Customer loss redemption method and device, computer equipment and storage medium

Abstract

The invention belongs to the field of artificial intelligence and financial science and technology, and relates to a customer loss redemption method, which comprises the following steps: performing prediction according to customer feature data of a target customer through a trained prediction model, the prediction model comprising a first prediction sub-model and a second prediction sub-model, the prediction sub-model comprising a first prediction sub-model and a second prediction sub-model comprising a first prediction sub-model and a second prediction sub-model; according to the method, client feature data is predicted through a first prediction sub-model and a second prediction sub-model, a first prediction result and a second prediction result are obtained, a client retrieval success rate is obtained through weighted summation, and then a decision is made according to client state information and the client retrieval success rate through a trained Markov decision model. And an optimal customer retrieval strategy is obtained. The invention further provides a customer loss recovery device, computer equipment and a storage medium. In addition, the invention also relates to a block chain technology, and client feature data can be stored in a block chain. According to the method, the prediction accuracy of the client retrieval success rate can be improved, and an accurate retrieval strategy is provided for the client.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and financial technology, and in particular to a method, device, computer equipment and storage medium for recovering customer churn. Background Art

[0002] In recent years, with the rapid popularization of mobile Internet, customer behavior habits have accelerated their migration to online, and the rapid development of financial technology, artificial intelligence, and big data application technology has made it possible to provide intelligent services for individual customers. For banks, customer resources are one of the most important resources for commercial banks, and preventing customer loss has become an important measure for commercial banks in the face of fierce competition. With the changes in the population structure and the rapid development of Internet quasi-credit card products such as Ant Huabei and JD Baitiao, the traditional credit card market has shifted from an incremental market to a stock market, which has put forward higher requirements for the refined operation of commercial banks, among which the effective recovery of lost customers is particularly important.

[0003] Traditional customer recovery mainly takes the form of phone calls, text messages, letters, and home visits by account managers. Its strategies mainly rely on manual experience, and its level of automation and intelligence is low. There are problems such as indiscriminate recovery, uncontrollable marketing costs, and excessive disturbance of users. As a result, the positioning of recovered customers is inaccurate, and it is impossible to provide accurate recovery strategies. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a method, apparatus, computer device and storage medium for recovering lost customers, so as to solve the technical problem that the prior art cannot accurately locate the recovered customers during the customer recovery process and cannot provide accurate recovery strategies.

[0005] In order to solve the above technical problems, the embodiment of the present application provides a method for recovering lost customers, which adopts the following technical solution:

[0006] Acquire customer feature data of target customers, and construct a customer feature vector according to the customer feature data;

[0007] Inputting the customer feature vector into a trained prediction model, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model;

[0008] Processing the customer feature data through the first prediction sub-model to obtain customer status information and a first prediction result, and inputting the customer status information into the second prediction sub-model;

[0009] Processing the customer feature vector and the customer status information by the second prediction sub-model to obtain a second prediction result;

[0010] Taking a weighted sum of the first prediction result and the second prediction result to obtain a customer recovery success rate;

[0011] The customer status information and the customer recovery success rate are input into a trained Markov decision model to make a decision, thereby obtaining an optimal recovery strategy for the target customer.

[0012] In order to solve the above technical problems, the embodiment of the present application further provides a device for recovering lost customers, which adopts the following technical solution:

[0013] An acquisition module, used to acquire customer feature data of target customers and construct a customer feature vector according to the customer feature data;

[0014] An input module, used for inputting the customer feature vector into a trained prediction model, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model;

[0015] A first prediction module, configured to process the customer feature data through the first prediction sub-model to obtain customer status information and a first prediction result, and input the customer status information into the second prediction sub-model;

[0016] A second prediction module, configured to process the customer feature vector and the customer status information through the second prediction sub-model to obtain a second prediction result;

[0017] A weighting module, used for performing a weighted summation of the first prediction result and the second prediction result to obtain a customer recovery success rate;

[0018] The decision module is used to input the customer status information and the customer recovery success rate into a trained Markov decision model to make a decision and obtain the optimal recovery strategy for the target customer.

[0019] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0020] The computer device includes a memory and a processor, wherein the memory stores computer-readable instructions, and when the processor executes the computer-readable instructions, the steps of the method for recovering customer churn as described above are implemented.

[0021] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0022] The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the method for recovering customer churn as described above are implemented.

[0023] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0024] The present application provides a method for recovering customer churn, which uses a trained prediction model to perform predictions based on customer feature data of target customers, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model, and the customer feature data is predicted by the first prediction sub-model and the second prediction sub-model respectively to obtain a first prediction result and a second prediction result, and a weighted sum is performed to obtain a customer recovery success rate, and then a decision is made based on customer status information and the customer recovery success rate through a trained Markov decision model to obtain an optimal customer recovery strategy, which can effectively improve the prediction accuracy of the customer recovery success rate, and in combination with the customer status information and the customer recovery success rate obtained by the prediction model, it can more accurately locate the recovered customers and provide them with accurate recovery strategies, making customer recovery more efficient, intelligent, and controllable. At the same time, it can promote business personnel to recover target customers and improve the work efficiency of business personnel. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0026] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0027] Figure 2 is a flow chart of an embodiment of a method for recovering customer churn according to the present application;

[0028] Figure 3 It is a structural schematic diagram of an embodiment of a device for recovering lost customers according to the present application;

[0029] Figure 4 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION

[0030] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0031] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0032] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0033] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0034] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0035] This application provides a method for recovering lost customers, which can be applied to Figure 1 In the system architecture 100 shown, the system architecture 100 may include terminal devices 101, 102, 103, a network 104 and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0036] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0037] Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 (Moving Picture Experts Group Audio Layer IV) players, laptop computers, desktop computers, etc.

