Information storage method and device based on distributed cluster, equipment and medium

Abstract

The embodiment of the invention discloses an information storage method and device based on a distributed cluster, equipment and a medium. A specific embodiment of the method comprises the following steps: executing a storage step: performing model training on at least one initial risk control detection model to obtain at least one risk control detection model; executing a second storage step: in response to the condition that at least one piece of model performance information meets a target performance condition, storing the residual feature set and risk control detection task information into a target model feature database; in response to the situation that the target performance condition is not met, executing an execution step: screening out features to be removed; removing corresponding feature data from the training data set; and taking the first removed data set as a training data set, taking the residual feature set as a full-quantity feature set, and executing a storage step. According to the embodiment, the feature set corresponding to the risk control detection task is efficiently and accurately generated on the basis of fully utilizing computing resources by utilizing the master node set and the slave node set corresponding to the distributed cluster.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to information storage methods, apparatuses, devices, and media based on distributed clusters. Background Technology

[0002] Currently, with the continuous development of artificial intelligence, deep learning networks are being applied in various fields to perform various tasks (e.g., risk control detection tasks). The common approach to determining the feature set for a target task is to manually select features based on the model's training performance.

[0003] However, when using the above method, the following technical problems often arise: The model needs to be trained multiple times sequentially, resulting in a large number of training iterations to determine which features are strongly correlated with the task. This leads to long training times, consumes a lot of computing resources, and results in low efficiency and inaccurate subjective judgments from users.

[0004] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0006] Some embodiments of this disclosure propose information storage methods, apparatuses, electronic devices, and computer-readable media based on distributed clusters to solve one or more of the technical problems mentioned in the background section above.

[0007] In a first aspect, some embodiments of this disclosure provide an information storage method based on a distributed cluster, comprising: obtaining a training dataset for at least one initial risk control detection model from a target model database; for the training dataset, performing a storage step: training the at least one initial risk control detection model using the slave node set corresponding to the distributed cluster based on the training dataset to generate at least one risk control detection model, wherein each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node; in response to at least one model performance information corresponding to the at least one risk control detection model satisfying the target performance condition, storing the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information to the target model feature database using the master node corresponding to the distributed cluster; in response to not satisfying the target performance condition, performing an execution step using the master node: filtering out features to be removed from the remaining feature set based on at least one model performance information to obtain at least one first target feature; removing feature data corresponding to at least one first target feature from the training dataset to obtain a first removed dataset; using the first removed dataset as the training dataset and the remaining feature set as the full feature set, and continuing to perform the above storage step.

[0008] Secondly, some embodiments of this disclosure provide an information storage device based on a distributed cluster, comprising: an acquisition unit configured to acquire a training dataset for at least one initial risk control detection model from a target model database; and a first execution unit configured to perform a storage step for the training dataset: training the at least one initial risk control detection model using a set of slave nodes corresponding to the distributed cluster, based on the training dataset, to generate at least one risk control detection model, wherein each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node; and responding to the performance of at least one model corresponding to the at least one risk control detection model. If the information meets the target performance conditions, the master node corresponding to the above-mentioned distributed cluster stores the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database. The second execution unit is configured to, in response to the failure to meet the above-mentioned target performance conditions, use the master node to execute the following steps: based on at least one model performance information, select the features to be removed from the remaining feature set to obtain at least one first target feature; remove the feature data corresponding to at least one first target feature from the training dataset to obtain the first removed dataset; use the first removed dataset as the training dataset and the remaining feature set as the full feature set, and continue to execute the above storage steps.

[0009] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, such that when the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any implementation of the first aspect.

[0010] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method as described in any implementation of the first aspect.

[0011] The above embodiments of this disclosure have the following beneficial effects: Through the information storage method based on a distributed cluster in some embodiments of this disclosure, the feature set corresponding to the risk control detection task is generated efficiently and accurately by utilizing the master and slave node sets corresponding to the distributed cluster, while fully utilizing computing resources. Specifically, the reason for the low efficiency and inaccuracy in related determinations is that the model needs to be trained multiple times sequentially, resulting in a large number of training iterations to determine which features are strongly correlated with the task. This leads to long training times, long-term consumption of computing resources, low determination efficiency, and inaccurate subjective judgments by users. Based on this, the information storage method based on a distributed cluster in some embodiments of this disclosure first obtains a training dataset for at least one initial risk control detection model from the target model database, using it as the basic dataset to filter features by training at least one initial risk control detection model. Then, for the training dataset, a storage step is performed: First, based on the training dataset, the at least one initial risk control detection model is trained using the slave node set corresponding to the distributed cluster to generate at least one risk control detection model. Each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node. Here, through the master and slave node sets in the distributed cluster, at least one initial risk control detection model can be trained simultaneously under different feature inputs, so as to determine the importance of each feature in the risk control detection process based on subsequent model performance information. By setting at least one corresponding slave node for each risk control detection model, the normal training of the risk control detection model can be guaranteed. Second, in response to the at least one model performance information corresponding to the above-mentioned at least one risk control detection model meeting the target performance condition, the master node corresponding to the above-mentioned distributed cluster stores the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database, so as to be called and used in subsequent risk control detection scenarios, and to determine which of the remaining feature sets are key features based on the performance information of at least one model. Next, in response to the failure to meet the above target performance condition, the master node executes the following steps: First, based on the performance information of at least one model, the features to be removed can be accurately selected from the remaining feature set as the first target feature, obtaining at least one first target feature. Second, the feature data corresponding to at least one first target feature is removed from the training dataset, obtaining the first removed data, so as to facilitate further model training and further filtering of useless features. The third step is to use the first removed dataset as the training dataset and the remaining feature set as the full feature set, and then continue to perform the above storage steps.In summary, by scheduling node tasks in the master and slave node sets of a distributed cluster and by using model performance information to determine useful and useless features, not only is the efficiency of feature selection improved, but also a large amount of computing resources can be avoided for extended periods. Attached Figure Description

