Loss customer identification method and device, and medium

Abstract

The invention relates to the field of data analysis, and discloses a lost customer identification method, equipment and a medium. The lost customer identification method comprises the following steps: acquiring customer feature data of a target customer; respectively inputting the customer feature data into a plurality of first prediction models to obtain a plurality of preliminary prediction results; inputting the plurality of preliminary prediction results into a second prediction model to obtain a final prediction result; and based on the final prediction result, determining high-loss-risk customers in the target customers. According to the method, potential customers with high loss risks can be efficiently and accurately identified, prospective prediction of customer loss tendencies is realized, financial institutions are helped to timely take effective measures to retain customers, user viscosity is enhanced, and operation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and in particular to a lost customer identification method, device and medium. Background Art

[0002] In the securities market, customer churn has always been a major challenge facing financial institutions. Customer churn not only means direct economic losses, but may also have a negative impact on the brand image and market position of financial institutions. Traditional customer churn early warning methods mainly rely on simple statistical analysis, such as basic information such as customer transaction frequency and account balance changes, but these methods often lack in-depth understanding of customer behavior patterns and forward-looking prediction capabilities.

[0003] With the increasing complexity of the financial market and the continuous changes in customer needs, traditional customer churn warning methods can no longer meet the precision marketing and risk management needs of financial institutions. Therefore, a more advanced and accurate method for identifying lost customers is urgently needed to achieve forward-looking predictions of customer churn tendencies and help financial institutions take effective measures to retain customers in a timely manner. Summary of the invention

[0004] The main purpose of the present invention is to provide a lost customer identification method, device and medium, aiming to solve the technical problem of low lost customer identification accuracy.

[0005] A first aspect of the present invention provides a lost customer identification method, the lost customer identification method comprising: Obtain customer characteristic data of target customers; Inputting the customer characteristic data into a plurality of first prediction models respectively to obtain a plurality of preliminary prediction results; Inputting the multiple preliminary prediction results into a second prediction model to obtain a final prediction result; Based on the final prediction result, high churn risk customers among the target customers are determined.

[0006] Optionally, in a first implementation of the first aspect of the present invention, the customer characteristic data includes the customer age and customer balance of the target customer, and the number of transactions, average position ratio and relative rate of return of the target customer in each sub-period of a preset period; The step of obtaining the target customer's customer characteristic data includes: Obtaining the target customer's transaction records and account information, as well as market data within each of the sub-periods; Based on the transaction records, obtaining the number of transactions and customer yield of the target customer in each sub-period; Based on the account information, obtaining the customer age, the customer balance and the average position ratio; Based on the market data, obtaining the market rate of return in each of the sub-periods; Based on the customer rate of return and the market rate of return, the relative rate of return in each sub-period is determined.

[0007] Optionally, in a second implementation of the first aspect of the present invention, the customer characteristic data further includes at least one of a first loss aversion index, a second loss aversion index, a risk preference index, an investment confidence index, an overconfidence index, a user's recent rate of return percentile, and a user's market rate of return difference; The step of obtaining the target customer's customer characteristic data further includes: Determine the ratio of the total customer yield rate of the loss-making sub-period to the total customer yield rate of the profit-making sub-period for the target customer within the preset period, and obtain the first loss aversion index; Determine the ratio of the total number of transactions in the sub-period with losses to the total number of transactions in the sub-period with profits for the target customer within the preset period, and obtain the second loss aversion index; Determine the ratio of the sum of squares of the customer's rate of return in each sub-period to the sum of absolute values ​​of the target customer within the preset period, and obtain the risk preference index; The average value of the product of the number of transactions and the average position of the target customer in each sub-period within the preset period is determined to obtain the investment confidence index.

[0008] Optionally, in a third implementation of the first aspect of the present invention, the multiple first prediction models include an extreme random tree regressor, a guided clustering algorithm regressor combined with linear regression, and a multilayer perceptron regressor; each of the first prediction models is obtained by training an initial regression model corresponding to each of the first prediction models based on multiple sample feature data at different times and the loss probability corresponding to each of the sample feature data.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, the second prediction model is obtained by training a secondary regression model corresponding to the second prediction model based on the preliminary prediction feature matrix output by the multiple first prediction models and the corresponding churn probability.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, before inputting the customer feature data into a plurality of first prediction models respectively to obtain a plurality of preliminary prediction results, the method further includes: Based on the sample characteristic data at different times, determining the churn type of the sample customers at each time; Determine that the churn type at the first moment is explicit churn, and mark the churn probability of the sample customer at the first moment as 1; Determine that the churn type at the second moment is normal use, and mark the churn probability of the sample customer at the second moment as 0; Determine that the churn type at the third moment is implicit churn, and mark the churn probability of the sample customer at the third moment as a preset maximum implicit churn probability; It is determined that the third moment exists after the second moment, and based on the time interval between the second moment and the corresponding third moment, the churn probability of the sample customer at the second moment is updated.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, determining the churn type of the sample customer at each moment based on the sample feature data at different moments includes: Determining that the sample characteristic data of the sample customer indicates that the sample customer is in an account cancellation state at a target time, and determining that the churn type of the sample customer at the target time is the explicit churn; Determining that the sample characteristic data of the sample customer indicates that the sample customer is in a long-term low balance state at the target time, and determining that the churn type of the sample customer at the target time is the hidden churn; Otherwise, it is determined that the churn type of the sample customer at the target time is the normal usage.

[0012] Optionally, in a seventh implementation manner of the first aspect of the present invention, the churn probability of the sample customer at the second moment is updated based on the time interval between the second moment and the corresponding third moment, and is calculated based on the following formula: ; in, is the updated churn probability of the sample customer at the second moment, To preset the initial probability, is the preset maximum implicit loss probability, is the preset decay rate parameter, is the time interval between the second moment and the corresponding third moment.

