System resource adjustment method, device and equipment

Abstract

The embodiments of this specification relate to the field of artificial intelligence technology, and disclose a system resource adjustment method, device and equipment, the method comprising obtaining a specified information set and a label set having feature data for characterizing user churn characteristics; the specified information set at least comprises a prediction sample set and a benchmark sample set; the label set comprises user churn results corresponding to each benchmark sample in the benchmark sample set; a plurality of sub-classifiers are constructed corresponding to the benchmark sample set and the label set; a similarity matrix corresponding to the prediction sample set is retrieved, so as to construct a sub-classifier using the similarity matrix to perform similarity loss on the prediction samples of the prediction sample set for user churn prediction; in combination with the similarity loss corresponding to the sub-classifier, the prediction sample set is used to optimize the corresponding sub-classifier to obtain an optimized sub-classifier, so as to evaluate the stability value obtained by the target user based on the optimized sub-classifier, and then adjust the system resources provided to the target user.

Description

Technical Field

[0001] The present specification relates to the field of artificial intelligence technology, and in particular, to a system resource adjustment method, device and equipment. Background Art

[0002] With the rapid development of big data service platform technology, financial service types and optional service channels are becoming more and more diverse and convenient, giving users more and more choices. Correspondingly, the users of financial institutions are becoming more and more unstable. In order to effectively retain users, the service systems of financial institutions usually need to perform a large amount of data analysis and resource adjustments to make the resources provided to users more in line with their needs. On this basis, user stability prediction in various application scenarios is very important for the accuracy of system resource adjustment.

[0003] At present, the commonly used user stability assessment method is mainly a classification method based on a supervised learning model. By modeling and analyzing the existing customer churn information, the trained model is used to predict the churn probability of new samples to determine the stability of each user. However, the classification method using a supervised learning model requires the use of information on known user churn results. However, with the diversification of financial service types, it is difficult to clearly define the user's churn results in many cases, resulting in the difficulty in efficiently and accurately selecting the sample data based on the prediction, which affects the accuracy of the user stability prediction, and then affects the accuracy and efficiency of the system resource adjustment. Therefore, a more accurate and efficient system resource adjustment method is urgently needed. Summary of the invention

[0004] The purpose of the embodiments of this specification is to provide a system resource adjustment method, device and equipment, which can improve the accuracy and efficiency of system resource adjustment.

[0005] This specification provides a system resource adjustment method, device and equipment which are implemented in the following ways:

[0006] A system resource adjustment method is applied to a server, the method comprising: obtaining a specified information set having feature data for characterizing user churn characteristics and a label set; the specified information set at least comprises a prediction sample set and a benchmark sample set; the label set comprises user churn results corresponding to each benchmark sample in the benchmark sample set; a plurality of sub-classifiers are constructed correspondingly using the benchmark sample set and the label set; wherein the number of sub-classifiers is an odd number greater than or equal to 3; retrieving a similarity matrix corresponding to the prediction sample set, so as to construct a similarity loss of user churn prediction by the sub-classifier for the prediction samples of the prediction sample set using the similarity matrix; wherein the similarity matrix is ​​used to characterize the distribution of the proximity of each prediction sample of the prediction sample set in the user churn feature space; combining the similarity loss corresponding to the sub-classifier, optimizing the corresponding sub-classifier using the prediction sample set to obtain an optimized sub-classifier, so as to adjust the system resources provided to the target user when the stability value obtained by evaluating the target user based on the optimized sub-classifier is lower than the preset stability value.

[0007] In some other embodiments of the method provided in this specification, the method further includes: optimizing each sub-classifier using the prediction sample set in combination with the difference loss between the sub-classifiers; wherein the difference loss is used to characterize the difference characteristics of any two sub-classifiers in predicting user churn for the prediction samples of the prediction sample set.

[0008] In some other embodiments of the method provided in this specification, the similarity matrix is ​​constructed in the following manner:

[0009]

[0010] s i,j =exp(-λ||x i -x j || 2 )

[0011] Among them, S U is the similarity matrix, whose dimension is |X U |×|X U |,|X U | is the prediction sample set X U The number of samples in , s i,j Used to characterize the prediction sample x i and x j The similarity between them is λ, and λ is a hyperparameter for adjusting the similarity measure.

[0012] In some other embodiments of the method provided in this specification, the similarity loss is constructed in the following manner:

[0013]

[0014] Among them, R simi (f k (X U ),S U ) represents the kth subclassifier for predicting the sample set X U The similarity loss of each prediction sample for user churn prediction, f k (·) is the discriminant function of the kth sub-classifier for predicting the prediction sample, S U is the similarity matrix, x m 、x n For the prediction sample set X U The prediction samples in .

[0015] In some other embodiments of the method provided in this specification, the difference loss is constructed in the following manner:

[0016]

[0017] Among them, R div (f p (X U ),f q (X U )) represents the prediction of sample set X by the p-th sub-classifier and the q-th sub-classifier U Differentiation loss for user churn prediction, f p (·),f q (·) are the discriminant functions of the p-th and q-th sub-classifiers, respectively, |X U |Prediction sample set X U The number of samples in, x is the predicted sample set X U The prediction samples in .

