High-quality customer churn prediction method, device and storage medium based on supervised learning

Abstract

The present invention discloses a method, device and storage medium for predicting the loss of high-quality customers based on supervised learning, wherein the method includes: obtaining high-quality customers to be predicted to be lost, and extracting the loss feature data of the high-quality customers to be predicted to be lost within a preset time period; inputting the loss feature data of the high-quality customers to be predicted to be lost into a pre-trained high-quality customer loss prediction model, and outputting the high-quality customer loss prediction result; wherein the high-quality customer loss prediction model is obtained by training the prediction model constructed based on the XGBoost algorithm using the constructed high-quality customer loss prediction sample data set. Selecting high-quality customer groups as the research object reduces the category imbalance of training data, improves the training speed and accuracy of the model, and can also grasp the core customers of the bank; using the XGBoost algorithm, the algorithm is insensitive to missing values ​​and does not need to interpolate the missing values; and has strong learning ability, fast training speed, and fast convergence speed.

Description

Technical Field

[0001] The present invention relates to the technical field of commercial bank customer churn prediction, and in particular to a method, device and storage medium for predicting high-quality customer churn based on supervised learning. Background Art

[0002] The current market competition is becoming increasingly fierce, and the differences in products or services among banks are becoming smaller and smaller. More and more banks are shifting from "product-centric" to "customer-centric" and are adopting information systems such as customer relationship management (CRM) to improve the system service level. Customer churn is the focus of CRM and one of the core issues of concern to the banking industry.

[0003] Relevant research has found that the cost of winning a new customer is 5 to 6 times that of retaining an old customer. Therefore, retaining old customers, predicting potential lost customers in advance, and preventing business crises caused by customer loss are of strategic significance for improving the competitiveness of enterprises. Research shows that in China, 10% of high-quality individual customers contribute at least 90% of profits. Therefore, compared with the total number of customers, the work of maintaining and retaining such customers and preventing such customers from losing is more meaningful and valuable.

[0004] The current customer churn prediction research is divided into three main steps, namely: 1. Define churn, 2. Select features, and 3. Establish a churn prediction model. In common research, churn is often defined for individual products, such as deposits, financial management, or credit cards, etc. The definition of churn is generally that existing customers stop purchasing the bank's products or services and instead purchase the products or services of its competitors, which is actually manifested as card cancellation and account cancellation. When selecting features, it is generally believed that there is a correlation between customers at risk of churn and the number of products held by customers, their age and gender; customers with different products have different churn risks; the interval time between the last transaction reflects the customer's activity level, which is also related to churn behavior; in addition, the discounts received by customers, the channels used, the length of service time, and customer complaints are also variables that help predict churn. When building a model, machine learning-based methods such as SVM, decision tree, Boosting, etc., and statistical analysis-based methods such as Cox proportional risk model, logistic regression, discriminant analysis, etc. are usually selected.

[0005] Each of the above steps determines the effectiveness of the churn prediction model in practical applications: using card and account cancellation actions to define churn is often not convincing. For example, some customers have lost contact with the bank, but they may not cancel their accounts. Therefore, under this definition, the scope of the target group is often reduced; choosing a single product to judge churn is often inaccurate. For example, although a customer has reduced or cancelled a time deposit, he or she has purchased a corresponding or higher amount of investment products. In this case, using the cancellation of a single product to judge churn will result in a definition of churn that is inconsistent with reality; at the same time, using features that do not meet the basic assumptions of the model as input will lead to incorrect conclusions, such as the Cox proportional hazard model requires that the variables used are all independent of time; finally, the learning ability of the model also determines the final prediction effect. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, an object of the present invention is to provide a method, device and storage medium for predicting high-quality customer churn based on supervised learning, so as to improve the accuracy of high-quality customer churn prediction.

[0007] First, a high-quality customer churn prediction method based on supervised learning is provided, including:

[0008] Obtain high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period;

[0009] The churn feature data of high-quality customers to be predicted are input into a pre-trained high-quality customer churn prediction model, and a high-quality customer churn prediction result is output; wherein the high-quality customer churn prediction model is trained on a constructed high-quality customer churn prediction data set using an XGBoost algorithm.