[0012] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0013] Figure 1 This is a flowchart of some embodiments of the information storage method based on a distributed cluster according to the present disclosure; Figure 2 This is a schematic diagram of the structure of some embodiments of the information storage device based on a distributed cluster according to the present disclosure; Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0014] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0015] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0016] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0017] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0018] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0019] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] refer to Figure 1 The diagram illustrates a flow 100 of some embodiments of the information storage method based on a distributed cluster according to the present disclosure. This information storage method based on a distributed cluster includes the following steps: Step 101: Obtain the training dataset for at least one initial risk control detection model from the target model database.

[0021] In some embodiments, in response to receiving session communication request information for a first user terminal and a second user terminal, the execution entity of the above-described distributed cluster-based information storage method (i.e., the distributed cluster) can obtain a training dataset for at least one initial risk control detection model from the target model database. The target model database can be a database storing model-related data. In practice, model-related data can include: at least one initial model for various task information and training datasets for various task information. In practice, task information can be task identifiers for tasks to be executed. For example, each task information can include, but is not limited to, at least one of the following: image recognition task information, risk control detection task information, and credit prediction task information. Each task information has a corresponding pre-acquired training dataset to facilitate model training for at least one pair of initial models. The initial model can be a neural network model that has not yet been fully trained. For the same task information, the model structures of the initial models in the at least one initial model can be different or partially the same. The initial risk control detection model can be a neural network model for risk control detection that has not yet been fully trained. In practice, for the credit reporting field, the corresponding risk control detection can be user risk control detection for pre-loan products. At least one initial risk control detection model may have some structural similarities among its constituent models. The training dataset may be a dataset used to train at least one initial risk control detection model.

[0022] In some optional implementations of certain embodiments, after step 101, the steps further include: The first step involves the aforementioned execution entity performing data preprocessing on each training data point in the training dataset to generate a processed dataset. This processed data includes the feature dataset corresponding to the full feature set. Data preprocessing may include, but is not limited to, one of the following: data cleaning, data standardization, data normalization, class encoding, data augmentation, and data balancing.

[0023] The second step is for the aforementioned executing entity to determine the features to be removed from the full feature set based on information theory indicators and statistical verification, and to obtain at least one second target feature.

[0024] In practice, information theory metrics here can include: Mutual Information (MI), Conditional Mutual Information (MI), and Symmetrical Uncertainty (SU). Mutual Information measures the non-linear dependency between features and the target variable, such as `mutual_info_classif`. Conditional Mutual Information can evaluate the new information content of candidate features given the selected features (e.g., the CMIM algorithm). Symmetrical Uncertainty can normalize mutual information and address the problem of multi-valued feature bias (e.g., the FCBF algorithm). Statistical validation here can be one of the following: Chi-square test or F-test. The Chi-square test assesses the independence of discrete features from classification labels and is suitable for classification problems. The F-test analyzes the linear correlation between features and continuous labels using ANOVA and is suitable for regression or classification tasks.

[0025] Third, the aforementioned execution entity can remove the feature data corresponding to at least one of the second target features from the aforementioned processed dataset to obtain the removed dataset, which can then be used as the training dataset.

[0026] In some optional implementations of certain embodiments, the executing entity may obtain a training dataset for at least one initial risk control detection model from a target model database, including the following steps: The first step is to determine the model application scenario group for each of the at least one initial risk control detection model, thus obtaining at least one model application scenario group. The model application scenario can be an application scenario of the initial risk control detection model. For example, for at least one initial risk control detection model including initial risk control detection model A, initial risk control detection model B, and initial risk control detection model C, the model application scenario group corresponding to initial risk control detection model A can include: credit risk management scenario, fraudulent transaction identification scenario, and account security protection scenario. The model application scenario group corresponding to initial risk control detection model B can include: investment risk early warning scenario. The model application scenario group corresponding to initial risk control detection model C can include: equipment security operation and maintenance scenario.

[0027] The second step involves deduplicating each model application scenario in the at least one model application scenario group to obtain a model application scenario set. No two scenarios are identical within this set.

[0028] Third, for each model application scenario in the model application scenario set, select data groups from the target model database that are associated with the scenario identifier corresponding to the model application scenario. The scenario identifier can be the identity information of the model application scenario. The scenario identifier can be in numerical form.

[0029] As an example, the aforementioned execution entity can query and filter data identifier groups that are associated with the scenario identifiers corresponding to the application scenario of the model from the data association table. Then, it can filter data groups that are associated with the data identifier groups from the target model database.

[0030] The fourth step is to deduplicate the obtained dataset to obtain a deduplicated dataset, which will serve as the training dataset.

[0031] Step 102, for the training dataset, perform the storage step: Step 1021: Based on the training dataset, use the slave node set corresponding to the distributed cluster to train at least one initial risk control detection model to generate at least one risk control detection model.