[0013] The second aspect of the present invention also provides a computer device, which includes: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the computer device executes the lost customer identification method as described above.

[0014] A third aspect of the present invention further provides a computer-readable storage medium, on which instructions are stored, and when the instructions are executed by a processor, the lost customer identification method as described above is implemented.

[0015] The embodiment of the present invention provides a method, device, equipment and storage medium for identifying lost customers. The method first obtains the customer feature data of the target customer, then performs preliminary prediction through multiple first prediction models, and inputs the preliminary prediction results into the second prediction model to obtain the final prediction results, and identifies the target customers with high risk of loss based on the final prediction results. This method comprehensively considers the various feature data of the customer, and uses multiple prediction models for cascading and progressive multi-level prediction, thereby improving the accuracy and efficiency of lost customer identification. Compared with the traditional customer loss warning method, the present invention can deeply explore customer behavior patterns and realize forward-looking prediction of customer loss tendency, which not only improves the accuracy of lost customer identification, but also provides forward-looking prediction capabilities for financial institutions, so that they can take effective retention measures in time, thereby reducing customer churn rate, enhancing user stickiness, improving operational efficiency, reducing operating costs, and improving customer satisfaction and loyalty. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of a flow chart of an embodiment of a method for identifying lost customers in an embodiment of the present invention; Figure 2 for Figure 1 A schematic flow chart of an embodiment of step 101 in the embodiment; Figure 3 for Figure 1 A flow chart of another embodiment of step 101 in the embodiment; Figure 4 for Figure 1 A schematic diagram of a flow chart of an embodiment before step 102 in the embodiment; Figure 5 The figure is a functional module diagram of an embodiment of a computer device in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] For ease of understanding, the specific process of the lost customer identification method in the embodiment of the present invention is described below. The lost customer identification method of this embodiment is applied to electronic devices, which can be terminals, such as smart phones, laptops, tablet computers, desktop computers, etc., or development boards, chip systems, etc. Please refer to Figure 1 , Figure 1 This is a flow chart of a first embodiment of a method for identifying lost customers in an embodiment of the present invention. In this embodiment, the method for identifying lost customers includes: 101. Obtain customer characteristic data of target customers; In this embodiment, the target customer refers to the customer that the financial institution hopes to warn of churn, usually all securities customers, but also non-member customers, customers with few transactions and small account balances who are easy to churn, etc. Customer feature data is key information used to analyze and predict customer churn tendency, which can fully reflect the customer's behavior pattern, financial status and investment preferences, such as the customer's transaction record, position, login frequency, capital flow, age, gender, occupation, income level, investment history, risk tolerance, etc. By collecting and integrating these diverse feature data, rich input information can be provided for the subsequent prediction model, and certain weights and biases can be configured for these feature data, and a series of related parameters can be adjusted to improve the accuracy and robustness of the prediction model, thereby improving the accuracy and efficiency of churn customer identification.

[0019] In this embodiment, these data can be obtained through various channels, for example, they can be obtained through the transaction system of financial institutions, customer relationship management systems, market research reports, etc. After obtaining these customer characteristic data, they can be pre-processed, for example, through missing value filling, outlier processing, standardization and normalization processing and other data cleaning methods to improve the quality and availability of the data and provide an accurate basis for subsequent analysis and prediction. Among them, missing value processing can be done by filling the mean, median or adjacent unit values, or by a deletion strategy to make the data more complete and reliable. Outlier processing can identify and process outliers through statistical methods such as box plots or machine learning algorithms to make the data more consistent with the actual distribution and rules. Standardization and normalization processing is to scale the data according to a certain ratio so that it falls into a small specific interval, such as through the formula Convert data of different dimensions to the same scale, where is the original value, is the average value, is the standard deviation, That is, it is the converted value, so that feature data of different magnitudes or units can be compared and analyzed on the same scale to eliminate the dimensional differences between different features, thereby improving the convergence speed and prediction performance of the model.

[0020] Optionally, in some embodiments, the customer characteristic data includes the customer age and customer balance of the target customer, as well as the number of transactions, average position ratio and relative rate of return of the target customer in each sub-period of the preset period. Figure 2 , obtain the customer feature data source of the target customers, including: 1011. Obtain the target customer's transaction records and account information, as well as market data in each sub-period; 1012. Based on the transaction records, obtain the number of transactions and customer yield of the target customer in each sub-period; 1013. Based on account information, obtain customer age, customer balance and average position ratio; 1014. Based on market data, obtain the market return rate in each sub-period; 1015. Based on the customer's rate of return and the market's rate of return, determine the relative rate of return in each sub-period.

[0021] In this optional embodiment, the customer age and customer balance of the target customer are basic user personal information and financial information, which can reflect the basic situation and financial health of the customer. The customer age may affect his investment preference and risk tolerance, while the customer balance directly reflects the customer's financial strength and investment scale. The number of transactions of the target customer in each sub-period of the preset period reflects the customer's trading activity and investment behavior pattern. Frequent transactions may mean that the customer is sensitive to market dynamics, while fewer transactions may indicate that the customer prefers long-term investment or conservative strategies. The average position ratio shows the customer's position in a specific period, which helps to assess the customer's investment style and risk preference. A high position may represent the customer's optimism about the market or a strong willingness to invest, while a low position may mean that the customer is more cautious or has a reserved attitude towards the market. The relative rate of return is a comparison between the customer's rate of return and the market rate of return, which reflects the customer's investment ability and market performance. A positive relative rate of return indicates that the customer's investment return is better than the market average, while a negative relative rate of return may mean that the customer's investment strategy or market judgment is insufficient. These customer characteristic data together constitute an important basis for analyzing and predicting the target customer's tendency to churn.