[0010] Furthermore, the high-quality customer churn prediction sample data set is constructed by the following method:

[0011] Obtain the customer's assets in the bank and compare them with the preset threshold to screen out multiple high-quality customers;

[0012] An observation period is selected to determine the churn labels and observation time of multiple high-quality customers; for high-quality customers that churn during the observation period, the observation time is the final churn time of the high-quality customers, and for high-quality customers that do not churn during the observation period, the observation time is the end time of the observation period;

[0013] For multiple high-quality customers, based on their corresponding observation time, we extract churn feature data from their natural attributes, product status information, in-bank asset information, and transaction behavior to obtain a sample data set for high-quality customer churn prediction;

[0014] The step of obtaining high-quality customers to be predicted to be lost and extracting churn feature data of the high-quality customers to be predicted to be lost within a preset time period includes:

[0015] Filter out high-quality customers to be predicted to churn at time t;

[0016] The loss characteristic data within a preset time period before time t is extracted from its natural attributes, product status information, in-bank asset information and transaction behavior.

[0017] Furthermore, the selected observation period determines the churn labels and observation time of multiple high-quality customers, including:

[0018] Select an observation period [t0,t n ];

[0019] Determine the churn label and churn time based on the three types of products of each high-quality customer: fixed deposit, current deposit and wealth management;

[0020] Verify the accuracy of the churn label of each high-quality customer to ensure that its churn label is consistent with the churn in the actual scenario;

[0021] The churn time of each lost high-quality customer is adjusted forward by a preset time as the final churn time; for high-quality customers lost during the observation period, the final churn time is used as the corresponding observation time; for high-quality customers who have not lost during the observation period, the end time of the observation period t n as its corresponding observation time.

[0022] Furthermore, the determination of the churn label and churn time based on the three types of products, namely, fixed deposit, current deposit and financial management, of each high-quality customer includes:

[0023] Obtain information about each high-quality customer's fixed deposit, current deposit, and financial products, and determine their churn labels and churn time based on the table below;

[0024] current regular Financial Management Churn Tags Lost Date * 1 1 1 (fx_dt, inv_dt) * ^1 1 1 inv_dt * 1 ^1 1 fx_dt 1 -1 -1 1 sv_dt 0 -1 -1 -1 * ^1 0 0 * 0 -1 0 -1 -1 -1 NULL

[0025] Among them, the three types of products, namely, current, fixed deposit and wealth management, have three status marks during the observation period: account closure, account opening and no account establishment, which are recorded as 1, 0 and -1 respectively. "*" represents any category among the three status marks, and "^" represents "non". When the churn label is 1, it means that the high-quality customer has been lost during the observation period. When the churn label is 0, it means that the high-quality customer has not been lost during the observation period. NULL means that there is no such high-quality customer. When the churn label is -1, it means that a churn judgment needs to be made for this type of high-quality customer at this time. Specifically, the continuous dates {t1, t2, …, t m Current account balance are all smaller than the corresponding [t0,ti ], the average daily demand balance in the interval i∈{1,2,…,m} Right now The churn label of this high-quality customer is 1, and its corresponding churn time is t1, otherwise the churn label of this high-quality customer is 0; sv_dt represents the latest account cancellation time of the current product during the observation period, fx_dt represents the latest account cancellation time of the fixed-term product during the observation period, and inv_dt represents the latest account cancellation time of the wealth management product during the observation period.

[0026] Furthermore, verifying the accuracy of the churn label of each high-quality customer includes:

[0027] When the status mark of one of the three types of products of a high-quality customer is 1 during the observation period, observe whether the customer re-opens an account within a preset period of time after the observation period, and finally determine the corresponding status mark and churn label;

[0028] When a high-quality customer's fixed-term product account is cancelled but the wealth management product account is not, or when the wealth management product account is cancelled but the fixed-term product account is not, a circulation verification is performed, i.e., it is verified whether the high-quality customer purchased the wealth management product when the fixed-term product account was cancelled or whether the high-quality customer purchased the fixed-term product when the wealth management product account was cancelled, and then the churn label is finally determined; when both the fixed-term product and the wealth management product account are cancelled, they are decomposed into two samples, which correspond to the churn of the fixed-term product and the churn of the wealth management product respectively, and have the churn time when the corresponding products are cancelled, and then they are subjected to circulation verification in the above-mentioned circulation verification method.

