Business data processing method and device

Abstract

The present application relates to the field of machine learning technology, and specifically discloses a business data processing method and device, wherein the method includes: obtaining a full feature sample set and a label set having feature data for characterizing business risk characteristics; splitting the full feature sample set into multiple partial feature sample sets according to preset business logic; constructing a sub-classifier using the partial feature sample sets and label sets of the multiple partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set; generating a multi-perspective comparative learning constraint based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample by each sub-classifier; optimizing each sub-classifier based on the multi-perspective comparative learning constraint, and using the optimized sub-classifier to process target business data. The above scheme can improve the efficiency and accuracy of business data processing.

Description

Technical Field

[0001] This specification relates to the field of machine learning technology, and in particular to a business data processing method and device. Background Art

[0002] With the rapid development of online business data processing technology, online business data processing is becoming increasingly diverse and convenient. Currently, in many business scenarios, it is often necessary to use known business data to predict the type of target business data and then perform business data processing based on the predicted type. For example, in electronic payment scenarios, it is often necessary to build a corresponding prediction model and use this prediction model to predict whether the transaction corresponding to the business data has a fraud risk based on the business data.

[0003] However, when existing business data processing methods are implemented, there is often a technical problem of low accuracy of the established prediction model, which makes it impossible to make accurate predictions before customer assets are likely to be lost. Business personnel need to make auxiliary judgments based on experience, resulting in low business data processing efficiency and the inability to properly and accurately process business data, resulting in a poor user experience.

[0004] To address the above issues, no effective solutions have been proposed so far. Summary of the Invention

[0005] The embodiments of this specification provide a business data processing method and device to improve the efficiency and accuracy of business data processing.

[0006] An embodiment of the present specification provides a business data processing method, including: obtaining a full feature sample set and a label set having feature data for characterizing business risk features, wherein the label set includes the risk category corresponding to each full feature sample in the full feature sample set; splitting the full feature sample set into multiple partial feature sample sets according to preset business logic; constructing a sub-classifier using the partial feature sample sets and label sets in the multiple partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set; generating multi-perspective comparative learning constraints based on the similarities and differences between risk classification results of partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to each partial feature sample set; optimizing the sub-classifiers corresponding to each partial feature sample set based on the multi-perspective comparative learning constraints, wherein the optimized sub-classifiers are used to process target business data.

[0007] In one embodiment, the risk categories include positive and negative categories, and accordingly, the full feature sample set includes positive samples and negative samples, wherein the number of positive samples in the full feature sample set is less than the number of negative samples in the full feature sample set, the label corresponding to the positive sample is the positive class, and the label corresponding to the negative sample is the negative class; after obtaining the full feature sample set and label set with feature data for characterizing business risk characteristics, it also includes: randomly oversampling the full feature sample set to increase the number of positive samples in the full feature sample set to obtain the oversampled full feature sample set; randomly undersampling the oversampled full feature sample set to reduce the number of negative samples in the full feature sample set.

[0008] In one embodiment, after constructing a sub-classifier using each partial feature sample set and a label set in a plurality of partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set, it also includes: using the sub-classifier corresponding to each partial feature sample set to classify each partial feature sample in each partial feature sample set; extracting the neighboring partial feature samples of each partial feature sample in each partial feature sample set, and calculating the inter-class similarity matrix corresponding to each partial feature sample set based on the similarities and differences between the categories of each partial feature sample in each partial feature sample set and the categories of its neighboring partial feature samples; accordingly, optimizing the sub-classifier corresponding to each partial feature sample set based on multi-view contrast learning constraints, including: generating inter-class similarity constraints based on the inter-class similarity matrix corresponding to each partial feature sample set; and optimizing the sub-classifier corresponding to each partial feature sample set based on the multi-view contrast learning constraints and the inter-class similarity constraints.

[0009] In one embodiment, the multi-view contrastive learning constraints are:

[0010]

[0011] Among them, R vcl is the multi-view contrast learning constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, and f p is the subclassifier corresponding to the pth partial feature sample set, f q is the sub-classifier corresponding to the qth partial feature sample set, is the i-th partial feature sample in the p-th partial feature sample set, is the i-th partial feature sample in the q-th partial feature sample set, and They all correspond to the i-th full-feature sample in the full-feature sample set.

[0012] In one embodiment, the elements in the inter-class similarity matrix are:

[0013]

[0014] in, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them, v = 1, 2, ... V, V is the number of multiple partial feature sample sets, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, express The neighboring feature samples of express The neighboring partial feature samples of , i, j = 1, 2, ... N, N is the number of partial feature samples in each partial feature sample set.

[0015] In one embodiment, the inter-class similarity constraint is:

[0016]

[0017] Among them, R bcs is the inter-class similarity constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them.

[0018] In one embodiment, based on the multi-view contrast learning constraint and the inter-class similarity constraint, the sub-classifiers corresponding to each partial feature sample set are optimized, including minimizing the objective function until a preset number of iterations is reached or the difference between the loss values ​​of two loss functions is less than a preset threshold, wherein the objective function is:

[0019] L=R emp +αR bcs +γR vcl

[0020]

[0021] Among them, L is the objective function, R emp is the experience loss, R vcl For multi-view contrast learning constraints, R bcsis the inter-class similarity constraint, α and γ are hyperparameters, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, For label collection The corresponding label.

[0022] In one embodiment, target business data is processed based on the optimized sub-classifier, including: generating a target classifier based on the sub-classifiers corresponding to the optimized feature sample sets; processing the target business data using the target classifier to determine the risk type of the business corresponding to the target business data; and processing the target business data based on the risk type of the business corresponding to the target business data.

[0023] An embodiment of the present specification also provides a business data processing device, including: an acquisition module, used to obtain a full feature sample set and a label set having feature data for characterizing business risk features, wherein the label set includes the risk category corresponding to each full feature sample in the full feature sample set; a splitting module, used to split the full feature sample set into multiple partial feature sample sets according to preset business logic; a construction module, used to construct a sub-classifier using the partial feature sample sets and label sets of the multiple partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set; a generation module, used to generate multi-perspective comparative learning constraints based on the differences and similarities between the risk classification results of the partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to the partial feature sample sets; an optimization module, used to optimize the sub-classifiers corresponding to the partial feature sample sets based on the multi-perspective comparative learning constraints, wherein the optimized sub-classifiers are used to process target business data.

