Customer loss risk prediction method and device, storage medium and electronic equipment

Abstract

The invention provides a customer loss risk prediction method and device, a storage medium and electronic equipment, and the method comprises the steps: obtaining a multi-source data set of each historical customer, and dividing each historical customer into a first customer and a second customer according to the data completeness condition of the multi-source data set; performing missing modal data completion on the multi-source data set of the first customer to obtain an original data set of the customer; taking the multi-source data set of the second customer as an original data set of the customer; on the basis of the original data set, multiple pieces of feature data of the historical customers are obtained through data cleaning and feature extraction, and loss state labels and the feature data of the historical customers serve as training samples; constructing a loss risk prediction model based on each training sample and a machine learning algorithm; and predicting the customer loss risk by using the loss risk prediction model. According to the method, personalized analysis can be performed on the customer data based on the learning ability of the loss risk prediction model, and the accuracy of loss risk prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data mining, and in particular to a method and device for predicting customer churn risk, a storage medium and an electronic device. Background Art

[0002] For banks and other corporate institutions, customer churn is one of the key factors affecting operations. In the process of corporate operations, in order to reduce customer churn, the risk of customer churn is usually predicted so that corresponding measures can be taken.

[0003] At present, the method of predicting customer churn risk is mainly to set churn risk assessment rules in advance according to actual business needs, such as setting risk levels corresponding to different transaction frequency value ranges, and then regularly counting customer behavior data such as customer transaction frequency, and evaluating customer behavior data through churn risk assessment rules to obtain the risk level of customer churn risk.

[0004] In actual business scenarios, there are usually some special customers in the customer group, such as new customers or customers with special operating habits. In the existing customer churn risk prediction method, different customers are analyzed based on a unified evaluation rule, but the customer churn tendency reflected by the customer behavior of special customers is often different from the normal situation, resulting in low accuracy in the prediction of customer churn risk by the existing prediction method, which is not conducive to accurately identifying customer churn risk. Summary of the invention

[0005] In view of this, an embodiment of the present invention provides a customer churn risk prediction method to solve the problem that in the existing customer churn risk prediction method, different types of customers are predicted based on a unified evaluation rule, and customers cannot be evaluated individually, resulting in low prediction accuracy.

[0006] The embodiment of the present invention also provides a customer churn risk prediction device to ensure the actual implementation and application of the above method.

[0007] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0008] A customer churn risk prediction method, comprising:

[0009] Determine churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes behavior data, user feature data, and emotional feature data corresponding to the historical customer;

[0010] For each multi-source data set corresponding to the historical customer, determine whether the multi-source data set meets a preset data completeness condition; if the multi-source data set does not meet the data completeness condition, treat the historical customer as the first customer; if the multi-source data set meets the data completeness condition, treat the historical customer as the second customer;

[0011] For each of the first customers, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer, and combining each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer;

[0012] taking the multi-source data set corresponding to each of the second customers as the original data set of the second customers;

[0013] Cleaning the original data set of each of the historical customers to obtain a feature data set of each of the historical customers;

[0014] For each of the historical customers, feature extraction is performed based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as training samples corresponding to the historical customer;

[0015] Performing model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model;

[0016] The churn risk prediction model is applied to predict the churn risk of target customers.

[0017] In the above method, optionally, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer includes:

[0018] Determine a user profile corresponding to the first customer and a user profile corresponding to each of the second customers;

[0019] Based on the user portraits corresponding to the respective second customers, the respective second customers are divided into groups to obtain respective customer groups; each of the customer groups includes at least one of the second customers;

[0020] Based on the user portrait corresponding to the first customer, determining a target customer group corresponding to the first customer from among the customer groups;

[0021] Determine each missing modal data category corresponding to the first customer;

[0022] Based on the multi-source data sets corresponding to each second customer in the target customer group and the multi-source data sets corresponding to the first customer, a preset multimodal variational autoencoder is applied to perform data prediction on each of the missing modal data categories, and synthetic data corresponding to each of the missing modal data categories generated by the multimodal variational autoencoder is obtained, and the synthetic data corresponding to each of the missing modal data categories is used as the modal prediction data corresponding to the first customer.

[0023] In the above method, optionally, the data cleaning of the original data set of each historical customer to obtain the characteristic data set of each historical customer includes:

[0024] For each original data set of the historical customer, performing abnormal data processing on the original data set, and using the original data set after the abnormal data processing as the first data set of the historical customer;

[0025] For each of the first data sets of the historical customers, performing missing value processing on the first data set, and using the first data set after the missing value processing as the second data set of the historical customer;

[0026] For each of the second data sets of the historical customers, performing redundancy cleaning on the second data sets, and using the second data sets after the redundancy cleaning as the third data set of the historical customers;

[0027] For each of the third data sets of the historical customers, data standardization is performed on the third data set, and the third data set after data standardization is used as the characteristic data set of the historical customer.

[0028] In the above method, optionally, the feature extraction is performed based on the feature data set of the historical customer to obtain a plurality of feature data corresponding to the historical customer, including:

[0029] Determining a first characteristic indicator based on the characteristic data set of the historical customer; the first characteristic indicator represents the business transaction frequency of the historical customer;

[0030] Determining a second characteristic indicator based on the characteristic data set of the historical customer; the second characteristic indicator represents the business transaction amount of the historical customer;

[0031] Determining a third characteristic indicator based on the characteristic data set of the historical customer; the third characteristic indicator represents the activity level of the historical customer;

[0032] Based on the feature data set of the historical customer, determining a fourth feature indicator; the fourth feature indicator represents the historical churn of the historical customer;

[0033] The first characteristic indicator, the second characteristic indicator, the third characteristic indicator and the fourth characteristic indicator are used as respective characteristic data corresponding to the historical customer.

[0034] The above method may optionally include performing model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model, including:

[0035] Constructing an initial model; the initial model is a machine learning model constructed based on the extreme gradient boosting tree;

[0036] The initial model is trained based on the training samples corresponding to each of the historical customers, and during the model training process, the expected maximum profit is used as the optimization target, and the Bayesian optimization algorithm is applied to optimize the hyperparameters of the initial model;

[0037] The initial model that completes the model training process is used as the churn risk prediction model.