[0029] Furthermore, the extraction of loss feature data from natural attributes, product status information, in-bank asset information and transaction behavior based on the corresponding observation time includes:

[0030] For natural attributes, the extracted churn feature data include gender, age, and household age of high-quality customers;

[0031] For the product status information, the extracted churn feature data includes the new additions, expiration and holding data of fixed-term products and wealth management products within the preset period before the observation time corresponding to high-quality customers;

[0032] For in-bank asset information, the extracted churn feature data include the average daily balances of current products and fixed-term products corresponding to high-quality customers in the preset period before the observation time;

[0033] For transaction behavior, the extracted churn feature data includes the income and expenditure data of high-quality customers in the preset period before the observation time;

[0034] Perform logarithmic processing on the amount features in the churn feature data of each high-quality customer.

[0035] Furthermore, the XGBoost algorithm is used to train the constructed high-quality customer churn prediction data set to obtain a high-quality customer churn prediction model, which specifically includes:

[0036] Assign corresponding weights to each sample in the high-quality customer churn prediction sample dataset based on its churn label and the assets held in the bank;

[0037] Select the XGBoost algorithm to build a prediction model;

[0038] Let y be the churn label of the high-quality customer churn prediction sample data set, X be the feature matrix of the high-quality customer churn prediction sample data set, θ be the hyperparameter of the XGBoost algorithm, and combine the grid search method to perform k-fold cross validation using the high-quality customer churn prediction sample data set to obtain the experimental optimal hyperparameters of the prediction model. Among them, w is the weight corresponding to the sample in the high-quality customer churn prediction sample data set, xgb(·) is the prediction model built based on the XGBoost algorithm, represents the predicted label corresponding to the sample data of xgb(·) under the hyperparameter function Represents the loss label y and the predicted label The matching degree, Represents the parameter θ corresponding to the maximum objective function;

[0039] Divide the high-quality customer churn prediction sample data set into training set and test set, and use the optimal parameters According to the early stop strategy, the optimal number of iterations m is determined in the model training to obtain a high-quality customer churn prediction model.

[0040] Furthermore, the method of assigning corresponding weights according to the churn label of each sample in the high-quality customer churn prediction sample data set and the situation of assets held in the bank includes:

[0041] The weight of each sample is determined by the following formula:

[0042]

[0043] Where a=(a1,a2,…,a n ) represents the asset of the sample in the high-quality customer churn prediction sample dataset, a i represents the asset of the i-th sample, y i represents the loss label of the i-th sample, α>0.

[0044] In the second aspect, a high-quality customer churn prediction device based on supervised learning is provided, comprising:

[0045] Feature extraction module: used to obtain high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period;

[0046] Churn prediction module: used to input the churn feature data of high-quality customers to be predicted into a pre-trained high-quality customer churn prediction model, and output high-quality customer churn prediction results; wherein, the high-quality customer churn prediction model is trained on the constructed high-quality customer churn prediction data set using the XGBoost algorithm.

[0047] In a third aspect, a computer-readable storage medium is provided, which includes a stored computer program, wherein the computer program is suitable for being loaded by a processor and executing the high-quality customer churn prediction method based on supervised learning as described above.

[0048] Beneficial Effects

[0049] The present invention proposes a method, device and storage medium for predicting high-quality customer churn based on supervised learning, which has the following advantages:

[0050] 1) The present invention selects high-quality customer groups as the research object. Compared with setting all customers as the target group, the customer research of high-quality customers not only reduces the category imbalance of training data, improves the training speed and accuracy of the model, but also can grasp the "core" customers of the bank, making the model more practical.

[0051] 2) The present invention starts from the perspective of multiple products. Compared with studying the churn of customers under a single product, considering multiple products at the same time not only reduces the modeling cost required to achieve the same prediction effect, achieving twice the result with half the effort, but also reduces the "false churn" generated from the perspective of a single product, that is, the transfer of customer assets between different products.

[0052] 3) The present invention has a flexible churn judgment method. Compared with using account cancellation to rigidly judge customer churn, it is more practical to identify whether a customer has been cancelled by combining flexible indicators. And by combining the churn status of customers on multiple products to obtain the final churn label, the accuracy of the churn label is enhanced.