[0024] An embodiment of this specification also provides a computer device, including a processor and a memory for storing processor-executable instructions, wherein when the processor executes the instructions, the steps of the business data processing method described in any of the above embodiments are implemented.

[0025] The embodiments of this specification also provide a computer-readable storage medium having computer instructions stored thereon, which, when executed, implement the steps of the business data processing method described in any of the above embodiments.

[0026] In an embodiment of the present specification, a business data processing method is provided, and a server can obtain a full feature sample set and a corresponding label set having feature data for characterizing business risk features, and can split the full feature sample set into multiple partial feature sample sets according to preset business logic. After obtaining multiple partial feature sample sets, sub-classifiers can be constructed using the partial feature sample sets and label sets in the multiple partial feature sample sets to obtain sub-classifiers corresponding to each partial feature sample set. Thereafter, multi-perspective comparative learning constraints can be generated based on the differences and similarities between risk classification results of partial feature samples corresponding to the same full feature sample based on the sub-classifiers corresponding to each partial feature sample set, so that the sub-classifiers corresponding to each partial feature sample set can be optimized based on the multi-perspective comparative learning constraints, and the optimized sub-classifiers can be used to process target business data. In the above scheme, the full feature sample set is split into multiple partial feature sample sets according to business logic, which can be learned from multiple perspectives. Compared with splicing all categories of features together, splitting according to business logic in this scheme can prevent the dimensional disaster problem caused by too many types of features during the training process; moreover, since the meanings of features from different perspectives are often quite different, direct splicing makes it difficult to analyze the features. By adopting different perspectives for modeling and learning, the accuracy of business risk classification can be improved. In addition, by generating multi-perspective comparative learning constraints based on the differences and similarities between the risk classification results of partial feature samples corresponding to the same full feature sample according to the sub-classifiers corresponding to each partial feature sample set, the sub-classifiers corresponding to each partial feature sample set are optimized. This can constrain the output space of the same training sample in different perspectives to be as close as possible. By fixing one perspective and optimizing the other, and optimizing two by two, it is possible to continuously iterate and optimize between multiple perspectives, thereby improving the effect of multi-perspective learning and thus improving the prediction effect of the sub-classifier. Furthermore, by determining the target risk category corresponding to the target business data based on the optimized sub-classifier and processing the business data based on the target risk category, the efficiency and accuracy of business data processing can be improved, the loss of customer assets can be avoided, losses can be reduced, and the user experience can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings described herein are used to provide a further understanding of this specification, constitute a part of this specification, and do not constitute a limitation of this specification. In the accompanying drawings:

[0028] Figure 1 A flowchart of a method for processing business data in one embodiment of this specification is shown;

[0029] Figure 2 A flowchart of a business data processing method according to an embodiment of the present specification is provided;

[0030] Figure 3A schematic diagram of a business data processing device in one embodiment of this specification is shown;

[0031] Figure 4 FIG1 is a schematic diagram of a computer device in an embodiment of the present specification. DETAILED DESCRIPTION

[0032] The principles and spirit of this specification will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement this specification, and are not intended to limit the scope of this specification in any way. Rather, these embodiments are provided to make this specification more thorough and complete, and to fully convey the scope of this disclosure to those skilled in the art.

[0033] Those skilled in the art will appreciate that the embodiments of this specification may be implemented as a system, device, method, or computer program product. Therefore, the disclosure herein may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0034] In a scenario example provided in an embodiment of this specification, the business data processing method can be applied to a device that performs business data processing, and the device may include a server or a server cluster consisting of multiple servers. For the target user, the server can extract feature data from various information of the target business data as the feature data of the target business data, and then use a pre-configured algorithm or model to perform risk prediction on the target business data to obtain a risk prediction result of the target business data, so as to process the target business data based on the risk prediction result. By accurately identifying the risk of the target business data, the business data can be processed more accurately, the efficiency and accuracy of business data processing can be improved, the loss of customer assets can be avoided, and losses can be reduced.

[0035] The embodiments of this specification provide a method for processing business data. Figure 1A schematic diagram of an application scenario of the business data processing method in one embodiment of this specification is shown. Although this specification provides method operation steps or device structures as shown in the following embodiments or drawings, more or fewer operation steps or module units may be included in the method or device based on routine or no creative labor. In the steps or structures where there is no necessary causal relationship logically, the execution order of these steps or the module structure of the device is not limited to the execution order or module structure described in the embodiment of this specification and shown in the drawings. When the method or module structure is applied to an actual device or terminal product, it can be connected in accordance with the method or module structure shown in the embodiment or drawings for sequential execution or parallel execution (for example, a parallel processor or multi-threaded processing environment, or even a distributed processing environment).

[0036] Specifically, if Figure 1 As shown, a business data processing method provided in one embodiment of this specification may include the following steps:

[0037] Step S101: Obtain a full feature sample set and a label set having feature data for characterizing business risk features, wherein the label set includes the risk category corresponding to each full feature sample in the full feature sample set.

[0038] Specifically, the method in this embodiment can be applied to a server. The server can obtain a full feature sample set and a label set. The full feature samples in the full feature sample set may include feature data used to characterize business risk characteristics. The feature data may, for example, be feature data extracted from business data stored in a financial institution's business system. Feature extraction can be performed through feature engineering. The feature data extraction method and feature type can be set according to the actual application scenario and are not limited here. Of course, it can also include feature data extracted by the server from business data obtained from a platform associated with the financial institution.

[0039] For example, in online transactions, business data may include transaction-related data, including: identity information characteristics of users accessing the network, such as the protocol stack, OS, browser, and hardware characteristics of the terminal feature information detection device; identity characteristics of both parties to the transaction, such as IP address anomaly detection, whether there is a correlation, etc.; whether there are suspicious or illegal business operations, such as whether there is a transaction between the two parties based on the historical transactions of the user, and analysis of changes in the balance of the accounts of both parties to the transaction in the past year, changes in the amount of funds transferred into and out of the accounts, etc.