[0018] In some other embodiments of the method provided in this specification, each sub-classifier is optimized based on the following objective function:

[0019]

[0020] Among them, L(f,X U ,X,Y) is the objective function, f represents each sub-classifier, T is the number of sub-classifiers, j, k, p, q are the labels of sub-classifiers, f j (·),f k (·),f p (·),f q (·) are the discriminant functions of the jth, kth, pth, and qth sub-classifiers, respectively. X is the benchmark sample set, Y is the label set corresponding to the benchmark sample set, and X U is the prediction sample set, SU is the similarity matrix, R emp (f j (X), Y) is the experience loss, R simi (f k (X U ),S U ) represents the similarity loss, R div (f p (X U ),f q (X U )) represents the difference loss, and α and λ are hyper parameters.

[0021] In some other embodiments of the method provided in this specification, the feature data includes time series aggregation features and time series historical features; wherein the time series aggregation features refer to data obtained by extracting features of the user's specified information based on different time dimensions and a time series feature extraction algorithm; the time series historical features include time series distribution data obtained by statistics of the user's specified information based on different time dimensions.

[0022] On the other hand, the embodiment of the present specification also provides a system resource adjustment device, which is applied to a server, and the device includes: a data acquisition module, which is used to obtain a specified information set and a label set having feature data for characterizing user churn characteristics; the specified information set includes at least a predicted sample set and a benchmark sample set; the label set includes user churn results corresponding to each benchmark sample in the benchmark sample set; a classifier construction module, which is used to construct a plurality of sub-classifiers using the benchmark sample set and the label set; wherein the number of sub-classifiers is an odd number greater than or equal to 3; a distribution feature retrieval module, which is used to retrieve similarity corresponding to the predicted sample set. A similarity matrix is ​​used to construct a similarity loss of a sub-classifier for predicting user churn of the prediction samples of the prediction sample set by using the similarity matrix; wherein the similarity matrix is ​​used to characterize the proximity distribution of each prediction sample of the prediction sample set in the user churn feature space; a classifier optimization module is used to optimize the corresponding sub-classifier using the prediction sample set in combination with the similarity loss corresponding to the sub-classifier to obtain an optimized sub-classifier; a resource adjustment module is used to adjust the system resources provided to the target user when the stability value obtained by evaluating the target user based on the optimized sub-classifier is lower than the preset stability value.

[0023] In some other embodiments of the device provided in this specification, the classifier optimization module is also used to optimize each sub-classifier using the prediction sample set in combination with the difference loss between sub-classifiers; wherein the difference loss is used to characterize the difference characteristics of any two sub-classifiers in predicting user churn for the prediction samples of the prediction sample set.

[0024] On the other hand, an embodiment of the present specification also provides a system resource adjustment device, which is applied to a server. The device includes at least one processor and a memory for storing processor executable instructions. When the instructions are executed by the processor, the steps of the method described in any one or more of the above embodiments are implemented.

[0025] The system resource adjustment method, device and equipment provided by one or more embodiments of this specification initializes T sub-classifiers by first using sample data of known user churn results, and then optimizes the classifier by using the potential data distribution information between the prediction samples of a large number of unknown user churn results, calculates the similarity matrix between the prediction samples, and uses the similarity matrix to construct the similarity loss between the prediction samples, so that the prediction samples with similar distribution in the feature space are as similar as possible in the output space, so that the classifier learns the overall distribution information of the prediction samples, and makes the classification boundary pass through the low-density area in the sample space as much as possible. Furthermore, the prediction samples can also be used to construct the difference loss to maximize the diversity of the sub-classifiers on the prediction samples, so as to construct a high-diversity sub-classifier, prevent the results of all sub-classifiers from being consistent, and avoid the classifier being disturbed by noise samples. Thereby, the user churn prediction results can be made more in line with the user churn law, thereby improving the accuracy of user resource adjustment, and achieving the purpose of effectively retaining users and reducing resource losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art description. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative labor. In the drawings:

[0027] Figure 1 A flowchart of an embodiment of a system resource adjustment method provided in this specification;

[0028] Figure 2 A schematic diagram of constructing a user churn prediction model in one embodiment provided in this specification;

[0029] Figure 3 A schematic diagram of a flow chart of a system resource adjustment method in an embodiment provided in this specification;

[0030] Figure 4 This is a schematic diagram of the module structure of another system resource adjustment device provided in this specification. DETAILED DESCRIPTION

[0031] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below in conjunction with the drawings in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all of the embodiments. Based on one or more embodiments of the specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the embodiments of this specification.

[0032] In a scenario example provided in an embodiment of this specification, the system resource adjustment method can be applied to a device for performing system resource adjustment, and the device may include a server or a server cluster composed of multiple servers. For the target user, the server can extract feature data from various information of the target user as the test data of the target user, and then use a pre-configured algorithm or model to perform a stability assessment on the target user to obtain a stability assessment result of the target user, so as to adaptively adjust the resources of the financial institution based on the stability assessment result. The system resources may include data resources such as services and products provided or recommended to users. Usually, the above-mentioned data resources associated with each user will also occupy certain system hardware resources. By reasonably allocating the data resources associated with the user, the rationality of data resource allocation can be further improved. While retaining users, the overall performance of the financial institution service system can also be further improved.