[0053] 4) The present invention adopts a flexible feature extraction interval and a time-controllable loss warning method. Since the observation time of each sample is corresponding to each other, the feature extraction of the training sample is allowed to be completed in an independent time range without losing feature information due to the unified time range. In addition, by appropriately advancing the observation time of the loss sample, the warning time of the loss can be flexibly controlled.

[0054] 5) The present invention uses the XGBoost algorithm, which is insensitive to missing values ​​and does not require interpolation of missing values. Moreover, the XGBoost algorithm constrains the complexity of the model through an explicit regularization method, which can effectively avoid overfitting. It also has the characteristics of strong learning ability, fast training speed, and fast convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 This is an overall flow chart of a method for predicting high-quality customer churn based on supervised learning provided by an embodiment of the present invention;

[0057] Figure 2 is a flow chart of a method for constructing a high-quality customer churn prediction sample data set provided by an embodiment of the present invention;

[0058] Figure 3 is a flow chart of a method for constructing a high-quality customer churn prediction model provided by an embodiment of the present invention;

[0059] Figure 4 It is a flow chart of performing churn prediction on high-quality customers to be churned provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0060] To make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be described in detail below. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other implementation methods obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.

[0061] like Figure 1 As shown, the embodiment of the present invention provides a method for predicting high-quality customer churn based on supervised learning. Before prediction, the XGBoost algorithm is used to train the constructed high-quality customer churn prediction data set to obtain a high-quality customer churn prediction model, such as Figure 2 and Figure 3 As shown, including:

[0062] A1: Construct a sample data set for predicting high-quality customer churn. The specific steps are as follows:

[0063] A1.1: According to the customer's assets in the bank, compare with the pre-set threshold to screen out high-quality customers. Compared with all customers, studying high-quality customers with high asset value is more conducive to improving the bank's operating efficiency, while simplifying the calculation amount of establishing a churn prediction model. Therefore, when screening high-quality customers, the appropriate number of high-quality customers can be screened out according to the distribution of bank customer assets and marketing strategies, generally controlled between 3% and 10% of all customers;

[0064] A1.2: Select an observation period and determine the churn labels and observation time of multiple high-quality customers. For high-quality customers that churn during the observation period, the observation time is the final churn time of the high-quality customers. For high-quality customers that do not churn during the observation period, the observation time is the end time of the observation period. Specifically, it includes:

[0065] A1.2.1: Select an observation period [t0,t n ], used to observe the loss of high-quality customers during the observation period, t0 is the observation start time, t n is the cut-off time for observation;

[0066] A1.2.2: Determine the churn label and churn time based on the three types of products of each high-quality customer: fixed deposit, current deposit and wealth management; including:

[0067] Obtain information about each high-quality customer's fixed deposit, current deposit, and financial products, and determine their churn labels and churn time based on the table below;

[0068] current regular Financial Management Loss category Churn Tags Number of combinations Lost Date * 1 1 1 1 3 (fx_dt,inv_dt) * ^1 1 3 1 6 inv_dt * 1 ^1 5 1 6 fx_dt 1 -1 -1 7 1 1 sv_dt 0 -1 -1 -1 -1 1 * ^1 0 2 0 6 * 0 -1 4 0 3 -1 -1 -1 NULL NULL 1