[0040] A label set can include the risk categories corresponding to multiple full-feature samples in a full-feature sample set. For example, a full-feature sample can include a business data identifier and feature data, while a label set can include a business data identifier and a label, with the two being associated through the business data identifier.

[0041] Step S102: split the full feature sample set into multiple partial feature sample sets according to preset business logic.

[0042] After obtaining the full feature sample set, the full feature sample set can be split into multiple partial feature sample sets according to the preset business logic. Each partial feature sample in each partial feature sample set can include a business identifier, and the business identifier is consistent with the business identifier in the corresponding full feature sample in the full feature sample set. That is, the partial feature samples in each partial feature sample set of the multiple partial feature sample sets can correspond one-to-one to the full feature samples in the full feature sample set. According to the preset business logic, the multiple features corresponding to the same partial feature sample belong to the same perspective or belong to the same major category.

[0043] Exemplarily, the types of features in the full feature sample set T include A, B, C, D, E, F, G, H, I, J, and K. The full feature sample set T can be split into three partial feature sample sets T1, T2, and T3 according to business logic, wherein the types of features in T1 can include A, C, and D, the types of features in T2 can include B, E, F, and K, and the types of features in T3 can include G, H, I, and J. In this embodiment, the full feature sample set is exemplarily split into three partial feature sample sets. It is understandable that the full feature sample set can be divided into other numbers of partial feature sample sets according to preset business logic.

[0044] For example, for business data related to online transactions, the full feature sample set can be divided into three partial feature sample sets, namely, the basic network access information sample set, the two-party transaction feature information sample set, and the sample set for whether the business operation is suspicious or in violation of regulations. Among them, the features in the basic network access information sample set may include the basic information of the user's network access device, mobile phone operating system, protocol stack, login time, gender, age group, marital status, housing, vehicle, occupation and industry. The features in the two-party transaction feature information sample set may include the account fund information of the transaction parties, relevant information about the user of the fund outflow, etc. The features in the sample set for whether the business operation is suspicious or in violation of regulations may include the historical information of the user's transaction, including the historical amount of the user's account, the change of the amount, and whether the operation is suspicious.

[0045] Step S103 : constructing a sub-classifier using each partial feature sample set and the label set in the plurality of partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set.

[0046] After obtaining multiple partial feature sample sets, subclassifiers can be constructed using each of the partial feature sample sets and the label set, resulting in a subclassifier corresponding to each partial feature sample set. The label corresponding to a partial feature sample is the label of the full feature sample corresponding to that partial feature sample. For example, a partial feature sample may include a business identifier, a full feature sample may include a business identifier, and the label set may also include a business identifier. The partial feature sample can be associated with the label using the business identifier.

[0047] Step S104 , generating multi-view comparative learning constraints based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to the partial feature sample sets.

[0048] Step S105 , optimizing the sub-classifiers corresponding to each part of the feature sample set based on the multi-view contrast learning constraint, wherein the optimized sub-classifiers are used to process the target business data.

[0049] After obtaining the subclassifiers corresponding to each partial feature sample set, multi-perspective comparative learning constraints can be generated based on the similarities and differences between the risk classification results of the subclassifiers corresponding to each partial feature sample set for the same full feature sample in the full feature sample set. The multiple perspectives here can be several perspectives obtained according to the splitting criteria specified in the preset business logic, corresponding to each partial feature sample set. For example, the full feature sample set T is split into three partial feature sample sets T1, T2, and T3. The subclassifier corresponding to T1 is f1, the subclassifier corresponding to T2 is f2, and the subclassifier corresponding to T1 is f3. Multi-perspective comparative learning constraints are generated based on the similarities and differences between the classification results of the first partial feature sample x1 corresponding to the full feature sample x1, the classification results of the second partial feature sample x2 corresponding to the full feature sample x2 by f2, and the classification results of the third partial feature sample x3 corresponding to the full feature sample x3 by f3. The full feature sample x is any full feature sample in the full feature sample set T.

[0050] After generating the multi-view contrastive learning constraints, the sub-classifiers corresponding to each partial feature sample set can be optimized based on the multi-view contrastive learning constraints to obtain optimized sub-classifiers. The optimized sub-classifiers corresponding to each partial feature sample set can be used to process the target business data. For example, the target risk category corresponding to the target business data can be determined based on the optimized classifiers. After receiving the target business data, feature engineering can be performed on the target business data to extract multiple target partial feature data corresponding to the target business data. Then, based on the multiple target partial feature data and the optimized sub-classifiers, the target risk category can be determined. The server can then process the target business data based on the target risk category. For example, if the risk category of the target business data is determined to be positive, indicating the presence of risk, a first predetermined process can be executed. For example, a reminder message can be sent to the payer of the transaction corresponding to the business data, indicating a possible anomaly. Another example is a reminder message can be sent to the business management system to terminate the transaction corresponding to the business data. If the risk category of the business data is negative, standard processing can be performed on the business data.

[0051] In the above embodiment, the full feature sample set is split into multiple partial feature sample sets according to business logic, and learning can be performed from multiple perspectives. Compared with splicing all categories of features together, splitting according to business logic in this solution can prevent the dimensional disaster problem caused by too many types of features during the training process; moreover, since the meanings of features from different perspectives are often quite different, direct splicing makes it difficult to analyze the features. By adopting different perspectives for modeling and learning, the accuracy of business risk classification can be improved. In addition, by generating multi-perspective comparative learning constraints based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample according to the sub-classifiers corresponding to each partial feature sample set, the sub-classifiers corresponding to each partial feature sample set are optimized. This can constrain the output space of the same training sample in different perspectives to be as close as possible. By fixing one perspective and optimizing the other, and optimizing two by two, it is possible to continuously iterate and optimize between multiple perspectives, thereby improving the effect of multi-perspective learning and thereby improving the prediction effect of the sub-classifier. Furthermore, by determining the target risk category corresponding to the target business data based on the optimized sub-classifier and processing the business data based on the target risk category, the efficiency and accuracy of business data processing can be improved, the loss of customer assets can be avoided, losses can be reduced, and the user experience can be improved.