[0032] In some embodiments, the aforementioned execution entity can train at least one initial risk control detection model based on the training dataset and utilizing the set of slave nodes corresponding to the distributed cluster, thereby generating at least one risk control detection model. The distributed cluster can be a server in a distributed structure. The distributed cluster has a corresponding set of master nodes (i.e., master service nodes) and slave nodes (slave service nodes). Each node in the distributed cluster has a deployed Graphics Processing Unit (GPU), and each risk control detection model has at least one corresponding slave node. The input features corresponding to each risk control detection model are different.

[0033] As an example, the aforementioned execution entity can train at least one initial risk control detection model based on the training dataset and the corresponding slave node set of the distributed cluster, using conventional model training methods, to generate at least one risk control detection model. In this model, the input features of each initial risk control detection model increase sequentially.

[0034] In some optional implementations of certain embodiments, training the at least one initial risk control detection model using the set of slave nodes corresponding to the distributed cluster based on the training dataset to generate at least one risk control detection model may include the following steps: The first step is to obtain the feature description information set corresponding to the full feature set and the risk control detection task information. There is a one-to-one correspondence between the features in the full feature set and the feature description information in the feature description information set. The feature description information can be information that describes various aspects of a feature. For example, the feature description information may include: the feature calculation process, the feature function, the feature identifier, and the feature complexity.

[0035] The second step involves generating prompts that determine the importance of features in the full feature set relative to the corresponding risk control detection task information, based on the aforementioned feature description information set. Importance represents the contribution of the corresponding feature content to the risk control detection and prediction process (i.e., whether the corresponding feature content is valid content). Importance can be a value between 0 and 1; a higher value indicates a greater contribution of the corresponding feature content to the risk control detection and prediction process. The generated prompts can be keywords indicating the importance level.

[0036] As an example, the aforementioned execution entity can input the feature description information set and the full feature set into the initial generation prompt information to obtain the generation prompt information.

[0037] The third step involves inputting the generated prompt information into the large language model deployed on the master node to generate the feature importance set corresponding to the full feature set. The large language model can be a pre-deployed large language model on the master node. For example, the large language model can be an LLM model. In practice, the large language model deployed on the master node can be a lightweight large language model. This large language model can be a pre-trained large language model based on feature processing knowledge. In practice, the large language model can be a Transformer-based model. There is a one-to-one correspondence between the features in the full feature set and the feature importance set. The feature importance can represent the importance of a feature in the risk control detection and prediction process.

[0038] Fourth, based on the aforementioned feature importance set, the entire feature set is divided into a necessary feature set, a pending feature set, and a non-necessary feature set. The feature importance of necessary features is higher than that of pending features. The feature importance of pending features is higher than that of non-necessary features. Necessary features are those that must be considered in the risk control detection and prediction process. Pending features are those that require further evaluation in the risk control detection and prediction process, but are likely to be considered. Non-necessary features are those that require further evaluation in the risk control detection and prediction process, but are likely to be considered.

[0039] As an example, the aforementioned implementing entity can divide the entire feature set into multiple feature subsets according to the feature importance intervals corresponding to necessary features, undetermined features, and unnecessary features, respectively. These subsets can be designated as the necessary feature set, the undetermined feature set, and the unnecessary feature set. The feature importance intervals corresponding to necessary features, undetermined features, and unnecessary features can be numerical intervals set based on historical experience.

[0040] The fifth step involves dividing the set of undetermined features into at least one group according to their importance from highest to lowest, resulting in a sequence of undetermined feature groups. Each subsequent undetermined feature group includes the preceding undetermined feature group, and also includes the undetermined features following the least important feature from the preceding undetermined feature group. For example, consider the undetermined feature set as: Feature 1, Feature 2, Feature 3, Feature 4, Feature 5, Feature 6, and Feature 7. Dividing them according to their importance from highest to lowest yields undetermined feature group A, undetermined feature group B, and undetermined feature group C. Undetermined feature group A can be {Feature 1, Feature 2}. Undetermined feature group B can be {Feature 1, Feature 2, Feature 3, Feature 4}. Undetermined feature group C can be {Feature 1, Feature 2, Feature 3, Feature 4, Feature 5, Feature 6, Feature 7}.

[0041] Step 6: Extract the feature dataset corresponding to the necessary feature set from the training dataset above, as the necessary feature dataset. There is a one-to-one correspondence between the necessary features in the necessary feature set and the feature data in the feature dataset. The necessary feature data can be the feature content corresponding to the necessary features.

[0042] Step 7: Extract the feature data sequence corresponding to the above-mentioned undetermined feature group sequence from the above training dataset. There is a one-to-one correspondence between the undetermined features in the undetermined feature group sequence and the feature data in the feature data sequence.

[0043] Step 8: Extract the feature dataset corresponding to the aforementioned non-essential feature set from the training dataset, and use it as the non-essential feature dataset. There is a one-to-one correspondence between the non-essential features in the non-essential feature set and the feature data in the feature dataset.

[0044] Step 9: Using the aforementioned necessary feature dataset as the first training dataset, train the first initial risk control detection model in at least one of the aforementioned initial risk control detection models to generate the first risk control detection model.

[0045] Step 10: For each feature data group in the above feature data group sequence, use the above feature data group and the above necessary feature dataset as the second training dataset, and train the second initial risk control detection model in the above at least one initial risk control detection model to generate the second risk control detection model.

[0046] Step 11: Using the aforementioned necessary feature dataset and the aforementioned non-necessary feature dataset as the third training dataset, train the third initial risk control detection model in the above at least one initial risk control detection model to generate the third risk control detection model.