[0033] Figure 1 It is a flowchart of an embodiment of the system resource adjustment method provided in this specification. Although this specification provides the method operation steps or device structure shown in the following embodiments or drawings, the method or device may include more or fewer operation steps or module units after partial merger 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 shown in the embodiments or drawings of this specification. When the method or module structure is applied in an actual device, server or terminal product, it can be executed sequentially or in parallel according to the method or module structure shown in the embodiments or drawings (for example, in an environment of parallel processors or multi-threaded processing, or even including distributed processing, server cluster implementation environment). A specific implementation example Figure 1 As shown, in one embodiment of the system resource adjustment method provided in this specification, the method can be applied to the data processing device, and the method may include the following steps:

[0034] S20: Acquire a specified information set and a label set having feature data for characterizing user churn characteristics; the specified information set includes at least a prediction sample set and a benchmark sample set; the label set includes user churn results corresponding to each benchmark sample in the benchmark sample set.

[0035] The server can obtain a specified information set and a label set. The specified information set may include multiple sample data. The sample data may include feature data for characterizing user churn characteristics. Accordingly, the specified information set may be a data set composed of feature data for characterizing user churn characteristics. The feature data may be feature data extracted from business data of users stored in a business system of a financial institution. Feature extraction may be performed through feature engineering. The feature data extraction method and feature type may be set according to the actual application scenario, which is not limited here. Of course, it may also include feature data extracted from user information obtained by the server from a platform associated with a financial institution.

[0036] In some embodiments, the sample data may be a prediction sample or a benchmark sample. The prediction sample may be sample data of unknown user churn results. The benchmark sample may be sample data of known user churn results. Accordingly, the specified information set may include at least a prediction sample set and a benchmark sample set. The label set may include user churn results corresponding to each benchmark sample. The user churn results may include, for example, user churn, user not churn, etc. For ease of processing, a single sample data may be set to correspond to a single user. Accordingly, feature data of users with known churn results and feature data of users with unknown churn results may be extracted respectively to construct benchmark samples and prediction samples. After associating the feature data corresponding to the benchmark sample and the prediction sample with the user identifier, they are stored in the benchmark sample set and the prediction sample set of the specified information set respectively. After associating the user churn result corresponding to the benchmark sample with the user identifier, they are stored in the label set.

[0037] In actual application scenarios, users may have applied for more than one business product or business service in the current financial institution. The types of business products or business services are also diverse, some of which may be continuous, such as deposits, etc., while some may be one-time applications, such as loans, financial management, etc. For different business products or business services, different user churn result determination methods may need to be formulated. For example, for deposits, if the balance in the user's account is lower than the preset balance threshold, and there is no capital flow in the user's account after a specified length of time, it can be considered that for deposit business, the user is a lost user. If the frequency of capital flow in the user's account is greater than the frequency threshold or the balance is greater than the balance threshold, the user can be considered as a non-lost user. For financial products, if the user's financial products in the current financial institution have all expired, and the user has not applied for any financial products after a period of time, it can be considered that for financial products, the user is a lost user. Alternatively, if the user's financial products in the current financial institution have not expired, the user can be considered as a non-lost user. Alternatively, the user's churn results can also be determined by combining multiple business products or business scenarios. Of course, the above-mentioned churn result determination method is only an example of preferred examples. In actual application scenarios, it can be flexibly configured as needed, and is not limited here.

[0038] The information characteristics of users corresponding to different products or services usually have great differences. It is also possible to construct a designated information set by distinguishing products or services, and then perform user stability prediction and system resource adjustment based on the corresponding designated information set, so that the prediction results can be more in line with the actual application scenario, thereby improving the prediction accuracy. For some new products or services, the number of corresponding users may be small. Accordingly, products or services with similar characteristics to the product or service can be obtained as designated products or designated services, and then the information of users corresponding to the designated product or designated service is obtained to construct a designated information set. Of course, the above implementation is only an example of preferred examples, and other methods of constructing designated information sets can also be used in specific implementations.

[0039] The pre-built information set can be stored locally or in a database. The server can extract the specified information set when adjusting system resources or building a prediction model. If the constructed information set refers to an information set composed of information about users corresponding to a specified product or a specified service scenario, an information set identifier can be set for each specified information set. Accordingly, the server can obtain the specified information set corresponding to the corresponding information set identifier according to the needs of the current test scenario for use in system resource adjustment under the current test scenario. Most of the business data in the business system is updated at a faster speed. Accordingly, the specified information set and label set can be dynamically updated at intervals to ensure the accuracy of the information in the information set.

[0040] S22: constructing a plurality of sub-classifiers using the benchmark sample set and the label set; wherein the number of the sub-classifiers is an odd number greater than or equal to 3.