[0052] In some embodiments of the present specification, risk categories may include positive and negative categories, and accordingly, the full feature sample set may include positive samples and negative samples, wherein the number of positive samples in the full feature sample set is less than the number of negative samples in the full feature sample set, the label corresponding to the positive sample is the positive class, and the label corresponding to the negative sample is the negative class; after obtaining the full feature sample set and label set with feature data for characterizing business risk characteristics, it may also include: randomly oversampling the full feature sample set to increase the number of positive samples in the full feature sample set to obtain the oversampled full feature sample set; randomly undersampling the oversampled full feature sample set to reduce the number of negative samples in the full feature sample set.

[0053] Considering that the number of positive samples in the full feature sample set is generally far less than the number of negative samples, there is an extreme imbalance, so it is necessary to perform necessary sampling processing on the data. After obtaining the full feature sample set with feature data used to characterize business risk characteristics, the full feature sample set can be randomly oversampled to increase the number of positive samples in the full feature sample set to obtain the oversampled full feature sample set. Afterwards, the oversampled full feature sample set can be randomly undersampled to reduce the number of negative samples in the full feature sample set. In the above manner, the number of positive samples in the full feature sample set can be increased, the separability of positive samples in the classification space can be increased, the prediction accuracy of the classifier can be improved, and the accuracy and efficiency of business data processing can be improved, thereby reducing customer asset loss.

[0054] In some embodiments of the present specification, the full feature sample set can be randomly undersampled to reduce the number of negative samples in the full feature sample set, and then the positive samples in the undersampled full feature sample set can be oversampled to increase the number of positive samples in the full feature sample set. For example, among all financial transaction data, there is an extreme imbalance in financial anti-fraud data, and anti-fraud sample data is very scarce. Therefore, it is necessary to perform necessary sampling processing on the data. First, Tomeklink is used as an undersampling technology to address the imbalance problem of positive and negative samples in anti-fraud financial data. It can mark out the positive sample pairs that are close to the negative samples in the sample, and then eliminate the negative samples with a large number of them. At the same time, the positive samples are oversampled to increase the separability of positive samples in the classification space.

[0055] In some embodiments of the present specification, after constructing a sub-classifier using each partial feature sample set and a label set in a plurality of partial feature sample sets to obtain a sub-classifier corresponding to each partial feature sample set, the method may also include: classifying each partial feature sample in each partial feature sample set using the sub-classifier corresponding to each partial feature sample set; extracting neighboring partial feature samples of each partial feature sample in each partial feature sample set, and calculating the inter-class similarity matrix corresponding to each partial feature sample set based on the similarities and differences between the categories of each partial feature sample in each partial feature sample set and the categories of its neighboring partial feature samples; accordingly, optimizing the sub-classifier corresponding to each partial feature sample set based on multi-view contrast learning constraints, including: generating inter-class similarity constraints based on the inter-class similarity matrix corresponding to each partial feature sample set; and optimizing the sub-classifier corresponding to each partial feature sample set based on the multi-view contrast learning constraints and the inter-class similarity constraints.

[0056] Considering that business data often has a certain degree of temporal dependence, similarities inevitably exist between different transaction features of samples under the same identity verification. Therefore, the similarity between neighboring samples must be fully considered. Spatially similar samples should be closer in the output space than spatially distant samples. Spatially similar heterogeneous samples contain more accurate and rich discriminant information than spatially distant heterogeneous samples. Inter-class similarity constraints can be designed based on the relationships between similar and heterogeneous neighbors of samples, allowing the model to fully learn the discriminant information between similar and heterogeneous neighboring samples, thereby optimizing the model's generalization effect. For example, the k closest neighboring samples are determined based on Euclidean distance.

[0057] The server can use the subclassifier corresponding to each partial feature sample set to classify each partial feature sample in each partial feature sample set. The server can then extract the neighboring partial feature samples of each partial feature sample in each partial feature sample set and calculate the inter-class similarity matrix corresponding to each partial feature sample set based on the similarities and differences between the categories of each partial feature sample in each partial feature sample set and the categories of its neighboring partial feature samples. After obtaining the inter-class similarity matrix, the server can generate inter-class similarity constraints based on the inter-class similarity matrix. The subclassifier corresponding to each partial feature sample set can then be optimized based on the multi-view comparative learning constraints and the inter-class similarity constraints. By further optimizing the classifier based on the similarities and differences between the categories of each partial feature sample and its neighboring partial feature samples, the prediction performance of the classifier can be further improved, including the accuracy, precision, and recall rates of the predictions.

[0058] In some embodiments of this specification, the multi-view contrastive learning constraints may be:

[0059]

[0060] Among them, R vcl is the multi-view contrast learning constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, and f p is the subclassifier corresponding to the pth partial feature sample set, f q is the sub-classifier corresponding to the qth partial feature sample set, is the i-th partial feature sample in the p-th partial feature sample set, is the i-th partial feature sample in the q-th partial feature sample set, and In this way, we can construct multi-view contrastive learning constraints based on the classification results of partial feature samples corresponding to the same full feature sample by different sub-classifiers.

[0061] In some embodiments of this specification, the elements in the inter-class similarity matrix may be:

[0062]

[0063] in, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them, v = 1, 2, ... V, V is the number of multiple partial feature sample sets, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, express The neighboring feature samples of express The neighboring partial feature samples, i, j = 1, 2, ... N, N is the number of partial feature samples in each partial feature sample set. In the above manner, the inter-class similarity matrix corresponding to each partial feature sample set can be calculated based on the similarity between the partial feature sample and its neighboring partial feature samples.

[0064] In some embodiments of this specification, the inter-class similarity constraint may be:

[0065]

[0066] Among them, R bcs is the inter-class similarity constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, is the subclassifier corresponding to the vth partial feature sample set, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and Through the above method, the inter-class similarity constraint can be constructed based on the inter-class similarity matrix corresponding to each part of the feature sample set.

