Important blank certificate usage amount prediction method and system, electronic equipment and storage medium

Abstract

The invention provides an important blank certificate usage amount prediction method and system, an electronic device and a storage medium, and the method comprises the steps: obtaining an initial training data set, and carrying out the preprocessing of the initial training data set, and obtaining a training data set; using each training sample and a preset limiting condition to construct a multi-task gradient boosting tree model, the multi-task gradient boosting tree model comprising a plurality of decision tree models arranged in sequence; when a prediction instruction of a to-be-predicted bank outlet is detected, obtaining current data under the to-be-predicted bank outlet, and preprocessing the current data to obtain target current data; for each decision-making tree model, performing prediction by using target current data through the decision-making tree model to obtain an important blank voucher usage amount of each important blank voucher type; and for each important blank voucher type, calculating the final demand quantity of the important blank vouchers of the important blank voucher type according to the usage quantity of each important blank voucher corresponding to the important blank voucher type.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method, system, electronic device, and storage medium for predicting the usage of important blank vouchers. Background Technology

[0002] When conducting business at a bank, important blank documents are frequently required. These include deposit slips, passbooks, transfer checks, cash checks, fixed-amount checks, bank drafts, promissory notes, and interbank statements. To ensure smooth operations, bank branches typically maintain a sufficient stock of important blank documents. However, due to the diverse types of blank documents, there is often an oversupply of some, consuming storage resources and increasing management costs, while the failure to replenish some blank documents in a timely manner hinders normal business operations.

[0003] To avoid problems arising from reserving too many or too few important blank vouchers, the required usage of these vouchers can be predicted in advance. Currently, this can be predicted based on human experience. However, relying on human experience is susceptible to subjective factors, resulting in poor prediction accuracy. This can still lead to issues such as business disruptions due to a shortage of important blank vouchers or resource waste due to excessive storage. Summary of the Invention

[0004] In view of this, the present invention provides a method, system, electronic device and storage medium for predicting the usage of important blank vouchers, so as to improve the prediction effect of the usage of important blank vouchers and avoid the problems of business processing being affected by the shortage of important blank vouchers or resource waste caused by excessive storage.

[0005] The first aspect of this application provides a method for predicting the usage of important blank vouchers, the method comprising:

[0006] Obtain an initial training dataset and preprocess the initial training dataset to obtain a training dataset; wherein, the training dataset includes multiple training samples, and each training sample includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers;

[0007] Using the training samples and preset constraints, a multi-task gradient boosting tree model is constructed, wherein the multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence.

[0008] When a prediction instruction for a bank branch to be predicted is detected, the current data of the bank branch to be predicted is obtained, and the current data is preprocessed to obtain the target current data.

[0009] For each decision tree model, the current target data is used to make predictions to obtain the usage of important blank vouchers for each important blank voucher type.

[0010] For each of the aforementioned important blank voucher types, the final demand for important blank vouchers of that important blank voucher type is calculated based on the usage of each important blank voucher corresponding to that important blank voucher type.

[0011] Optionally, obtaining the initial training dataset and preprocessing the initial training dataset to obtain the training dataset includes:

[0012] Obtain an initial training dataset, wherein the initial training data includes multiple initial training samples, and the initial training samples include first historical feature data and important blank voucher usage values ​​for multiple important blank voucher types; the first historical feature data includes initial historical data for multiple feature types;

[0013] For each important blank feature type in each initial training sample, determine whether there is missing data in the first historical feature data;

[0014] If so, determine that there is missing data in the first historical feature data, and calculate the ratio of the missing data to the first historical feature data; wherein, the missing data includes at least one missing record;

[0015] Determine whether the percentage is less than a preset percentage threshold;

[0016] If the value is less than the first historical feature data, the missing data is deleted to obtain the second historical feature data; wherein, the second historical feature data includes historical data for each of the aforementioned feature types;

[0017] If it is not less than, each missing record in the first historical feature data is supplemented according to the feature type of each missing record to obtain the second historical feature data;

[0018] Calculate the correlation coefficient between every two historical data of the aforementioned feature types in the second historical feature data;

[0019] The second historical feature data is processed based on the correlation coefficient of historical data for each pair of the aforementioned feature types to obtain historical feature data;

[0020] Training samples are generated based on the historical feature data of each of the aforementioned important blank voucher types and the actual usage values ​​of important blank vouchers.

[0021] Optionally, supplementing each missing record in the first historical feature data according to the feature type of each missing record includes:

[0022] For each missing record in the first historical feature data, determine the feature type of the missing record;

[0023] If the feature type is a time-series feature, perform linear interpolation on the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record;

[0024] If the feature type is not a time series feature and the data type is a numerical type, calculate the mean or median of the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to supplement the missing records.

[0025] If the feature type is not a time-series feature and the data type is a non-numerical type, the missing part in the missing record is supplemented with the mode.

[0026] Optionally, constructing a multi-task gradient boosting tree model using each of the training samples and preset constraints includes:

[0027] In the first iteration, a decision tree model is constructed based on preset constraints, historical feature data of each important blank voucher type in each training sample, and the target value of important blank vouchers. The target value of important blank vouchers for each important blank voucher type in each training sample used in the first iteration is the actual usage of important blank vouchers for each important blank voucher type in each training sample. The decision tree consists of multiple non-leaf nodes and multiple leaf nodes, with each non-leaf node configured with a corresponding loss value.

[0028] In the Nth iteration, a decision tree model is constructed based on the preset constraints, historical feature data of each important blank voucher type in each training sample, and the actual target value of the important blank voucher. Wherein, N is greater than 1 and increments by 1 until it equals M. The target value of the important blank voucher for each important blank voucher type in each training sample used in the Nth iteration is determined by the predicted usage and target value of the important blank voucher for each important blank voucher type in each training sample, as predicted by the decision tree model obtained in the (N-1)th iteration.

[0029] By integrating the decision tree models obtained from all iterations, a multi-task gradient boosting tree model is obtained.

[0030] Optionally, after obtaining the multi-task gradient boosting tree model, the method further includes:

[0031] New training data is periodically acquired, and the training dataset is updated using the new training data to obtain the target training dataset;

[0032] The loss value of each non-leaf node in each decision tree model in the multi-task gradient boosting tree model is updated using the target training dataset to update the multi-task gradient boosting tree model.

[0033] A second aspect of this application provides a system for predicting the usage of important blank vouchers, the system comprising:

[0034] The acquisition module is used to acquire an initial training dataset and preprocess the initial training dataset to obtain a training dataset; wherein, the training dataset includes multiple training samples, and each training sample includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers;

[0035] A construction module is used to construct a multi-task gradient boosting tree model using the various training samples and preset constraints, wherein the multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence.