[0047] Step 12: The first risk control detection model, the obtained second risk control detection model sequence, and the third risk control detection model are identified as at least one of the above-mentioned risk control detection models.

[0048] In some optional implementations of certain embodiments, the executing entity may train the at least one initial risk control detection model based on the training dataset and utilizing the set of slave nodes corresponding to the distributed cluster, to generate at least one risk control detection model, including the following steps: The first step, for each of the at least one initial risk control detection model, is to perform the following training steps: Sub-step 1: Select a subset of training data from the training dataset that corresponds to the initial risk control detection model.

[0049] As an example, the aforementioned execution entity can select a subset of training data corresponding to the initial risk control detection model from the training dataset based on the correspondence between training data and model application scenarios, and the correspondence between model application scenarios and the initial risk control detection model.

[0050] Sub-step 2: Determine the slave node corresponding to the initial risk control detection model as the target slave node.

[0051] Sub-step 3 involves retrieving the data backup information corresponding to the target slave node from the cache of the master node in the distributed cluster. The data backup information may be data description information of each piece of data stored on the target slave node. The master node in the distributed cluster is a server node used to schedule and manage the data or tasks corresponding to each slave node.

[0052] Sub-step 4: Select data from the training data subset that does not exist in the data backup information as target training data to obtain the target training data subset.

[0053] Sub-step 5: For each target training data point in the subset of target training data, perform the following data processing steps: The first sub-step involves determining, in response to the determination that the target training data is in text format, whether there exists data in the dataset corresponding to the backup data that has a data similarity higher than the target training data. Here, data similarity can characterize the content differences between data items. In practice, data similarity can be determined through character comparison or cosine similarity between semantic vectors. Target similarity can be a numerical value used to measure whether data items are similar. That is, a similarity higher than the target indicates a larger semantic difference between the two data items, while a similarity lower than the target indicates a smaller semantic difference between the two data items.

[0054] The second sub-step, in response to the determination of existence, involves identifying textual difference information between the target training data and data with a similarity higher than the target. This textual difference information can be information about differences in text content. Specifically, the textual difference information may include: data identifiers and content difference information of the target training data.

[0055] As an example, the aforementioned execution entity can utilize a lightweight large language model deployed on the master node to determine textual difference information between the target training data and data with higher similarity to the target. Here, the lightweight large language model can be a large language model with relatively few training iterations and a relatively simple network structure.

[0056] The third sub-step involves determining the target training data as textual difference information in response to the determination that it does not exist.

[0057] Sub-step 6 involves sending the obtained text difference information subset to the target slave node, so that the target slave node can obtain the target training data subset based on the text difference information subset. Here, the target slave node recovers the target training data subset based on the text difference information subset.

[0058] Sub-step 7 instructs the target slave node to train the initial risk control detection model based on the dataset corresponding to the data backup information and the target training data subset, so as to generate a risk control detection model.

[0059] In some optional implementations of certain embodiments, after step 1021, the steps further include: Using the master node corresponding to the above distributed cluster, perform the following generation steps: Sub-step 1: Based on the training dataset corresponding to the aforementioned full feature set, train the initial full feature risk control detection model corresponding to the aforementioned master node to generate the full feature risk control detection model. The training dataset corresponds to each feature data within the full feature set.

[0060] Sub-step 2 involves determining at least one remaining feature set under the current storage step. Each remaining feature set corresponds one-to-one with a risk control detection model within the at least one risk control detection model. The remaining feature sets can be those suitable for training the risk control detection model at the current time step of the storage step. That is, subsequent feature data generation and risk control detection model training are based on the remaining feature sets.

[0061] Sub-step 3 involves extracting the corresponding model parameter information from the full-feature risk control detection model for each of the at least one remaining feature group. The model parameter information can be the processing parameters used in the full-feature risk control detection model to process each feature in the remaining feature group.

[0062] As an example, the aforementioned execution entity can extract parameter information groups whose keys are the remaining feature groups from the full-feature risk control detection model, and use them as model parameter information.

[0063] Sub-step 4: Send each of the at least one model parameter information obtained to the corresponding slave node.

[0064] Optionally, the aforementioned execution entity can train at least one initial risk control detection model based on the training dataset and the corresponding slave node set of the distributed cluster, to generate at least one risk control detection model, including the following steps: For each of the at least one initial risk control detection model mentioned above, perform the following training steps: Sub-step 1: Determine at least one slave node and the remaining feature group corresponding to the above initial risk control detection model, as the target remaining feature group.

[0065] Sub-step 2 involves generating an inbound risk control detection model corresponding to the initial risk control detection model using at least one model parameter information received from at least one slave node. The inbound risk control detection model can be a risk control detection model with the model parameter information replaced.

[0066] As an example, the aforementioned execution entity can utilize at least one model parameter information received from at least one slave node to generate the migration risk control detection model corresponding to the initial risk control detection model, based on the parameter migration method corresponding to transfer learning.

[0067] Sub-step 3: Select the feature dataset corresponding to the remaining feature group of the target from the above training dataset as the target feature dataset.

[0068] Sub-step 4: Select the subset of feature data corresponding to the current storage step from the above target feature dataset.

[0069] Sub-step 5 involves training the migration risk control detection model based on the aforementioned subset of feature data to obtain the risk control detection model. The specific implementation details are omitted here.