[0041] The benchmark samples in the benchmark sample set can be randomly divided to obtain multiple subsets. The number of subsets can be an odd number greater than or equal to 3. The benchmark samples in the subsets are guaranteed to be non-overlapping as much as possible. Then, each subset and the label set can be used to construct a sub-classifier to obtain multiple sub-classifiers, and accordingly, the number of sub-classifiers is an odd number greater than or equal to 3. Of course, other sub-classifier construction methods can also be used. For example, in addition to the subsets used to construct the sub-classifiers, a part of the benchmark samples can be separately divided out as a verification sample set to verify and preliminarily optimize the sub-classifiers.

[0042] In the process of building sub-classifiers, you can choose the classification algorithm according to your needs, such as Bayesian, support vector machine, neural network, etc.

[0043] By using sample data of known user churn results to preliminarily construct a sub-classifier, the degree of fit of the sub-classifier on the sample data of known classification results can be maximized, preliminarily ensuring the accurate guidance of the constructed sub-classifier in classification.

[0044] S24: Retrieve the similarity matrix corresponding to the prediction sample set, and use the similarity matrix to construct a sub-classifier to perform similarity loss on the prediction samples of the prediction sample set for user churn prediction; wherein the similarity matrix is ​​used to characterize the proximity distribution of each prediction sample of the prediction sample set in the user churn feature space.

[0045] The server may also retrieve a pre-constructed similarity matrix corresponding to the prediction sample set. The similarity matrix may be used to characterize the distribution of the proximity of each prediction sample in the prediction sample set in the user churn feature space. The similarity matrix may be used to construct a subclassifier to perform similarity loss of user churn prediction on the prediction samples in the prediction sample set.

[0046] In practical applications, most user data do not have churn results, or it is difficult to determine whether there is churn. However, this part of data contains data distribution characteristics of a large number of users. By analyzing the data distribution characteristics of this part of users and optimizing the classifier based on its distribution characteristics, the final constructed prediction model can be more generalized and the classification results are closer to the actual user churn law. At present, most prediction methods usually use user data based on known user churn results to assign pseudo labels to this part of data when combining this part of data. Then, the user data assigned with pseudo labels is regarded as sample data of known user churn results to construct the prediction model. When building the model based on this method, the assignment of pseudo labels is overly dependent on user data with known user churn results; at the same time, the difference between pseudo labels and actual user churn results is usually difficult to guarantee, which makes the noise when building the prediction model relatively large, affecting the accuracy of the prediction results.

[0047] This embodiment utilizes the potential data distribution information between a large number of sample data of unknown user churn results, calculates the similarity matrix between each prediction sample, and uses the similarity matrix to construct the similarity loss between the unpredicted samples, so that the prediction samples with similar distribution in the user churn feature space are as similar as possible in the output space, so that the classifier can learn the overall distribution information of the prediction samples, and the classification boundary can pass through the low-density area in the sample space as much as possible, thereby improving the user churn prediction accuracy of the final model.

[0048] In some embodiments, the similarity matrix can be constructed in the following manner:

[0049]

[0050] s i,j =exp(-λ||x i -x j || 2 )

[0051] Among them, S U is the similarity matrix, whose dimension is |X U |×|X U |,|X U | is the prediction sample set X U The number of samples in , s i,j Used to characterize the prediction sample x i and x jThe similarity between them is λ, and λ is a hyperparameter for adjusting the similarity metric. By constructing the similarity matrix in this way, the accuracy of determining the proximity distribution of the predicted samples in the user churn space can be further improved, thereby improving the accuracy of user churn prediction. Of course, the above similarity matrix construction method is only an explanation of the preferred implementation method. In actual applications, other methods can also be used to construct the similarity matrix, such as discrete degree, central tendency, etc.

[0052] In some embodiments, the similarity loss may be constructed in the following manner:

[0053]

[0054] Among them, R simi (f k (X U ),S U ) represents the kth subclassifier for predicting the sample set X U The similarity loss of each prediction sample for user churn prediction, f k (·) is the discriminant function of the kth subclassifier for predicting the prediction sample. Correspondingly, f k (X U ) indicates that the prediction processing is performed on each prediction sample in the prediction sample set by using the kth subclassifier. The explanation of the discriminant function in other expressions in the embodiments of this specification can refer to this explanation and will not be repeated. U is the similarity matrix, x m 、x n For the prediction sample set X U The prediction samples in . By constructing the similarity loss through the above calculation model, the model optimization process can be made more accurate and efficient. Of course, the above similarity loss construction method is only an explanation of the preferred implementation method. In actual applications, the similarity loss can also be constructed according to the adopted classifier construction algorithm, or a simple deformation can be performed on the basis of the above calculation model.

[0055] S26: In combination with the similarity loss corresponding to the sub-classifier, the corresponding sub-classifier is optimized using the predicted sample set to obtain an optimized sub-classifier, so as to adjust the system resources provided to the target user when a stability value obtained by evaluating the target user based on the optimized sub-classifier is lower than a preset stability value.

[0056] The server can optimize the corresponding sub-classifier using the predicted sample set in combination with the similarity loss corresponding to the sub-classifier to obtain an optimized sub-classifier. For example, the similarity loss can be combined with the objective function of the classifier, and the optimization problem of the objective function can be solved by using methods such as gradient descent until a preset number of iterations is reached or the difference between the output values ​​of two objective functions is less than a preset threshold, and the corresponding optimized sub-classifier is output.