[0038] The server 105 may be a server that provides various services, such as a background server that provides support for web pages displayed on the terminal devices 101 , 102 , and 103 .

[0039] It should be noted that the method for recovering lost customers provided in the embodiments of the present application is generally executed by a server / terminal device, and accordingly, the device for recovering lost customers is generally arranged in the server / terminal device.

[0040] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0041] Continue to refer Figure 2 , showing a flow chart of an embodiment of a method for recovering customer churn according to the present application, comprising the following steps:

[0042] Step S201, obtaining customer feature data of target customers, and constructing a customer feature vector according to the customer feature data.

[0043] The target customers are the customers who need to be recovered. Customer characteristic data include but are not limited to customer information, account information, consumption records, repayment records, credit status, historical customer contacts and other time series data. Among them, customer information includes age, gender, education, occupation, group value level, etc.; account information includes card opening time, card level, credit limit, account opening method, etc.; consumption records include consumption time, number of consumption, consumption amount, credit limit utilization rate, etc.; credit status includes credit score, number of valid credit cards, valid credit limit, number of inquiries, etc.; historical customer contacts include customer contact date, customer contact time, customer contact method, whether connected, whether in person, whether agree to participate, whether to receive the task, final task completion rate, etc.; other time series data include latitude and longitude information, app login, wifi information, etc.

[0044] In this embodiment, the method for recovering lost customers is executed on an electronic device (eg Figure 1 The server / terminal device shown in the figure can obtain the customer characteristic data of the target customer through a wired connection or a wireless connection. It should be noted that the above wireless connection method may include but is not limited to 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other wireless connection methods currently known or to be developed in the future.

[0045] In this embodiment, after obtaining the customer feature data of the target customer, the customer feature data is cleaned and preprocessed to ensure the integrity and consistency of the data. Through data cleaning and preprocessing, standardized customer feature data is obtained. Feature extraction is performed based on the standardized customer feature data to extract features related to the customer churn recovery problem, so as to construct a customer feature vector suitable for the input model.

[0046] For example, by using the Pandas library in Python, missing values ​​are detected and processed, and duplicate data is deleted; then, data standardization is performed, such as Z-score standardization of numerical features such as age and consumption amount to eliminate the impact of dimension.

[0047] It should be emphasized that in order to further ensure the privacy and security of customer characteristic data, the above customer characteristic data can also be stored in a blockchain node.

[0048] The blockchain referred to in this application is a new application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, encryption algorithm, etc. Blockchain is essentially a decentralized database, a string of data blocks generated by cryptographic methods. Each data block contains a batch of network transaction information, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block. Blockchain can include the underlying blockchain platform, platform product service layer, and application service layer.

[0049] Step S202: input the customer feature vector into a trained prediction model, where the prediction model includes a first prediction sub-model and a second prediction sub-model.

[0050] The prediction model is a trained model used to predict the target user's success rate of recovery. Among them, the first prediction sub-model can use the LSTM model. LSTM (long short term memory network) is a time recursive neural network that solves the gradient vanishing and explosion problems of RNN and can effectively process long-term time series data. LSTM will record the last customer characteristics and affect subsequent model predictions. The second prediction sub-model can use the XGBoost model. XGBoost is an optimization of the Boosting algorithm, which integrates weak classifiers into a strong classifier. The XGBoost algorithm generates a new tree to fit the residual of the previous tree through continuous iteration. As the number of iterations increases, the accuracy continues to improve. Finally, a linear search method is used to determine an optimal weak classifier, namely the XGBoost model.

[0051] Step S203, processing the customer feature vector through the first prediction sub-model to obtain customer status information and a first prediction result, and inputting the customer status information into the second prediction sub-model.

[0052] Among them, the first prediction sub-model includes at least two LSTM layers, a fully connected layer and an output layer. The customer feature vector is input into the first prediction sub-model, and the customer feature vector is processed by at least two connected LSTM layers, the fully connected layer and the output layer to capture the timing characteristics and change trends of the target customers, and the hidden state representation of the customer features and the first prediction result are output. The hidden state representation of the customer features is the customer state information, and the first prediction result is the customer recovery success rate predicted by the first prediction sub-model according to the customer feature vector.

[0053] In some optional implementations of this embodiment, the step of processing the customer feature vector by the first prediction sub-model to obtain the customer status information and the first prediction result includes:

[0054] Through each LSTM layer, the customer feature vector is extracted in turn to obtain the feature hidden states of different scales;

[0055] The hidden states of features at different scales are input into the fully connected layer for concatenation to obtain customer status information;

[0056] The customer status information is input into the output layer for prediction to obtain the first prediction result.

[0057] In this embodiment, each LSTM layer processes information of different time scales to capture features of different time scales, and the input of the next LSTM layer is the output of the previous LSTM layer and the input customer feature vector. For example, the first prediction sub-model may include two LSTM layers, three LSTM layers, or four LSTM layers, which may be set according to actual needs.

[0058] By processing the customer feature vector through the first prediction sub-model, it is possible to capture the long-term dependencies between the input data, improve the efficiency and accuracy of time series feature extraction, and improve the prediction efficiency and accuracy.

[0059] In some optional implementations, different LSTM layers are connected through residual connections, which can help gradients flow more effectively in deep networks and prevent the gradient vanishing problem.