[0067] In some embodiments of the present specification, optimizing the sub-classifiers corresponding to each partial feature sample set according to the multi-view contrast learning constraint and the inter-class similarity constraint may include: minimizing the objective function until a preset number of iterations is reached or the difference between the loss values ​​of two loss functions is less than a preset threshold, wherein the objective function is:

[0068] L=R emp +αR bcs +γR vcl

[0069]

[0070] Among them, L is the objective function, R emp is the experience loss, R vcl For multi-view contrast learning constraints, R bcs is the inter-class similarity constraint, α and γ are hyperparameters, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, For label collection The corresponding label.

[0071] Specifically, this optimization problem can be solved using gradient descent. By minimizing the model's objective function until a preset number of iterations is reached or the difference between the loss values ​​of two loss functions is less than a preset threshold, the final sub-classifier is obtained. This approach can improve the model's prediction accuracy, thereby improving the accuracy and efficiency of business data processing.

[0072] In some embodiments of the present specification, a target classifier can be generated based on the sub-classifiers corresponding to the optimized feature sample sets of each part; the target classifier is used to process the target business data to determine the risk type of the business corresponding to the target business data; and the target business data is processed based on the risk type of the business corresponding to the target business data.

[0073] After obtaining the sub-classifiers corresponding to the optimized feature sample sets, a target classifier can be generated based on each sub-classifier. In one embodiment, the final target classifier can be obtained by integrating the three sub-classifiers. The specific discriminant function is as follows:

[0074]

[0075] In one embodiment, the three sub-classifiers can be connected together for training using the AdaBoost algorithm. By initializing the sub-classifiers with equal weight coefficients, the loss function is continuously optimized through training. After training, the weights of each sub-classifier can be obtained, and the target classifier can be obtained. The specific loss function is as follows:

[0076]

[0077] After obtaining the target classifier, the target business data can be processed using the target classifier. Specifically, based on the target business data, partial feature data corresponding to each sub-classifier can be generated, and each generated partial feature data can be input into each classifier. After obtaining the classification result, the final classification result can be obtained based on the target classifier. After obtaining the final classification result, the target business data can be processed based on the risk type of the target business data. In this way, a final target classifier can be generated based on multiple sub-classifiers to predict the risk type of the target business data, and then the business data can be processed. This can improve the efficiency and accuracy of business data processing and avoid the loss of customer assets.

[0078] The above method is described below with reference to a specific embodiment. However, it should be noted that this specific embodiment is only for better illustrating this specification and does not constitute an improper limitation to this specification.

[0079] Please refer to Figure 2 , which shows a flow chart of the business data processing method in this specific embodiment. The business data processing method in this specific embodiment can be applied to anti-fraud detection. In this specific embodiment, the model training process combines inter-class similarity constraints and a customer asset churn prediction model using multi-perspective learning to improve the accuracy of anti-fraud predictions.

[0080] like Figure 2As shown, the following steps are included: first, transaction data is pulled, data preprocessing and feature engineering are performed on samples, and the unbalanced sample classes are detected using Tomeklink technology (for example, after obtaining Tomeklink pairs, positive samples are upsampled and negative samples are eliminated for downsampling), a training set is constructed, and samples are input into the model. Different feature subspaces are learned through multi-perspective comparison, and sub-classifiers are obtained by constraining inter-class similarity and minimizing empirical loss. The anti-fraud prediction model is then obtained through AdaBoost classifier ensemble training.

[0081] The detailed training process is mainly divided into three parts: Tomeklink data processing, feature engineering, model building and training. The following is an explanation of each part:

[0082] Tomeklink's data processing uses sample customer selection criteria based on the customer's online identity information, funds flow information, historical transaction data, and behavioral analysis. This information consists of two parts: login device information and basic customer information, including age, gender, and region. Funds flow information includes both the customer's own funds and information about the user to whom the funds flow. Historical transaction data is also analyzed to identify suspicious behavior.

[0083] Observe the data columns in the data table that contain basic customer information, customer asset information, customer holding information, and customer behavior information. For columns with missing values, fill them in a specific way. For example, missing values ​​of numerical features are filled with the column value '0', and missing values ​​of non-numerical features are filled with "unknown".

[0084] Since the positive samples in the dataset are very few and precious, we use Tomeklink technology to detect the data in the samples. We use the nearest neighbor algorithm between the data to find Tomeklink pairs of similar categories. After finding them, we remove the negative samples (noise) in the data, leaving a small number of positive sample data for subsequent training. We use a random 3 to 5 times upsampling for the positive samples.

[0085] In the feature engineering phase, we can construct a basic network access information perspective. Features within this basic information perspective include basic information about the user's device, mobile operating system, protocol stack, login time, gender, age group, marital status, housing, vehicle, occupation, and industry. We can also construct a transaction feature perspective. This includes information about the account balance of the transacting parties and information about the user who is withdrawing funds. We can also construct a perspective to determine whether any business operations are suspicious or illegal. This information includes historical user transaction history, including the total amount in the user's account, fluctuations in the amount, and whether the operation is suspicious. We can then construct label information, defining positive samples as fraudulent and labeling them as 1, representing the first category of samples (ω1). Negative samples are defined as normal transactions and labeled as -1, representing the second category of samples (ω2).

[0086] After model building and training, the inter-class similarity matrix can be constructed. Within the perspective, the similarity relationship between neighboring samples is fully considered. For samples of the same class, the closer the spatial distribution, the closer the output space. For heterogeneous samples, the closer the spatial distribution, the more accurate and rich the discriminant information they contain. The inter-class similarity matrix is ​​constructed based on the relationship between the samples' similar and heterogeneous neighbors. The elements in the inter-class similarity matrix are:

[0087]

[0088] in, is the element in the inter-class similarity matrix corresponding to the partial feature sample set under the v-th perspective, indicating and The similarity between them, v = 1, 2, ... V, V is the number of perspectives, represents the i-th partial feature sample in the partial feature sample set corresponding to the v-th perspective, represents the jth partial feature sample in the partial feature sample set corresponding to the vth perspective, express The neighboring feature samples of express The neighboring partial feature samples, i, j = 1, 2, ... N, N is the number of partial feature samples in the partial feature sample set corresponding to each perspective.