[0070] Here, the above content serves as another inventive point of this disclosure, solving another technical problem: "Each model training session wastes a significant amount of training resources and has a long training cycle. If each slave node trains the model separately, retraining based on the training data is required, leading to a significant waste of training resources." Based on this, this application first uses the master node to train the model under the full feature set to obtain a full-feature risk control detection model including all features. On this basis, model parameter information corresponding to each of at least one remaining feature group is determined. Thus, slave nodes can perform preliminary model adjustments based on the model parameter information, enabling subsequent model training based on the adjusted model (i.e., the imported risk control detection model), which can significantly shorten training time and reduce the waste of training resources.

[0071] Step 1022: Utilize the master node corresponding to the aforementioned distributed cluster to execute the second storage step: Step 10221: In response to the fact that the performance information of at least one model corresponding to at least one risk control detection model meets the target performance condition, the master node corresponding to the distributed cluster is used to store the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database.

[0072] In some embodiments, in response to the at least one model performance information corresponding to the at least one risk control detection model satisfying the target performance condition, the executing entity can utilize the master node corresponding to the distributed cluster to store the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database. There is a one-to-one correspondence between the risk control detection model in the at least one risk control detection model and the model performance information in the at least one model performance information. The model performance information can characterize the detection performance corresponding to the risk control detection model. In practice, model performance information may include, but is not limited to, at least one of the following: accuracy, precision, recall, F1 score, ROC (Receiver Operating Characteristic curve) - AUC (Area Under Curve). The target performance condition can vary depending on the settings of the model input features corresponding to the at least one risk control detection model. For example, for risk control detection models where the model input features increase sequentially, the corresponding target performance condition could be that the model performance of the at least one model performance information improves with the increase of these input features. For example, for risk control detection models where the input features increase sequentially, the corresponding target performance condition could be a condition where, with the increase of input features, the performance of at least one model performance information set exhibits a linear relationship. Here, the remaining feature set is the feature set after removing the first target feature from the full feature set. In this case, during the initial execution of the second storage step, the first target feature is empty. In subsequent executions of the second storage step, the first target feature is the feature determined in the previous second storage step to be removed. The full feature set can be the complete feature set of the entire risk control detection task.

[0073] As an example, firstly, the executing entity can obtain the corresponding validation dataset from the target model database. Then, based on the validation dataset, it generates model performance information for each of the at least one risk control detection model, thus obtaining at least one model performance information.

[0074] As an example, the aforementioned executing entity can store the remaining feature set and the corresponding risk control detection task information in the target model feature database in key-value pairs.

[0075] In some optional implementations of certain embodiments, the above-mentioned full feature set is generated through the following steps: The first step is to obtain at least one similar task information corresponding to the aforementioned risk control detection task information. This similar task information can be task information with similar relationships in terms of corresponding task content or task type.

[0076] As an example, the aforementioned executing entity can obtain at least one similar task information corresponding to the aforementioned risk control detection task information by querying the task association knowledge graph.

[0077] The second step is to query at least one similar feature set corresponding to the at least one similar task information from the aforementioned target model feature database. There is a one-to-one correspondence between the similar task information in the at least one similar task information and the similar feature set in the at least one similar feature set. The similar feature set can be any feature that must be considered during the task execution process corresponding to the similar task information.

[0078] The third step is to obtain the first initial full feature set corresponding to the aforementioned risk control detection task information. This first initial full feature set is based on data collected and uploaded manually by experienced personnel. Specifically, the first initial full feature set can be obtained by experienced personnel through research and summarization of various materials.

[0079] The fourth step is to generate feature generation prompts that represent the full feature set of the risk control detection task information described above. These prompts can be keywords used to generate the full feature set.

[0080] The fifth step is to input the aforementioned feature generation prompts into the large language model deployed on the aforementioned master node to generate the second initial full feature set.

[0081] Step 6: Generate feature recommendation suggestions for the first initial full feature set, at least one task information difference between the risk control detection task information and at least one similar task information, and risk control detection task scenario information. The feature recommendation suggestions may be feature sets recommended from at least one initial full feature set that are suitable for the risk control detection scenario. The task information difference may be the task differences between the risk control detection task information and similar task information. The risk control detection task scenario information may be scenario-related content within the risk control detection task scenario. For example, the risk control detection task scenario information may include: basic knowledge corpus of the risk control detection scenario.

[0082] Step 7: Input the above feature filtering prompts, the differences in at least one task information, and the risk control detection task scenario information into the large language model deployed on the main node to generate the full set of recommended features.

[0083] Step 8: Display the aforementioned first initial full feature set in the first area of ​​the target feature selection interface for feature selection. This first area contains information about the feature sources corresponding to the first initial full feature set. The first area is positioned above the corresponding area in the second area. The target feature selection interface can allow users to select features related to risk control detection. The feature source information can describe the origin of the features. In practice, the feature source information can be one of the following: large model recommendation source, user collection, or similar task source.

[0084] Step nine involves displaying the aforementioned second initial full feature set in the second region of the target feature selection interface for feature selection. This second region includes information about the feature sources corresponding to the second initial full feature set. The second region is positioned above the corresponding region of the third region.

[0085] Step 10: Display the recommended full feature set in the third area of ​​the target feature selection interface for feature selection. This third area includes information about the source of the features corresponding to the recommended full feature set.

[0086] In the eleventh step, in response to receiving the selection information in each area of ​​the target feature selection interface, the selected features are deduplicated to obtain the full feature set.