[0038] In the above method, optionally, the applying the churn risk prediction model to predict the churn risk of target customers includes:

[0039] Acquire a multi-source data set corresponding to the target customer;

[0040] Determining a feature data set corresponding to the target customer based on the multi-source data set corresponding to the target customer;

[0041] Based on the feature data set corresponding to the target customer, determining each feature data corresponding to the target customer;

[0042] Inputting each characteristic data corresponding to the target customer into the churn risk prediction model, using the churn risk prediction model to predict the churn risk of the target customer, and obtaining the churn probability output by the churn risk prediction model;

[0043] Performing score conversion on the churn probability to obtain a churn risk score corresponding to the churn probability;

[0044] Based on the churn risk score, a churn warning report of the target customer is generated.

[0045] In the above method, optionally, generating a churn warning report of the target customer based on the churn risk score includes:

[0046] Determining a target risk level corresponding to the churn risk score;

[0047] Determine the target customer maintenance strategy corresponding to the target risk level among various pre-set customer maintenance strategies;

[0048] Determine characteristic change trend information corresponding to the target customers;

[0049] The target customer loss warning report is composed of the loss risk score, the target risk level, the target customer maintenance strategy and the characteristic change trend information.

[0050] A customer churn risk prediction device, comprising:

[0051] A first determination unit is used to determine the churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes the behavior data, user feature data and emotional feature data corresponding to the historical customer;

[0052] a data judgment unit, configured to judge, for each multi-source data set corresponding to the historical customer, whether the multi-source data set meets a preset data completeness condition, and if the multi-source data set does not meet the data completeness condition, regard the historical customer as the first customer; if the multi-source data set meets the data completeness condition, regard the historical customer as the second customer;

[0053] a data prediction unit, configured to perform missing modal data prediction processing on the multi-source data set corresponding to each of the first customers, obtain each modal prediction data corresponding to the first customer, and combine each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer;

[0054] A second determining unit, configured to use the multi-source data set corresponding to each of the second customers as the original data set of the second customer;

[0055] A data cleaning unit, used to clean the original data set of each of the historical customers to obtain a feature data set of each of the historical customers;

[0056] A feature extraction unit is used to extract features for each of the historical customers based on the feature data set of the historical customer, obtain multiple feature data corresponding to the historical customer, and use the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer as a training sample corresponding to the historical customer;

[0057] A model training unit, used to perform model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model;

[0058] The risk prediction unit is used to apply the churn risk prediction model to perform churn risk prediction on target customers.

[0059] A storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the customer churn risk prediction method as described above.

[0060] An electronic device includes a memory and one or more instructions, wherein the one or more instructions are stored in the memory and configured to be executed by one or more processors to perform the customer churn risk prediction method as described above.

[0061] A customer churn risk prediction method provided based on the above-mentioned embodiment of the present invention includes: determining churn status labels corresponding to multiple historical customers, and obtaining a multi-source data set corresponding to each historical customer; the multi-source data set corresponding to each historical customer includes behavioral data, user feature data and emotional feature data corresponding to the historical customer; for each multi-source data set corresponding to a historical customer, judging whether the multi-source data set meets a preset data completeness condition, if the multi-source data set does not meet the data completeness condition, the historical customer is taken as a first customer, and if the multi-source data set meets the data completeness condition, the historical customer is taken as a second customer; for each first customer, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer According to the present invention, each modal prediction data and the multi-source data set corresponding to the first customer are combined into the original data set of the first customer; the multi-source data set corresponding to each second customer is used as the original data set of the second customer; the original data set of each historical customer is cleaned to obtain the feature data set of each historical customer; for each historical customer, feature extraction is performed based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as the training sample corresponding to the historical customer; model training is performed based on the training samples corresponding to each historical customer and the preset machine learning algorithm to obtain a churn risk prediction model; and the churn risk prediction model is used to predict the churn risk of the target customer. By using the method provided in the embodiment of the present invention, corresponding training samples can be constructed based on the multi-source data of historical customers with clear churn status, and the corresponding machine learning model can be obtained as the churn risk prediction model through training with the training samples. The churn risk prediction model can learn and extract useful information from the relevant data of various real customers. When it is necessary to predict the churn risk of the target customer, the churn risk prediction model can be applied to analyze the relevant data of the target customer, so as to predict the churn risk of the target customer. For different types of customers, the churn risk prediction model can perform personalized analysis of customers based on the knowledge it has learned, which is conducive to improving the prediction accuracy of customer churn risk. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0063] Figure 1 A method flow chart of a customer churn risk prediction method provided by an embodiment of the present invention;

[0064] Figure 2 A method flow chart of another method for predicting customer churn risk provided by an embodiment of the present invention;

[0065] Figure 3 Another method flow chart of a method for predicting customer churn risk provided by an embodiment of the present invention;

[0066] Figure 4 A schematic diagram of the structure of a customer churn risk prediction device provided by an embodiment of the present invention;

[0067] Figure 5 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.

[0070] The embodiment of the present invention provides a method for predicting customer churn risk. The method can be applied to a customer churn early warning system of a bank. The execution subject can be a processor of the system. The method flow chart of the method is as follows: Figure 1 As shown, including:

[0071] S101: Determine churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes behavior data, user feature data, and emotional feature data corresponding to the historical customer;

[0072] In the method provided by the embodiment of the present invention, customers with a clear churn status can be screened from existing customers as historical customers, and a churn status label corresponding to each historical customer can be set according to the churn status of the customer, and the churn status label is a label that characterizes whether the customer has churned. Based on the customer identification of each historical customer, the relevant data of each historical customer is obtained from various pre-set data sources. Various data sources can include various business systems within the enterprise, and can also include third-party information channels. The acquired data can include user transaction data, account data, customer interaction data, product usage data, user basic information, user social media data, and third-party credit data, etc. In terms of the acquisition of transaction data, edge computing and Internet of Things (IoT) technologies can be introduced to capture transaction behavior data in real time through bank terminal equipment, and the real-time data collection capability can be enhanced in combination with IoT technology. For some important customer data, real-time data streaming technology can be used to ensure timely acquisition and processing of key information.