[0089] Within the view, the inter-class similarity matrix is ​​used to construct the inter-class similarity constraint, which constrains the outputs of the neighboring samples of the same class to be as similar as possible, and the outputs of the neighboring samples of different classes to be as different as possible. The purpose is to optimize the classifier by using the spatial distribution information of the neighbors of the same class and different classes, so that the classifier can learn the overall distribution information of the samples and make the classification boundary pass through the middle area of ​​the two classes of samples as much as possible. The inter-class similarity constraint term R bcsThe specific calculation method is shown as follows:

[0090]

[0091] Among them, R bcs is the inter-class similarity constraint, V is the number of perspectives, N is the number of partial feature samples in each partial feature sample set corresponding to each perspective, represents the i-th partial feature sample in the partial feature sample set corresponding to the v-th perspective, represents the jth partial feature sample in the partial feature sample set corresponding to the vth perspective, f v is the sub-classifier corresponding to the partial feature sample set corresponding to the v-th perspective, is the element in the inter-class similarity matrix corresponding to the partial feature sample set corresponding to the v-th perspective, indicating and The similarity between them.

[0092] Between different perspectives, we make full use of the advantages of multi-perspective learning and design multi-perspective comparative learning constraints to constrain the output space of the same training sample in different perspectives to be as close as possible, so that multiple perspectives can be optimized in pairs, and the collaborative learning between multiple perspectives can be used to improve the effect of multi-perspective learning. vcl The calculation is as follows:

[0093]

[0094] Among them, R vcl is the multi-view contrast learning constraint, V is the number of viewpoints, N is the number of partial feature samples in each partial feature sample set corresponding to each viewpoint, and f p is the subclassifier corresponding to the partial feature sample set corresponding to the p-th perspective, f q is the sub-classifier corresponding to the partial feature sample set corresponding to the q-th perspective, is the i-th partial feature sample in the partial feature sample set corresponding to the p-th perspective, is the i-th partial feature sample in the partial feature sample set corresponding to the q-th perspective, and They all correspond to the i-th full-feature sample in the full-feature sample set.

[0095] Afterwards, the gradient descent method can be used to solve this optimization problem by minimizing the model's objective function until the preset number of iterations is reached or the difference between the loss values ​​of the two loss functions is less than the preset threshold. The final sub-classifiers f1, f2, and f3 are obtained. The objective function can be:

[0096] L=R emp +αR bcs+γR vcl

[0097]

[0098] Among them, L is the objective function, R emp is the experience loss, R vcl For multi-view contrast learning constraints, R bcs is the inter-class similarity constraint, α and γ are hyperparameters, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, For label collection The corresponding label.

[0099] The objective function can be expanded as follows:

[0100]

[0101] The gradient descent method can be used to solve this optimization problem. By minimizing the objective function of the model until the preset number of iterations is reached or the difference between the loss values ​​of the two loss functions is less than the preset threshold, the final sub-classifiers f1, f2, and f3 are obtained.

[0102] The final anti-fraud scene detection application based on tomeklink technology and multi-view learning is obtained by integrating three sub-classifiers. The three sub-classifiers are connected together for training through the AdaBoost algorithm. By initializing the sub-classifiers with equal weight coefficients, the loss function is continuously optimized through training. The specific loss function is as follows:

[0103]

[0104] When testing the model, for the test sample x, the discriminant function of the classifier is input to obtain the discrimination result of the model.

[0105]

[0106] The model in the above embodiment outperforms traditional semi-supervised learning algorithms in terms of precision, recall, and comprehensive evaluation for predicting customer asset churn, enabling more accurate prediction of customer asset churn. This model can be applied to financial institutions such as banks to accurately predict potential customer asset churn before it occurs. Relationship managers can then use the model's predictions to take appropriate action, thereby preventing customer asset churn, reducing losses, and enhancing the financial institutions' competitiveness within their respective industries.

[0107] Based on the same inventive concept, a business data processing device is also provided in the embodiments of this specification, as described in the following embodiments. Since the principles of solving the problems of the business data processing device are similar to those of the business data processing method, the implementation of the business data processing device can refer to the implementation of the business data processing method, and the repeated parts will not be repeated. As used below, the terms "unit" or "module" can be a combination of software and / or hardware that implements the predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, the implementation of hardware, or a combination of software and hardware, is also possible and conceived. Figure 3 This is a structural block diagram of a business data processing device according to an embodiment of this specification. Figure 3 As shown, it includes: an acquisition module 301, a splitting module 302, a construction module 303, a generation module 304 and an optimization module 305. The structure is described below.

[0108] The acquisition module 301 is used to obtain a full feature sample set and a label set having feature data for characterizing business risk features, wherein the label set includes the risk category corresponding to each full feature sample in the full feature sample set.

[0109] The splitting module 302 is used to split the full feature sample set into multiple partial feature sample sets according to preset business logic.

[0110] The construction module 303 is used to construct a sub-classifier using each partial feature sample set and a label set in the plurality of partial feature sample sets, and obtain a sub-classifier corresponding to each partial feature sample set.

[0111] The generation module 304 is used to generate multi-view comparative learning constraints based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to the partial feature sample sets.

[0112] The optimization module 305 is used to optimize the sub-classifiers corresponding to each part of the feature sample set based on the multi-view contrast learning constraint, wherein the optimized sub-classifiers are used to process the target business data.