[0069] Among them, the three types of products, current, fixed and financial products, have three status marks during the observation period: account closure, account opening and no account establishment, and they are recorded as 1, 0 and -1 respectively. "*" represents any category of the three status marks, and "^" represents "non"; when the churn label is 1, it means that the high-quality customer has churned during the observation period, when the churn label is 0, it means that the high-quality customer has not churned during the observation period, and NULL means that there is no such high-quality customer. The churn category is judged based on the identifiers of the three accounts. Odd numbers represent churn, even numbers represent no churn, and -1 means that such high-quality customers only have current accounts and have not been closed. According to the actual situation, the churn label can be set to 0 or a churn judgment can be made for such customers separately, such as the consecutive dates {t1, t2,…, t m Current account balance are all smaller than the corresponding [t0,t i ], the average daily demand balance in the interval i∈{1,2,…,m} Right now The churn label of this high-quality customer is 1, and its corresponding churn time is t1. Otherwise, the churn label of this high-quality customer is 0. The churn category is an intermediate result, which helps to understand the number of samples in each churn category and guide the subsequent error analysis. The number of combinations indicates the number of combinations of identifiers that meet the corresponding rules. For example, the number of combinations of churn category 1 is 3, which includes 3 combinations of identifiers, namely (0, 1, 1), (1, 1, 1), and (-1, 1, 1). The total number of combinations is 3. 3 sv_dt indicates the latest account cancellation time of current products during the observation period, fx_dt indicates the latest account cancellation time of fixed-term products during the observation period, and inv_dt indicates the latest account cancellation time of financial products during the observation period. For customers who have not lost during the observation period, there is no loss time.

[0070] A1.2.3: Verify the accuracy of the churn label of each high-quality customer to ensure that its churn label is consistent with the churn in the actual scenario; including:

[0071] When the status mark of one of the three types of products of a high-quality customer is 1 during the observation period, observe whether the customer reopens the account within the preset time period after the observation period, and finally determine the corresponding status mark and churn label. This verification is particularly important for samples whose churn time is close to the end of the observation period, and can increase the accuracy of the churn label.

[0072] When a high-quality customer's fixed-term product account is cancelled but the wealth management product account is not, or when the wealth management product account is cancelled but the fixed-term product account is not, a circulation verification is performed, that is, verifying whether the high-quality customer purchased the wealth management product when the fixed-term product account was cancelled or whether the customer purchased the fixed-term product when the wealth management product account was cancelled, and then finally determining the churn label; when both the fixed-term product and the wealth management product of a high-quality customer are cancelled, it is decomposed into two samples, the two samples correspond to the churn of the fixed-term product and the churn of the wealth management product, and have the churn time when the corresponding products are cancelled, and then the circulation verification is performed in the above-mentioned circulation verification method. Through circulation verification, samples of "false churn" can be effectively avoided.

[0073] A1.2.4: The churn time of each lost high-quality customer is adjusted forward by a preset time as the final churn time to achieve the effect of advance warning; for high-quality customers lost during the observation period, the final churn time is used as the corresponding observation time. If the actual churn time is observed to be t i , then the loss time t i Appropriate advance, ready to use As the final churn time, it aims to give the churn prediction result before the high-quality customers churn. That is, the trained high-quality customer churn prediction model predicts the probability of the high-quality customers to be churned at time t+T, where t is the current time, so as to achieve the effect of churn warning. For the high-quality customers who have not churned during the observation period, the end time of the observation period tn as its corresponding observation time.

[0074] A1.3: For multiple high-quality customers, based on their corresponding observation time, churn feature data are extracted from their natural attributes, product status information, in-bank asset information, and transaction behavior to obtain a high-quality customer churn prediction sample data set. Each sample in the high-quality customer churn prediction sample data set contains its corresponding churn feature data and churn label; wherein, the extraction of churn feature data specifically includes:

[0075] For natural attributes, the extracted churn feature data include gender, age, and household age of high-quality customers;

[0076] For the product status information, the extracted churn feature data includes the new additions, expiration and holding data of fixed-term products and wealth management products within the preset period before the observation time corresponding to high-quality customers;

[0077] For in-bank asset information, the extracted churn feature data include the average daily balances of current products and fixed-term products corresponding to high-quality customers in the preset period before the observation time;

[0078] For transaction behavior, the extracted churn feature data includes the income and expenditure data of high-quality customers in the preset period before the observation time;

[0079] The above-mentioned scheme for extracting loss feature data can be specifically referred to in the following table:

[0080]

[0081] In the above table, according to the different loss labels, the observation time is the final loss time Or the observation period end time t n . In addition, in addition to the features of natural attributes, the calculation of other features is closely related to the observation time obtained in step A1.2. The calculation of features is often based on this time, and statistics are performed on a certain time dimension. At the same time, this time may vary from sample to sample. Therefore, each sample completes the feature calculation on a certain time dimension of "itself", which is the core of the model's ability to accurately predict lost customers. The certain time dimension refers to the length of time for extracting the lost features, such as the past 12 months, the past 6 months, the past 1 month, 7 days, etc.