[0060] In some optional implementations of this embodiment, each LSTM layer includes a forget gate, an input gate, and an output gate. The steps of extracting features from the customer feature vector in turn through each LSTM layer to obtain feature hidden states of different scales include:

[0061] Based on the feature hidden state at the previous moment and the customer feature vector at the current moment, the forget gate output feature is calculated through the forget gate;

[0062] The forget gate output feature, the feature hidden state at the previous moment, and the customer feature vector at the current moment are calculated through the input gate to obtain the input gating signal and the candidate memory cell state;

[0063] Update the memory cell state according to the forget gate output feature, the input gating signal and the candidate memory cell state;

[0064] The output gate calculates the output gating signal according to the feature hidden state at the previous moment and the client feature vector at the current moment, and outputs the feature hidden state at the current moment based on the output gating signal and the memory cell state;

[0065] The feature hidden states at the current moment output by each LSTM layer are combined into feature hidden states of different scales.

[0066] Among them, the calculation formula of the forget gate is as follows:

[0067] f t =σ(W xf x t +W hf h t-1 +b f );

[0068] Among them, f t is the output feature of the forget gate; σ is the sigmoid function, W xf , W hf and b f are the weight and bias of the forget gate respectively; x t is the customer feature vector at the current moment, h t-1 is the feature hidden state at the previous moment.

[0069] The calculation formula of the input gate is as follows:

[0070] i t =σ(W xi x t +W hi h t-1 +b i );

[0071] Among them, i t is the input gating signal; W xi , W hi and b i are the weight and bias of the input gate respectively; x t is the customer feature vector at the current moment, h t-1 is the feature hidden state at the previous moment.

[0072] The calculation formula for the candidate memory cell state is as follows:

[0073]

[0074] Among them, tanh is the hyperbolic tangent function, and the range of tanh is [-1, 1]; W xc , W hc and b c They are the weight and bias for calculating the candidate memory cell state; x t is the customer feature vector at the current moment, h t-1 is the feature hidden state at the previous moment.

[0075] The calculation formula for updating the memory cell state is as follows:

[0076]

[0077] The feature hidden state formula at the current moment is as follows:

[0078]

[0079] u t = σ(W0[h t-1 , x t + b0);

[0080] where is the candidate feature hidden state at the current moment.

[0081] Concatenate the feature hidden states at the current moment output by each LSTM layer to obtain feature hidden states of different scales.

[0082] In this embodiment, the LSTM layer overcomes the problems of gradient vanishing and gradient explosion existing in model prediction through a mechanism constructed by using an input gate, a forget gate, and an output gate, and can capture long-term dependencies in the sequence, achieve long-term memory, and improve the accuracy of prediction.

[0083] Step S204, process the customer feature vector and the customer status information through a second prediction sub-model to obtain a second prediction result.

[0084] In this embodiment, the second prediction sub-model adopts an XGBoost model. Feature extraction is performed on the customer feature vector and the customer status information through the XGBoost model to obtain M basic features, where M is a positive integer; according to the stepwise regression algorithm, the M basic features are added to the XGBoost model for iterative calculation to obtain an iterative calculation result, where the iterative calculation result includes contribution scores corresponding to the basic features; from the iterative calculation result, the contribution scores are sorted in descending order, the top N contribution scores are selected, and the basic features corresponding to the top N contribution scores are used as N target features, where N < M and N is a positive integer; based on the target features and the corresponding contribution scores, a second prediction result is output.

[0085] Among them, the second prediction result is the customer retention success rate predicted by the XGBoost model based on the customer feature vector and the customer status information.

[0086] By combining the customer status information and the customer feature vector output by the first prediction sub-model through the second prediction sub-model for prediction, the operation speed of the prediction model can be effectively improved, and at the same time, the accuracy and stability of the prediction are improved.

[0087] Step S205, perform weighted summation on the first prediction result and the second prediction result to obtain the customer retention success rate.

[0088] In this embodiment, the first prediction result and the second prediction result are combined with different weights through the inverse error method to construct a prediction model, which can combine the advantages of each model and reduce the fluctuation of the randomness of the sample on the single model, so that the error of the entire combined model tends to decrease and the overall prediction accuracy is improved.

[0089] Specifically, the customer recovery success rate Q t The calculation formula is as follows:

[0090] Q t =W1f 1t +W2f 2t ;

[0091]

[0092] Wherein, t represents the time, t = 1, 2, 3, ...; f 1t represents the first prediction result; f 2t represents the second prediction result; W1 and W2 represent the weight coefficients of the first prediction sub-model and the second prediction sub-model respectively; ω1 and ω2 represent the errors of the first prediction sub-model and the second prediction sub-model respectively.

[0093] Step S206, inputting the customer status information and the customer recovery success rate into the trained Markov decision model for decision making, and obtaining the optimal recovery strategy for the target customer.

[0094] The Markov decision model is based on the definition of the Markov decision process, which can be described by the tuple (S, A, P, Q, r). S is a finite state set, A is a finite strategy set, P is the state transition probability, Q is the corresponding value return (reward function), and r is a discount factor between 0 and 1 to ensure that the farther the reward is from the current moment, the greater the discount.

[0095] In some optional implementations of this embodiment, before the step of inputting the customer status information and the customer recovery success rate into the trained Markov decision model for decision making, the following steps are included:

[0096] Consider customer recovery as a multi-stage decision-making problem and define a Markov decision process, which includes a finite state set, a finite strategy set, state transition probability and a reward function.