[0073] It should be noted that before using the technical solutions disclosed in the embodiments of the present invention, the types, scope of use, usage scenarios, etc. of personal information involved in the embodiments of the present invention should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0074] In the method provided by the embodiment of the present invention, the relevant data of each historical customer obtained from various data sources are classified, and the various types of data obtained are divided into multimodal data such as behavioral data, user characteristic data, and emotional characteristic data. The behavioral data refers to the relevant data generated by the customer's transaction behavior or business handling behavior, the user characteristic data refers to the data reflecting the personal characteristics of the customer, such as the customer's personal information, and the emotional characteristic data refers to the data showing the customer's emotional tendency, such as the customer's evaluation of the business, etc. The behavioral data, user characteristic data, and emotional characteristic data corresponding to each historical customer form a multi-source data set corresponding to the historical customer.

[0075] S102: for each multi-source data set corresponding to the historical customer, determining whether the multi-source data set meets a preset data completeness condition; if the multi-source data set does not meet the data completeness condition, treating the historical customer as a first customer; if the multi-source data set meets the data completeness condition, treating the historical customer as a second customer;

[0076] In the method provided in the embodiment of the present invention, various types of data are obtained in a pre-set manner during the data acquisition process to obtain a multi-source data set corresponding to historical customers. However, in actual scenarios, not all actual customers have corresponding data of various types. For example, new customers or customers with low activity usually lack data of some modalities, that is, some modal data in the multi-source data set corresponding to historical customers may be null values, and not all are valid data.

[0077] In the method provided by the embodiment of the present invention, a data completeness condition can be set in advance, that is, a condition for determining whether the data in the multi-source dataset is complete. For example, a complete modal data type can be set, and whether the multi-source dataset meets the data completeness condition is identified by determining whether the multi-source dataset contains valid modal data of each type of modal data type. If the multi-source dataset does not contain valid modal data of all modal data types, that is, the data of the multi-source dataset is incomplete, then the multi-source dataset is considered to not meet the data completeness condition; if the multi-source dataset already contains valid modal data of all modal data types, that is, the data of the multi-source dataset is complete, then the multi-source dataset is considered to meet the data completeness condition.

[0078] In the embodiment of the present invention, it is judged whether the multi-source data set corresponding to each historical customer meets the data completeness condition to determine whether the data in the multi-source data set corresponding to each historical customer is complete, and the historical customer whose data in the multi-source data set is incomplete is marked as the first customer, and the historical customer whose data in the multi-source data set is complete is marked as the second customer.

[0079] S103: for each of the first customers, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer, and combining each modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer;

[0080] In the method provided by the embodiment of the present invention, the missing modal data in the multi-source data set corresponding to each first customer can be predicted by a preset modal data prediction method to complete the missing modal data, and the modal data obtained by modal data prediction for the first customer and the original multi-source data set of the first customer are integrated into the original data of the first customer. That is, the missing modal data in the multi-source data set corresponding to each first customer is completed, and the multi-source data set after the modal data is completed is used as the original data of the corresponding first customer.

[0081] S104: taking the multi-source data set corresponding to each of the second customers as the original data set of the second customers;

[0082] In the method provided by the embodiment of the present invention, for each second customer, the data in the multi-source data set corresponding to the second customer is complete, and the multi-source data set corresponding to the second customer is used as the original data set of the second customer. In this way, the original data set of each historical customer can be determined, that is, if the data in the multi-source data set corresponding to the historical customer is complete, then the original data set of the historical customer is the multi-source data set corresponding to the second customer. If the data in the multi-source data set corresponding to the historical customer is incomplete, then the missing modal data in the multi-source data set corresponding to the historical customer is predicted, and the modal data in the multi-source data set is supplemented with the predicted data, and the multi-source data set after the modal data is supplemented is used as the original data set of the historical customer.

[0083] S105: performing data cleaning on the original data set of each of the historical customers to obtain a feature data set of each of the historical customers;

[0084] In the method provided in the embodiment of the present invention, the original data set of each historical customer can be cleaned according to the preset data cleaning method, and the original data set of the historical customer after data cleaning is used as the feature data set of the historical customer. The specific operation of data cleaning can be set according to actual needs, for example, missing values, abnormal values ​​and duplicate values ​​can be processed.

[0085] S106: For each of the historical customers, extract features based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and use the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer as a training sample corresponding to the historical customer;

[0086] In the method provided by the embodiment of the present invention, feature extraction is performed based on the feature data set of each historical customer through feature engineering. According to a predetermined feature extraction method, each feature data corresponding to each historical customer is extracted from the feature data set of each historical customer. The churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as a group of training samples, thereby obtaining training samples corresponding to each historical customer.

[0087] S107: Performing model training based on the training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model;

[0088] In the method provided in the embodiment of the present invention, the training samples corresponding to each historical customer are used as a training data set, and the training data set is used to train the machine learning model constructed based on the machine learning algorithm, so that the machine learning model can master the ability to predict customer churn risk based on various feature data, and the trained machine learning model is used as a churn risk prediction model. The specific machine learning algorithm can be selected based on actual needs, for example, machine learning algorithms such as extreme gradient boosting tree, logistic regression or decision tree can be selected.

[0089] S108: Apply the churn risk prediction model to predict the churn risk of target customers.

[0090] In the method provided by the embodiment of the present invention, when it is necessary to predict the churn risk of a customer whose churn status is unknown, the customer can be used as a target customer. The characteristic data of the target customer is determined, and then the churn risk prediction model is applied to analyze the characteristic data of the target customer, thereby predicting the churn risk of the target customer.