[0087] Here, the aforementioned "step one to step eleven" serves as another inventive point of this disclosure, solving another technical problem: "how to effectively determine all features is crucial." This disclosure uses database queries, manual queries, large model generation, and interactive interface selection to accurately, efficiently, and visually display all matched features, ensuring that the subsequent training of at least one initial risk control detection model has a precisely matching feature set.

[0088] Step 103: In response to the failure to meet the above target performance conditions, the following execution steps are performed using the master node: Step 1031: Based on at least one model performance information, select features to be removed from the remaining feature set to obtain at least one first target feature.

[0089] In some embodiments, the aforementioned execution entity may, based on at least one model performance information, filter out features to be removed from the remaining feature set to obtain at least one first target feature. The first target feature may be a feature to be removed that does not provide useful information during the risk control detection process.

[0090] As an example, firstly, the aforementioned execution entity can filter out a set of model performance information from at least one set of model performance information whose accuracy exhibits a linear relationship. Then, it removes the aforementioned set of model performance information from the at least one set of model performance information to obtain at least one first target feature.

[0091] In some optional implementations of certain embodiments, the aforementioned execution entity may, based on at least one model performance information, filter out features to be removed from the remaining feature set as first target features to obtain at least one first target feature, including the following steps: The first step is to generate a correlation table representing the relationship between model performance and features, which includes at least one model performance information and at least one set of remaining features. The remaining feature sets in the at least one set of remaining features correspond to the risk control detection models in the at least one risk control detection model. This correlation table represents the corresponding relationships between model performance and features.

[0092] The second step is to generate feature identification prompts based on the above association table and the above remaining feature set to determine important features.

[0093] The third step is to input the aforementioned feature determination prompts into the large language model deployed on the aforementioned master node to generate an important feature set.

[0094] The fourth step is to perform a union operation on the aforementioned important feature set and the aforementioned necessary feature set to generate a union feature set.

[0095] The fifth step is to remove the union feature set from the remaining feature set to generate at least one first target feature.

[0096] Step 1032: Remove at least one feature data corresponding to the first target feature from the training dataset to obtain the first removed dataset.

[0097] In some embodiments, the execution entity may remove feature data corresponding to at least one first target feature from the training dataset to obtain a first removed dataset. The first removed data may be the training data from which feature content corresponding to at least one first target feature has been removed.

[0098] Step 1033: Use the first removed dataset as the training dataset and continue with the above storage steps.

[0099] In some embodiments, the execution entity may use the first removed dataset as the training dataset and continue to perform the storage step described above.

[0100] In some optional implementations of certain embodiments, after step 103, the steps further include: The first step is to respond to the received model training request for the target risk control detection model by parsing the risk control detection task information from the model training request, which is then used as the target risk control detection task information. The model training request may be information requesting model training parameters.

[0101] The second step is to generate a query statement targeting the aforementioned target risk control detection task information. This query statement can be a database query that retrieves the feature set corresponding to the target risk control detection task information.

[0102] The third step is to execute the above query statement to retrieve the feature set corresponding to the target risk control detection task information from the target model feature database, which will be used as the third target feature set.

[0103] The fourth step is to extract the historical feature dataset corresponding to the third target feature set from the target model database.

[0104] The fifth step involves using the aforementioned historical feature dataset as the training dataset and, with the help of the corresponding slave node set of the distributed cluster, performing model processing on the aforementioned target risk control detection model to generate the processed risk control detection model.

[0105] The sixth step is to send the model parameters corresponding to the risk control detection model processed above to the training request terminal. The training request terminal can be the terminal that sent the model training request.

[0106] The above embodiments of this disclosure have the following beneficial effects: Through the information storage method based on a distributed cluster in some embodiments of this disclosure, the feature set corresponding to the risk control detection task is generated efficiently and accurately by utilizing the master and slave node sets corresponding to the distributed cluster, while fully utilizing computing resources. Specifically, the reason for the low efficiency and inaccuracy in related determinations is that the model needs to be trained multiple times sequentially, resulting in a large number of training iterations to determine which features are strongly correlated with the task. This leads to long training times, long-term consumption of computing resources, low determination efficiency, and inaccurate subjective judgments by users. Based on this, the information storage method based on a distributed cluster in some embodiments of this disclosure first obtains a training dataset for at least one initial risk control detection model from the target model database, using it as the basic dataset to filter features by training at least one initial risk control detection model. Then, for the training dataset, a storage step is performed: First, based on the training dataset, the at least one initial risk control detection model is trained using the slave node set corresponding to the distributed cluster to generate at least one risk control detection model. Each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node. Here, through the master and slave node sets in the distributed cluster, at least one initial risk control detection model can be trained simultaneously under different feature inputs, so as to determine the importance of each feature in the risk control detection process based on subsequent model performance information. By setting at least one corresponding slave node for each risk control detection model, the normal training of the risk control detection model can be guaranteed. Second, in response to the at least one model performance information corresponding to the above-mentioned at least one risk control detection model meeting the target performance condition, the master node corresponding to the above-mentioned distributed cluster stores the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database, so as to be called and used in subsequent risk control detection scenarios, and to determine which of the remaining feature sets are key features based on the performance information of at least one model. Next, in response to the failure to meet the above target performance condition, the master node executes the following steps: First, based on the performance information of at least one model, the features to be removed can be accurately selected from the remaining feature set as the first target feature, obtaining at least one first target feature. Second, the feature data corresponding to at least one first target feature is removed from the training dataset, obtaining the first removed data, so as to facilitate further model training and further filtering of useless features. The third step is to use the first removed dataset as the training dataset and the remaining feature set as the full feature set, and then continue to perform the above storage steps.In summary, by scheduling node tasks in the master and slave node sets of a distributed cluster and by using model performance information to determine useful and useless features, not only is the efficiency of feature selection improved, but also a large amount of computing resources can be avoided for extended periods.