[0036] The preprocessing module is used to obtain the current data of the bank branch to be predicted when a prediction instruction is detected, and to preprocess the current data to obtain the target current data.

[0037] The prediction module is used to predict the usage of important blank vouchers for each of the decision tree models by using the target current data.

[0038] The calculation module is used to calculate the final demand for important blank vouchers of each important blank voucher type based on the usage of each important blank voucher corresponding to that important blank voucher type.

[0039] Optionally, the acquisition module is specifically used for:

[0040] Obtain an initial training dataset, wherein the initial training data includes multiple initial training samples, and the initial training samples include first historical feature data and important blank voucher usage values ​​for multiple important blank voucher types; the first historical feature data includes initial historical data for multiple feature types;

[0041] For each important blank feature type in each initial training sample, determine whether there is missing data in the first historical feature data;

[0042] If so, determine that there is missing data in the first historical feature data, and calculate the ratio of the missing data to the first historical feature data; wherein, the missing data includes at least one missing record;

[0043] Determine whether the percentage is less than a preset percentage threshold;

[0044] If it is less than, delete the missing data from the first historical feature data to obtain the second historical feature data;

[0045] If the value is not less than the specified value, each missing record in the first historical feature data is supplemented according to the feature type of each missing record to obtain the second historical feature data; wherein, the second historical feature data includes historical data of each of the specified feature types;

[0046] Calculate the correlation coefficient between every two historical data of the aforementioned feature types in the second historical feature data;

[0047] The second historical feature data is processed based on the correlation coefficient of historical data for each pair of the aforementioned feature types to obtain historical feature data;

[0048] Training samples are generated based on the historical feature data of each of the aforementioned important blank voucher types and the actual usage values ​​of important blank vouchers.

[0049] Optionally, the step of supplementing each missing record in the first historical feature data according to the feature type of each missing record to obtain the second historical feature data is specifically used for:

[0050] For each missing record in the first historical feature data, determine the feature type of the missing record;

[0051] If the feature type is a time-series feature, perform linear interpolation on the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record;

[0052] If the feature type is not a time series feature and the data type is a numerical type, calculate the mean or median of the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to supplement the missing records.

[0053] If the feature type is not a time-series feature and the data type is a non-numerical type, the missing part in the missing record is supplemented with the mode.

[0054] A third aspect of this application provides an electronic device, comprising: a processor and a memory, the processor and the memory being connected via a bus; wherein the processor is configured to call and execute a program stored in the memory; and the memory is configured to store the program, the program being configured to implement the method for predicting the usage of important blank vouchers as provided in the first aspect of this application.

[0055] A fourth aspect of this application provides a computer-readable storage medium storing computer-executable instructions for performing a method for predicting the usage of important blank vouchers as provided in the first aspect of this application.

[0056] This application provides a method for predicting the usage of important blank vouchers, including a system, electronic device, and storage medium. The method involves acquiring an initial training dataset and preprocessing it to obtain a training dataset. The training dataset includes multiple training samples, each containing historical feature data and actual usage of important blank vouchers for multiple important blank voucher types. A multi-task gradient boosting tree model is constructed using the training samples and preset constraints. This model includes multiple decision tree models arranged in sequence. When a prediction instruction for a bank branch to be predicted is detected, the current data for that bank branch is acquired and preprocessed to obtain target current data. For each decision tree model, the target current data is used to predict the usage of important blank vouchers for each type. Finally, for each important blank voucher type, the final demand for important blank vouchers for that type is calculated based on the usage of each important blank voucher corresponding to that type. Therefore, this application constructs multiple decision tree models (multi-task gradient boosting tree models) that can simultaneously predict the actual usage of multiple important blank voucher types by utilizing training datasets related to various bank branches and preset constraints. When a prediction instruction for a bank branch to be predicted is detected, each decision tree model uses the target current data of the bank branch to predict the usage of important blank vouchers for each important blank voucher type, reducing reliance on manual experience and achieving accurate prediction of important blank voucher usage. This enables control over the allocation of bank branches, improves inventory turnover and resource utilization, and avoids problems such as business processing being affected by shortages of important blank vouchers or resource waste caused by excessive storage. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0058] Figure 1 Here are structural examples of a neural network model and the MT-GBM model;

[0059] Figure 2 A flowchart illustrating an important method for predicting the usage of blank vouchers provided in this application embodiment;

[0060] Figure 3 This application provides an example diagram of data extraction and multi-task gradient boosting tree model training;

[0061] Figure 4 An example diagram of a partitioning scheme for existing technologies when the prediction target is the usage of a single type of heavy air.

[0062] Figure 5 Example diagram of another data extraction and multi-task gradient boosting tree model training provided for embodiments of this application;

[0063] Figure 6 An example diagram illustrating a data update method provided in an embodiment of this application;

[0064] Figure 7 This is a schematic diagram of the structure of an important blank voucher usage prediction system provided in an embodiment of the present invention;

[0065] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] In this application, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0068] To better understand this application, the following explanations are provided regarding the technologies involved:

[0069] Important blank vouchers: Blank vouchers printed by banks without a face value, which, after being filled in and signed by the bank or an entity, have the effect of disbursement. Important blank vouchers are specific vouchers used by banks to process fund payments. Banks must strictly manage their printing, disbursement, safekeeping, sale, and destruction. In the following text, "important blank vouchers" refers to important blank vouchers.

[0070] Incremental learning: Incremental learning is a machine learning technique whose core lies in the system's ability to continuously absorb new knowledge from new data. It allows the model to gradually receive new data and update its parameters, adapting to changes in data distribution (such as time series data and streaming data). In contrast, traditional batch learning requires processing all data at once, making it impossible to absorb new data to adjust the model and adapt to changes in data distribution.

[0071] Experience replay is a technique used to improve training efficiency and stability. By reusing historical data, the model can balance new and old knowledge, avoid "catastrophic forgetting", reduce overfitting of the model to new data, and improve the model's generalization ability.

[0072] Catastrophic forgetting refers to the significant forgetting of old knowledge caused by parameter updates when a model learns new knowledge.

[0073] Neural network model: A classic multi-task learning model in which different tasks share hidden layer information and ultimately obtain the output of multiple tasks.

[0074] MT-GBM model: Multi-Task Gradient Boosting Machine, a gradient boosting decision tree model suitable for multi-task learning, borrowing the structure of neural network models (such as...). Figure 1The method employs a homogeneous heterogeneous tree (i.e., different tasks share the same split node structure, but each task has an independent leaf node output value) to achieve multi-task joint learning.