[0082] After the churn feature extraction is completed, the amount feature in the churn feature data of each high-quality customer is logarithmized. i The amount-based features are (F1, F2, ..., F n ), according to the conversion formula The new feature of the amount class is Reducing the dimension of features helps in searching for the optimal segmentation point when using a "tree" algorithm to improve accuracy.

[0083] A1.4: Divide the high-quality customer churn prediction sample data set into a training set D with an appropriate proportion Train and the test set D Test , which are used for model training and evaluation respectively. High-quality customer churn prediction sample data set D = D Train ∪D Test According to the actual situation of the data, the training set sample size N is usually divided Train and the sample size N of the test set Test The ratio is

[0084] A2: Use the constructed high-quality customer churn prediction sample data set to train the prediction model built based on the XGBoost algorithm to obtain a high-quality customer churn prediction model, including:

[0085] A2.1: Assigning corresponding weights based on the churn labels and assets held in the bank for each sample in the high-quality customer churn prediction sample dataset can help alleviate the impact of category imbalance and allow the model to pay more attention to the churn of high-asset customers, thereby improving the accuracy and prediction value of the model. Specifically,

[0086] The weight of each sample is determined by the following formula:

[0087]

[0088] Where a=(a1,a2,…,a n ) represents the asset of the sample in the high-quality customer churn prediction sample dataset, a i represents the asset of the i-th sample, y i represents the loss label of the i-th sample, α>0, α∈{1,2,…}, and α is appropriately adjusted according to the evaluation indicators of the model.

[0089] A2.2: Select the XGBoost algorithm to build a prediction model;

[0090] A2.3: Let y be the churn label of the high-quality customer churn prediction sample data set, X be the feature matrix of the high-quality customer churn prediction sample data set, and θ be the hyperparameter of the XGBoost algorithm. Combined with the grid search method, use the high-quality customer churn prediction sample data set for k-fold cross validation to obtain the experimental optimal hyperparameters of the prediction model. Among them, w is the weight corresponding to the sample in the high-quality customer churn prediction sample data set, xgb(·) is the prediction model built based on the XGBoost algorithm, represents the predicted label corresponding to the sample data of xgb(·) under the hyperparameter θ function Represents the loss label y and the predicted label The matching degree of , considering the sample weight w, that is is located in the calculation, that is where w i represents the weight of the i-th sample, y i represents the loss label of the i-th sample, represents the predicted label of the i-th sample, Represents the parameter θ corresponding to the maximum objective function;

[0091] A2.4: Use optimal parameters Combining the training set and the test set, according to the early stop strategy, the optimal number of iterations m is determined in the model training to obtain a high-quality customer churn prediction model.

[0092] After obtaining the trained high-quality customer churn prediction model, the high-quality customer churn prediction based on supervised learning can be performed, such as Figure 4 As shown, specifically including:

[0093] S01: Acquire high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period;

[0094] S02: Input the churn feature data of the high-quality customers to be predicted into a pre-trained high-quality customer churn prediction model, and output the high-quality customer churn prediction results.

[0095] More specifically, they include:

[0096] Screen out high-quality customers who are predicted to be lost at time t; time t may be the current time, and the method for screening high-quality customers who are predicted to be lost can refer to the method in step A1.1;

[0097] Extract the churn feature data within the preset time period before the current time t from its natural attributes, product status information, in-bank asset information and transaction behavior respectively; the method for extracting churn feature data can refer to step A1.3, the difference is that the observation time at this time is the current time t; the preset time period here can be set to 3 months, 6 months or 12 months, etc. according to needs.

[0098] The extracted churn feature data is input into the trained high-quality customer churn prediction model, and the probability of the predicted high-quality customer to be churned at time t+T can be obtained through calculation.

[0099] The embodiment of the present invention also provides a high-quality customer churn prediction device based on supervised learning, comprising:

[0100] Feature extraction module: used to obtain high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period;

[0101] Churn prediction module: used to input the churn feature data of high-quality customers to be predicted into a pre-trained high-quality customer churn prediction model, and output high-quality customer churn prediction results; wherein, the high-quality customer churn prediction model is trained on the constructed high-quality customer churn prediction data set using the XGBoost algorithm.