[0107] Further reference Figure 2 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of an information storage device based on a distributed cluster, these device embodiments being similar to... Figure 1 Corresponding to the method embodiments shown, this information storage device based on a distributed cluster can be specifically applied to various electronic devices.

[0108] like Figure 2 As shown, an information storage device 200 based on a distributed cluster includes: an acquisition unit 201, a first execution unit 202, and a second execution unit 203. The acquisition unit 201 is configured to acquire a training dataset for at least one initial risk control detection model from a target model database. The first execution unit 202 is configured to perform the following storage steps on the training dataset: Based on the training dataset, using the set of slave nodes corresponding to the distributed cluster, train the at least one initial risk control detection model to generate at least one risk control detection model. Each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node. In response to the at least one model performance information corresponding to the at least one risk control detection model satisfying the target performance condition, the first execution unit 202 uses the distributed cluster... The master node of the distributed cluster stores the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database. The second execution unit 203 is configured to, in response to the failure to meet the above target performance conditions, use the master node to execute the following steps: based on at least one model performance information, select the features to be removed from the remaining feature set to obtain at least one first target feature; remove the feature data corresponding to at least one first target feature from the training dataset to obtain the first removed dataset; use the first removed dataset as the training dataset and the remaining feature set as the full feature set, and continue to execute the above storage steps.

[0109] It is understandable that the units described in the distributed cluster-based information storage device 200 and the reference Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the information storage device 200 based on a distributed cluster and the units contained therein, and will not be repeated here.

[0110] The following is for reference. Figure 3It shows a schematic diagram of the structure of an electronic device (e.g., an electronic device) 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0111] like Figure 3 As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0112] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0113] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 309, or installed from storage device 308, or installed from ROM 302. When the computer program is executed by processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0114] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0115] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0116] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs. When the electronic device executes the aforementioned one or more programs, the electronic device causes the following actions: It retrieves a training dataset for at least one initial risk control detection model from a target model database; for the training dataset, it performs a storage step: based on the training dataset, it trains the at least one initial risk control detection model using the set of slave nodes corresponding to the distributed cluster to generate at least one risk control detection model, wherein each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node; in response to at least one model performance information corresponding to the at least one risk control detection model satisfying the target performance condition, it uses the master node corresponding to the distributed cluster to store the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database; in response to not satisfying the target performance condition, it uses the master node to perform an execution step: based on at least one model performance information, it filters out features to be removed from the remaining feature set to obtain at least one first target feature; it removes feature data corresponding to at least one first target feature from the training dataset to obtain a first removed dataset; it uses the first removed dataset as the training dataset and the remaining feature set as the full feature set, and continues to perform the aforementioned storage step. The remaining feature set is then used as the full feature set, and the above storage steps are continued.

[0117] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0118] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0119] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including an acquisition unit, a first execution unit, and a second execution unit. The names of these units do not necessarily limit the unit itself; for example, an acquisition unit may also be described as "a unit that acquires a training dataset for at least one initial risk control detection model from a target model database."

[0120] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0121] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. An information storage method based on a distributed cluster, comprising: Obtain a training dataset from the target model database for at least one initial risk control detection model; For the training dataset, perform the storage step: Based on the training dataset, the at least one initial risk control detection model is trained using the set of slave nodes corresponding to the distributed cluster to generate at least one risk control detection model. Each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node. In response to the fact that the performance information of at least one model corresponding to the at least one risk control detection model meets the target performance condition, the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information are stored in the target model feature database using the master node corresponding to the distributed cluster. In response to the failure to meet the target performance conditions, the following steps are executed using the master node: Based on at least one model performance information, features to be removed are selected from the remaining feature set to obtain at least one first target feature; Remove at least one feature data corresponding to the first target feature from the training dataset to obtain the first removed dataset; The dataset after the first removal is used as the training dataset, and the remaining feature set is used as the full feature set. The storage step is then performed.

2. The method according to claim 1, wherein, After obtaining the training dataset for at least one initial risk control detection model from the target model database, the method further includes: Data preprocessing is performed on each training data in the training dataset to generate a processed dataset, wherein the processed data includes: the feature dataset corresponding to the full feature set; Based on information theory indicators and statistical verification, the features to be removed from the full feature set are determined as the second target features, thus obtaining at least one second target feature. Remove the feature data corresponding to at least one second target feature from the processed dataset to obtain the removed dataset, which is used as the training dataset.

3. The method according to claim 1, wherein, The method further includes: In response to receiving a model training request for a target risk control detection model, risk control detection task information is parsed from the model training request and used as target risk control detection task information; Generate a query statement targeting the risk control detection task information of the target; Execute the query statement to retrieve the feature set corresponding to the target risk control detection task information from the target model feature database, as the third target feature set; Extract the historical feature dataset corresponding to the third target feature set from the target model database; The historical feature dataset is used as the training dataset, and the target risk control detection model is processed using the slave node set corresponding to the distributed cluster to generate the processed risk control detection model. The processed risk control detection model parameters are sent to the training request terminal.