[0075] See Figure 2 The diagram illustrates a flowchart of a method for predicting the usage of important blank vouchers according to an embodiment of this application. This method specifically includes the following steps:

[0076] S201: Obtain the initial training dataset and preprocess it to obtain the training dataset; the training dataset includes multiple training samples, each of which includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers.

[0077] In the embodiments of this application, see Figure 3 First, historical bank branch data is acquired, which may include historical business information related to the bank branches, historical branch information, and the actual usage of important blank vouchers for each important blank voucher type. Second, the historical bank branch data is stored in a relational database and deployed to a data lake warehouse. Finally, the actual demand of the bank branches to be predicted is determined, and features are extracted from the data lake warehouse based on the actual demand to extract an initial training dataset related to the actual demand. This initial training dataset is then preprocessed to obtain the corresponding training dataset, which is used to construct a model set using the various training samples and the multi-task gradient boosting tree model to be trained.

[0078] It should be noted that the initial training dataset includes first historical feature data and important blank voucher usage values ​​for multiple important blank voucher types; the first historical feature data includes historical data for multiple feature types; among them, feature types can be time-series features, institutional features, other features, etc.

[0079] It should also be noted that the actual demand of the bank branches to be predicted can be the usage of important blank vouchers for each important blank voucher type in week T.

[0080] For example, when the actual requirement is to predict the usage of important blank vouchers of various important blank voucher types in week T, the first historical feature data contained in each initial training sample in the extracted initial training dataset can be as shown in Table 1.

[0081] Table 1:

[0082]

[0083] It should be noted that the initial historical data of the time series features includes the usage of important blank vouchers of a certain institution in the same period last year, the usage last week, and the fluctuation ratio. Among them, it can be seen from the initial historical data of the time series features that the usage is relatively stable under the influence of factors such as no major events or holidays, and is used as an estimation parameter.

[0084] The initial historical data of an organization's characteristics can include the number of business types supported by the organization, which affects the types of important blank vouchers used by the organization. Specifically, the more business types supported by an organization, the more types of important blank vouchers are used.

[0085] The institution type can be a vault, a branch, etc.; when the institution type is a vault, predict the usage of various types of important blank vouchers for all branches within its jurisdiction; when the institution type is a branch, predict the usage of various types of important blank vouchers for the institution itself.

[0086] The GDP of the region can include developed regions, where the transaction volume is large and the corresponding usage of important blank vouchers is also large.

[0087] The customer base can include young people with high activity levels. In particular, regions with young people and high activity levels have a high usage of important blank vouchers.

[0088] In this embodiment, after extracting the initial training dataset, for each blank voucher type's first historical feature data in each initial training sample of the initial training dataset, it can first be determined whether there is missing data in the first historical feature data. If not, the correlation coefficient between the initial historical data of each pair of feature types in the first historical feature data can be directly calculated, so as to process the first historical feature data according to the correlation coefficient of the initial historical data of each pair of feature types to obtain the corresponding historical feature data. If there is missing data, the missing data in the first historical feature data is processed to obtain the second historical feature data, and the historical data of each feature type in the second historical feature data is processed to obtain the corresponding historical feature data. Finally, corresponding training samples are constructed based on the historical feature data of each important blank voucher and the actual usage of important blank vouchers. The training samples include the historical feature data of multiple important blank voucher types and the actual usage of important blank vouchers.

[0089] Optionally, the process of obtaining an initial training dataset and preprocessing it to obtain the training dataset can be as follows: Obtain the initial training dataset, which includes multiple initial training samples. Each initial training sample includes first historical feature data and usage values ​​of multiple important blank voucher types. The first historical feature data includes initial historical data of multiple feature types. For each important blank feature type's first historical feature data in each initial training sample, determine whether there is missing data. If so, determine the presence of missing data in the first historical feature data and calculate the ratio of missing data to the total first historical feature data. The missing data data... The process involves: including at least one missing record; determining whether the proportion is less than a preset proportion threshold; if less, deleting the missing data from the first historical feature data to obtain the second historical feature data; wherein the second historical feature data includes historical data for each feature type; if not less than the threshold, supplementing each missing record in the first historical feature data according to the feature type of each missing record to obtain the second historical feature data; calculating the correlation coefficient between historical data of every two feature types in the second historical feature data; processing the second historical feature data according to the correlation coefficient of historical data of every two feature types to obtain historical feature data; and generating corresponding training samples based on the historical feature data of each important blank voucher type and the actual usage values ​​of important blank vouchers.

[0090] As one implementation of this application, the process of supplementing each missing record in the first historical feature data according to the feature type of each missing record to obtain the second historical feature data can be as follows: For each missing record in the first historical feature data, determine the feature type of the missing record; if the feature type is a time-series feature, perform linear interpolation on the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record; if the feature type is not a time-series feature and the data type is a numeric type, calculate the mean or median of the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record; if the feature type is not a time-series feature and the data type is a non-numeric type, supplement the missing part of the missing record with the mode. Finally, generate the corresponding second historical feature data from each supplemented missing record and the records without missing data.

[0091] It should be noted that before generating the second historical feature data, for each record (each missing record after supplementation and each record without missing data), if the data type of the record is non-numeric, the non-numeric features in the record can be numerically encoded.

[0092] It should also be noted that the preset percentage threshold can be set to 5% in advance, and can be set according to actual application. This application embodiment does not limit it.

[0093] In practical applications, the initial training dataset directly extracted from the data lake warehouse may contain issues such as missing data and formatting problems. Using this initial training dataset directly for model training may affect the prediction accuracy of the resulting model. Therefore, after obtaining the initial training dataset, we can first determine whether there are missing data in the first historical feature data of each initial training sample. If so, we calculate the ratio between the missing data and the first historical feature data. If this ratio is less than a preset threshold, the proportion of missing data in the first historical feature data is considered extremely low, and the missing data can be directly deleted. If the ratio is not less than the preset threshold, the proportion of missing data in the first historical feature data is not low, and to ensure the prediction accuracy of the model, the missing data cannot be directly deleted. For each missing record in the first historical feature data, we can determine whether the feature type of the missing record is a time feature. For time-related features (such as the weekly usage of various important blank vouchers), linear interpolation can be used to estimate the missing record and fill it with the interpolated value. If the feature is not time-related (such as institutional or other features), it can be further determined whether the missing data is numerical. If it is numerical, the feature mean or median can be used to replace it. That is, the feature mean or median of the missing record is calculated and used to replace the missing value in the missing data. If it is not numerical, i.e., non-numerical, digital encoding technology can be used to process the missing record and the missing part of the missing record can be filled with the mode to obtain the target historical record that meets the input requirements of the model.