[0022] In this optional embodiment, these customer feature data can be obtained from different data sources, and the specific data sources include the target customer's transaction records and account information, as well as the market data in each sub-period. Among them, the target customer's transaction record is obtained from the securities company's trading system. The detailed information such as the number of transactions, transaction amount, and buying and selling direction of the customer in each sub-period can be obtained through the transaction record. The account information can be obtained from the customer management system of the financial institution, including the basic information of the customer's age, gender, occupation, income level, customer balance, etc. The market data in each sub-period is obtained from the public market data source or professional financial data provider. These data reflect the performance of the market in different time periods, including key indicators such as the market rate of return. After obtaining the customer rate of return and the market rate of return, the relative rate of return in each sub-period can be calculated based on the customer rate of return and the market rate of return. By integrating these data, the investment behavior and market performance of the target customer can be fully depicted, and the customer's behavior pattern and investment performance can be fully understood, providing strong data support for the identification and early warning of subsequent lost customers.

[0023] In this optional embodiment, after obtaining these data, they can usually be further processed to obtain some data that are more conducive to model training, such as the interactive characteristics of age and relative yield represented by the product of age and relative yield, and the importance characteristics of age represented by the square of age. After actual verification, the introduction of these data can make the trained model have higher accuracy and robustness, and ensure the accuracy and stability of lost customer identification.

[0024] Optionally, in some embodiments, the customer characteristic data further includes at least one of a first loss aversion index, a second loss aversion index, a risk preference index, an investment confidence index, an overconfidence index, a user's recent rate of return percentile, and a user's market rate of return difference. Figure 3 , obtain the customer feature data source of the target customers, including: 1016. Determine the ratio of the total customer rate of return of the target customer in the loss-making sub-period to the total customer rate of return of the target customer in the profit-making sub-period within the preset period, and obtain a first loss aversion index; 1017. Determine the ratio of the total number of transactions in the loss-making sub-period to the total number of transactions in the profit-making sub-period within the preset period for the target customer, and obtain a second loss aversion index; 1018. Determine the ratio of the sum of squares and the sum of absolute values ​​of the target customer's customer yields in each sub-period within a preset period to obtain a risk preference index; 1019. Determine the average value of the product of the number of transactions and the average position of the target customer in each sub-period within a preset period to obtain the investment confidence index.

[0025] In this optional embodiment, it is specified that the customer characteristic data may include a variety of psychological indexes, among which the Loss Aversion Index (LAI) refers to the index used to measure the customer's sensitivity to losses when facing the same amount of gains and losses, and the customer's aversion to losses is much higher than the degree of preference for gains. The first loss aversion index and the second loss aversion index are parameters used to evaluate loss aversion from two perspectives, namely, from the perspective of yield and number of transactions. The first loss aversion index is obtained by calculating the ratio of the total customer yield of the sub-period of losses to the total customer yield of the sub-period of profits within a preset period, reflecting the customer's aversion to losses at the yield level. The second loss aversion index is obtained by calculating the ratio of the total number of transactions in the sub-period of losses to the total number of transactions in the sub-period of profits, reflecting the customer's tendency to avoid losses in trading behavior. These two indexes jointly reveal the psychological reactions and behavioral characteristics of customers when facing investment losses, which helps to deeply understand the customer's investment psychology and risk preferences.

[0026] In this optional embodiment, specifically, the first loss aversion index can be based on the formula Get, where is the first loss aversion index, is the customer rate of return in the ith sub-period, and n is the preset period. For example, , that is, the preset cycle is 6, and the self-cycle is each month in the last six months, represents the customer rate of return in the i-th month. In this formula, It means that when a loss occurs in a sub-period, the loss value is taken into the numerator for summation, while when there is a profit or no change, it is not taken into account. Accordingly, It means that when there is a profit in the sub-period, the profit value is taken into the denominator for summation, while it is not taken into account when there is a loss or no change. Therefore, this formula can be used to calculate the relative importance of the target customer in the state of loss and profit, and then reflect their psychological characteristics of loss aversion. A higher first loss aversion index may mean that the customer is very sensitive to investment losses and tends to avoid high-risk investments, while a lower value may indicate that the customer has a higher tolerance for losses and is more willing to take risks in pursuit of higher returns.

[0027] In this optional embodiment, specifically, the second loss aversion index can be based on the formula Get, where is the second loss aversion index, The same is the customer rate of return in the ith sub-period, is the number of transactions in the ith sub-period, so , that is, when a loss occurs, the number of transactions in the sub-period is included in the numerator of the formula and added to the sum. , that is, when there is a profit, the number of transactions in the sub-period is included in the denominator of the formula and added to the sum, so that the target customer's loss aversion in trading behavior can be calculated through this formula, further revealing their loss aversion psychological tendency. A higher second loss aversion index may mean that the customer is more inclined to reduce transactions to avoid further losses when facing losses, while a lower value may indicate that the customer is still willing to maintain or increase trading activities in a loss state, trying to recover losses or find new investment opportunities.

[0028] In this optional embodiment, the Risk Preference Index (RPI) is used to measure the customer's attitude and tolerance to risk. In this embodiment, the risk preference index is obtained by calculating the ratio of the sum of squares of customer returns in each sub-period within a preset period to the sum of absolute values, that is, its formula is: The index reflects the client's sensitivity and tolerance to return fluctuations during the investment process. A higher risk appetite index means that the client is more willing to take greater risks in pursuit of higher returns, while a lower risk appetite index indicates that the client prefers a conservative strategy and avoids greater investment fluctuations.

[0029] In this optional embodiment, the Investment Confidence Index (ICI) is used to measure the customer's confidence in the market and investment. In this embodiment, the Investment Confidence Index is obtained by calculating the average product of the number of transactions in each sub-period and the average position within a preset period, that is, its formula is: A higher number of transactions and average position ratio may mean that the client is more confident in his or her investment judgment and is willing to actively participate in market transactions and positions, while a lower number of transactions and average position ratio may indicate that the client is more cautious in investment decisions and lacks sufficient confidence.