[0091] Based on the method provided by the embodiment of the present invention, the churn status labels corresponding to multiple historical customers are determined, and a multi-source data set corresponding to each historical customer is obtained; the multi-source data set corresponding to each historical customer includes behavioral data, user feature data and emotional feature data corresponding to the historical customer; for the multi-source data set corresponding to each historical customer, it is determined whether the multi-source data set meets the preset data completeness condition, if the multi-source data set does not meet the data completeness condition, the historical customer is taken as the first customer, if the multi-source data set meets the data completeness condition, the historical customer is taken as the second customer; for each first customer, the multi-source data set corresponding to the first customer is processed for missing modal data prediction, and each modal prediction data corresponding to the first customer is obtained, and each modal prediction data is processed. The predicted data and the multi-source data set corresponding to the first customer constitute the original data set of the first customer; the multi-source data set corresponding to each second customer is used as the original data set of the second customer; the original data set of each historical customer is cleaned to obtain the feature data set of each historical customer; for each historical customer, feature extraction is performed based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as the training sample corresponding to the historical customer; model training is performed based on the training samples corresponding to each historical customer and the preset machine learning algorithm to obtain a churn risk prediction model; and the churn risk prediction model is used to predict the churn risk of the target customer. By using the method provided in the embodiment of the present invention, corresponding training samples can be constructed based on the multi-source data of historical customers with clear churn status, and the corresponding machine learning model can be obtained as the churn risk prediction model through training with the training samples. The churn risk prediction model can learn and extract useful information from the relevant data of various real customers. When it is necessary to predict the churn risk of the target customer, the churn risk prediction model can be applied to analyze the relevant data of the target customer, thereby predicting the churn risk of the target customer. For different types of customers, the churn risk prediction model can perform personalized analysis of customers based on the knowledge it has learned, which is conducive to improving the prediction accuracy of customer churn risk.

[0092] exist Figure 1 Based on the method shown, Figure 2 As shown, in the method provided by the embodiment of the present invention, the process of performing missing modal data prediction processing on the multi-source data set corresponding to the first customer mentioned in step S103 to obtain each modal prediction data corresponding to the first customer includes:

[0093] S201: Determine a user profile corresponding to the first customer and a user profile corresponding to each of the second customers;

[0094] In the method provided by the embodiment of the present invention, modal data prediction is performed for the first customer with missing modal data using similar group characteristics. Specifically, for each historical customer, a corresponding user profile can be constructed for each historical customer using demographic information such as age, gender, and occupation. When predicting the missing modal data of the multi-source data set of the current first customer, the user profile corresponding to the current first customer and the user profiles corresponding to each second customer can be obtained.

[0095] S202: Divide each of the second customers into groups based on the user portraits corresponding to each of the second customers to obtain customer groups; each of the customer groups includes at least one of the second customers;

[0096] In the method provided in the embodiment of the present invention, each second customer can be divided into various customer groups based on the user portrait corresponding to each second customer and according to similar user characteristics. Each customer group contains at least one second customer. It can be understood that the second customers in each customer group have similar user characteristics of a certain type.

[0097] S203: Determine a target customer group corresponding to the first customer from among the customer groups based on the user portrait corresponding to the first customer;

[0098] In the method provided by the embodiment of the present invention, based on the user portrait of the current first customer, a customer group having similar characteristics to the current first customer can be determined from various customer groups, and the customer group can be used as the target customer group.

[0099] S204: Determine each missing modal data category corresponding to the first customer;

[0100] In the method provided by the embodiment of the present invention, the multi-source data set corresponding to the current first customer is analyzed to determine which type of modal data is missing in the multi-source data set, and the missing modal data category is used as the missing modal data category.

[0101] S205: Based on the multi-source data sets corresponding to each second customer in the target customer group and the multi-source data sets corresponding to the first customer, apply a preset multimodal variational autoencoder to perform data prediction on each of the missing modal data categories, obtain synthetic data corresponding to each of the missing modal data categories generated by the multimodal variational autoencoder, and use the synthetic data corresponding to each of the missing modal data categories as the modal prediction data corresponding to the first customer.

[0102] In the method provided by the embodiment of the present invention, the encoder, decoder and loss function in the multimodal variational autoencoder can be predefined. The multimodal variational autoencoder is trained using the multi-source data set corresponding to each second customer in the target customer group, so that the multimodal variational autoencoder learns the features of each modal data, and then the multi-source data set corresponding to the current first customer is input into the encoder in the multimodal variational autoencoder to generate a potential representation, and the missing modality is completed by the decoder, that is, the multimodal variational autoencoder is applied to learn the modal data features in the target customer group, and then on the basis of the multi-source data set corresponding to the first customer, the modal data of each missing modal data category is synthesized, thereby generating synthetic data corresponding to each missing modal data category, and each synthetic data generated by the multimodal variational autoencoder is used as the modal prediction data corresponding to the first customer.

[0103] exist Figure 1 On the basis of the method shown, in the method provided by the embodiment of the present invention, the process of performing data cleaning on the original data set of each of the historical customers mentioned in step S105 to obtain the feature data set of each of the historical customers includes:

[0104] For each original data set of the historical customer, performing abnormal data processing on the original data set, and using the original data set after the abnormal data processing as the first data set of the historical customer;

[0105] In the method provided by the embodiment of the present invention, after the original data set of each historical customer is obtained, a data warehouse is established to centrally store the original data sets of each historical customer for easy access, management and analysis.

[0106] In the method provided by the embodiment of the present invention, a multimodal anomaly detection method can be used to perform anomaly detection on the original data set of each historical customer, and then the abnormal data can be corrected. Specifically, the graph attention network and the time series model can be combined to detect cross-channel and cross-time abnormal data in the multi-source data set. The original data set of the historical customer after abnormal data processing is used as the first data set of the historical customer.

[0107] For each of the first data sets of the historical customers, performing missing value processing on the first data set, and using the first data set after the missing value processing as the second data set of the historical customer;

[0108] In the method provided by the embodiment of the present invention, for the first data set of historical customers, it can be detected whether there are missing values ​​in each type of modal data. If there are missing values, the missing values ​​are filled, for example, the mean, mode or the most recent observation value can be used for filling, and severely missing records can be discarded, etc. The first data set of the historical customer after the missing value processing is used as the second data set of the historical customer.