[0094] It should also be noted that if there are outliers in the missing data due to data recording errors or other reasons, the 3σ principle can be used to remove data in the record that differs from the mean by three times the standard deviation, thus obtaining the corresponding target historical record.

[0095] In some embodiments, although the meanings of the data (initial historical data or historical data) of each feature type are different, the content they reflect may be strongly correlated, and redundant data will result in low generalization ability of the model. Therefore, the Pearson method can be used to analyze the correlation between the data of each pair of feature types to obtain the correlation coefficient between the data of each pair of feature types, where the correlation coefficient from 0 to 1 indicates the correlation from weak to strong. When the correlation coefficient between the data of two feature types reaches 0.9 or higher, the data of either feature type can be removed. Finally, the historical data of the remaining feature types are determined as the historical feature data of the important blank voucher type.

[0096] S202: Construct a multi-task gradient boosting tree model using various training samples and preset constraints. The multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence.

[0097] In the specific execution step S202, corresponding preset constraints can be set in advance so that after obtaining the training dataset, the decision tree model in the first iteration can be constructed by further utilizing the preset constraints, the historical feature data of each important blank voucher type in each training sample and the target value of the important blank voucher. Then, the multi-task gradient boosting tree model in the Nth iteration can be constructed by continuing to use the preset constraints, the decision tree model obtained in the N-1th iteration and the parameters used in the N-1th iteration. Here, N is greater than 1 and increments by 1 until it equals M. Finally, the multi-task gradient boosting tree models obtained in each iteration are integrated into a model set.

[0098] It should be noted that in order to obtain a decision tree model with higher accuracy, corresponding preset constraints can be configured in advance. These preset constraints include multiple parameters, which may include: learning rate, number of iterations, minimum amount of data in a leaf node, maximum depth of a single decision tree, and loss function.

[0099] It should also be noted that the learning rate can be within the range (0, 1). Since a smaller learning rate helps to obtain stable model performance, the learning rate can be set to 0.1. The number of iterations can be set to 100-1000. To prevent underfitting or overfitting, the number of iterations can be set to 500. The minimum amount of data in the leaf nodes can be set to 20. When the training dataset is small, this parameter can be appropriately lowered. To avoid overfitting, a single decision tree based on gradient boosting should not be too deep. Therefore, the maximum depth of a single decision tree can be set to 5. Loss function: The mean squared error (MSE) is used to measure the loss. Its expression is as follows, where y represents the actual usage of important blank voucher types. This indicates the amount of predictions used in the model's output.

[0100] (1)

[0101] It should be noted that this application constructs corresponding training samples by collecting historical feature data of different important blank voucher types and actual usage of important blank vouchers, and constructs corresponding training datasets based on the constructed training samples. It then uses the machine learning algorithm MT-GBM and the constructed training datasets for iterative training, which can dynamically capture the usage patterns of each important blank voucher type. This allows for the construction of a decision tree model that can simultaneously predict the usage of multiple important blank voucher types, reducing reliance on human experience, achieving accurate prediction of important blank voucher usage, controlling the allocation of funds to branches, and improving inventory turnover and resource utilization.

[0102] Optionally, the process of constructing a multi-task gradient boosting tree model using various training samples and preset constraints can be as follows: In the first iteration, based on the preset constraints, historical feature data of each important blank voucher type in each training sample, and the target value of important blank vouchers, a decision tree model for the first iteration is constructed; wherein, the target value of important blank vouchers for each important blank voucher type in each training sample used in the first iteration is the actual usage of important blank vouchers for each important blank voucher type in each training sample; the decision tree consists of multiple non-leaf nodes and multiple leaf nodes, and each non-leaf node is configured with a corresponding loss. The loss value; in the Nth iteration, based on the preset constraints, the historical feature data of each important blank voucher type in each training sample, and the actual target value of the important blank voucher, a decision tree model for the Nth iteration is constructed; where N is greater than 1 and increments by 1 until it equals M, and the target value of the important blank voucher for each important blank voucher type in each training sample used in the Nth iteration is determined by the predicted usage and target value of the important blank voucher for each important blank voucher type in each training sample based on the decision tree model obtained in the N-1th iteration; the decision tree models obtained in all iterations are integrated to obtain a multi-task gradient boosting tree model.

[0103] In this embodiment, research has shown that traditional decision trees, through a greedy strategy, aim to minimize the loss of a single task by progressively splitting nodes to construct a tree model, i.e., building a corresponding prediction model. However, the prediction model constructed in this way can only solve the problem of predicting the usage of a single type of important blank voucher. For example, when the prediction target is the usage of important blank vouchers of type A, the historical feature data obtained is shown in Table 2. At this time, the initial root node includes all data. First, the root node can be split (for example, based on whether the usage of important blank vouchers of type A in week T-1 is greater than 20). Then, the sum of the variances M of the two target values ​​is used as a measure of the loss of this split. The split with the smallest M value is the optimal split for the current dataset. The two datasets after the optimal split are the left and right child nodes, and then a new split is performed on them respectively. This process is repeated until the maximum depth or other constraints are reached, and then the corresponding tree model is obtained.

[0104] Table 2:

[0105]

[0106] However, in real-world scenarios, bank branches use various types of important blank vouchers, and the training data obtained in this case can be shown in Table 3. If we want to simultaneously predict the usage of multiple types of important blank vouchers, such as type A and type B, then the traditional decision tree construction method is no longer applicable. Figure 4 As shown, in traditional decision tree models, the objective values ​​for the two tasks differ when calculating the loss. Split a minimizes the loss prediction for the usage of important blank vouchers of type A, while split b minimizes the loss prediction for the usage of important blank vouchers of type B. This reveals that the optimal split for a single task may not be applicable to other tasks. Therefore, building a decision tree considering only one task impairs the model's ability to handle other tasks. Existing technologies address this issue by training separate prediction models for different types of important blank vouchers. However, with numerous types of important blank vouchers, building a large number of models is not only inefficient but also extremely wasteful of resources. Therefore, to ensure the model's accuracy and efficiency, finding the optimal split point that considers all types of important blank vouchers is the best solution.

[0107] Table 3:

[0108]

[0109] In the embodiments of this application, it was found that the MT-GBM model borrows the structural pattern of neural networks, retains the high training efficiency of decision trees, and gains the ability to handle multiple tasks. It uses the loss gradient as the prediction target and uses gradient fusion to fuse the gradients of the prediction targets of different important blank certificate types to obtain the corresponding fused gradient. In this way, each node has a unified target when selecting the split point, which can minimize the loss of the fused gradient and thus take into account all important blank certificate types.