[0097] By defining the state space, the state space of the multi-stage decision problem is determined, and the finite state set of the customer at different stages is obtained;

[0098] According to the customer status characteristics, the action space is defined, and the selectable recovery strategies in each state are determined to obtain a finite set of strategies for customer recovery;

[0099] By analyzing customer status characteristics and recovery strategies, the state transition probability is defined and the state transition probability under the current state is obtained;

[0100] Define the reward function to obtain the reward values ​​for taking different recovery strategies in different states;

[0101] By setting the discount factor and using the strategy iteration algorithm, the Markov decision process established based on the finite state set, finite strategy set, state transition probability and reward function is solved to obtain the optimal customer recovery strategy and thus obtain a trained Markov decision model.

[0102] Among them, the state space is the set of all possible states, and the state space is obtained according to the customer state information output by the first prediction sub-model. Each state s corresponds to an action space a, that is, the recovery strategy.

[0103] The state transition probability is the probability of transitioning from state s to s' for each state s and action a after executing action a. The calculation formula is as follows:

[0104]

[0105] in, Indicates that given the current state S t =s, current recovery strategy A t =a, the probability of the state transitioning from s to s'.

[0106] The goal of reinforcement learning is to find a recovery strategy that maximizes the cumulative expected value return (reward value) under a given Markov process, which can be expressed as:

[0107]

[0108] Where k represents the number of iterations; Q * (s,a) means that based on state s, the recovery strategy a is adopted to maximize the expected sum of the reward values ​​of the next k iterations; π represents the customer recovery strategy, and when the expected sum is maximized, it is the optimal customer recovery strategy.

[0109] In this embodiment, the reward function is the customer recovery success rate output by the prediction model. For example, the customer recovery success rate output by the prediction model is introduced as the reward function, and the customer recovery strategy for day t+1 is adjusted on day t. Under the limited strategy set A (for example, customer contact time and customer contact method), the customer recovery success rate Q(a) corresponding to customer recovery strategy a can be predicted, and the customer recovery strategy corresponding to the maximum expected recovery success rate is the optimal customer recovery strategy recommended on day t+1.

[0110] Based on the trained prediction model, the corresponding customer recovery success rate can be predicted according to the customer feature vector. At the same time, by applying reinforcement learning theory, a Markov decision process is defined, and the prediction model is introduced to find the customer recovery success rate corresponding to the maximum value return (reward value). During the training process, the decision model can be continuously trained and optimized, and the theoretical optimal strategy can be continuously approached. The known optimal customer recovery operation is always executed, which greatly improves the customer recovery effect and reliability. In particular, after time accumulation, when the model training is very reliable, the execution strategy is optimal, and automatic intelligent customer recovery is fully realized.

[0111] In this embodiment, after obtaining the trained Markov decision model, the customer status information and the customer recovery success rate are input into the trained Markov decision model to obtain the optimal recovery strategy corresponding to the target customer.

[0112] This application uses a prediction model to make predictions based on customer feature vectors, which can effectively improve the prediction accuracy of customer recovery success rate. It also combines the Markov decision model to make decisions based on customer feature vectors and customer recovery success rate to obtain the optimal customer recovery strategy. It can more accurately locate customers who need to be recovered and provide them with accurate recovery strategies, making customer recovery more efficient, intelligent, and controllable. At the same time, it can promote business personnel to recover target customers and improve the work efficiency of business personnel.

[0113] In some optional implementations, before the step of inputting the customer feature vector into the trained prediction model, the method further includes:

[0114] Obtain customer characteristic factors in all business scenarios, and perform feature selection on customer characteristic factors to obtain key customer characteristics;

[0115] Perform feature engineering on key customer features to obtain normalized customer feature data, and randomly divide the normalized customer feature data into training sets and test sets;

[0116] Input the training set into the pre-built neural network models, including LSTM model and XGBoost model;

[0117] The LSTM model is trained using the training set to obtain the predicted time series features and the first predicted probability;

[0118] Input the predicted time series features and training set into the XGBoost model to obtain the second predicted probability;

[0119] The first predicted probability and the second predicted probability are weighted and summed to obtain the predicted recovery probability;

[0120] Based on the preset loss function, the loss value between the predicted recovery probability and the actual recovery probability is calculated;

[0121] Adjust the model parameters of the neural network model according to the loss value, continue iterative training until convergence, and obtain the model to be verified;

[0122] The test set is input into the model to be verified to obtain the verification result. When the verification result meets the preset conditions, the model to be verified is determined to be the final prediction model.

[0123] In this embodiment, the customer feature factors used for modeling are selected by comprehensive methods such as saturation, correlation, and PSI. Among them, saturation (Variance Inflation Factor, VIF) is an indicator to measure the multicollinearity between features. If there is a high correlation between features, that is, the information they provide has a large overlap, then they are considered saturated. The calculation formula of VIF is as follows:

[0124]

[0125] Where VIF j Indicates the saturation of the j-th customer characteristic factor; It represents the coefficient of determination (R-squared) of the multiple linear regression model when the j-th customer characteristic factor is used as the response variable and other characteristics are used as independent variables.

[0126] Correlation is to calculate the correlation coefficient between each customer feature factor, such as the Pearson correlation coefficient, etc., retain the features with higher correlation and remove the features with lower correlation. PSI (Population Stability Index) is used to evaluate the stability of customer feature factors in different sample subsets. The lower the PSI value, the more stable it is. Divide the data set into multiple sample subsets, calculate the importance score of each customer feature factor in different subsets, and then calculate the PSI value. The calculation formula is as follows:

[0127]

[0128] Among them, Importance i is the feature importance score in the i-th subset, μ is the average importance score; n is the number of subsets.