[0102] The specific implementation scheme in this embodiment refers to the high-quality customer churn prediction method based on supervised learning provided in the above embodiment, which will not be repeated here.

[0103] An embodiment of the present invention further provides a computer-readable storage medium, which includes a stored computer program, wherein the computer program is suitable for being loaded by a processor and executing the high-quality customer churn prediction method based on supervised learning as described above.

[0104] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0105] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0106] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0107] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0108] It can be understood that the same or similar parts of the above embodiments can be referenced to each other, and the contents not described in detail in some embodiments can refer to the same or similar contents in other embodiments.

[0109] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code that includes one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present invention includes alternative implementations in which functions may not be performed in the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present invention belong.

[0110] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A method for predicting high-quality customer churn based on supervised learning, characterized in that: include: Obtain high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period; Input the churn feature data of the high-quality customers to be predicted into a pre-trained high-quality customer churn prediction model, and output the high-quality customer churn prediction result; wherein the high-quality customer churn prediction model is trained on the constructed high-quality customer churn prediction data set using the XGBoost algorithm; The high-quality customer churn prediction sample dataset is constructed by the following method: Obtain the customer's assets in the bank and compare them with the preset threshold to screen out multiple high-quality customers; An observation period is selected to determine the churn labels and observation time of multiple high-quality customers; for high-quality customers that churn during the observation period, the observation time is the final churn time of the high-quality customers, and for high-quality customers that do not churn during the observation period, the observation time is the end time of the observation period; For multiple high-quality customers, based on their corresponding observation time, we extract churn feature data from their natural attributes, product status information, in-bank asset information, and transaction behavior to obtain a sample data set for high-quality customer churn prediction; The selected observation period determines the churn labels and observation time of multiple high-quality customers, including: Select an observation period [t0,t n ]; Determine the churn label and churn time based on the three types of products of each high-quality customer: fixed deposit, current deposit and wealth management; Verify the accuracy of the churn label of each high-quality customer to ensure that its churn label is consistent with the churn in the actual scenario; The churn time of each lost high-quality customer is adjusted forward by a preset time as the final churn time; for high-quality customers lost during the observation period, the final churn time is used as the corresponding observation time; for high-quality customers who have not lost during the observation period, the end time of the observation period t n As its corresponding observation time; The XGBoost algorithm is used to train the high-quality customer churn prediction model on the constructed high-quality customer churn prediction dataset, including: Assign corresponding weights to each sample in the high-quality customer churn prediction sample dataset based on its churn label and the assets held in the bank; Select the XGBoost algorithm to build a prediction model; Let y be the churn label of the high-quality customer churn prediction sample data set, X be the feature matrix of the high-quality customer churn prediction sample data set, θ be the hyperparameter of the XGBoost algorithm, and combine the grid search method to perform k-fold cross validation using the high-quality customer churn prediction sample data set to obtain the experimental optimal hyperparameters of the prediction model. Among them, w is the weight corresponding to the sample in the high-quality customer churn prediction sample data set, xgb(·) is the prediction model built based on the XGBoost algorithm, represents the predicted label corresponding to the sample data of xgb(·) under the hyperparameter θ function Represents the loss label y and the predicted label The matching degree, Represents the parameter θ corresponding to the maximum objective function; Divide the high-quality customer churn prediction sample data set into training set and test set, and use the optimal parameters According to the earlystop strategy, the optimal number of iterations m is determined in the model training to obtain a high-quality customer churn prediction model; The method of assigning corresponding weights to the churn labels of each sample in the high-quality customer churn prediction sample data set and the situation of the assets held in the bank includes: The weight of each sample is determined by the following formula: Where a=(a1,a2,…,a n ) represents the asset of the sample in the high-quality customer churn prediction sample dataset, a i represents the asset of the i-th sample, y i represents the loss label of the i-th sample, α>0.

2. The method for predicting high-quality customer churn based on supervised learning according to claim 1, characterized in that: The step of obtaining high-quality customers to be predicted to be lost and extracting churn feature data of the high-quality customers to be predicted to be lost within a preset time period includes: Filter out high-quality customers to be predicted to churn at time t; The loss characteristic data within a preset time period before time t is extracted from its natural attributes, product status information, in-bank asset information and transaction behavior.