[0110] Therefore, in this embodiment of the application, the selection of the target value of the important blank certificate type used in the process of constructing the corresponding decision tree model is different from the selection of the split point of the traditional single-task decision tree model. In any iteration other than the first iteration, the absolute value of the difference between the predicted usage of each important blank certificate type and the target value of the important blank certificate in each training sample predicted by the decision tree model obtained in the previous iteration is the prediction target (the target value of the important blank certificate of each important blank certificate type in each training sample of the current iteration). The fusion gradient of each training sample is calculated by the target value of the important blank certificate of each important blank certificate type in each training sample, so as to achieve the purpose of simultaneously taking into account multiple tasks.

[0111] In this invention, training data containing historical feature data of all important blank platform types and actual usage of important blank vouchers is used to train the corresponding model at once. The resulting decision tree model can predict the usage of important blank vouchers of all important blank voucher types at once.

[0112] In practical applications, during the first iteration, the training dataset is used as the root node (the first root node can be pre-configured). Assuming the training dataset is dataset A, it can be split based on a certain feature type, where the resulting datasets satisfy preset constraints. The two datasets obtained from splitting dataset A are then distributed to the two child nodes under the non-leaf node (root node). Specifically, dataset A can be split using the first data splitting method to obtain two datasets, dataset 1 and dataset 2. Assuming dataset 1 is the data allocated to the left node of the two child nodes under the non-leaf node (root node), dataset 1 can be as shown in Table 4. According to Table 4, each training sample includes three important empty data points. The target value of important blank vouchers for the white voucher type is defined as follows: the target value of important blank vouchers for the important blank voucher type is the actual number of important blank vouchers used for that type. For each important blank voucher type, the mean of the target value of important blank vouchers for that type in each training sample can be calculated. For each training sample, the absolute value of the difference between the mean and the target value of important blank vouchers for each important blank voucher type is calculated, and the mean of each absolute value is calculated to obtain the fusion gradient of that training sample. The sum of the squares of the fusion gradients of each training sample is calculated to obtain the loss value of the left node. The same method is used to continue calculating the loss value of the right node based on dataset 2 of the right node. Finally, the sum of the loss values ​​of the left and right nodes is calculated to obtain the loss value of the child node under the first data splitting method.

[0113] Continue calculating the loss value of the non-leaf node (root node) under each data splitting method using the above method, and take the minimum loss value among all loss values ​​under each data splitting method as the loss value of the non-leaf node (root node). Then, based on the determined loss value of the non-leaf node (root node), split the dataset A of the non-leaf node so that the two datasets obtained from splitting dataset A are actually distributed to the two child nodes under the non-leaf node (root node). If the current depth has not yet reached the maximum depth of a single decision tree, continue in this manner, taking each child node as the root node again, and continue splitting the dataset on each root node in the above way to construct the loss value of each root node until the maximum depth of a single decision tree is reached. Stop splitting then, and determine the child node corresponding to the maximum depth of a single decision tree as the leaf node to construct the decision tree model in the first iteration. Use the data corresponding to the leaf nodes in the decision tree model to predict the predicted usage of each important blank voucher type in each training sample, as shown in Table 5.

[0114] Table 4:

[0115]

[0116] Among them, target value 1 is the target value of important blank voucher type 1, target value 2 is the target value of important blank voucher type 2, target value 3 is the target value of important blank voucher type 3, mean value 1 is the mean of important blank voucher type 1, mean value 2 is the mean of important blank voucher type 2, and mean value 3 is the mean of important blank voucher type 3.

[0117] Table 5:

[0118]

[0119] Among them, forecast value 1 is the forecast usage of important blank voucher type 1, forecast value 2 is the forecast usage of important blank voucher type 2, and forecast value 3 is the forecast usage of important blank voucher type 3.

[0120] In the second iteration, the training dataset can also be used as the dataset of the first root node. Assuming the training dataset is dataset A, it can be split into two datasets, dataset 1 and dataset 2, according to the first data splitting method. Assuming dataset 1 is the data assigned to the left node of the two child nodes under the non-leaf node (root node), based on the predicted and target values ​​of the important blank voucher types for each training sample shown in Table 5, the difference between the predicted and target values ​​of the important blank voucher types for each training sample is calculated, and this difference is used as the data for each training sample in the second iteration. The target values ​​of important blank vouchers for each important blank voucher type are shown in Table 6. For each important blank voucher type, the mean of the target values ​​of important blank vouchers for that important blank voucher type in each training sample can be calculated. For each training sample, the absolute value of the difference between the mean of each important blank voucher type and the target value of the important blank voucher is calculated, and the mean of each absolute value is calculated to obtain the fusion gradient of the training sample. The sum of the squares of the fusion gradients of each training sample is calculated to obtain the loss value of the left node. The loss value of the right node is calculated based on the dataset 2 of the right node using the same method. Finally, the sum of the loss values ​​of the left and right nodes is calculated to obtain the loss value of the child node under the first data splitting method.

[0121] Continue calculating the loss value of the non-leaf node (root node) under each data splitting method as described above, and take the minimum loss value among all loss values ​​under each data splitting method as the loss value of the non-leaf node (root node). Then, based on the determined loss value of the non-leaf node (root node), split the dataset A of the non-leaf node (root node) so that the two datasets obtained from splitting dataset A are actually distributed to the two child nodes under the non-leaf node. If the current depth has not yet reached the maximum depth of a single decision tree, continue in this manner, taking each child node as the root node again, and continue splitting the dataset at each root node in the above way to construct the loss value of each root node until the maximum depth of a single decision tree is reached. Stop splitting when the maximum depth of a single decision tree is reached, and determine the child node corresponding to the maximum depth of a single decision tree as the leaf node to construct the decision tree model in the second iteration. Use the data corresponding to the leaf nodes in the decision tree model to predict the predicted usage of each important blank voucher type within each training sample.

[0122] Following the above iterative method, the decision tree model is constructed in each subsequent iteration until the number of iterations reaches the pre-configured number of iterations, resulting in a number of decision tree models. Finally, the obtained decision tree models are integrated into a multi-task gradient boosting tree model.

[0123] Table 6:

[0124]

[0125] It should be noted that multiple data splitting methods can be preset. One of these methods is to split the dataset based on whether the actual usage of important blank voucher type 1 last week was greater than 10.

[0126] It should be noted that non-leaf nodes are nodes in the decision tree model other than those in the last layer, while leaf nodes are nodes in the last layer of the decision tree model.

