature extraction functions, only the example context data and improvement instruction information are used to update the current state information of the target interface, without adjusting the model parameters. This achieves continuous learning and self-evolution without model fine-tuning, thereby reducing the dependence on human experience in constructing ticketing risk control feature extraction functions, improving the adaptability to business changes, and realizing the automated accumulation and evolution of feature extraction functions. In other words, this embodiment of the application can implement a context data information reinforcement learning mechanism, which does not perform gradient updates on model parameters, but achieves the updating and evolution of context data information through continuous iterative updates. On the other hand, the embodiments of this application can realize parallel exploration and reward-driven strategy attribution. That is, multiple candidate feature extraction functions can be generated in parallel for the same target interface, reward values ​​are obtained through a unified evaluation environment, and a meta-inference model compares and analyzes candidate feature extraction functions with high and low reward values ​​to summarize success and failure pattern information. Specifically, the feature extraction function generation model generates candidate feature extraction functions based on the target interface description information, current instruction information, and contextual data information. The meta-inference model analyzes the merits of the feature vector generation strategy based on the reward values ​​of the candidate feature extraction functions and generates improved instruction information in natural language form. This improved instruction information is used to improve the feature vector generation strategy, thereby regenerating candidate feature extraction functions, forming a closed loop of "retrieval → generation → evaluation → attribution → update → accumulation," as referenced. Figure 3 As shown.

[0089] In some embodiments, reference Figure 2 As shown, the above method also includes: Step S109: When the reward values ​​corresponding to multiple candidate feature extraction functions meet the preset conditions, the target feature extraction function and its corresponding reward value are stored as context data information.

[0090] In this embodiment of the application, in order to generate candidate feature extraction functions more accurately when performing risk control on other interfaces in the future, the target feature extraction function and its corresponding reward value can be stored as context data information when the reward values ​​corresponding to multiple candidate feature extraction functions meet the preset conditions.

[0091] Figure 4 This is a schematic diagram of a ticketing risk control processing device provided in an embodiment of this application. Figure 4 As shown, the ticketing risk control processing device includes: The acquisition module 401 is used to acquire the current status information of the target interface.

[0092] The target interface is the interface that requires risk control. The pre-state information includes the description information of the target interface, the instruction information used to describe the risk control characteristics, and the context data information corresponding to the description information.

[0093] The generation module 402 is used to generate multiple candidate feature extraction functions based on the current state information of the target interface.

[0094] Among them, the candidate feature extraction function is used to generate a set of candidate feature vectors; the set of candidate feature vectors is used to describe the risk control features of the request messages received by the target interface.

[0095] The processing module 403 is also used to evaluate multiple candidate feature extraction functions based on a preset dataset and obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0096] The processing module 403 is also used to determine the multiple candidate feature extraction functions as the target feature extraction function when the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions, or to determine the target feature extraction function according to the current state information; and to extract features from the request message received by the target interface through the target feature extraction function to obtain the feature vector corresponding to the request message.

[0097] In some embodiments, the processing module 403 is further configured to determine example context data information and improvement instruction information based on the multiple candidate feature extraction functions and their corresponding reward values ​​when the reward values ​​corresponding to the multiple candidate feature extraction functions do not meet the preset conditions.

[0098] Based on the example context data, update the context data in the current state information, and based on the improved instruction information, update the instruction information in the current state information to update the current state information of the target interface.

[0099] Based on the updated current state information of the target interface, the steps are re-executed. Based on the current state information of the target interface, multiple candidate feature extraction functions and steps are generated. Based on a preset dataset, the multiple candidate feature extraction functions are evaluated, and the reward values ​​corresponding to the multiple candidate feature extraction functions are obtained until the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions.

[0100] In some embodiments, the processing module 403 is specifically configured to, based on multiple candidate feature extraction functions and their corresponding reward values, identify candidate feature extraction functions and their corresponding reward values ​​whose reward values ​​are higher than the reward values ​​in the context data information of the current state information as example context data information. Based on the example context data information, the lower-level context data information is determined in the current state information. The lower-level context data information in the current state information is replaced with the example context data information to update the context data information in the current state information.

[0101] The reward value for the lower-level context data information is lower than the reward value for the example context data information.

[0102] In some embodiments, the processing module 403 is specifically used to obtain multiple sets of candidate risk control feature vectors based on a preset dataset and using multiple candidate feature extraction functions; and to perform reward analysis on the multiple candidate feature extraction functions based on the multiple sets of candidate risk control feature vectors to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0103] In some embodiments, the preset dataset includes a first dataset and a second dataset.