[0030] In this optional embodiment, the Overconfidence Index (OCI) is a parameter used to measure whether the client is overconfident in investment decisions. Overconfidence may cause the client to overestimate his investment ability and market judgment, thereby making irrational investment decisions. In this embodiment, the overconfidence index can be comprehensively evaluated by analyzing factors such as the client's trading behavior, position status, and market performance. The specific calculation method can be designed and adjusted according to the actual situation. For example, the overconfidence index can be the median of the current position ratio and the maximum position ratio, wherein the current position ratio refers to the proportion of the investment target currently held by the target client to the total investment amount, and the median of the maximum position ratio refers to the median of the highest value of the position ratio reached by the target client in the past period of time. When the current position ratio of the target client is significantly higher than the median of the maximum position ratio, it may mean that the client is too optimistic about his investment judgment and is overconfident. By analyzing the overconfidence index, the psychological state and investment behavior characteristics of the client can be further understood, providing a reference for formulating personalized investment strategies and risk management measures.

[0031] In this embodiment, the user's recent return percentile refers to comparing the target customer's return rate within a preset period with the return rate of all customers during the same period to determine its percentile position. This indicator can reflect the performance level of the target customer in the overall market. A higher return percentile means that the target customer's investment performance is better than most customers, while a lower return percentile may indicate that its investment strategy or market judgment needs to be improved. By analyzing the user's recent return percentile, the customer's investment ability and market performance can be evaluated, providing a basis for subsequent lost customer warning and retention strategies. The user's market return difference refers to the difference between the target customer's return rate within a preset period and the market return rate. This indicator is used to measure the customer's performance relative to the market as a whole. A positive return difference indicates that the customer's investment return is higher than the market average, while a negative return difference means that the customer's investment return is lower than the market average performance. When the user's recent return percentile is low or the user's market return difference is negative, the customer may doubt his or her investment ability, and then feel frustrated and anxious. This psychological state may cause the customer to lose confidence in investment and increase the risk of loss.

[0032] In this optional embodiment, the above-mentioned first loss aversion index, second loss aversion index, risk preference index, investment confidence index, overconfidence index, user recent rate of return percentile and user market rate of return difference are all effective psychological indexes, which can reveal the psychological characteristics and investment behavior patterns of target customers from different dimensions, and provide a strong basis for accurately identifying lost customers. By integrating and analyzing these customer characteristic data, the psychological characteristics and investment behavior patterns of target customers can be fully understood, and strong data support can be provided for subsequent identification and early warning of lost customers. At the same time, personalized investment strategies and risk management measures can be formulated according to the psychological characteristics and investment behavior patterns of customers to improve customer satisfaction and loyalty and reduce the risk of loss. Using these psychological indexes as input parameters of the prediction model can significantly improve the prediction accuracy and generalization ability of the model, build a comprehensive and accurate lost customer identification model, and provide strong decision support for financial institutions.

[0033] 102. Input the customer feature data into a plurality of first prediction models respectively to obtain a plurality of preliminary prediction results; In this embodiment, the first prediction model is a prediction model constructed based on a machine learning algorithm, such as a logistic regression model, a support vector machine model, a random forest model or a neural network model. These models can make a preliminary prediction of the target customer's tendency to churn based on various factors in the customer feature data. Each first prediction model may focus on different features or data combinations, thereby providing a variety of prediction perspectives and results. By inputting the customer feature data into multiple first prediction models respectively, multiple preliminary prediction results can be obtained, and these results will serve as the basis for subsequent comprehensive analysis and decision-making. Among them, the first prediction model preferably adopts a regression model, which is a statistical model that can predict continuous variables. It predicts the value of the target variable by analyzing the relationship between the independent variable and the dependent variable. In the identification of lost customers, the regression model can be used to predict the possibility of customer churn, that is, the churn probability. The probability value is usually between 0 and 1, indicating the possibility of customer churn in the future. By constructing a regression model, we can better understand the driving factors of customer churn and formulate effective retention strategies based on these factors. In this embodiment, using a regression model as the first prediction model can make full use of the continuous variable information in the customer feature data and improve the accuracy of the prediction. At the same time, the regression model also has good interpretability, which can help us deeply understand the logic and reasons behind the model's prediction results. By inputting customer feature data into the regression model, we can get the predicted probability of each customer churn in the future, providing a basis for subsequent decision-making.

[0034] Optionally, in some embodiments, multiple first prediction models include an extreme random tree regressor, a guided clustering algorithm regressor combined with linear regression, and a multilayer perceptron regressor; each first prediction model is obtained by training an initial regression model corresponding to each first prediction model based on multiple sample feature data at different times and the loss probability corresponding to each sample feature data.

[0035] In this optional embodiment, the sample feature data can be obtained in the same manner as the aforementioned customer feature data. The sample feature data is used here to distinguish the features used to train the model from the features used in actual prediction. When the model training is specifically implemented, the sample feature data can be screened from historical customer feature data. These feature data correspond to churn probabilities, which can be obtained through statistical historical data, expert annotations, or specific precise calculation methods, and represent the actual churn of customers under the corresponding feature data. These sample feature data and their corresponding churn probabilities are used as training sets and input into different initial regression models respectively. The model is trained through a machine learning algorithm so that it can learn the mapping relationship between feature data and churn probability.

[0036] In this optional embodiment, the Extra Trees Regressor is a regression model constructed using the Extremely Randomized Trees (ERT) algorithm. Extremely randomized trees are an integrated learning method that improves the stability and accuracy of the model by constructing multiple decision trees and integrating their prediction results, and introduces additional randomness in the process of its construction. In the extreme randomized trees regressor, the splitting nodes of each decision tree are selected randomly rather than based on the optimal splitting strategy. This randomness helps to reduce the risk of overfitting of the model and improve the generalization ability of the model. By inputting customer feature data into the extreme randomized trees regressor, the churn probability based on the comprehensive prediction of multiple decision trees can be obtained, providing a more robust prediction result for subsequent churn customer identification.