[0129] Perform feature engineering on the key customer features obtained by feature selection, including over-normalization, binning, one-hot / target coding conversion, etc. Among them, normalization is to scale the data to a specific range, usually 0 to 1, or convert it to a distribution with a specific mean and standard deviation. Common normalization methods include minimum-maximum normalization and Z-score normalization. Binning is to divide continuous numerical data into several intervals (boxes) and represent the values ​​in each interval with the mean or median of the box, which can reduce the impact of outliers and may improve the generalization ability of the model.

[0130] The normalized customer feature data obtained by feature engineering is randomly divided into training set and test set according to a preset ratio, such as training set: test set = 8:2.

[0131] In this embodiment, the preset loss function uses Mean Absolute Percentage Error (MAPE) and Root Mean Square Error (RMSE). MAPE is the average value of the percentage difference between the predicted value (predicted probability of recovery) and the actual value (actual probability of recovery). It is usually used to measure the accuracy of the prediction. The calculation formula is as follows:

[0132]

[0133] in, and t are the predicted probability of recovery and the actual probability of recovery at time t respectively; N is the total number of predictions.

[0134] RMSE is the square root of the average of the squares of the deviations between the observed value (predicted probability of recovery) and the true value (actual probability of recovery). It is a standard measure of prediction error and is calculated as follows:

[0135]

[0136] in, and t are the predicted probability of recovery and the actual probability of recovery at time t respectively; N is the total number of predictions.

[0137] The loss value is calculated based on the preset loss function, the model parameters are adjusted according to the loss value, and the iterative training continues until the model converges. The convergence condition can be that the loss value does not change significantly, or the number of iterations reaches the preset number.

[0138] The test set is input into the model to be verified to obtain the verification prediction probability. The prediction accuracy of the model is calculated based on the verification prediction probability and the actual recovery probability, and the prediction accuracy is used as the verification result. When the prediction accuracy is greater than or equal to the preset threshold, the model to be verified is output as the final prediction model; if the prediction accuracy is less than the preset threshold, it means that the prediction accuracy of the model is not high, and it is necessary to increase the number of samples, or modify the model parameters and retrain to improve the prediction accuracy.

[0139] This application can effectively improve the quality of data preparation by performing feature selection and feature engineering on the acquired customer characteristic factors and then training the model, which can increase the robustness of model learning and predicting the success rate of customer recovery, enhance the model's anti-interference ability and generalization ability, and thus improve prediction accuracy.

[0140] In some optional implementations of this embodiment, the step of inputting the predicted time series features and the training set into the XGBoost model to obtain the second prediction probability includes:

[0141] Initialize the model and set the basic parameters of the XGBoost model, including at least the learning rate, the maximum depth of the tree, and the objective function;

[0142] The first tree is constructed using a greedy algorithm. Each node of the tree is split according to the importance of the input prediction time series features and the normalized customer feature data in the training set to maximize the objective function of information gain and output the first prediction value.

[0143] A second tree is constructed based on the predicted time series features, the training set, and the new target variable, and the second predicted value is output; the new target variable is the residual of the first tree, that is, the difference between the true value and the first predicted value;

[0144] Iterate to build the remaining trees in the XGBoost model, using the residual of the previous tree as the new target variable in each iteration to build the next tree, until the last tree, i.e. the kth tree, gets the kth prediction value;

[0145] The first predicted value to the kth predicted value are weighted summed to obtain the second predicted probability of the XGBoost model.

[0146] The objective function of the XGBoost model is expressed as follows:

[0147]

[0148]

[0149] in, represents the loss function; and tare the predicted probability of recovery and the actual probability of recovery at time t; Ω(f k ) represents the regularization term; f k represents the kth tree; T represents the number of leaves in the tree; λ represents the regularization parameter; ω is the leaf weight.

[0150] During the training process of the XGBoost model, a new tree is constructed and added to the XGBoost model at each iteration so that the constructed tree can minimize the objective function of the XGBoost model. In the i-th iteration, let f i (x t ) is the tree generated by the t-th sample in the i-th iteration, then the objective function is:

[0151]

[0152] In the formula, It is the predicted recovery probability of the target time series (including the predicted time series features and the training set) at time t in the i-th iteration.

[0153] By training the XGBoost model, the characteristics of customer feature factors are fully utilized, key information related to customer recovery is effectively extracted, the characteristics of customer recovery are captured more comprehensively, and the accuracy and reliability of predictions are improved.

[0154] The present application can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc. The present application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present application can also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules can be located in local and remote computer storage media including storage devices.

[0155] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0156] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0157] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a device for recovering lost customers. Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0158] like Figure 3 As shown, the device 300 for recovering lost customers in this embodiment includes: an acquisition module 301, an input module 302, a first prediction module 303, a second prediction module 304, a weighting module 305 and a decision module 306. Among them:

[0159] The acquisition module 301 is used to acquire customer feature data of target customers and construct a customer feature vector according to the customer feature data;

[0160] The input module 302 is used to input the customer feature vector into a trained prediction model, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model;

[0161] The first prediction module 303 is used to process the customer feature data through the first prediction sub-model to obtain customer status information and a first prediction result, and input the customer status information into the second prediction sub-model;

[0162] The second prediction module 304 is used to process the customer feature vector and the customer status information through the second prediction sub-model to obtain a second prediction result;

[0163] The weighting module 305 is used to perform weighted summation on the first prediction result and the second prediction result to obtain a customer recovery success rate;

[0164] The decision module 306 is used to input the customer status information and the customer recovery success rate into the trained Markov decision model to make a decision and obtain the optimal recovery strategy for the target customer.