[0057] The server can evaluate the stability of the target user based on the optimized sub-classifier and output the stability value of the target user. The stability value can be used to characterize the possibility of the target user losing in the future. In the case where the stability value of the target user evaluated based on the optimized sub-classifier is lower than the preset stability value, the system resources provided to the target user can be adjusted. The preset stability value can be pre-set according to the needs of the actual application scenario. If it is predicted that the user is likely to lose in the future, the system resources provided to the target user can be adjusted to try to provide some resources that can retain the user.

[0058] The system resources may include data resources such as services and products provided or recommended to users. Usually, the above data resources associated with each user will also occupy certain system hardware resources. By reasonably allocating the data resources associated with the user, the rationality of data resource allocation can be further improved, which can retain users and further improve the overall performance of the financial institution service system.

[0059] In some other embodiments, the server can also optimize each sub-classifier using the prediction sample set in combination with the difference loss between the sub-classifiers. The difference loss can be used to characterize the difference characteristics of any two sub-classifiers in predicting user churn for the prediction samples of the prediction sample set. By constructing the difference loss using the prediction samples, the diversity of the sub-classifiers on the prediction samples can be maximized, thereby constructing a highly diverse sub-classifier, preventing the results of each sub-classifier from being consistent, and at the same time preventing the sub-classifiers from being disturbed by noise samples, thereby improving the accuracy of the overall prediction of the classifier.

[0060] In some embodiments, the difference loss can be constructed in the following manner:

[0061]

[0062] Among them, R div (f p (X U ),f q (X U )) represents the prediction of sample set X by the p-th sub-classifier and the q-th sub-classifier UDifferentiation loss for user churn prediction, f p (·),f q (·) are the discriminant functions of the p-th and q-th sub-classifiers, respectively, |X U |Prediction sample set X U The number of samples in, x is the predicted sample set X U The prediction samples in. By constructing the similarity loss through the above calculation model, the model optimization process can be made more accurate and efficient. Of course, the above difference loss construction method is only an explanation of the preferred implementation method. In actual applications, the difference loss can also be constructed according to the adopted classifier construction algorithm, or a simple deformation can be performed on the basis of the above calculation model.

[0063] In some implementations, the sub-classifiers may be optimized by combining the difference loss and the similarity loss. Preferably, each sub-classifier may be optimized based on the following objective function:

[0064]

[0065] Among them, L(f,X U ,X,Y) is the objective function, f represents each sub-classifier, T is the number of sub-classifiers, j, k, p, q are the labels of sub-classifiers, and f j (·),f k (·),f p (·),f q (·) are the discriminant functions of the jth, kth, pth, and qth sub-classifiers, respectively. X is the benchmark sample set, Y is the label set corresponding to the benchmark sample set, and X U is the prediction sample set, S U is the similarity matrix, R emp (f j (X), Y) is the experience loss, R simi (f k (X U ),S U ) represents the similarity loss, R div (f p (X U ),f q (X U )) represents the difference loss, and α and λ are hyper parameters.

[0066] In this embodiment, the T sub-classifiers can be iteratively optimized by minimizing the experience loss, similarity loss, and difference loss. Then, a user churn prediction model can be constructed based on the optimized T sub-classifiers. Accordingly, the discriminant function f of the user churn prediction model is * It can be expressed as,

[0067] f* =argmin f L(f,X U ,X,Y)

[0068] Among them, argmin means solving the objective function L(f,X U ,X,Y).

[0069] If there are two user churn prediction results, churn ω1 or no churn ω2, then the prediction result of the test data x corresponding to the target user can be expressed as,

[0070]

[0071] Among them, sign is the sign function. That is, the output result of the final model is obtained by voting by T sub-classifiers.

[0072] The above embodiment, by comprehensively considering the spatial distribution information between a large number of prediction samples, makes the classification boundary of the model have better generalization ability. At the same time, the prediction samples can also be used to improve the diversity of sub-classifiers, so that the final integrated model has better classification performance, and the user churn prediction results are more in line with the user churn law, thereby improving the accuracy of user resource adjustment and achieving the purpose of effectively retaining users.

[0073] With the development of Internet finance, the cost for corporate clients to reselect financial service institutions is getting lower and lower. If the loss of corporate clients becomes more serious, it will have an adverse impact on financial institutions, causing the reputation of financial institutions to decline and profits to decrease. At the same time, financial institutions may need to conduct large-scale service and product analysis and adjust data resources to come up with strategies to retain users, which will further lead to a waste of hardware resources and costs of financial institutions. Figure 2 As shown, in an implementation scenario provided by an embodiment of this specification, taking a corporate client as an example, the solution provided by the above embodiment is described as follows.

[0074] First, characteristic information related to legal person customer churn prediction is obtained from the data warehouse, including legal person basic information, legal person asset information, legal person loan information, and legal person transaction information. Data preprocessing and feature extraction are performed on the test samples, and the specified information set is constructed using the legal person's basic information characteristics and deposit and loan time series information characteristics.