[0037] In this optional embodiment, in the process of training to obtain the extreme random tree regressor, the sample feature data is input into the initial regression model of the extreme random tree regressor as input, and the corresponding churn probability is used as the target output variable, and the input sample feature data is divided into a training set and a validation set. In the training process, multiple random trees are first constructed through the training set, and for each tree, features and feature values ​​are randomly selected to split nodes until a stopping condition is met, such as reaching a maximum depth or a minimum number of samples, to obtain a preliminary extreme random tree regressor, wherein, in the process of splitting nodes, for each node, features and thresholds that minimize the Gini impurity or variance can be selected to split the node, and these features and thresholds are randomly selected from the candidate sample feature data, rather than based on a global optimal selection. For the newly input sample set, each tree will give a prediction value. During the prediction process, the target variable mean of the node is output as the prediction value according to which node its feature value falls into. Finally, the average value or weighted average value of each tree output is used as the prediction value, and the prediction value is compared with the churn probability of the target output. By calculating the difference between the prediction value and the actual churn probability, the model parameters are continuously adjusted to optimize the prediction performance of the model and obtain the final extreme random tree regressor. This iterative training process based on the training set and the validation set can ensure that the extreme random tree regressor has good generalization ability while maintaining high prediction accuracy, and can be applied to customer data with different feature distributions.

[0038] In this optional embodiment, the Bagging Regressor with Linear Regression combined with linear regression is a regressor constructed using a Bagging algorithm and a linear regression model. The Bagging algorithm is an integrated learning method that improves the stability and accuracy of the model by constructing multiple models, such as weak learners of the same type, using linear regression (LinearRegression) as a base learner on these models, and synthesizing the prediction results of these base learners. In the process of training the Bagging Regressor combined with linear regression, sample feature data and its corresponding loss probability can also be used for training, and the training set data is randomly divided into multiple subsets, such as 50 subsets, each of which is used to train a linear regression model. These linear regression models are used as base learners to learn and optimize on their respective training subsets, and each linear model is trained on data from different subsets extracted by the bootstrap method to obtain preliminary prediction capabilities. Then, the prediction results of multiple linear regression models are integrated through the guided aggregation algorithm, and the final prediction results can be obtained by averaging or weighted averaging. This ensemble learning method helps improve the stability and accuracy of the model and reduce the risk of overfitting that may be caused by a single model. At the same time, the linear regression model, as a base learner, has the advantages of being simple, easy to understand and implement, and can quickly adapt to the training needs of large-scale data. In the guided aggregation algorithm regressor combined with linear regression, by continuously optimizing the parameters and ensemble strategies of the base learner, the prediction performance of the model can be further improved, providing more reliable prediction results for churn identification.

[0039] In this optional embodiment, the multilayer perceptron regressor (Multilayer Perceptron Regressor) is a regression model built based on an artificial neural network. The multilayer perceptron (Multilayer Perceptron, MLP) is a feedforward neural network, which consists of an input layer, a hidden layer and an output layer. Each hidden layer contains multiple neurons. Each neuron applies an activation function, and a nonlinear transformation is performed after summing the weighted input. The network adjusts the weights through the back propagation algorithm to minimize the loss function, such as the mean square error, so as to learn the mapping relationship between the input and the output. For example, the parameters of the model can set the hidden layers to 64 layers and 32 layers respectively, the activation function can use the ReLU function, the optimization algorithm can use the Adam optimization algorithm, the size of each batch is set to 64, the learning rate is set to 0.001, and the loss function uses the mean square error. In the process of training to obtain the multilayer perceptron regressor, the sample feature data and its corresponding loss probability can also be used for training. The data of the training set is input into the initial model of the multilayer perceptron regressor, the predicted value is calculated by forward propagation, and compared with the loss probability of the target output, and the loss function is calculated. Then, the back propagation algorithm is used to adjust the weights and bias parameters of each layer of neurons in the model according to the gradient information of the loss function to minimize the loss function. Through multiple iterations of training, until the prediction performance of the model on the validation set reaches stability or a preset stop condition, such as the loss function converges or reaches the maximum number of iterations, a trained multi-layer perceptron regressor is obtained. This regression model based on artificial neural networks has powerful nonlinear mapping and adaptive learning capabilities. It can automatically learn the potential relationship between customer churn and features from complex sample feature data, and provide high-precision prediction results for churn customer identification. By inputting customer feature data into the trained multi-layer perceptron regressor, the churn probability based on neural network prediction can be obtained. This probability value can more accurately reflect the customer's churn tendency and provide strong support for the subsequent formulation of customer retention strategies.

[0040] Optional, see Figure 4 In order to obtain the churn probability as the target output variable of the prediction model, before respectively inputting the customer feature data into a plurality of first prediction models to obtain a plurality of preliminary prediction results, the method further includes: 1001. Based on the sample characteristic data at different times, determine the churn type of the sample customers at each time; 1002. Determine that the churn type at the first moment is explicit churn, and mark the churn probability of the sample customer at the first moment as 1; 1003. Determine that the churn type at the second moment is normal use, and mark the churn probability of the sample customer at the second moment as 0; 1004. Determine that the churn type at the third moment is implicit churn, and mark the churn probability of the sample customer at the third moment as a preset maximum implicit churn probability; 1005. Determine that there is a third moment after the second moment, and update the churn probability of the sample customer at the second moment based on the time interval between the second moment and the corresponding third moment.