[0109] For each of the second data sets of the historical customers, performing redundancy cleaning on the second data sets, and using the second data sets after the redundancy cleaning as the third data set of the historical customers;

[0110] In the method provided by the embodiment of the present invention, customer information and transaction data may be repeatedly entered during the data collection process. Therefore, for the second data set of historical customers, it is necessary to identify whether there is redundant data, that is, repeated data, and clean up the redundant data. The second data set after the redundant data is cleaned up is used as the third data set of the historical customer.

[0111] For each of the third data sets of the historical customers, data standardization is performed on the third data set, and the third data set after data standardization is used as the characteristic data set of the historical customer.

[0112] In the method provided by the embodiment of the present invention, all data in the third data set of historical customers are subjected to data standardization processing in a unified format so that various types of data are stored in a unified format, such as standardization of date format, amount unit and other elements. The third data set of historical customers that has undergone data standardization processing is used as the characteristic data set of the historical customers. In general, in the data cleaning process, for each original data set of historical customers, operations such as abnormal data processing, missing value processing, redundant data cleaning and data standardization are sequentially performed, and the original data set that has undergone the above operations is used as the characteristic data set of the corresponding historical customer.

[0113] exist Figure 1 On the basis of the method shown, in the method provided by the embodiment of the present invention, the process of extracting features based on the feature data set of the historical customer mentioned in step S106 to obtain multiple feature data corresponding to the historical customer includes:

[0114] Determining a first characteristic indicator based on the characteristic data set of the historical customer; the first characteristic indicator represents the business transaction frequency of the historical customer;

[0115] In the method provided by the embodiment of the present invention, characteristic data related to the customer's business transactions can be obtained from the characteristic data set of the current historical customer, and the obtained characteristic data can be analyzed according to the predetermined analysis method of the business transaction characteristics to obtain a first characteristic indicator that characterizes the business transaction frequency of the current historical customer. The first characteristic indicator can be calculated based on the purchase frequency, which refers to the number of times a customer makes purchases or transactions within a specific time period. This feature can reflect the level of interaction activity between the customer and the bank or a certain type of product. By counting the number of purchases or transactions made by the customer in a day, a week, or a month, the changing trend of the purchase frequency is calculated, and the purchase frequency or the changing trend of the purchase frequency is used as the first characteristic indicator. If a customer's transaction frequency has shown a significant downward trend in the past few months, it may be an early sign of loss.

[0116] Determining a second characteristic indicator based on the characteristic data set of the historical customer; the second characteristic indicator represents the business transaction amount of the historical customer;

[0117] In the method provided by the embodiment of the present invention, characteristic data related to the customer's business transaction amount can be obtained from the characteristic data set of the current historical customer, and the relevant characteristic data can be analyzed according to the predetermined analysis method of the business transaction amount characteristics to obtain a second characteristic indicator that characterizes the size of the current historical customer's business transaction amount. The second characteristic indicator can be calculated based on the customer's purchase amount. The purchase amount refers to the total amount spent by the customer in each transaction. This feature can indicate the customer's capital flow and the degree of use of bank services. Significant fluctuations in the purchase amount may indicate a change in the customer's dependence on bank products. The average transaction amount of the customer over a period of time can be counted, and the fluctuation range of the purchase amount can be calculated. For example, the standard deviation or coefficient of variation can be used to measure the degree of fluctuation of the customer's transaction amount. The fluctuation range of the calculated purchase amount is used as the second characteristic indicator. If the volatility increases, it may mean the instability of the customer.

[0118] Determining a third characteristic indicator based on the characteristic data set of the historical customer; the third characteristic indicator represents the activity level of the historical customer;

[0119] In the method provided by the embodiment of the present invention, characteristic data related to customer activity can be obtained from the characteristic data set of current historical customers, and the relevant characteristic data can be analyzed according to the predetermined analysis method of activity characteristics to obtain a third characteristic indicator characterizing the current historical customer activity. The third characteristic indicator can be obtained based on the activity analysis. Customer activity can be obtained by measuring various types of interaction behaviors between customers and banks, for example, by extracting user login frequency, transaction activity, customer service interaction frequency, etc., converting the above data into continuous values, expressed as a monthly total activity score, and combining the activity change trend in the past few months to identify the fluctuation of activity, and use the fluctuation of activity as the third characteristic indicator. Low activity is usually a precursor to customer churn.

[0120] Based on the feature data set of the historical customer, determining a fourth feature indicator; the fourth feature indicator represents the historical churn of the historical customer;

[0121] In the method provided by the embodiment of the present invention, feature data related to the customer churn history can be obtained from the feature data set of the current historical customers, and a fourth feature indicator characterizing the historical churn situation of the current historical customers can be extracted therefrom. The historical churn situation refers to whether the customer has ever shown signs of churn in the past, such as suspending the use of banking services or reducing transaction activities for a period of time. Specifically, it is possible to count whether the customer has suspended the use of services, cancelled accounts, etc., record whether the customer has become active again after churn, and whether the behavior pattern has changed after recovery, and convert the churn history into a binary feature (0 means no churn history, 1 means churn history), which can be combined with the recovery situation after churn to measure the customer's potential churn risk. This historical behavior feature has a direct reference value for predicting future churn.

[0122] The first characteristic indicator, the second characteristic indicator, the third characteristic indicator and the fourth characteristic indicator are used as respective characteristic data corresponding to the historical customer.

[0123] In the method provided by the embodiment of the present invention, through feature engineering, multidimensional features that have a significant impact on customer churn can be extracted from the feature data set of historical customers, specifically including various feature indicators related to factors such as purchase frequency, purchase amount, customer activity and historical churn, and these feature indicators are used as various feature data corresponding to historical customers.

[0124] exist Figure 1 On the basis of the method shown, in the method provided by the embodiment of the present invention, the process of performing model training based on the training samples corresponding to each of the historical customers and the preset machine learning algorithm to obtain the churn risk prediction model mentioned in step S107 includes:

[0125] Constructing an initial model; the initial model is a machine learning model constructed based on the extreme gradient boosting tree;

[0126] In the method provided in the embodiment of the present invention, a Bayesian optimization algorithm is used to optimize the extreme gradient boosting tree to construct a churn risk prediction model. This combined method can optimize model performance while being profit-oriented to construct a churn risk prediction model suitable for complex business scenarios.