[0075] Data selection. The relevant characteristics of corporate customer company deposits can be divided into four categories: basic information of legal persons, legal person asset information, legal person loan information, and legal person transaction information. The data scope can be determined by category, thereby determining the data table involved.

[0076] Data preprocessing. Observe the data columns in the data table that involve corporate client company deposit and loan information. Concatenate the data columns in different tables that involve corporate client company deposit information according to corporate client ID and time to form the original features. For columns with incorrect data types, convert them to the correct data types first. For example, the data type should be numeric, but a pseudo-string type is set in the data table. You can determine whether it is wrong based on the meaning of the data column name and convert the wrong ones. For columns with missing values, fill them in a certain way, such as missing values ​​of numerical features, fill them with "0", and missing values ​​of non-numeric features, fill them with "-1".

[0077] Then, feature extraction can be performed. For categorical features, such as economic nature and enterprise scale, they are One-Hot encoded, and for some numerical features with particularly large ranges, they are bucketed.

[0078] In some embodiments, the feature data may include time series aggregation features and time series historical features. The time series aggregation features may refer to data obtained by extracting features from the user's specified information based on different time dimensions and time series feature extraction algorithms. The time series historical features may include time series distribution data obtained by statistics of the user's specified information based on different time dimensions. The time dimensions may include, for example, the previous month, the previous two months, the previous three months, etc., as well as the previous second month, the previous third month, the previous fourth month, etc. The time series feature extraction algorithm may include, for example, mean value, variance, standard deviation, etc. By further combining the time series feature information to construct feature data, the features of users of different churn types can be more accurately characterized, thereby improving the accuracy of system resource allocation. By performing time series feature analysis on information in the user's information that fluctuates more significantly over time, a horizontal analysis of user features can be achieved, thereby greatly improving the accuracy of user stability prediction.

[0079] In some implementations, the time series aggregation feature F agg It can be extracted in the following way:

[0080] F agg =[f(feature) time ,time=1,2,3,4,5,6,1-2,1-3,1-4,1-5,1-6]

[0081] f() takes Mean() average, Max() maximum, Min() minimum, Std() standard deviation respectively, and the time periods are the previous month, previous two months, previous three months, previous four months, previous five months, previous six months, previous second month, previous third month, previous fourth month, previous fifth month, previous sixth month. Accordingly, each deposit and loan feature derives 44-dimensional time series aggregation features.

[0082] Time series historical features F his It can be extracted in the following way:

[0083] F his =[feature time ,time=1,2,3,4,5,6]

[0084] The time periods are the first month, the second month, the third month, the fourth month, the fifth month, and the sixth month. Accordingly, each deposit and loan feature can derive a 6-dimensional time series historical feature.

[0085] By the above method, based on the feature information at different time nodes, the time series feature information is constructed, so that the model can better take into account the previous feature information when learning the features of the current time node. Then, the feature data extracted by the above method can be used as feature data to construct a specified information set. The specified information set can include a benchmark sample and a prediction sample. Then, based on the above steps S20-S26, a user churn prediction model can be constructed. Figure 3 As shown, after that, the churn prediction of the target users can be performed based on the constructed user churn prediction model, and the resources can be adjusted based on the churn prediction results.

[0086] The solution provided in the above embodiment can maximize the fitting degree of the sub-classifiers and ensure the correct classification orientation of the sub-classifiers by initializing T sub-classifiers with sample data of known user churn results. Afterwards, the classifier is optimized using the potential data distribution information between the prediction samples of a large number of unknown user churn results, and the similarity matrix between the prediction samples is calculated. The similarity loss between the prediction samples is constructed using the similarity matrix, so that the prediction samples with similar distribution in the feature space are as similar as possible in the output space, so that the classifier learns the overall distribution information of the prediction samples and the classification boundary passes through the low-density area in the sample space as much as possible.

[0087] Furthermore, the prediction samples are used to construct the difference loss to maximize the diversity of the sub-classifiers on the prediction samples, thereby constructing high-diversity sub-classifiers to prevent the results of all sub-classifiers from becoming consistent and avoid the classifiers being disturbed by noise samples.

[0088] After that, the T sub-classifiers can be iteratively optimized by minimizing the experience loss, similarity loss, and difference loss. The final model is obtained by voting by the T sub-classifiers, so that the final integrated model has better prediction performance, and the user churn prediction results are more in line with the user churn law, thereby improving the accuracy of user resource adjustment and achieving the purpose of effectively retaining users.

[0089] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. For details, please refer to the description of the above-mentioned related processing related embodiments, and no further description is given here.

[0090] The above is a description of a specific embodiment of the specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0091] Based on the system resource adjustment method described above, one or more embodiments of this specification also provide a system resource adjustment device. The device may include a system, software (application), module, component, server, etc. that uses the method described in the embodiments of this specification and is combined with necessary implementation hardware. Based on the same innovative concept, the device in one or more embodiments provided in the embodiments of this specification is as described in the following embodiments. Since the implementation scheme of the device to solve the problem is similar to the method, the implementation of the specific device in the embodiments of this specification can refer to the implementation of the aforementioned method, and the repetitions will not be repeated. As used below, the term "unit" or "module" can be a combination of software and / or hardware that implements 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. Specifically, Figure 4 A schematic diagram of the module structure of an embodiment of a system resource adjustment device provided in the specification is shown as follows: Figure 4 As shown, applied to a server, the device may include:

[0092] The data acquisition module 102 can be used to obtain a specified information set and a label set having feature data for characterizing user churn characteristics; the specified information set includes at least a predicted sample set and a benchmark sample set; the label set includes user churn results corresponding to each benchmark sample in the benchmark sample set.