[0041] In this optional embodiment, the sample customer refers to the target customer group corresponding to the sample characteristic data. After obtaining the sample characteristic data, the sample customers need to be labeled with the churn type in order to construct a training set and a validation set. First, based on the sample characteristic data at different times, the churn type of the sample customers at each time can be determined. The churn types are specifically divided into three types: explicit churn, normal use, and implicit churn. Explicit churn refers to the situation where a customer has clearly expressed or demonstrated the intention or behavior to leave or stop using the service; normal use means that the customer is still using the service at the current moment and shows no signs of churn; implicit churn means that although the customer is still using the service, its frequency of use, activity, or consumption amount and other indicators have shown a significant decline, indicating that the customer may be about to churn.

[0042] In this optional embodiment, specifically, based on the sample feature data at different times, determining the churn type of the sample customer at each time includes: (1) Determine that the sample characteristic data of the sample customers indicates that the sample customers are in the account cancellation state at the target time, and determine that the churn type of the sample customers at the target time is explicit churn; (2) Determine that the sample characteristic data of the sample customers indicates that the sample customers are in a state of long-term low balance at the target time, and determine that the churn type of the sample customers at the target time is hidden churn; (3) Otherwise, determine that the churn type of the sample customer at the target time is normal usage.

[0043] In this optional embodiment, the account cancellation status refers to the state in which the customer's account in the customer relationship management (CRM) system of a bank or financial institution has been cancelled or closed, and there is a record of account cancellation in the past three months, which usually means that the customer has clearly stated that he wants to stop using the service of the institution, so the churn type of such customers at the corresponding time can be marked as explicit churn. The long-term low balance state means that the customer's account balance continues to remain at a low level for a period of time, and there is no obvious transaction activity or capital flow. For example, when the current user's assets are less than 100 yuan for more than three consecutive months, and the account assets were greater than 5,000 yuan three months ago, it can be judged that the user is in a long-term low balance state. This state may indicate that the customer's frequency of use or dependence on the service has been significantly reduced, indicating that the customer may be about to churn, so the churn type of such customers at the corresponding time can be marked as implicit churn. For customers in normal use, their sample feature data are usually characterized by stable account balances, frequent transaction activities, etc., indicating that the customer is still using the service and has no signs of churn.

[0044] In this optional embodiment, the same customer will show different types of churn at different times, where the first moment refers to the moment when the customer first shows explicit churn behavior. At this moment, since the customer has clearly expressed or shown the intention or behavior to leave or stop using the service, it can be determined that the churn type is explicit churn, and the churn probability of the sample customer at the first moment is marked as 1, indicating that the customer has completely churned. Explicit churn is the most direct and obvious type of churned customer. For financial institutions, once a customer has explicit churn, it often means that the customer has been lost. Therefore, it is necessary to take timely measures to understand the reasons for customer churn in order to improve services and avoid similar situations from happening again.

[0045] In this optional embodiment, the second moment refers to the moment when the customer is using the service normally. At this moment, the customer is still using the service and shows no signs of churn, so it can be determined that the churn type is normal use, and the churn probability of the sample customer at the second moment is initially marked as 0, indicating that the customer is currently not at risk of churn. Normal use is the norm for customers to use services, and it is also the state expected by financial institutions. By paying attention to the normal usage behavior of customers, financial institutions can understand the needs and preferences of customers, further optimize services, and improve customer satisfaction and loyalty. However, since normal customers may still be at risk of churn, this embodiment will subsequently update and optimize the preliminary marking results of the second moment according to the smoothness of the change in the churn probability.

[0046] In this optional embodiment, the third moment refers to the moment when it is determined that the customer has hidden churn behavior. At this moment, although the customer is still using the service, its usage frequency, activity or consumption amount and other indicators have shown a significant decline, indicating that the customer may be about to churn. Therefore, it can be determined that the churn type is hidden churn, and the churn probability of the sample customer at the third moment is marked as the preset maximum hidden churn probability. The preset maximum hidden churn probability is an indicator set based on historical data and experience. For example, the maximum hidden churn probability can usually be set to 0.7, which is used to indicate the possibility of customer churn in a hidden churn state. By marking the hidden churn probability, financial institutions can promptly identify potential churn customers and take corresponding retention measures to reduce customer churn rates.

[0047] In this optional embodiment, after determining the churn type and churn probability of the sample customer at each moment, it is also necessary to consider the effect of the time interval on the churn probability. In particular, when there is a third moment after the second moment, since the customer's usage behavior may change over time, it is necessary to update the churn probability of the sample customer at the second moment based on the time interval between the second moment and the corresponding third moment. This update can take into account the effect of the time factor on the customer churn tendency, making the prediction result more accurate and reliable.

[0048] Optionally, in one embodiment, based on the time interval between the second moment and the corresponding third moment, the probability of churn of the sample customer at the second moment is updated based on the following formula: Calculated; among them, is the probability of churn of the updated sample customers at the second moment, To preset the initial probability, To preset the maximum hidden loss probability, is the preset decay rate parameter, is the time interval between the second moment and the corresponding third moment In this optional embodiment, in order to smooth the change of churn probability, a calculation method of time decay function is introduced. When it is detected for the first time that a customer meets the churn condition, such as the above-mentioned judgment standard of just reaching the long-term low balance state, the formula can be used to gradually reduce the churn probability from that moment forward until it reaches the minimum value, so that the corresponding churn probability at the second moment is more in line with the actual behavior change trend of the customer. Among them, the preset initial probability It can be set according to the actual situation. It can usually be set to a smaller value, such as 0.2, which indicates the initial risk of customer churn under normal use. As before, it can be set based on historical data and experience, such as 0.7. Preset decay rate parameters It determines the speed at which the probability of churn decays with time interval t. The larger it is, the faster the probability of churn decays. The smaller it is, the slower the probability of churn decays. By adjusting the preset random rate parameter to control the growth rate of the probability, such as 0.01. By adjusting these parameters, the updated churn probability can be made more consistent with the actual situation and the accuracy of the prediction can be improved. For example, if a customer is determined to be a hidden churn on June 10, 2024, the probability on that day is marked as 70, and the churn probability is marked forward by date, such as: churn probability on June 9, 2024 = 20 + (70-20) * e -0.01*1 ≈69.5.