[0104] The processing module 403 is specifically used to obtain multiple sets of candidate risk control feature vectors based on the first dataset using multiple candidate feature extraction functions.

[0105] Based on the second dataset, multiple candidate feature extraction functions are used to obtain multiple sets of risk control feature vectors for training; based on the multiple sets of risk control feature vectors for training, multiple risk control models are trained and obtained respectively; based on the multiple sets of candidate risk control feature vectors and multiple risk control models, the risk control result information of each risk control model is obtained; based on the risk control result information corresponding to the multiple sets of candidate risk control feature vectors, reward analysis is performed to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

[0106] In some embodiments, the acquisition module 401 is specifically used to acquire description information of the target interface; based on the description information of the target interface, if first context data information exists in the stored context data information, then the first context data information is determined as the context data information corresponding to the description information. Preset instruction information is determined as the instruction information corresponding to the target interface.

[0107] The first context data information is the context data information that has the highest similarity to the description information of the target interface among the stored context data information, and whose similarity value is greater than a preset similarity threshold.

[0108] In some embodiments, the processing module 403 is further configured to determine the preset second context data information as the context data information corresponding to the description information if the first context data information does not exist in the stored context data information.

[0109] In some embodiments, the preset conditions include: the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface, and the number of times the highest reward value corresponding to multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface reaches a preset number threshold.

[0110] In some embodiments, the processing module 403 is further configured to store the target feature extraction function and its corresponding reward value as context data information when the reward values ​​corresponding to multiple candidate feature extraction functions meet preset conditions.

[0111] Corresponding to the above embodiments, this application also provides an electronic device. Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 500 may include a processor 501, a memory 502, and a communication unit 503. These components communicate through one or more buses. Those skilled in the art will understand that the structure of the server shown in the figure does not constitute a limitation on the embodiment of the present invention. It may be a bus topology or a star topology, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0112] The communication unit 503 is used to establish a communication channel, enabling the storage device to communicate with other devices. It receives user data from other devices or sends user data to other devices.

[0113] The processor 501 serves as the control center of the storage device, connecting various parts of the electronic device via various interfaces and lines. It executes software programs and / or modules stored in the memory 502, and calls data stored in the memory to perform various functions of the electronic device and / or process data. The processor can be composed of integrated circuits (ICs), such as a single packaged IC or multiple packaged ICs with the same or different functions connected together. For example, the processor 501 may consist only of a central processing unit (CPU). In this embodiment of the invention, the CPU may have a single processing core or include multiple processing cores.

[0114] The memory 502 is used to store the execution instructions of the processor 501. The memory 502 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0115] When the execution instructions in memory 502 are executed by processor 501, the electronic device 500 is able to perform some or all of the steps in the above embodiments.

[0116] In a specific implementation, the present invention also provides a computer storage medium, wherein the computer storage medium may store a program, which, when executed, may include some or all of the steps of the automated mining method for ticket risk control features provided by the present invention. The storage medium may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0117] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions in the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present invention.

[0118] The same or similar parts between the various embodiments in this specification can be referred to mutually. In particular, the device embodiments and terminal embodiments are basically similar to the method embodiments, so the description is relatively simple, and the relevant parts can be referred to the description in the method embodiments.

Claims

1. An automated method for mining ticketing risk control features, characterized in that, include: Obtain the current status information of the target interface; the target interface is the interface that needs to be subject to risk control; the current status information includes the description information of the target interface, the instruction information used to describe the risk control characteristics of the target interface, and the context data information corresponding to the description information; Based on the current state information of the target interface, multiple candidate feature extraction functions are generated; the candidate feature extraction functions are used to extract feature vectors; wherein, the feature vectors are used to describe the risk characteristics of the request messages received by the target interface; Based on a preset dataset, the multiple candidate feature extraction functions are evaluated, and the reward values ​​corresponding to the multiple candidate feature extraction functions are obtained; When the reward values ​​corresponding to the plurality of candidate feature extraction functions meet the preset conditions, the plurality of candidate feature extraction functions are determined as the target feature extraction function or determined as the target feature extraction function according to the current state information; The target feature extraction function is used to extract features from the request message received by the target interface to obtain the feature vector corresponding to the request message.

2. The method according to claim 1, characterized in that, The method further includes: When the reward values ​​corresponding to the multiple candidate feature extraction functions do not meet the preset conditions, the example context data information and improvement instruction information are determined based on the multiple candidate feature extraction functions and their corresponding reward values. Based on the example context data information, update the context data information in the current state information, and based on the improved instruction information, update the instruction information in the current state information to update the current state information of the target interface; Based on the updated current state information of the target interface, the steps are re-executed: generating multiple candidate feature extraction functions and steps based on the current state information of the target interface; evaluating the multiple candidate feature extraction functions based on a preset dataset; obtaining the reward values ​​corresponding to the multiple candidate feature extraction functions; until the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions.