[0093] The classifier construction module 104 can be used to construct a plurality of sub-classifiers using the benchmark sample set and the label set; wherein the number of the sub-classifiers is an odd number greater than or equal to 3.

[0094] The distribution feature retrieval module 106 can be used to retrieve the similarity matrix corresponding to the prediction sample set, so as to use the similarity matrix to construct a sub-classifier to perform similarity loss on the prediction samples of the prediction sample set for user churn prediction; wherein the similarity matrix is ​​used to characterize the distribution of the degree of proximity of each prediction sample of the prediction sample set in the user churn feature space.

[0095] The classifier optimization module 108 may be used to optimize the corresponding sub-classifier by combining the similarity loss corresponding to the sub-classifier and utilizing the predicted sample set to obtain an optimized sub-classifier.

[0096] The resource adjustment module 110 may be configured to adjust the system resources provided to the target user when a stability value obtained by evaluating the target user based on the optimized sub-classifier is lower than a preset stability value.

[0097] In some other embodiments, the classifier optimization module 108 can also be used to optimize each sub-classifier using the prediction sample set in combination with the difference loss between the sub-classifiers. The difference loss can be used to characterize the difference characteristics of any two sub-classifiers in predicting user churn for the prediction samples of the prediction sample set.

[0098] It should be noted that the above-mentioned device may also include other implementations according to the description of the method embodiment. The specific implementation methods can refer to the description of the relevant method embodiment, and will not be described one by one here.

[0099] This specification also provides a system resource adjustment device, which can be applied to a separate system resource adjustment system or to a variety of computer data processing systems. The system can be a separate server, or it can include a server cluster, system (including distributed system), software (application), actual operation device, logic gate circuit device, quantum computer, etc. that uses one or more of the methods or one or more embodiments of this specification and a terminal device combined with necessary implementation hardware. In some embodiments, the device may include at least one processor and a memory for storing processor executable instructions, and when the instructions are executed by the processor, the steps of the method described in any one or more of the above embodiments are implemented.

[0100] The memory may include a physical device for storing information, which is usually to digitize the information and then store it in a medium using electrical, magnetic or optical means. The storage medium may include: a device that uses electrical energy to store information, such as various memories, such as RAM, ROM, etc.; a device that uses magnetic energy to store information, such as a hard disk, a floppy disk, a magnetic tape, a magnetic core memory, a magnetic bubble memory, a USB flash drive; a device that uses optical means to store information, such as a CD or a DVD. Of course, there are other types of readable storage media, such as quantum memory, graphene memory, etc.

[0101] It should be noted that the above-mentioned device may also include other implementation methods according to the description of the method or device embodiment. The specific implementation method can refer to the description of the relevant method embodiment, which will not be described one by one here.

[0102] It should be noted that the embodiments of this specification are not limited to the situations that must comply with the standard data model / template or the embodiments of this specification. Certain industry standards or slightly modified implementation plans based on the implementation described in the custom method or the embodiment can also achieve the same, equivalent or similar, or predictable implementation effects after deformation of the above-mentioned embodiments. The embodiments obtained by using these modified or deformed data acquisition, storage, judgment, processing methods, etc. can still fall within the scope of the optional implementation plans of this specification.

[0103] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. In the description of this specification, the description of the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of this specification. In this specification, the schematic representation of the above terms does not necessarily target the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, in the absence of contradiction, a person skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0104] The above description is only an embodiment of the present specification and is not intended to limit the present specification. For those skilled in the art, the present specification may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present specification shall be included in the scope of the claims of the present specification.