[0127] S203: When a prediction instruction for a bank branch to be predicted is detected, the current data of the bank branch to be predicted is obtained, and the current data is preprocessed to obtain the target current data.

[0128] In the specific execution step S203, the existence of a prediction instruction for the bank branch to be predicted can be detected in real time. This prediction instruction can specify the usage of each important blank document type for the bank branch to be predicted. When this prediction instruction is detected, the current data of the bank branch to be predicted can be obtained. This current data includes feature data for each important blank document type, and each important blank document type's feature data includes data from multiple feature types. First, it can be determined whether there is missing data in the feature data of the important blank document types. If not, the correlation coefficient between each pair of feature types in the current data can be directly calculated to process the current data based on the correlation coefficient between each pair of feature types, obtaining the target feature data for the important blank document types. If missing data exists, the missing data in the feature data is processed, and the correlation coefficient between each pair of feature types in the obtained data is calculated to process the data based on the correlation coefficient between each pair of feature types, obtaining the target feature data. Finally, the target current data for the bank branch to be predicted is generated based on the target feature data for each important blank document type. The target feature data includes target data for at least one feature type. The process of processing the current data to obtain the target current data is the same as the process of processing the initial training samples to obtain the training samples. Please refer to the corresponding content disclosed above. This application embodiment does not limit it.

[0129] S204: For each decision tree model, use the current target data to make predictions and obtain the usage of important blank vouchers for each important blank voucher type.

[0130] In the specific execution step S204, after obtaining the target current data, the target current data can be input into each decision tree model so that each decision tree model can make predictions based on the input target current data and obtain the usage of important blank vouchers for each important blank voucher type.

[0131] S205: For each important blank voucher type, calculate the final demand for important blank vouchers of that type based on the usage of each important blank voucher corresponding to that type.

[0132] In the specific execution step S205, after obtaining the usage amount of important blank vouchers for each important blank voucher type predicted by each decision tree model, for each important blank voucher type, the weight of each decision tree model and the usage amount of important blank vouchers for that important blank voucher type predicted by each decision tree model can be weighted and calculated to obtain the final demand amount of important blank vouchers for that important blank voucher type.

[0133] It should be noted that after obtaining the final demand for important blank vouchers of each important blank voucher type, the final demand for important blank vouchers of each important blank voucher type can be output, so that the tellers of the bank branches to be predicted can calculate the difference between the final demand for important blank vouchers of each important blank voucher type and the current stock, and obtain the final demand for important blank vouchers actually used for each important blank voucher type. At the same time, in order to cope with small fluctuations in demand, a margin of 5% is added as a safety measure. Assuming that for a certain important blank voucher type, the current stock of the institution is N, and the predicted final demand for important blank vouchers is P, then the formula for calculating the quantity D to be delivered (the actual final usage) can be shown in formula (2):

[0134] (2)

[0135] Furthermore, in the embodiments of this application, it was found that traditional batch learning requires loading all data at once. However, in this problem, data related to important blank voucher types is dynamically generated in the form of a stream. Over time, the distribution of this data and the importance of some features will change, and outdated experience contained in the old data will no longer be suitable for the new environment. Therefore, the model should periodically forget some of the old data and grasp the patterns in the new data. The model update method in this invention is as follows.

[0136] Optionally, after obtaining the model set, new training data can be periodically acquired and the training dataset can be updated using the new training data to obtain the target training dataset; the loss value of each non-leaf node in each decision tree model in the multi-task gradient boosting tree model can be updated using the target training dataset to achieve the update of each target gradient boosting decision tree model.

[0137] In some embodiments, in this application, combined with Figure 3 See Figure 5 After reaching the upper bound of each decision tree model in the multi-task gradient boosting tree model, incremental learning and experience replay training can be used to incrementally train each decision tree model after deployment. Specifically, after obtaining new training data, incremental learning can be used to update the training dataset, that is, removing an equal amount of the oldest data from the training dataset as the new training data, such as... Figure 6 As shown, the updated target training dataset is reused for experience replay training of each decision tree model in the model set, thereby enabling online updates of each decision model.

[0138] Therefore, it is evident that after the initial training and deployment of the multi-task gradient boosting tree model, its structure remains unchanged. By periodically (monthly) performing incremental learning and experience replay training on each decision model within the deployed multi-task gradient boosting tree model, the weight information of the leaf nodes in each decision model is updated only. This dynamically adjusts the model rather than rebuilding it, thereby improving training efficiency and saving resources. Furthermore, before each incremental learning iteration, the training data is updated by adding recently added data and removing an equal amount of the oldest data, enabling each model to adapt to changes in various real-world factors and improve the corresponding prediction accuracy.

[0139] This application provides a method for predicting the usage of important blank vouchers, including a system, electronic device, and storage medium. The method involves acquiring an initial training dataset and preprocessing it to obtain a training dataset. The training dataset includes multiple training samples, each containing historical feature data and actual usage of important blank vouchers for multiple important blank voucher types. A multi-task gradient boosting tree model is constructed using the training samples and preset constraints. This model comprises multiple decision tree models arranged in sequence. When a prediction instruction for a bank branch to be predicted is detected, the current data for that bank branch is acquired and preprocessed to obtain target current data. For each decision tree model, the usage of important blank vouchers for each important blank voucher type is predicted using the target current data. Finally, for each important blank voucher type, the final demand for important blank vouchers for that type is calculated based on the usage of each important blank voucher corresponding to that type. Therefore, this application constructs multiple decision tree models (multi-task gradient boosting tree models) that can simultaneously predict the actual usage of multiple important blank voucher types by utilizing training datasets related to various bank branches and preset constraints. When a prediction instruction for a bank branch to be predicted is detected, each decision tree model uses the target current data of the bank branch to predict the usage of important blank vouchers for each important blank voucher type, reducing reliance on manual experience and achieving accurate prediction of important blank voucher usage. This enables control over the allocation of bank branches, improves inventory turnover and resource utilization, and avoids problems such as business processing being affected by shortages of important blank vouchers or resource waste caused by excessive storage.

[0140] Based on the method for predicting the usage of important blank vouchers shown in the above embodiments of the present invention, correspondingly, the embodiments of the present invention also show a system for predicting the usage of important blank vouchers, such as... Figure 7 As shown, the system includes:

[0141] The acquisition module 701 is used to acquire the initial training dataset and preprocess the initial training dataset to obtain the training dataset; wherein, the training dataset includes multiple training samples, and each training sample includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers;

[0142] Module 702 is used to construct a multi-task gradient boosting tree model using various training samples and preset constraints. The multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence.