[0049] 103. Input the multiple preliminary prediction results into the second prediction model to obtain the final prediction result; In this embodiment, after obtaining multiple preliminary prediction results, the prediction results of each basic model will be used as the input features of the secondary second prediction model. The second prediction model is usually a linear regression or other simple model, which is based on the prediction results of multiple first prediction models to make a further prediction to obtain a more stable and accurate final prediction result. In this way, the prediction capabilities of each first prediction model can be fully utilized while avoiding the deviation and overfitting problems that may be caused by a single model. In a specific implementation, the second prediction model can use a weighted average and bias processing method to perform weighted summation and bias on the prediction results of each first prediction model to obtain the final prediction result. Among them, the determination of weights and biases can be comprehensively considered based on factors such as the performance and relevance of the first prediction model. In this way, the accuracy and reliability of the prediction results can be further improved, providing more powerful support for the identification of lost customers.

[0050] Optionally, in one embodiment, the second prediction model is obtained by training a secondary regression model corresponding to the second prediction model based on the preliminary prediction feature matrices and corresponding churn probabilities output by the multiple first prediction models.

[0051] In this optional embodiment, the preliminary prediction feature matrix is ​​obtained based on the preliminary prediction results output by the multiple first prediction models. Specifically, the first prediction models are respectively based on the training set ( , ) is trained and the corresponding test set ( , ) is iteratively optimized, and each first prediction model predicts the training set and the test set to generate a new feature matrix and , that is, the preliminary feature matrix, and the churn probability is the target variable corresponding to the output of the second prediction model, which can use the churn probability value consistent with the aforementioned training of each first prediction model, that is, ,in is the algorithm corresponding to the second prediction model, and its parameters are determined by minimizing the corresponding output and the loss function of the churn probability. The second prediction model uses the trained parameters to predict the prediction results of the first prediction model on the test set, thereby obtaining the final prediction value For example, the three first prediction models respectively obtain three prediction values ​​of 24.5, 26.0 and 27.3, which constitute the feature vector =[24.5,26.0,27.3], then the secondary regression model corresponding to the second prediction model will be based on the feature matrix in the training set and the original target variable To learn weights and perform bias processing, the second prediction model is in the form of ,in, , , is the weight, b is the bias term, , , are the predicted values ​​of the three first prediction models respectively. Then for a new input sample , the secondary model will be based on the formula , prediction is made using the trained weights and bias terms.

[0052] The second prediction model obtained by the training in this optional embodiment can more accurately reflect the customer's churn tendency when predicting the churn probability of new sample customers. This comprehensive prediction method based on multiple prediction models not only improves the accuracy of the prediction, but also enhances the generalization ability of the model, making the prediction results more reliable and stable. At the same time, by carefully dividing and marking the churn types of sample customers and considering the impact of time intervals on the churn probability, the accuracy and practicality of the prediction model are further improved, and it is possible to accurately learn how to best combine the prediction values ​​obtained by these first prediction models to produce a more accurate final prediction. It is beneficial for financial institutions to promptly discover potential churn risks and take corresponding retention measures, thereby effectively reducing customer churn rates and improving customer satisfaction and loyalty. In addition, this comprehensive prediction method can also provide financial institutions with more refined customer management strategies, such as formulating personalized services and marketing strategies for customers of different churn types to better meet customer needs and enhance customer stickiness.

[0053] 104. Based on the final prediction results, identify target customers with high churn risk.

[0054] In this embodiment, after obtaining the final prediction result, the high churn risk customers among the target customers can be determined according to the result. The final prediction result is actually a relative value, or a churn coefficient, which reflects the churn tendency of a user relative to other users, rather than a probability in a strict statistical sense. Therefore, in practical applications, this value is more suitable for user stratification and priority sorting. High churn risk customers refer to customers who are at a higher level in user stratification and priority and have a relatively large churn coefficient. These customers have a higher churn tendency and are the objects that financial institutions need to focus on and retain.

[0055] In this embodiment, high churn risk customers among target customers can be determined in a variety of ways. For example, high churn risk customers can be determined by setting a threshold. Specifically, a threshold of a churn coefficient can be set according to actual business needs. When the churn coefficient of a customer exceeds the threshold, it can be determined as a high churn risk customer. At the same time, the final predicted values ​​can be sorted in reverse order and a preset number or preset ratio of the target customers with the highest final predicted values, such as the top 100 or top 1%, can be taken as high churn risk customers. The corresponding preset number and preset ratio can be set according to the resources and retention capabilities of the financial institution to ensure that these high-risk customers can be effectively paid attention to and dealt with. In actual operation, financial institutions can also flexibly select and adjust methods for identifying high churn risk customers in combination with their own business characteristics and customer data. In addition, in order to continuously improve the accuracy and efficiency of churn customer identification, financial institutions should also continuously optimize prediction models, update training data, introduce new feature variables, etc., to adapt to market changes and the dynamic development of customer needs, so as to better understand and grasp the customer's churn tendency and provide strong support for formulating effective customer management strategies.

[0056] In this embodiment, through the above-mentioned method, financial institutions can more accurately identify potential lost customers, so as to take targeted retention measures and focus limited resources on those users who need the most attention, thereby improving operational efficiency and ensuring customer retention rate. For example, personalized retention activities or preferential plans can be designed for high-loss risk users, customer service can be strengthened, potential problems can be solved in advance, and user satisfaction can be improved. Human and financial resources can be reasonably allocated to ensure that the most efforts are invested in the most critical places. At the same time, this method can also help financial institutions make more data-driven decisions, reduce the impact of subjective judgments, and improve the scientificity and rationality of decisions. For example, by analyzing the prediction results, better understand the user's behavior patterns and preferences, take preventive measures in advance, and accurately locate high-risk users to avoid unnecessary large-scale marketing activities and reduce operating costs. Through this method, high-loss risk users can be effectively and accurately paid special attention to in the future, improve user experience, enhance user stickiness, and thus improve overall user satisfaction and loyalty.