[0165] It should be emphasized that in order to further ensure the privacy and security of customer characteristic data, the above customer characteristic data can also be stored in a blockchain node.

[0166] Based on the above-mentioned customer churn recovery device 300, predictions are made based on customer feature vectors through a prediction model, which can effectively improve the prediction accuracy of the customer recovery success rate. In combination with the Markov decision model, decisions are made based on customer feature vectors and customer recovery success rates to obtain the optimal customer recovery strategy. Customers who need to be recovered can be located more accurately and accurate recovery strategies can be provided for them, making customer recovery more efficient, intelligent, and controllable. At the same time, it can promote business personnel to recover target customers and improve the work efficiency of business personnel.

[0167] In some optional implementations, the first prediction sub-model includes at least two LSTM layers, a fully connected layer, and an output layer; the first prediction module 303 includes a feature extraction sub-module, a splicing sub-module, and a prediction sub-module, wherein:

[0168] The feature extraction submodule is used to extract features from the customer feature vector in turn through each LSTM layer to obtain feature hidden states of different scales;

[0169] The splicing submodule is used to input the feature hidden states of different scales into the fully connected layer for splicing to obtain customer state information;

[0170] The prediction submodule is used to input the customer status information into the output layer for prediction to obtain a first prediction result.

[0171] By processing the customer feature vector through the first prediction sub-model, it is possible to capture the long-term dependencies between the input data, improve the efficiency and accuracy of time series feature extraction, and improve the prediction efficiency and accuracy.

[0172] In some optional implementations of this embodiment, each LSTM layer includes a forget gate, an input gate, and an output gate; and the feature extraction submodule is further used for:

[0173] Based on the feature hidden state at the previous moment and the customer feature vector at the current moment, calculate the forget gate output feature through the forget gate;

[0174] Calculate the input gate control signal and the candidate memory cell state by calculating the output feature of the forget gate, the feature hidden state at the previous moment, and the customer feature vector at the current moment through the input gate;

[0175] Update the memory cell state according to the output feature of the forget gate, the input gate control signal, and the candidate memory cell state;

[0176] Calculate the output gate control signal through the output gate according to the feature hidden state at the previous moment and the customer feature vector at the current moment, and output the feature hidden state at the current moment based on the output gate control signal and the memory cell state;

[0177] Concatenate the feature hidden states at the current moment output by each LSTM layer to obtain feature hidden states of different scales.

[0178] In this embodiment, the mechanism constructed by the LSTM layer using the input gate, forget gate, and output gate overcomes the problems of gradient disappearance and gradient explosion existing in model prediction, and can capture long-term dependencies in the sequence, realize long-term memory, and improve the prediction accuracy.

[0179] In some optional implementation manners, the second prediction sub-model adopts an XGBoost model, and the second prediction module 304 is further configured to:

[0180] Extract features from the customer feature vector and the customer state information through the XGBoost model to obtain M basic features, where M is a positive integer;

[0181] According to the stepwise regression algorithm, add the M basic features to the XGBoost model for iterative calculation to obtain an iterative calculation result, where the iterative calculation result includes the contribution scores corresponding to the basic features;

[0182] From the iterative calculation result, arrange the contribution scores in descending order, select the top N contribution scores, and use the basic features corresponding to the top N contribution scores as the N target features, where N < M and N is a positive integer;

[0183] Output a second prediction result based on the target features and the corresponding contribution scores.

[0184] By combining the customer status information and customer feature vector output by the first prediction sub-model for prediction through the second prediction sub-model, the calculation speed of the prediction model can be effectively improved, while improving the accuracy and stability of the prediction.

[0185] In some optional implementations, the device 300 for recovering lost customers further includes defining a decision module, which is used to:

[0186] Considering customer recovery as a multi-stage decision-making problem, a Markov decision process is defined, which includes a finite state set, a finite strategy set, a state transition probability and a reward function;

[0187] By defining a state space, determining the state space of the multi-stage decision problem, and obtaining a finite state set of the customer at different stages;

[0188] According to the customer status characteristics, the action space is defined, and the selectable recovery strategies in each state are determined to obtain a finite set of strategies for customer recovery;

[0189] By analyzing customer status characteristics and recovery strategies, the state transition probability is defined and the state transition probability under the current state is obtained;

[0190] Define the reward function to obtain the reward values ​​for taking different recovery strategies in different states;

[0191] By setting a discount factor and using a policy iteration algorithm, the Markov decision process established according to the finite state set, the finite strategy set, the state transition probability and the reward function is solved to obtain the optimal customer recovery strategy, so as to obtain a trained Markov decision model.

[0192] By defining the Markov decision process and making decisions based on customer feature vectors and customer recovery success rates to obtain the optimal customer recovery strategy, we can more accurately locate customers who need to be recovered and provide them with precise recovery strategies, making customer recovery more efficient, intelligent and controllable.

[0193] In some optional implementations, the device 300 for recovering lost customers further includes a training module, including:

[0194] The feature selection submodule is used to obtain customer feature factors in all business scenarios and perform feature selection on the customer feature factors to obtain key customer features;

[0195] A feature engineering submodule, used to perform feature engineering on the key customer features to obtain normalized customer feature data, and randomly divide the normalized customer feature data into a training set and a test set;

[0196] An input submodule, used to input the training set into a pre-built neural network model, wherein the neural network model includes an LSTM model and an XGBoost model;

[0197] A first training submodule, used to train the LSTM model using the training set to obtain a predicted time series feature and a first predicted probability;

[0198] A second training submodule, used for inputting the predicted time series features and the training set into the XGBoost model to obtain a second prediction probability;

[0199] A weighted summation submodule, configured to perform a weighted summation of the first predicted probability and the second predicted probability to obtain a predicted probability of recovery;

[0200] A calculation submodule, used for calculating the loss value between the predicted probability of recovery and the actual probability of recovery based on a preset loss function;

[0201] An iterative submodule, used for adjusting the model parameters of the neural network model according to the loss value, continuing the iterative training until convergence, and obtaining a model to be verified;

[0202] The verification submodule is used to input the test set into the model to be verified to obtain a verification result, and when the verification result meets a preset condition, determine that the model to be verified is the final prediction model.