[0127] In the method provided by the embodiment of the present invention, an initial model is firstly constructed based on the extreme gradient boosting tree.

[0128] The initial model is trained based on the training samples corresponding to each of the historical customers, and during the model training process, the expected maximum profit is used as the optimization target, and the Bayesian optimization algorithm is applied to optimize the hyperparameters of the initial model;

[0129] In the method provided by the embodiment of the present invention, the binary label of whether the customer has churned is used as the target variable, and the various feature data of historical customers are prioritized by feature importance analysis, and the training samples of each historical customer are used as the training data set. At the same time, the hyperparameters of Bayesian optimization are defined, including the category weights that adjust the weight ratio of lost customers to non-lost customers, the subsampling rate that specifies the proportion of randomly selected training samples in each iteration, the learning rate that controls the step length of weight update in each iteration, and the maximum depth of the tree that determines the complexity of each decision tree. Then, with the expected maximum profit as the optimization goal, several sets of hyperparameter combinations are randomly selected, and the initial model is trained and evaluated based on the training data set. The objective function is fitted based on the Gaussian process to predict the performance of each set of hyperparameters. The expected improvement strategy is used to select the hyperparameter combination that is most likely to improve the objective function, train the new model and update the probability distribution, gradually approach the global optimal solution, and stop model training when the preset maximum number of iterations is 50. The trained and optimized model uses the optimal hyperparameter combination determined by the Bayesian optimization algorithm as the model parameter.

[0130] The initial model that completes the model training process is used as the churn risk prediction model.

[0131] In the method provided in the embodiment of the present invention, the trained initial model is used as the churn risk prediction model.

[0132] Based on the method provided in the embodiment of the present invention, the category weights and other hyperparameters of the tree are optimized through a Bayesian optimization algorithm based on the profit maximization criterion. Compared with the traditional model, the computing cost can be significantly reduced, and the profitability of the model is jointly improved through Bayesian optimization and category weight adjustment.

[0133] exist Figure 1 Based on the method shown, Figure 3 As shown, in the method provided by the embodiment of the present invention, the process of applying the churn risk prediction model to predict the churn risk of target customers mentioned in step S108 includes:

[0134] S301: Acquire a multi-source data set corresponding to the target customer;

[0135] In the method provided by the embodiment of the present invention, when it is necessary to predict the loss risk of target customers, a multi-source data set corresponding to the target customers can be obtained from multiple data sources. Similar to the multi-source data set corresponding to historical customers, the multi-source data set corresponding to the target customers also includes multimodal data such as behavioral data, user feature data, and emotional feature data corresponding to the target customers.

[0136] S302: Determine a feature data set corresponding to the target customer based on the multi-source data set corresponding to the target customer;

[0137] In the method provided in the embodiment of the present invention, it can be determined whether the multi-source data set corresponding to the target customer meets the preset data completeness conditions. If so, the multi-source data set is data cleaned, and the multi-source data set after data cleansing is used as the feature data set corresponding to the target customer. If the multi-source data set corresponding to the target customer does not meet the data completeness conditions, the multi-source data set is predicted for missing modal data to complete the modal data in the multi-source data set, and the multi-source data set with the completed modal data is data cleaned, and the multi-source data set after data cleansing is used as the feature data set corresponding to the target customer. The method for completing the modal data in the multi-source data set of the target customer can be implemented by using the method mentioned in the previous embodiment, which uses a multi-modal variational autoencoder to synthesize the missing modal data based on the feature data of similar customer groups.

[0138] S303: determining each feature data corresponding to the target customer based on the feature data set corresponding to the target customer;

[0139] In the method provided by the embodiment of the present invention, feature extraction is performed on the basis of a feature data set corresponding to the target customer, and each extracted feature data is used as the feature data corresponding to the target customer.

[0140] S304: Inputting each characteristic data corresponding to the target customer into the churn risk prediction model, and using the churn risk prediction model to predict the churn risk of the target customer to obtain a churn probability output by the churn risk prediction model;

[0141] In the method provided by the embodiment of the present invention, each characteristic data of the target customer is loaded into the input layer of the churn risk prediction model, and the churn risk prediction model analyzes the characteristic data and outputs the corresponding churn probability of the target customer.

[0142] S305: Perform score conversion on the churn probability to obtain a churn risk score corresponding to the churn probability;

[0143] In the method provided in the embodiment of the present invention, the churn probability corresponding to the target customer can be converted into a score through a preset score conversion method, that is, the churn probability is converted into a corresponding score, and the customer churn risk is quantified by the scoring system.

[0144] S306: Generate a churn warning report for the target customer based on the churn risk score.

[0145] In the method provided in the embodiment of the present invention, a target customer churn warning report can be generated based on the churn risk score corresponding to the target customer. The churn warning report includes the churn risk score and other data analyzed based on the churn risk score. The target customer churn warning report can be automatically sent to the account manager or other relevant staff corresponding to the target customer.

[0146] exist Figure 3 On the basis of the method shown, in the method provided by the embodiment of the present invention, the process of generating the target customer churn warning report based on the churn risk score mentioned in step S306 includes:

[0147] Determining a target risk level corresponding to the churn risk score;

[0148] In the method provided in the embodiment of the present invention, the scoring range corresponding to different risk levels can be preset, for example, the scoring value is a value between 0 and 100, and the higher the score, the greater the risk of churn. Customers can be divided into different risk levels according to the score, for example, 0-40 points represent low-risk customers, 40-70 points represent medium-risk customers, and 70-100 points represent high-risk customers.

[0149] According to the matching of the churn risk score corresponding to the target customer and the score range, the target risk level corresponding to the churn risk score can be determined from various pre-set risk levels.

[0150] Determine the target customer maintenance strategy corresponding to the target risk level among various pre-set customer maintenance strategies;

[0151] In the method provided in the embodiment of the present invention, customer maintenance strategies corresponding to different risk levels can be pre-set. For example, for high-risk customers, the customer maintenance strategy is to assign a dedicated account manager, who will contact the customer directly by phone or email to understand the customer's specific needs and problems, provide one-on-one consultation, and solve the customer's possible pain points. For medium-risk customers, the customer maintenance strategy is to automatically remind customers of account usage and recommend current promotions via text messages. If the customer does not complete the operation within a specific time, a follow-up reminder will be sent. For low-risk customers, the customer maintenance strategy is to regularly send customized financial advice to consolidate customer trust.