4. The method according to claim 1, wherein, The step of obtaining the training dataset for at least one initial risk control detection model from the target model database includes: Determine the model application scenario group for each initial risk control detection model in the at least one initial risk control detection model to obtain at least one model application scenario group; Deduplication is performed on each model application scenario in the at least one model application scenario group to obtain a model application scenario set. For each model application scenario in the model application scenario set, data groups that are associated with the scenario identifier corresponding to the model application scenario are selected from the target model database; The obtained dataset is deduplicated to obtain a deduplicated dataset, which is used as the training dataset.

5. The method according to claim 4, wherein, The step of training the at least one initial risk control detection model using the training dataset and the corresponding slave node set of the distributed cluster to generate at least one risk control detection model includes: For each of the at least one initial risk control detection model, perform the following training steps: Select a subset of training data corresponding to the initial risk control detection model from the training dataset; The slave node corresponding to the initial risk control detection model is determined as the target slave node; Retrieve the data backup information corresponding to the target slave node from the cache of the master node in the distributed cluster; Data that does not exist in the data backup information is selected from the training data subset and used as the target training data subset; For each target training data point in the subset of target training data, perform the following data processing steps: In response to determining that the target training data is in text form, determine whether there is data in the dataset corresponding to the data backup data that has a higher data similarity to the target training data than the target similarity; In response to the determination of existence, textual difference information between the target training data and data with higher similarity to the target is determined; In response to the determination that the target training data does not exist, the target training data is identified as textual difference information; The obtained subset of text difference information is sent to the target slave node so that the target slave node can obtain the target training data subset based on the subset of text difference information. The target slave node is instructed to train the initial risk control detection model based on the dataset corresponding to the data backup information and the target training data subset to generate the risk control detection model.

6. The method according to claim 1, wherein, The step of training the at least one initial risk control detection model using the training dataset and the corresponding slave node set of the distributed cluster to generate at least one risk control detection model includes: Obtain the feature description information set corresponding to the full feature set and the risk control detection task information; Generate prompt information to determine the importance of features in the full feature set relative to the corresponding task of the risk control detection task information based on the feature description information set; The generated prompt information is input into the large language model deployed on the master node to generate the feature importance set corresponding to the full feature set; Based on the set of feature importance, the full feature set is divided into a set of necessary features, a set of undetermined features, and a set of unnecessary features. The feature importance of necessary features is higher than that of undetermined features, and the feature importance of undetermined features is higher than that of unnecessary features. The set of undetermined features is divided into at least one part according to the importance of the features from high to low, to obtain a sequence of undetermined feature groups. The latter undetermined feature group includes the former undetermined feature group, and the latter undetermined feature group also includes the undetermined feature after the undetermined feature with the lowest importance in the former undetermined feature group. Extract the feature dataset corresponding to the necessary feature set from the training dataset, and use it as the necessary feature dataset; Extract the feature data group sequence corresponding to the undetermined feature group sequence from the training dataset; Extract the feature dataset corresponding to the unnecessary feature set from the training dataset, and use it as the unnecessary feature dataset; The necessary feature dataset is used as the first training dataset to train the first initial risk control detection model in the at least one initial risk control detection model, so as to generate the first risk control detection model. For each feature data group in the feature data group sequence, the feature data group and the necessary feature dataset are used as the second training dataset to train the second initial risk control detection model in the at least one initial risk control detection model, so as to generate the second risk control detection model. The necessary feature dataset and the non-necessary feature dataset are used as the third training dataset to train the third initial risk control detection model in the at least one initial risk control detection model, so as to generate the third risk control detection model. The first risk control detection model, the obtained second risk control detection model sequence, and the third risk control detection model are determined as at least one risk control detection model.

7. The method according to claim 6, wherein, The step of filtering out features to be removed from the remaining feature set based on at least one model performance information to obtain at least one first target feature includes: Generate an association table representing the relationship between model performance and features between the at least one model performance information and at least one remaining feature group, wherein the remaining feature set in the at least one remaining feature group has a feature correspondence with the risk control detection model in the at least one risk control detection model; Generate feature determination prompts to determine important features based on the association table and the remaining feature set; The feature determination prompts are input into the large language model deployed on the master node to generate an important feature set; The important feature set and the necessary feature set are combined to form a union feature set. Remove the union feature set from the remaining feature set to generate at least one first target feature.

8. An information storage device based on a distributed cluster, comprising: The acquisition unit is configured to acquire a training dataset for at least one initial risk control detection model from the target model database; The first execution unit is configured to perform the following storage steps for the training dataset: Based on the training dataset, using the set of slave nodes corresponding to the distributed cluster, train the at least one initial risk control detection model to generate at least one risk control detection model, wherein each node in the distributed cluster has a deployed graphics processor, and each risk control detection model has at least one corresponding slave node; in response to the at least one model performance information corresponding to the at least one risk control detection model satisfying the target performance condition, using the master node corresponding to the distributed cluster, store the remaining feature set after removing the first target feature from the full feature set and the corresponding risk control detection task information into the target model feature database; The second execution unit is configured to, in response to the failure to meet the target performance condition, use the master node to execute the following steps: based on at least one model performance information, select features to be removed from the remaining feature set to obtain at least one first target feature; remove feature data corresponding to at least one first target feature from the training dataset to obtain a first removed dataset; use the first removed dataset as the training dataset and the remaining feature set as the full feature set to continue executing the storage step.

9. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images