[0143] The preprocessing module 703 is used to obtain the current data of the bank branch to be predicted when a prediction instruction for the bank branch to be predicted is detected, and to preprocess the current data to obtain the target current data.

[0144] The prediction module 704 is used to predict the usage of important blank vouchers for each important blank voucher type by using the current target data through the decision tree model for each decision tree model.

[0145] The calculation module 705 is used to calculate the final demand for important blank vouchers of each important blank voucher type based on the usage of each important blank voucher corresponding to the important blank voucher type.

[0146] The specific principles and execution processes of each unit in the important blank voucher usage prediction system disclosed in the above embodiments of this application are the same as those of the important blank voucher usage prediction method disclosed in the above embodiments of this application. Please refer to the corresponding parts of the important blank voucher usage prediction method disclosed in the above embodiments of this application, and they will not be repeated here.

[0147] This application provides a method for predicting the usage of important blank vouchers, including a system, electronic device, and storage medium. The method involves acquiring an initial training dataset and preprocessing it to obtain a training dataset. The training dataset includes multiple training samples, each containing historical feature data and actual usage of important blank vouchers for multiple important blank voucher types. A multi-task gradient boosting tree model is constructed using the training samples and preset constraints. This model comprises multiple decision tree models arranged in sequence. When a prediction instruction for a bank branch to be predicted is detected, the current data for that bank branch is acquired and preprocessed to obtain target current data. For each decision tree model, the usage of important blank vouchers for each important blank voucher type is predicted using the target current data. Finally, for each important blank voucher type, the final demand for important blank vouchers for that type is calculated based on the usage of each important blank voucher corresponding to that type. Therefore, this application constructs multiple decision tree models (multi-task gradient boosting tree models) that can simultaneously predict the actual usage of multiple important blank voucher types by utilizing training datasets related to various bank branches and preset constraints. When a prediction instruction for a bank branch to be predicted is detected, each decision tree model uses the target current data of the bank branch to predict the usage of important blank vouchers for each important blank voucher type, reducing reliance on manual experience and achieving accurate prediction of important blank voucher usage. This enables control over the allocation of bank branches, improves inventory turnover and resource utilization, and avoids problems such as business processing being affected by shortages of important blank vouchers or resource waste caused by excessive storage.

[0148] Optional, the acquisition module is specifically used for:

[0149] Obtain the initial training dataset, which includes multiple initial training samples. The initial training samples include the first historical feature data and the usage values ​​of multiple important blank voucher types. The first historical feature data includes the initial historical data of multiple feature types.

[0150] For each important blank feature type in each initial training sample, determine whether there is missing data in the first historical feature data.

[0151] If so, identify missing data in the first historical feature data and calculate the ratio of missing data to the first historical feature data; wherein, missing data includes at least one missing record;

[0152] Determine whether the percentage is less than a preset percentage threshold;

[0153] If it is less than, delete the missing data from the first historical feature data to obtain the second historical feature data;

[0154] If the value is not less than the specified value, each missing record in the first historical feature data is supplemented according to the feature type of each missing record to obtain the second historical feature data; wherein, the second historical feature data includes historical data of each feature type;

[0155] Calculate the correlation coefficient between historical data of every two feature types in the second historical feature data;

[0156] The second historical feature data is processed based on the correlation coefficient between the historical data of each pair of feature types to obtain the historical feature data;

[0157] Training samples are generated based on the historical feature data of each important blank voucher type and the actual usage values ​​of important blank vouchers.

[0158] Optionally, a module for obtaining second historical feature data is used to supplement each missing record in the first historical feature data according to the feature type of each missing record, specifically for:

[0159] For each missing record in the first historical feature data, determine the feature type of the missing record;

[0160] If the feature type is a time series feature, perform linear interpolation on the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to fill in the missing records.

[0161] If the feature type is not a time series feature and the data type is a numeric type, calculate the mean or median of the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to fill in the missing records.

[0162] If the feature type is not a time-series feature and the data type is a non-numeric type, the missing part of the missing record is supplemented with the mode.

[0163] Optional, building modules, specifically used for:

[0164] In the first iteration, a decision tree model is constructed based on preset constraints, historical feature data of each important blank voucher type in each training sample, and the target value of important blank vouchers. The target value of important blank vouchers for each important blank voucher type in each training sample used in the first iteration is the actual amount of important blank vouchers used for each important blank voucher type in each training sample. The decision tree consists of multiple non-leaf nodes and multiple leaf nodes, and each non-leaf node is configured with a corresponding loss value.

[0165] In the Nth iteration, a decision tree model is constructed based on preset constraints, historical feature data of each important blank voucher type in each training sample, and actual target values ​​of important blank vouchers. Here, N is greater than 1 and increments by 1 until it equals M. The target value of important blank vouchers for each important blank voucher type in each training sample used in the Nth iteration is determined by the predicted usage and target value of important blank vouchers for each important blank voucher type in each training sample, as predicted by the decision tree model obtained in the N-1th iteration.

[0166] By integrating the decision tree models obtained from all iterations, a multi-task gradient boosting tree model is obtained.

[0167] Optionally, building modules are also used for:

[0168] New training data is periodically acquired and used to update the training dataset to obtain the target training dataset.

[0169] The loss value of each non-leaf node in each decision tree model of the multi-task gradient boosting tree model is updated using the target training dataset to achieve the updating of the multi-task gradient boosting tree model.

[0170] This application also provides a storage medium storing program instructions that, when loaded and executed by a processor, implement any of the above-described embodiments of the method for predicting the usage of important blank vouchers.

[0171] This application also provides an electronic device, such as Figure 8 As shown, the device includes a processor 801 and a memory 802, which are connected via a bus; the memory stores program instructions; the processor calls the program instructions in the memory to execute any of the above-described embodiments of the method for predicting the usage of important blank vouchers.

[0172] The processor mentioned in this article can be the terminal's CPU, an integrated MCU within the terminal, or a combination of a CPU and an MCU. Furthermore, the processor contains a kernel that retrieves the corresponding program from memory; one or more kernels can be configured.

[0173] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0174] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0175] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0176] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0177] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the usage of important blank vouchers, characterized in that, The method includes: Obtain an initial training dataset and preprocess the initial training dataset to obtain a training dataset; wherein, the training dataset includes multiple training samples, and each training sample includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers; Using the training samples and preset constraints, a multi-task gradient boosting tree model is constructed, wherein the multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence. When a prediction instruction for a bank branch to be predicted is detected, the current data of the bank branch to be predicted is obtained, and the current data is preprocessed to obtain the target current data. For each decision tree model, the current target data is used to make predictions to obtain the usage of important blank vouchers for each important blank voucher type. For each of the aforementioned important blank voucher types, the final demand for important blank vouchers of that important blank voucher type is calculated based on the usage of each important blank voucher corresponding to that important blank voucher type.