[0152] For target customers, you can find a customer maintenance strategy that matches them among various customer maintenance strategies based on their corresponding target risk levels as the target customer maintenance strategy.

[0153] Determine characteristic change trend information corresponding to the target customers;

[0154] In the method provided in the embodiment of the present invention, based on the characteristic data set of the target customers, the changing trends of the corresponding indicator data can be statistically analyzed to obtain the dynamic changing characteristics of customer behavior, such as the changing trends of indicators such as customer activity, transaction volume, and consumption frequency, and the dynamic changing characteristics of customer behavior can be used as the characteristic changing trend information of the target customers.

[0155] The target customer loss warning report is composed of the loss risk score, the target risk level, the target customer maintenance strategy and the characteristic change trend information.

[0156] In the method provided by the embodiment of the present invention, the target customer's corresponding churn risk score, target risk level, target customer maintenance strategy and characteristic change trend information are used as report content, thereby generating a target customer churn warning report.

[0157] Based on the method provided in the embodiment of the present invention, the customer's behavior change trend and suggestions for customer maintenance strategies can be displayed in the customer churn warning report, which is helpful for staff to better deal with the customer churn problem.

[0158] and Figure 1 Corresponding to the customer churn risk prediction method shown in FIG. 1 , the embodiment of the present invention further provides a customer churn risk prediction device for predicting customer churn risk. Figure 1 The specific implementation of the method shown in is shown in the structural diagram Figure 4 As shown, including:

[0159] The first determination unit 401 is used to determine the churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes the behavior data, user feature data and emotional feature data corresponding to the historical customer;

[0160] A data judgment unit 402 is used to judge whether the multi-source data set corresponding to each of the historical customers meets a preset data completeness condition, and if the multi-source data set does not meet the data completeness condition, the historical customer is regarded as the first customer; if the multi-source data set meets the data completeness condition, the historical customer is regarded as the second customer;

[0161] The data prediction unit 403 is used to perform missing modal data prediction processing on the multi-source data set corresponding to each of the first customers to obtain each modal prediction data corresponding to the first customer, and combine each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer;

[0162] A second determining unit 404 is configured to use the multi-source data set corresponding to each of the second customers as the original data set of the second customer;

[0163] A data cleaning unit 405 is used to clean the original data set of each of the historical customers to obtain a feature data set of each of the historical customers;

[0164] A feature extraction unit 406 is used to extract features for each of the historical customers based on the feature data set of the historical customer, obtain multiple feature data corresponding to the historical customer, and use the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer as a training sample corresponding to the historical customer;

[0165] A model training unit 407 is used to perform model training based on the training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model;

[0166] The risk prediction unit 408 is used to apply the churn risk prediction model to perform churn risk prediction on target customers.

[0167] By using the device provided by the embodiment of the present invention, corresponding training samples can be constructed based on multi-source data of historical customers with clear churn status, and a corresponding machine learning model can be obtained through training of the training samples as a churn risk prediction model. The churn risk prediction model can learn and extract useful information from the relevant data of various types of real customers. When it is necessary to predict the churn risk of a target customer, the churn risk prediction model can be used to analyze the relevant data of the target customer, thereby predicting the churn risk of the target customer. For different types of customers, the churn risk prediction model can perform personalized analysis of the customer based on the knowledge it has learned, which is conducive to improving the prediction accuracy of customer churn risk.

[0168] exist Figure 4 Based on the device shown, the device provided by the embodiment of the present invention can be further expanded into multiple units. The function of each unit can be found in the description of each embodiment provided in the previous text for the customer churn risk prediction method, and no further examples will be given here.

[0169] An embodiment of the present invention further provides a storage medium, which includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the customer churn risk prediction method as described above.

[0170] The embodiment of the present invention further provides an electronic device, the structural diagram of which is shown in FIG. Figure 5 As shown, it specifically includes a memory 501 and one or more instructions 502, wherein the one or more instructions 502 are stored in the memory 501 and are configured to be executed by one or more processors 503 to perform the following operations:

[0171] Determine churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes behavior data, user feature data, and emotional feature data corresponding to the historical customer;

[0172] For each multi-source data set corresponding to the historical customer, determine whether the multi-source data set meets a preset data completeness condition; if the multi-source data set does not meet the data completeness condition, treat the historical customer as the first customer; if the multi-source data set meets the data completeness condition, treat the historical customer as the second customer;

[0173] For each of the first customers, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer, and combining each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer;

[0174] taking the multi-source data set corresponding to each of the second customers as the original data set of the second customers;

[0175] Performing data cleaning on the original data set of each of the historical customers to obtain a feature data set of each of the historical customers;

[0176] For each of the historical customers, feature extraction is performed based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as training samples corresponding to the historical customer;

[0177] Performing model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model;

[0178] The churn risk prediction model is applied to predict the churn risk of target customers.

[0179] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0180] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0181] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting customer churn risk, characterized in that: include: Determine the churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes the behavior data, user feature data, and emotional feature data corresponding to the historical customer; For each multi-source data set corresponding to the historical customer, determine whether the multi-source data set meets a preset data completeness condition; if the multi-source data set does not meet the data completeness condition, treat the historical customer as the first customer; if the multi-source data set meets the data completeness condition, treat the historical customer as the second customer; For each of the first customers, performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer, and combining each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer; taking the multi-source data set corresponding to each of the second customers as the original data set of the second customers; Cleaning the original data set of each of the historical customers to obtain a feature data set of each of the historical customers; For each of the historical customers, feature extraction is performed based on the feature data set of the historical customer to obtain multiple feature data corresponding to the historical customer, and the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer are used as training samples corresponding to the historical customer; Performing model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model; The churn risk prediction model is applied to predict the churn risk of target customers.