Claims

1. A system resource adjustment method, characterized in that: Applied to a server, the method comprises: Acquire a specified information set having characteristic data for characterizing user churn characteristics and a label set; the specified information set at least includes a prediction sample set and a benchmark sample set; the label set includes user churn results corresponding to each benchmark sample in the benchmark sample set; A plurality of sub-classifiers are constructed correspondingly using the benchmark sample set and the label set; wherein the number of sub-classifiers is an odd number greater than or equal to 3; Retrieving the similarity matrix corresponding to the prediction sample set, and using the similarity matrix to construct a sub-classifier to perform similarity loss of user churn prediction on the prediction samples of the prediction sample set; wherein the similarity matrix is ​​used to characterize the distribution of the closeness of each prediction sample of the prediction sample set in the user churn feature space; In combination with the similarity loss corresponding to the sub-classifier and the difference loss between the sub-classifiers, the corresponding sub-classifier is optimized using the prediction sample set; The similarity loss is constructed in the following way: Among them, R simi (f k (X U ),S U ) represents the kth subclassifier for predicting the sample set X U The similarity loss of each prediction sample for user churn prediction, f k (·) is the discriminant function of the kth sub-classifier for predicting the prediction sample, S U is the similarity matrix, x m 、x n For the prediction sample set X U The prediction samples in ; The difference loss is constructed in the following way: Among them, R div (f p (X U ),f q (X U )) represents the prediction of sample set X by the p-th sub-classifier and the q-th sub-classifier U Differentiation loss for user churn prediction, f p (g), f q (g) are the discriminant functions of the p-th and q-th sub-classifiers, respectively, |X U |Prediction sample set X U The number of samples in, x is the predicted sample set X U The prediction samples in ; Each sub-classifier is optimized based on the following objective function to obtain the optimized sub-classifier: Among them, L(f,X U ,X,Y) is the objective function, f represents each sub-classifier, T is the number of sub-classifiers, j, k, p, q are the labels of sub-classifiers, f j (g), f k (g), f p (g), f q (g) are the discriminant functions of the jth, kth, pth, and qth sub-classifiers, respectively. X is the benchmark sample set, Y is the label set corresponding to the benchmark sample set, and X U is the prediction sample set, S U is the similarity matrix, R emp (f j (X), Y) is the experience loss, R simi (f k (X U ),S U ) represents the similarity loss, R div (f p (X U ),f q (X U )) represents the difference loss, α and λ are hyper parameters; When a stability value of a target user obtained by evaluating the target user based on the optimized sub-classifier is lower than a preset stability value, the system resources provided to the target user are adjusted.

2. The method according to claim 1, characterized in that The similarity matrix is ​​constructed in the following way: s i,j =exp(-λ||x i -x j || 2 ) Among them, S U is the similarity matrix, whose dimension is |X U |×|X U |,|X U | is the prediction sample set X U The number of samples in , s i,j Used to characterize the prediction sample x i and x j The similarity between them is λ, and λ is a hyperparameter for adjusting the similarity measure.

3. The method according to claim 1, characterized in that The feature data includes time series aggregation features and time series historical features; wherein the time series aggregation features refer to data obtained by extracting features from the user's specified information based on different time dimensions and a time series feature extraction algorithm; the time series historical features include time series distribution data obtained by statistics of the user's specified information based on different time dimensions.

4. A system resource adjustment device, characterized in that: Applied to a server, the device comprises: A data acquisition module, used to acquire a specified information set having characteristic data for characterizing user churn characteristics, and a label set; the specified information set at least includes a prediction sample set and a benchmark sample set; the label set includes user churn results corresponding to each benchmark sample in the benchmark sample set; A classifier construction module, used to construct a plurality of sub-classifiers using the benchmark sample set and the label set; wherein the number of sub-classifiers is an odd number greater than or equal to 3; A distribution feature retrieval module, used to retrieve the similarity matrix corresponding to the prediction sample set, so as to use the similarity matrix to construct a sub-classifier to perform similarity loss on the prediction samples of the prediction sample set for user churn prediction; wherein the similarity matrix is ​​used to characterize the distribution of the closeness of each prediction sample of the prediction sample set in the user churn feature space; A classifier optimization module, used to optimize the corresponding sub-classifier by combining the similarity loss corresponding to the sub-classifier and the difference loss between the sub-classifiers and using the prediction sample set; The similarity loss is constructed in the following way: Among them, R simi (f k (X U ),S U ) represents the kth subclassifier for predicting the sample set X U The similarity loss of each prediction sample for user churn prediction, f k (·) is the discriminant function of the kth sub-classifier for predicting the prediction sample, S U is the similarity matrix, x m 、x n For the prediction sample set X U The prediction samples in ; The difference loss is constructed in the following way: Among them, R div (f p (X U ),f q (X U )) represents the prediction of sample set X by the p-th sub-classifier and the q-th sub-classifier U Differentiation loss for user churn prediction, f p (g), f q (g) are the discriminant functions of the p-th and q-th sub-classifiers, respectively, |X U |Prediction sample set X U The number of samples in, x is the predicted sample set X U The prediction samples in ; Each sub-classifier is optimized based on the following objective function to obtain the optimized sub-classifier: Among them, L(f,X U ,X,Y) is the objective function, f represents each sub-classifier, T is the number of sub-classifiers, j, k, p, q are the labels of sub-classifiers, f j (g), f k (g), f p (g), f q (g) are the discriminant functions of the jth, kth, pth, and qth sub-classifiers, respectively. X is the benchmark sample set, Y is the label set corresponding to the benchmark sample set, and X U is the prediction sample set, S U is the similarity matrix, R emp (f j (X), Y) is the experience loss, R simi (f k (X U ),S U ) represents the similarity loss, R div (f p (X U ),f q (X U )) represents the difference loss, α and λ are hyper parameters; The resource adjustment module is used to adjust the system resources provided to the target user when the stability value obtained by evaluating the target user based on the optimized sub-classifier is lower than the preset stability value.

5. A system resource adjustment device, characterized in that: Applied to a server, the device includes at least one processor and a memory for storing processor executable instructions, and the instructions, when executed by the processor, implement the steps of the method described in any one of claims 1 to 3 above.

Images