[0057] The present invention also provides a computer device, which includes a memory and a processor. The memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes the lost customer identification method as described above. Figure 5 2 is a functional module diagram of a computer device provided by an embodiment of the present invention. The computer device 200 may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 210 (for example, one or more processors) and a memory 220, and one or more storage media 230 (for example, one or more mass storage devices) storing application programs 233 or data 232. Among them, the memory 220 and the storage medium 230 may be temporary storage or permanent storage. The program stored in the storage medium 230 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the computer device 200. Furthermore, the processor 210 may be configured to communicate with the storage medium 230 to execute a series of instruction operations in the storage medium 230 on the computer device 200.

[0058] The computer device 200 may also include one or more power supplies 240, one or more wired or wireless network interfaces 250, one or more input and output interfaces 260, and / or one or more operating systems 231, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 5The illustrated computer device structure does not constitute a limitation on the computer device, and may include more or fewer components than illustrated, or combine certain components, or arrange the components differently.

[0059] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the lost customer identification method as described above.

[0060] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. If the integrated module or unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0061] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying lost customers, characterized in that: include: Obtain customer characteristic data of target customers; Inputting the customer characteristic data into a plurality of first prediction models respectively to obtain a plurality of preliminary prediction results; Inputting the multiple preliminary prediction results into a second prediction model to obtain a final prediction result; Based on the final prediction result, high churn risk customers among the target customers are determined.

2. The lost customer identification method according to claim 1, characterized in that: The customer characteristic data includes the customer age and customer balance of the target customer, as well as the number of transactions, average position ratio and relative rate of return of the target customer in each sub-period of the preset period; The step of obtaining the target customer's customer characteristic data includes: Obtaining the target customer's transaction records and account information, as well as market data within each of the sub-periods; Based on the transaction records, obtaining the number of transactions and customer yield of the target customer in each sub-period; Based on the account information, obtaining the customer age, the customer balance and the average position ratio; Based on the market data, obtaining the market rate of return in each of the sub-periods; Based on the customer rate of return and the market rate of return, the relative rate of return in each sub-period is determined.

3. The lost customer identification method according to claim 2, characterized in that: The customer characteristic data also includes at least one of a first loss aversion index, a second loss aversion index, a risk preference index, an investment confidence index, an overconfidence index, a user's recent rate of return percentile, and a user's market rate of return difference; The step of obtaining the target customer's customer characteristic data further includes: Determine the ratio of the total customer yield rate of the loss-making sub-period to the total customer yield rate of the profit-making sub-period for the target customer within the preset period, and obtain the first loss aversion index; Determine the ratio of the total number of transactions in the sub-period with losses to the total number of transactions in the sub-period with profits for the target customer within the preset period, and obtain the second loss aversion index; Determine the ratio of the sum of squares of the customer's rate of return in each sub-period to the sum of absolute values ​​of the target customer within the preset period, and obtain the risk preference index; The average value of the product of the number of transactions and the average position of the target customer in each sub-period within the preset period is determined to obtain the investment confidence index.

4. The lost customer identification method according to claim 1, characterized in that: The multiple first prediction models include an extreme random tree regressor, a guided clustering algorithm regressor combined with linear regression, and a multilayer perceptron regressor; each of the first prediction models is obtained by training an initial regression model corresponding to each of the first prediction models based on multiple sample feature data at different times and the loss probability corresponding to each of the sample feature data.

5. The lost customer identification method according to claim 4, characterized in that: The second prediction model is obtained by training a secondary regression model corresponding to the second prediction model based on the preliminary prediction feature matrix output by the multiple first prediction models and the corresponding churn probabilities.

6. The lost customer identification method according to claim 4, characterized in that: Before the customer feature data is input into a plurality of first prediction models respectively to obtain a plurality of preliminary prediction results, the method further includes: Based on the sample characteristic data at different times, determining the churn type of the sample customers at each time; Determine that the churn type at the first moment is explicit churn, and mark the churn probability of the sample customer at the first moment as 1; Determine that the churn type at the second moment is normal use, and mark the churn probability of the sample customer at the second moment as 0; Determine that the churn type at the third moment is implicit churn, and mark the churn probability of the sample customer at the third moment as a preset maximum implicit churn probability; It is determined that the third moment exists after the second moment, and based on the time interval between the second moment and the corresponding third moment, the churn probability of the sample customer at the second moment is updated.

7. The lost customer identification method according to claim 6, characterized in that: The determining the churn type of the sample customer at each moment based on the sample characteristic data at different moments includes: Determining that the sample characteristic data of the sample customer indicates that the sample customer is in an account cancellation state at a target time, and determining that the churn type of the sample customer at the target time is the explicit churn; Determining that the sample characteristic data of the sample customer indicates that the sample customer is in a long-term low balance state at the target time, and determining that the churn type of the sample customer at the target time is the hidden churn; Otherwise, it is determined that the churn type of the sample customer at the target time is the normal usage.

8. The lost customer identification method according to claim 6, characterized in that: Based on the time interval between the second moment and the corresponding third moment, the churn probability of the sample customer at the second moment is updated and calculated based on the following formula: ； in, is the updated churn probability of the sample customer at the second moment, To preset the initial probability, is the preset maximum implicit churn probability, is the preset decay rate parameter, is the time interval between the second moment and the corresponding third moment.

9. A computer device, characterized in that: The computer device comprises: a memory and at least one processor, wherein instructions are stored in the memory; The at least one processor calls the instructions in the memory to enable the computer device to execute the lost customer identification method according to any one of claims 1 to 8.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the lost customer identification method according to any one of claims 1 to 8 is implemented.

Images