2. The customer churn risk prediction method according to claim 1, characterized in that: The performing missing modal data prediction processing on the multi-source data set corresponding to the first customer to obtain each modal prediction data corresponding to the first customer includes: Determine a user profile corresponding to the first customer and a user profile corresponding to each of the second customers; Based on the user portraits corresponding to the respective second customers, the respective second customers are divided into groups to obtain respective customer groups; each of the customer groups includes at least one of the second customers; Based on the user portrait corresponding to the first customer, determining a target customer group corresponding to the first customer from among the customer groups; Determine each missing modal data category corresponding to the first customer; Based on the multi-source data sets corresponding to each second customer in the target customer group and the multi-source data sets corresponding to the first customer, a preset multimodal variational autoencoder is applied to perform data prediction on each of the missing modal data categories, and synthetic data corresponding to each of the missing modal data categories generated by the multimodal variational autoencoder is obtained, and the synthetic data corresponding to each of the missing modal data categories is used as the modal prediction data corresponding to the first customer.

3. The customer churn risk prediction method according to claim 1, characterized in that: The step of performing data cleaning on the original data set of each of the historical customers to obtain the characteristic data set of each of the historical customers includes: For each original data set of the historical customer, performing abnormal data processing on the original data set, and using the original data set after the abnormal data processing as the first data set of the historical customer; For each of the first data sets of the historical customers, performing missing value processing on the first data set, and using the first data set after the missing value processing as the second data set of the historical customer; For each of the second data sets of the historical customers, performing redundancy cleaning on the second data sets, and using the second data sets after the redundancy cleaning as the third data set of the historical customers; For each of the third data sets of the historical customers, data standardization is performed on the third data set, and the third data set after data standardization is used as the characteristic data set of the historical customer.

4. The customer churn risk prediction method according to claim 1, characterized in that: The feature extraction is performed based on the feature data set of the historical customer to obtain a plurality of feature data corresponding to the historical customer, including: Determining a first characteristic indicator based on the characteristic data set of the historical customer; the first characteristic indicator represents the business transaction frequency of the historical customer; Determining a second characteristic indicator based on the characteristic data set of the historical customer; the second characteristic indicator represents the business transaction amount of the historical customer; Determining a third characteristic indicator based on the characteristic data set of the historical customer; the third characteristic indicator represents the activity level of the historical customer; Based on the feature data set of the historical customer, determining a fourth feature indicator; the fourth feature indicator represents the historical churn of the historical customer; The first characteristic indicator, the second characteristic indicator, the third characteristic indicator and the fourth characteristic indicator are used as respective characteristic data corresponding to the historical customer.

5. The customer churn risk prediction method according to claim 1, characterized in that: The model training is performed based on the training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model, including: Constructing an initial model; the initial model is a machine learning model constructed based on the extreme gradient boosting tree; The initial model is trained based on the training samples corresponding to each of the historical customers, and during the model training process, the expected maximum profit is used as the optimization target, and the Bayesian optimization algorithm is applied to optimize the hyperparameters of the initial model; The initial model that completes the model training process is used as the churn risk prediction model.

6. The customer churn risk prediction method according to claim 1, characterized in that: The applying the churn risk prediction model to predict the churn risk of target customers includes: Acquire a multi-source data set corresponding to the target customer; Determining a feature data set corresponding to the target customer based on the multi-source data set corresponding to the target customer; Based on the feature data set corresponding to the target customer, determining each feature data corresponding to the target customer; Inputting each characteristic data corresponding to the target customer into the churn risk prediction model, using the churn risk prediction model to predict the churn risk of the target customer, and obtaining the churn probability output by the churn risk prediction model; Performing score conversion on the churn probability to obtain a churn risk score corresponding to the churn probability; Based on the churn risk score, a churn warning report of the target customer is generated.

7. The customer churn risk prediction method according to claim 6, characterized in that: The generating of the target customer churn warning report based on the churn risk score includes: Determining a target risk level corresponding to the churn risk score; Determine the target customer maintenance strategy corresponding to the target risk level among various pre-set customer maintenance strategies; Determine characteristic change trend information corresponding to the target customers; The target customer loss warning report is composed of the loss risk score, the target risk level, the target customer maintenance strategy and the characteristic change trend information.

8. A customer churn risk prediction device, characterized in that: include: A first determination unit is used to determine the churn status labels corresponding to multiple historical customers, and obtain a multi-source data set corresponding to each of the historical customers; the multi-source data set corresponding to each of the historical customers includes the behavior data, user feature data and emotional feature data corresponding to the historical customer; a data judgment unit, configured to judge, for each multi-source data set corresponding to the historical customer, whether the multi-source data set meets a preset data completeness condition, and if the multi-source data set does not meet the data completeness condition, regard the historical customer as the first customer; if the multi-source data set meets the data completeness condition, regard the historical customer as the second customer; a data prediction unit, configured to perform missing modal data prediction processing on the multi-source data set corresponding to each of the first customers, obtain each modal prediction data corresponding to the first customer, and combine each of the modal prediction data and the multi-source data set corresponding to the first customer to form an original data set of the first customer; A second determining unit, configured to use the multi-source data set corresponding to each of the second customers as the original data set of the second customer; A data cleaning unit, used to clean the original data set of each of the historical customers to obtain a feature data set of each of the historical customers; A feature extraction unit is used to extract features for each of the historical customers based on the feature data set of the historical customer, obtain multiple feature data corresponding to the historical customer, and use the churn status label corresponding to the historical customer and each feature data corresponding to the historical customer as a training sample corresponding to the historical customer; A model training unit, used to perform model training based on training samples corresponding to each of the historical customers and a preset machine learning algorithm to obtain a churn risk prediction model; The risk prediction unit is used to apply the churn risk prediction model to perform churn risk prediction on target customers.

9. A storage medium, characterized in that: The storage medium includes stored instructions, wherein when the instructions are executed, the device where the storage medium is located is controlled to execute the customer churn risk prediction method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: It comprises a memory and one or more instructions, wherein the one or more instructions are stored in the memory and are configured to be executed by one or more processors to perform the customer churn risk prediction method as described in any one of claims 1 to 7.

Images