[0113] In some embodiments of the present specification, the risk categories include positive and negative categories, and accordingly, the full feature sample set includes positive samples and negative samples, wherein the number of positive samples in the full feature sample set is less than the number of negative samples in the full feature sample set, the label corresponding to the positive sample is the positive class, and the label corresponding to the negative sample is the negative class; the acquisition module can also be used to: randomly oversample the full feature sample set to increase the number of positive samples in the full feature sample set to obtain the oversampled full feature sample set; randomly undersample the oversampled full feature sample set to reduce the number of negative samples in the full feature sample set

[0114] In some embodiments of the present specification, the device may further include a computing module, which may be specifically used to: construct a sub-classifier using each partial feature sample set and a label set in a plurality of partial feature sample sets, and after obtaining a sub-classifier corresponding to each partial feature sample set, classify each partial feature sample in each partial feature sample set using the sub-classifier corresponding to each partial feature sample set; extract neighboring partial feature samples of each partial feature sample in each partial feature sample set, and calculate the inter-class similarity matrix corresponding to each partial feature sample set based on the similarities and differences between the categories of each partial feature sample in each partial feature sample set and the categories of its neighboring partial feature samples; accordingly, the optimization module may be specifically used to: generate an inter-class similarity constraint based on the inter-class similarity matrix corresponding to each partial feature sample set; and optimize the sub-classifier corresponding to each partial feature sample set based on the multi-perspective comparative learning constraint and the inter-class similarity constraint.

[0115] In some embodiments of this specification, the multi-view contrastive learning constraints are:

[0116]

[0117] Among them, R vcl is the multi-view contrast learning constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, and f p is the subclassifier corresponding to the pth partial feature sample set, f q is the sub-classifier corresponding to the qth partial feature sample set, is the i-th partial feature sample in the p-th partial feature sample set, is the i-th partial feature sample in the q-th partial feature sample set, and They all correspond to the i-th full-feature sample in the full-feature sample set.

[0118] In some embodiments of this specification, the elements in the inter-class similarity matrix are:

[0119]

[0120] in, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them, v = 1, 2, ... V, V is the number of multiple partial feature sample sets, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, express The neighboring feature samples of express The neighboring partial feature samples of , i, j = 1, 2, ... N, N is the number of partial feature samples in each partial feature sample set.

[0121] In some embodiments of this specification, the inter-class similarity constraint is:

[0122]

[0123] Among them, R bcs is the inter-class similarity constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them.

[0124] In some embodiments of the present specification, the optimization module may be specifically configured to minimize an objective function until a preset number of iterations is reached or the difference between the loss values ​​of two loss functions is less than a preset threshold, wherein the objective function is:

[0125] L=R emp +αR bcs +γR vcl

[0126]

[0127] Among them, L is the objective function, R emp is the experience loss, R vcl For multi-view contrast learning constraints, R bcs is the inter-class similarity constraint, α and γ are hyperparameters, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, For label collection The corresponding label.

[0128] From the above description, it can be seen that the embodiments of this specification achieve the following technical effects: by splitting the full feature sample set into multiple partial feature sample sets according to business logic, learning can be performed from multiple perspectives. Compared with splicing all categories of features together, splitting according to business logic in this solution can prevent the dimensional disaster problem caused by too many types of features during the training process; moreover, since the meanings of features from different perspectives are often quite different, direct splicing makes it difficult to analyze the features. By adopting different perspectives for modeling and learning, the accuracy of business risk classification can be improved. In addition, by generating multi-perspective comparative learning constraints based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample according to the sub-classifiers corresponding to each partial feature sample set, the sub-classifiers corresponding to each partial feature sample set are optimized. This can constrain the output space of the same training sample in different perspectives to be as close as possible. By fixing one perspective and optimizing the other, and optimizing two by two, it is possible to continuously iterate and optimize between multiple perspectives, thereby improving the effect of multi-perspective learning and thereby improving the prediction effect of the sub-classifier. Furthermore, by determining the target risk category corresponding to the target business data based on the optimized sub-classifier and processing the business data based on the target risk category, the efficiency and accuracy of business data processing can be improved, the loss of customer assets can be avoided, losses can be reduced, and the user experience can be improved.

[0129] This specification also provides a computer device. Figure 4 The diagram shows the structure of a computer device based on the business data processing method provided in the embodiments of this specification. The computer device may include an input device 41, a processor 42, and a memory 43. The memory 43 is used to store processor-executable instructions. When the processor 42 executes these instructions, the steps of the business data processing method described in any of the above embodiments are implemented.

[0130] In this embodiment, the input device can specifically be one of the primary devices for exchanging information between a user and a computer system. The input device can include a keyboard, mouse, camera, scanner, light pen, handwriting input tablet, voice input device, etc.; the input device is used to input raw data and programs for processing these data into the computer. The input device can also receive data transmitted from other modules, units, and devices. The processor can be implemented in any appropriate manner. For example, the processor can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (such as software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers, etc. The memory can specifically be a memory device used to store information in modern information technology. The memory can include multiple levels. In digital systems, anything that can store binary data can be considered a memory device. In integrated circuits, a circuit with storage functionality that does not have a physical form is also called a memory device, such as a RAM or FIFO. In systems, a physical storage device is also called a memory device, such as a memory stick or a TF card.

[0131] In this embodiment, the specific functions and effects achieved by the computer device can be explained in comparison with other embodiments and will not be repeated here.

[0132] The embodiments of this specification also provide a computer storage medium based on the business data processing method, wherein the computer storage medium stores computer program instructions, and when the computer program instructions are executed, the steps of the business data processing method described in any of the above embodiments are implemented.

[0133] In this embodiment, the storage medium includes, but is not limited to, random access memory (RAM), read-only memory (ROM), cache, hard disk drive (HDD), or memory card. The memory can be used to store computer program instructions. The network communication unit can be an interface configured in accordance with the standards specified by the communication protocol for network connection communication.

[0134] In this embodiment, the functions and effects specifically implemented by the program instructions stored in the computer storage medium can be explained in comparison with other embodiments and will not be repeated here.

[0135] Obviously, those skilled in the art should understand that the various modules or steps of the above-mentioned embodiments of this specification can be implemented using a general-purpose computing device, they can be concentrated on a single computing device, or distributed across a network composed of multiple computing devices. Alternatively, they can be implemented using program code executable by the computing device, so that they can be stored in a storage device and executed by the computing device. In some cases, the steps shown or described can be performed in a different order than herein, or they can be made into separate integrated circuit modules, or multiple modules or steps can be made into a single integrated circuit module for implementation. Thus, the embodiments of this specification are not limited to any specific combination of hardware and software.