3. The method for predicting high-quality customer churn based on supervised learning according to claim 2, characterized in that: The churn labels and churn time are determined based on the three types of products, namely, fixed deposit, current deposit and financial management products of each high-quality customer, including: Obtain information about each high-quality customer's fixed deposit, current deposit, and financial products, and determine their churn labels and churn time based on the table below; Among them, the three types of products, namely, current, fixed deposit and wealth management, have three status marks during the observation period: account closure, account opening and no account establishment, which are recorded as 1, 0 and -1 respectively. "*" represents any category among the three status marks, and "^" represents "not"; when the churn label is 1, it means that the high-quality customer is lost during the observation period, when the churn label is 0, it means that the high-quality customer is not lost during the observation period, NULL means that there is no such high-quality customer, and when the churn label is -1, it means that a churn judgment needs to be made for this type of high-quality customer at this time, specifically, the continuous dates {t1, t2, …, t m Current account balance are all smaller than the corresponding [t0,t i ], the average daily demand balance in the interval i∈{1,2,…,m} Right now The churn label of this high-quality customer is 1, and its corresponding churn time is t1, otherwise the churn label of this high-quality customer is 0; sv_dt represents the latest account cancellation time of the current product during the observation period, fx_dt represents the latest account cancellation time of the fixed-term product during the observation period, and inv_dt represents the latest account cancellation time of the wealth management product during the observation period.

4. The method for predicting high-quality customer churn based on supervised learning according to claim 3, characterized in that: Verifying the accuracy of the churn label of each high-quality customer includes: When the status mark of one of the three types of products of a high-quality customer is 1 during the observation period, observe whether the customer re-opens an account within a preset period of time after the observation period, and finally determine the corresponding status mark and churn label; When a high-quality customer's fixed-term product account is cancelled but the wealth management product account is not, or when the wealth management product account is cancelled but the fixed-term product account is not, a circulation verification is performed, i.e., it is verified whether the high-quality customer purchased the wealth management product when the fixed-term product account was cancelled or whether the high-quality customer purchased the fixed-term product when the wealth management product account was cancelled, and then the churn label is finally determined; when both the fixed-term product and the wealth management product account are cancelled, they are decomposed into two samples, which correspond to the churn of the fixed-term product and the churn of the wealth management product respectively, and have the churn time when the corresponding products are cancelled, and then they are subjected to circulation verification in the above-mentioned circulation verification method.

5. The method for predicting high-quality customer churn based on supervised learning according to claim 2, characterized in that: The extraction of loss feature data from their natural attributes, product status information, in-bank asset information and transaction behavior based on their corresponding observation time includes: For natural attributes, the extracted churn feature data include gender, age, and household age of high-quality customers; For the product status information, the extracted churn feature data includes the new additions, expiration and holding data of fixed-term products and wealth management products within the preset period before the observation time corresponding to high-quality customers; For in-bank asset information, the extracted churn feature data include the average daily balances of current products and fixed-term products corresponding to high-quality customers in the preset period before the observation time; For transaction behavior, the extracted churn feature data includes the income and expenditure data of high-quality customers in the preset period before the observation time; Perform logarithmic processing on the amount features in the churn feature data of each high-quality customer.

6. A device for predicting high-quality customer churn based on supervised learning, characterized in that: The method for predicting high-quality customer churn based on supervised learning as claimed in any one of claims 1 to 5 comprises: Feature extraction module: used to obtain high-quality customers to be predicted to churn, and extract churn feature data of the high-quality customers to be predicted to churn within a preset time period; Churn prediction module: used to input the churn feature data of high-quality customers to be predicted into a pre-trained high-quality customer churn prediction model, and output high-quality customer churn prediction results; wherein, the high-quality customer churn prediction model is trained on the constructed high-quality customer churn prediction data set using the XGBoost algorithm.

7. A computer-readable storage medium comprising a stored computer program, characterized in that: The computer program is suitable for being loaded by a processor and executing the method for predicting high-quality customer churn based on supervised learning as described in any one of claims 1 to 5.

Images