[0203] By performing feature selection and feature engineering on the acquired customer characteristic factors and then training the model, the quality of data preparation can be effectively improved, the robustness of model learning and predicting customer recovery success rate can be increased, and the model's anti-interference and generalization capabilities can be enhanced, thereby improving prediction accuracy.

[0204] In this embodiment, the second training submodule is further used for:

[0205] Initialize the model and set the basic parameters of the XGBoost model, including at least the learning rate, the maximum depth of the tree, and the objective function;

[0206] A greedy algorithm is used to construct a first tree, and each node of the tree is split according to the importance of the input prediction time series features and the normalized customer feature data in the training set to maximize the objective function of information gain and output a first prediction value;

[0207] Building a second tree based on the predicted time series features, the training set and a new target variable, and outputting a second predicted value; the new target variable is the residual of the first tree, that is, the difference between the true value and the first predicted value;

[0208] Iteratively build the remaining trees in the XGBoost model, using the residual of the previous tree as the new target variable in each iteration to build the next tree, until the last tree, i.e., the kth tree, obtains the kth prediction value;

[0209] The first predicted value to the kth predicted value are weightedly summed to obtain the second predicted probability of the XGBoost model.

[0210] By training the XGBoost model, the characteristics of customer feature factors are fully utilized, key information related to customer recovery is effectively extracted, the characteristics of customer recovery are captured more comprehensively, and the accuracy and reliability of predictions are improved.

[0211] To solve the above technical problems, the present application also provides a computer device. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0212] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected through a system bus. It should be noted that the figure only shows a computer device 4 with components 41-43, but it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to microprocessors, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable gate arrays (Field-Programmable Gate Array, FPGA), digital processors (Digital Signal Processor, DSP), embedded devices, etc.

[0213] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0214] The memory 41 includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 41 can be an internal storage unit of the computer device 4, such as a hard disk or memory of the computer device 4. In other embodiments, the memory 41 can also be an external storage device of the computer device 4, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (FlashCard), etc. equipped on the computer device 4. Of course, the memory 41 can also include both the internal storage unit of the computer device 4 and its external storage device. In this embodiment, the memory 41 is generally used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions of a method for recovering customer churn, etc. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or are to be output.

[0215] The processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 42 is generally used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to run computer-readable instructions stored in the memory 41 or process data, such as computer-readable instructions for running the customer churn recovery method.

[0216] The network interface 43 may include a wireless network interface or a wired network interface. The network interface 43 is generally used to establish a communication connection between the computer device 4 and other electronic devices.

[0217] This embodiment implements the steps of the customer churn recovery method as in the above embodiment when the processor executes computer-readable instructions stored in the memory. The prediction model is used to make predictions based on the customer feature vector, which can effectively improve the prediction accuracy of the customer recovery success rate. In combination with the Markov decision model, decisions are made based on the customer feature vector and the customer recovery success rate to obtain the optimal customer recovery strategy. It can more accurately locate the customers who need to be recovered and provide them with accurate recovery strategies, making customer recovery more efficient, intelligent, and controllable. At the same time, it can promote business personnel to recover target customers and improve the work efficiency of business personnel.

[0218] The present application also provides another implementation mode, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by at least one processor so that the at least one processor executes the steps of the customer churn recovery method as described above, and the prediction accuracy of the customer recovery success rate can be effectively improved by predicting according to the customer feature vector through the prediction model, and making decisions according to the customer feature vector and the customer recovery success rate in combination with the Markov decision model to obtain the optimal customer recovery strategy, which can more accurately locate the customers who need to be recovered and provide them with accurate recovery strategies, making customer recovery more efficient, intelligent and controllable, and at the same time, can promote business personnel to recover target customers and improve the work efficiency of business personnel.

[0219] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0220] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for recovering lost customers, characterized in that: The steps include: Acquire customer feature data of target customers, and construct a customer feature vector according to the customer feature data; Inputting the customer feature vector into a trained prediction model, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model; Processing the customer feature data through the first prediction sub-model to obtain customer status information and a first prediction result, and inputting the customer status information into the second prediction sub-model; Processing the customer feature vector and the customer status information by the second prediction sub-model to obtain a second prediction result; Taking a weighted sum of the first prediction result and the second prediction result to obtain a customer recovery success rate; The customer status information and the customer recovery success rate are input into a trained Markov decision model to make a decision, thereby obtaining an optimal recovery strategy for the target customer.

2. The method for recovering lost customers according to claim 1, characterized in that: The first prediction sub-model includes at least two LSTM layers, a fully connected layer and an output layer; the step of processing the customer feature vector by the first prediction sub-model to obtain customer status information and a first prediction result includes: Extracting features from the customer feature vector in turn through each LSTM layer to obtain feature hidden states of different scales; Inputting the feature hidden states of different scales into the fully connected layer for splicing to obtain customer status information; The customer status information is input into the output layer for prediction to obtain a first prediction result.