3. The method according to claim 2, characterized in that, The determination of example context data information based on the multiple candidate feature extraction functions and their corresponding reward values ​​includes: Based on the multiple candidate feature extraction functions and their corresponding reward values, the candidate feature extraction functions and their corresponding reward values ​​that have a reward value higher than the reward value in the context data information of the current state information are determined as example context data information. The step of updating the context data information in the current state information based on the example context data information includes: Based on the example context data information, lower-level context data information is determined from the current state information; wherein, the reward value of the lower-level context data information is lower than the reward value of the example context data information; The example context data information is used to replace the lower-level context data information in the current state information to update the context data information in the current state information.

4. The method according to claim 1, characterized in that, The step of evaluating the multiple candidate feature extraction functions based on a preset dataset and obtaining the reward values ​​corresponding to the multiple candidate feature extraction functions includes: Based on a preset dataset, multiple sets of candidate risk control feature vectors are obtained using the multiple candidate feature extraction functions. Based on the multiple sets of candidate risk control feature vectors, reward analysis is performed on the multiple candidate feature extraction functions to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

5. The method according to claim 3 or 4, characterized in that, The preset dataset includes a first dataset and a second dataset; The step of obtaining multiple sets of candidate risk control feature vectors based on a preset dataset and using the multiple candidate feature extraction functions includes: Based on the first dataset, multiple candidate feature extraction functions are used to obtain multiple sets of candidate risk control feature vectors; The step of performing reward analysis on the multiple candidate feature extraction functions based on the multiple sets of candidate risk control feature vectors, and obtaining the reward values ​​corresponding to the multiple candidate feature extraction functions, includes: Based on the second dataset, multiple sets of risk control feature vectors for training are obtained using the multiple candidate feature extraction functions. Based on the multiple sets of training risk control feature vectors, multiple risk control models are trained and obtained respectively. Based on the multiple sets of candidate risk control feature vectors and the multiple risk control models, obtain the risk control result information of each of the risk control models; Based on the risk control results, a reward analysis is performed to obtain the reward values ​​corresponding to the multiple candidate feature extraction functions.

6. The method according to claim 1, characterized in that, The process of obtaining the current status information of the target interface includes: Obtain the description information of the target interface; Based on the description information of the target interface, if there is first context data information in the stored context data information, then the first context data information is determined as the context data information corresponding to the description information; wherein, the first context data information is the context data information in the stored context data information that has the highest similarity to the description information of the target interface and the similarity value is greater than a preset similarity threshold. The preset instruction information is determined as the instruction information corresponding to the target interface.

7. The method according to claim 6, characterized in that, Also includes: If the first context data information is not present in the stored context data information, then the preset second context data information is determined as the context data information corresponding to the description information.

8. The method according to claim 1, characterized in that, The preset conditions include: the highest reward value corresponding to the multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface, and the number of times the highest reward value corresponding to the multiple candidate feature extraction functions does not exceed the highest reward value in the context data information of the current state information of the target interface reaches a preset number threshold.

9. The method according to claim 1, characterized in that, The method further includes: When the reward values ​​corresponding to the multiple candidate feature extraction functions meet the preset conditions, the target feature extraction function and its corresponding reward value are stored as context data information.

10. An automated device for mining ticketing risk control features, characterized in that, include: The acquisition module is used to acquire the current status information of the target interface; the target interface is the interface that needs to be subject to risk control; the current status information includes the description information of the target interface, the instruction information for describing the risk control characteristics, and the context data information corresponding to the description information; The generation module is used to generate multiple candidate feature extraction functions based on the current state information of the target interface; the candidate feature extraction functions are used to generate feature vectors. The feature vector is used to describe the risk control characteristics of the request message received by the target interface; The processing module is used to evaluate the multiple candidate feature extraction functions based on a preset dataset and obtain the reward values ​​corresponding to the multiple candidate feature extraction functions; The processing module is further configured to determine the multiple candidate feature extraction functions as target feature extraction functions when the reward values ​​corresponding to the multiple candidate feature extraction functions meet preset conditions, and to extract features from the request message received by the target interface through the target feature extraction function to obtain the feature vector corresponding to the request message.

11. An electronic device, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device performs the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the method according to any one of claims 1 to 9.

Images