[0136] It should be understood that the above description is intended to be illustrative and not limiting. Numerous embodiments and applications beyond the examples provided will be readily apparent to those skilled in the art upon reading the above description. Therefore, the scope of this specification should not be determined with reference to the above description, but rather with reference to the preceding claims, along with the full scope of equivalents to which such claims are entitled.

[0137] The above description is merely a preferred embodiment of this specification and is not intended to limit this specification. Those skilled in the art will readily appreciate that various modifications and variations to the embodiments of this specification are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this specification shall be within the scope of protection of this specification.

Claims

1. A business data processing method, characterized in that: include: Obtaining a full-feature sample set and a label set having feature data for characterizing business risk features, wherein the label set includes a risk category corresponding to each full-feature sample in the full-feature sample set; Splitting the full feature sample set into multiple partial feature sample sets according to preset business logic; Constructing a sub-classifier using each partial feature sample set in the multiple partial feature sample sets and the label set to obtain a sub-classifier corresponding to each partial feature sample set; Generate a multi-perspective comparative learning constraint based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to the partial feature sample sets; classify the partial feature samples in the partial feature sample sets using the sub-classifiers corresponding to the partial feature sample sets; calculate an inter-class similarity matrix based on the similarities and differences between the categories of the partial feature samples in the partial feature sample sets and the categories of their neighboring partial feature samples; and generate an inter-class similarity constraint based on the inter-class similarity matrix; Optimizing the sub-classifiers corresponding to the feature sample sets based on the multi-view contrast learning constraint and the inter-class similarity constraint, wherein the optimized sub-classifiers are used to process the target business data; The multi-view contrastive learning constraints are: Among them, R vcl is the multi-view contrast learning constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, and f p 、f q are the sub-classifiers corresponding to the p-th and q-th partial feature sample sets respectively, are the i-th partial feature samples in the p-th and q-th partial feature sample sets respectively, and They all correspond to the i-th full-feature sample in the full-feature sample set.

2. The method according to claim 1, characterized in that The risk categories include positive and negative categories. Correspondingly, the full feature sample set includes positive samples and negative samples, wherein the number of positive samples in the full feature sample set is less than the number of negative samples in the full feature sample set, the labels corresponding to the positive samples are positive, and the labels corresponding to the negative samples are negative; After obtaining a full feature sample set and label set containing feature data for characterizing business risk characteristics, the following is also included: Randomly oversampling the full feature sample set to increase the number of positive samples in the full feature sample set to obtain an oversampled full feature sample set; Random undersampling is performed on the oversampled full feature sample set to reduce the number of negative samples in the full feature sample set.

3. The method according to claim 1, characterized in that The elements in the inter-class similarity matrix are: in, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them, v = 1, 2, ... V, V is the number of the plurality of partial feature sample sets, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, express The neighboring feature samples of express The neighboring partial feature samples, i, j = 1, 2, ... N, N is the number of partial feature samples in each partial feature sample set.

4. The method according to claim 1, wherein The inter-class similarity constraint is: Among them, R bcs is the inter-class similarity constraint, V is the number of the plurality of partial feature sample sets, N is the number of partial feature samples in each of the partial feature sample sets, represents the i-th partial feature sample in the v-th partial feature sample set, represents the jth partial feature sample in the vth partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, is the element in the inter-class similarity matrix corresponding to the vth partial feature sample set, indicating and The similarity between them.

5. The method according to claim 1, wherein Optimizing the sub-classifiers corresponding to the respective feature sample sets according to the multi-view contrast learning constraint and the inter-class similarity constraint, including: By minimizing the objective function until a preset number of iterations is reached or the difference between the loss values ​​of two loss functions is less than a preset threshold, wherein the objective function is: L=R emp +αR bcs +γR vcl Wherein, L is the objective function, R emp is the experience loss, R vcl is the multi-view contrastive learning constraint, R bcs is the inter-class similarity constraint, α and γ are hyperparameters, V is the number of the multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, represents the i-th partial feature sample in the v-th partial feature sample set, f v is the subclassifier corresponding to the vth partial feature sample set, For the tag set The corresponding label.

6. A business data processing device, characterized in that: include: An acquisition module, configured to acquire a full feature sample set and a label set having feature data for characterizing business risk characteristics, wherein the label set includes a risk category corresponding to each full feature sample in the full feature sample set; A splitting module, configured to split the full feature sample set into multiple partial feature sample sets according to preset business logic; A construction module, configured to construct a sub-classifier using each partial feature sample set in the plurality of partial feature sample sets and the label set, to obtain a sub-classifier corresponding to each partial feature sample set; A generation module is configured to generate a multi-perspective comparative learning constraint based on the similarities and differences between the risk classification results of the partial feature samples corresponding to the same full feature sample by the sub-classifiers corresponding to the partial feature sample sets; classify the partial feature samples in the partial feature sample sets using the sub-classifiers corresponding to the partial feature sample sets; calculate an inter-class similarity matrix based on the similarities and differences between the categories of the partial feature samples in the partial feature sample sets and the categories of their neighboring partial feature samples; and generate an inter-class similarity constraint based on the inter-class similarity matrix; an optimization module, configured to optimize the sub-classifiers corresponding to the respective feature sample sets based on the multi-view contrast learning constraints and the inter-class similarity constraints, wherein the optimized sub-classifiers are used to process the target business data; The multi-view contrastive learning constraints are: Among them, R vcl is the multi-view contrast learning constraint, V is the number of multiple partial feature sample sets, N is the number of partial feature samples in each partial feature sample set, and f p 、f q are the sub-classifiers corresponding to the p-th and q-th partial feature sample sets respectively, are the i-th partial feature samples in the p-th and q-th partial feature sample sets respectively, and They all correspond to the i-th full-feature sample in the full-feature sample set.

7. A computer device, characterized in that: The method comprises a processor and a memory for storing processor-executable instructions, wherein the processor implements the steps of the method according to any one of claims 1 to 5 when executing the instructions.

8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the instructions are executed, the steps of the method according to any one of claims 1 to 5 are implemented.

Images