3. The method for recovering lost customers according to claim 2, characterized in that: Each LSTM layer includes a forget gate, an input gate, and an output gate; the step of extracting features from the customer feature vector in turn through each LSTM layer to obtain feature hidden states of different scales includes: Based on the feature hidden state at the previous moment and the customer feature vector at the current moment, the forget gate output feature is calculated through the forget gate; The forget gate output feature, the feature hidden state at the previous moment, and the customer feature vector at the current moment are calculated through the input gate to obtain an input gating signal and a candidate memory cell state; Update the memory cell state according to the forget gate output feature, the input gating signal and the candidate memory cell state; Calculating an output gating signal through the output gate according to the feature hidden state at the previous moment and the client feature vector at the current moment, and outputting the feature hidden state at the current moment based on the output gating signal and the memory cell state; The feature hidden states at the current moment output by each LSTM layer are concatenated to obtain feature hidden states of different scales.

4. The method for recovering lost customers according to claim 1, characterized in that: The second prediction sub-model adopts an XGBoost model; the step of processing the customer feature vector and the customer status information by the second prediction sub-model to obtain a second prediction result comprises: Performing feature extraction on the customer feature vector and the customer status information through the XGBoost model to obtain M basic features, where M is a positive integer; According to the stepwise regression algorithm, M of the basic features are added to the XGBoost model for iterative calculation to obtain an iterative calculation result, where the iterative calculation result includes contribution scores corresponding to the basic features; From the iterative calculation result, the contribution scores are sorted in descending order, the top N contribution scores are selected, and the basic features corresponding to the top N contribution scores are used as N target features, where N < M and N is a positive integer; Based on the target features and the corresponding contribution scores, a second prediction result is output.

5. The method for recovering lost customers according to claim 1, characterized in that: Before the step of inputting the customer status information and the customer retention success rate into the trained Markov decision model for decision-making, it includes: Regarding customer retention as a multi-stage decision-making problem, a Markov decision process is defined, and the Markov decision process includes a finite state set, a finite policy set, a state transition probability, and a reward function; By defining the state space, the state space of the multi-stage decision-making problem is determined to obtain a finite state set of the customer at different stages; According to the customer status characteristics, an action space is defined to determine the retention strategies that can be selected in each state, and a finite policy set for customer retention is obtained; By analyzing the customer status characteristics and the retention strategies, the state transition probability is defined to obtain the state transition probability in the current state; Define a reward function to obtain the reward values for taking different retention strategies in different states; By setting a discount factor and using the policy iteration algorithm, the established Markov decision process is solved according to the finite state set, the finite policy set, the state transition probability, and the reward function to obtain the optimal customer retention strategy, so as to obtain a trained Markov decision model.

6. The method for recovering lost customers according to claim 1, characterized in that: Before the step of inputting the customer feature vector into the trained prediction model, it further includes: Obtain customer feature factors in all business scenarios, and perform feature selection on the customer feature factors to obtain key customer features; Perform feature engineering on the key customer features to obtain normalized customer feature data, and randomly divide the normalized customer feature data into a training set and a test set; Input the training set into a pre-constructed neural network model, and the neural network model includes an LSTM model and an XGBoost model; Train the LSTM model with the training set to obtain predicted time series features and a first prediction probability; Input the predicted time series features and the training set into the XGBoost model to obtain a second prediction probability; Perform weighted summation on the first prediction probability and the second prediction probability to obtain a predicted retention probability; Based on a preset loss function, calculate the loss value between the predicted retention probability and the actual retention probability; Adjust the model parameters of the neural network model according to the loss value, and continue iterative training until convergence to obtain a model to be verified; Input the test set into the model to be verified to obtain a verification result. When the verification result meets the preset conditions, determine the model to be verified as the final prediction model.

7. The method for recovering lost customers according to claim 6, characterized in that: The step of inputting the predicted time series feature and the training set into the XGBoost model to obtain a second predicted probability comprises: Initialize the model and set the basic parameters of the XGBoost model, including at least the learning rate, the maximum depth of the tree, and the objective function; A greedy algorithm is used to construct a first tree, and each node of the tree is split according to the importance of the input prediction time series features and the normalized customer feature data in the training set to maximize the objective function of information gain and output a first prediction value; Building a second tree based on the predicted time series features, the training set and the new target variable, and outputting a second predicted value; the new target variable is the residual of the first tree, that is, the difference between the true value and the first predicted value; Iteratively build the remaining trees in the XGBoost model, using the residual of the previous tree as the new target variable in each iteration to build the next tree, until the last tree, i.e., the kth tree, obtains the kth prediction value; The first predicted value to the kth predicted value are weightedly summed to obtain the second predicted probability of the XGBoost model.

8. A device for recovering lost customers, characterized in that: include: An acquisition module, used to acquire customer feature data of target customers and construct a customer feature vector according to the customer feature data; An input module, used for inputting the customer feature vector into a trained prediction model, wherein the prediction model includes a first prediction sub-model and a second prediction sub-model; A first prediction module, configured to process the customer feature data through the first prediction sub-model to obtain customer status information and a first prediction result, and input the customer status information into the second prediction sub-model; A second prediction module, configured to process the customer feature vector and the customer status information through the second prediction sub-model to obtain a second prediction result; A weighting module, used for performing a weighted summation of the first prediction result and the second prediction result to obtain a customer recovery success rate; The decision module is used to input the customer status information and the customer recovery success rate into a trained Markov decision model to make a decision and obtain the optimal recovery strategy for the target customer.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the method for recovering customer churn as claimed in any one of claims 1 to 7 when executing the computer-readable instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the method for recovering customer churn according to any one of claims 1 to 7.

Images