2. The method according to claim 1, characterized in that, The process of obtaining an initial training dataset and preprocessing the initial training dataset to obtain a training dataset includes: Obtain an initial training dataset, wherein the initial training data includes multiple initial training samples, and the initial training samples include first historical feature data and important blank voucher usage values ​​for multiple important blank voucher types; the first historical feature data includes initial historical data for multiple feature types; For each important blank feature type in each initial training sample, determine whether there is missing data in the first historical feature data; If so, determine that there is missing data in the first historical feature data, and calculate the ratio of the missing data to the first historical feature data; wherein, the missing data includes at least one missing record; Determine whether the percentage is less than a preset percentage threshold; If the value is less than the first historical feature data, the missing data is deleted to obtain the second historical feature data; wherein, the second historical feature data includes historical data for each of the aforementioned feature types; If it is not less than, each missing record in the first historical feature data is supplemented according to the feature type of each missing record to obtain the second historical feature data; Calculate the correlation coefficient between every two historical data of the aforementioned feature types in the second historical feature data; The second historical feature data is processed based on the correlation coefficient of historical data for each pair of the aforementioned feature types to obtain historical feature data; Training samples are generated based on the historical feature data of each of the aforementioned important blank voucher types and the actual usage values ​​of important blank vouchers.

3. The method according to claim 2, characterized in that, The step of supplementing each missing record in the first historical feature data according to the feature type of each missing record includes: For each missing record in the first historical feature data, determine the feature type of the missing record; If the feature type is a time-series feature, perform linear interpolation on the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record; If the feature type is not a time series feature and the data type is a numerical type, calculate the mean or median of the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to supplement the missing records. If the feature type is not a time-series feature and the data type is a non-numerical type, the missing part in the missing record is supplemented with the mode.

4. The method according to claim 1, characterized in that, The construction of a multi-task gradient boosting tree model using the training samples and preset constraints includes: In the first iteration, a decision tree model is constructed based on preset constraints, historical feature data of each important blank voucher type in each training sample, and the target value of important blank vouchers. The target value of important blank vouchers for each important blank voucher type in each training sample used in the first iteration is the actual usage of important blank vouchers for each important blank voucher type in each training sample. The decision tree consists of multiple non-leaf nodes and multiple leaf nodes, with each non-leaf node configured with a corresponding loss value. In the Nth iteration, a decision tree model is constructed based on the preset constraints, historical feature data of each important blank voucher type in each training sample, and the actual target value of the important blank voucher. Wherein, N is greater than 1 and increments by 1 until it equals M. The target value of the important blank voucher for each important blank voucher type in each training sample used in the Nth iteration is determined by the predicted usage and target value of the important blank voucher for each important blank voucher type in each training sample, as predicted by the decision tree model obtained in the (N-1)th iteration. By integrating the decision tree models obtained from all iterations, a multi-task gradient boosting tree model is obtained.

5. The method according to claim 4, characterized in that, After obtaining the multi-task gradient boosting tree model, the method further includes: New training data is periodically acquired, and the training dataset is updated using the new training data to obtain the target training dataset; The loss value of each non-leaf node in each decision tree model in the multi-task gradient boosting tree model is updated using the target training dataset to update the multi-task gradient boosting tree model.

6. A system for predicting the usage of important blank vouchers, characterized in that, The system includes: The acquisition module is used to acquire an initial training dataset and preprocess the initial training dataset to obtain a training dataset; wherein, the training dataset includes multiple training samples, and each training sample includes historical feature data of multiple important blank voucher types and actual usage of important blank vouchers; A construction module is used to construct a multi-task gradient boosting tree model using the various training samples and preset constraints, wherein the multi-task gradient boosting tree model includes multiple decision tree models arranged in sequence. The preprocessing module is used to obtain the current data of the bank branch to be predicted when a prediction instruction is detected, and to preprocess the current data to obtain the target current data. The prediction module is used to predict the usage of important blank vouchers for each of the decision tree models by using the target current data. The calculation module is used to calculate the final demand for important blank vouchers of each important blank voucher type based on the usage of each important blank voucher corresponding to that important blank voucher type.

7. The system according to claim 6, characterized in that, The acquisition module is specifically used for: Obtain an initial training dataset, wherein the initial training data includes multiple initial training samples, and the initial training samples include first historical feature data and important blank voucher usage values ​​for multiple important blank voucher types; the first historical feature data includes initial historical data for multiple feature types; For each important blank feature type in each initial training sample, determine whether there is missing data in the first historical feature data; If so, determine that there is missing data in the first historical feature data, and calculate the ratio of the missing data to the first historical feature data; wherein, the missing data includes at least one missing record; Determine whether the percentage is less than a preset percentage threshold; If it is less than, delete the missing data from the first historical feature data to obtain the second historical feature data; If the value is not less than the specified value, each missing record in the first historical feature data is supplemented according to the feature type of each missing record to obtain the second historical feature data; wherein, the second historical feature data includes historical data of each of the specified feature types; Calculate the correlation coefficient between every two historical data of the aforementioned feature types in the second historical feature data; The second historical feature data is processed based on the correlation coefficient of historical data for each pair of the aforementioned feature types to obtain historical feature data; Training samples are generated based on the historical feature data of each of the aforementioned important blank voucher types and the actual usage values ​​of important blank vouchers.

8. The system according to claim 7, characterized in that, The step of supplementing each missing record in the first historical feature data according to the feature type of each missing record to obtain the second historical feature data is specifically used for: For each missing record in the first historical feature data, determine the feature type of the missing record; If the feature type is a time-series feature, perform linear interpolation on the missing record to obtain the missing value of the missing record, and use the missing value to supplement the missing record; If the feature type is not a time series feature and the data type is a numerical type, calculate the mean or median of the missing records to obtain the missing values ​​of the missing records, and use the missing values ​​to supplement the missing records. If the feature type is not a time-series feature and the data type is a non-numerical type, the missing part in the missing record is supplemented with the mode.

9. An electronic device, characterized in that, include: A processor and a memory are connected via a bus; wherein the processor is used to call and execute a program stored in the memory; The memory is used to store a program for implementing the method for predicting the usage of important blank vouchers as described in any one of claims 1-5.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions for performing the method for predicting the usage of important blank vouchers as described in any one of claims 1-5.

Images