Customer value prediction and stratification management method and system, and storage medium

Abstract

This invention relates to a customer value prediction and tiered management method, system, and storage medium. The method includes: constructing a comprehensive feature vector based on raw customer data; inputting the comprehensive feature vector into a customer value prediction engine to perform a hybrid prediction using an XGBoost model and an LSTM model to determine the customer's lifetime value; determining the customer tiering result based on the lifetime value; determining the customer's churn risk score, membership level transition probability, and transition time window based on the raw data; and generating customer intervention strategy identifiers using a decision tree algorithm based on the tiering result, churn risk score, membership level transition probability, and transition time window. This invention enables refined processing and intelligent analysis of customer data, improving the targeting and effectiveness of customer tiered management.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a customer value prediction and tiered management method, system, and storage medium. Background Technology

[0002] In data processing systems, customer value assessment and tiered management are important application areas of machine learning technology. By analyzing and processing customer behavior data, the system can output quantitative assessment results of customer value and classify customer data accordingly. Customer Lifetime Value (CLV), as a predictive metric, is widely used in customer value assessment systems.

[0003] Existing technologies employ statistically-based feature construction methods, primarily relying on statistics such as customer spending amounts and frequencies to construct features. This approach suffers from the problem of ineffectively incorporating behavioral data from other dimensions into the feature system, resulting in limited information dimensions carried by the constructed feature vectors. Furthermore, the use of a single-model structure for uniformly processing feature data presents challenges in obtaining differentiated processing paths for different types of data within the model, impacting the model's ability to extract data features. Using analysis methods based on overall distribution characteristics to process customer behavior data results in weak detection capabilities for localized abnormal fluctuations in the data sequence, making it difficult to promptly identify abnormal change patterns. Employing state determination methods based on preset thresholds to process customer level-related data results in weak predictive capabilities for future state transition trends. Finally, using data processing methods based on single-dimensional results to generate output results presents the challenge of effectively integrating analysis results from multiple dimensions, affecting the comprehensiveness and accuracy of the output.

[0004] Therefore, how to construct a technical solution that can integrate multi-dimensional behavioral data, realize dynamic prediction of customer value, real-time detection of abnormal behavior, and quantitative analysis of grade change trends is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, it is necessary to provide a customer value prediction and tiered management method, system and storage medium to solve the technical problems of existing technologies such as single evaluation dimensions, insufficient prediction capabilities, lagging risk management and rigid membership management.

[0006] To address the aforementioned problems, in a first aspect, the present invention provides a customer value prediction and tiered management method, comprising: Based on the customer's original data, a comprehensive feature vector is constructed, which includes historical behavioral features and time-series features; The comprehensive feature vector is input into the customer value prediction engine to perform hybrid prediction using the XGBoost model and LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, while the LSTM model processes the time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor. The customer segmentation results are determined based on the stated lifetime value; Based on the raw data, the customer's churn risk score, membership level upgrade probability, and upgrade time window are determined. Based on the stratification results, the churn risk score, the membership level transition probability and transition time window, the intervention strategy identifier for the customer is generated using a decision tree algorithm.

[0007] In one possible implementation, determining the customer's lifetime value based on the base value score and the time-series adjustment factor includes: The Lifetime Value (CLV) is calculated using the following formula: CLV = V_base×(1+α)×β Where V_base is the base value score, α is the time series adjustment factor, β is the time decay factor, and β = exp(-λt), where λ is the time decay parameter and t is the prediction time span.

[0008] In one possible implementation, determining the customer's churn risk score based on the raw data includes: Based on the original data, determine the current monthly consumption, historical average consumption, current monthly active days, historical average active days, current quarterly consumption growth rate, historical average growth rate, and number of days since the last interaction; The consumption decline rate is determined based on the current monthly consumption and the historical average consumption. The consumption decline rate is calculated by dividing the difference between the current monthly consumption and the historical average consumption by the historical average consumption. The activity decay rate is determined based on the current monthly active days and the historical average active days. The activity decay rate is calculated by dividing the difference between the current monthly active days and the historical average active days by the historical average active days. The consumption growth rate is determined based on the current quarterly consumption growth rate and the historical average growth rate. The consumption growth rate is calculated by dividing the difference between the current quarterly consumption growth rate and the historical average growth rate by the historical average growth rate. The consumption decline rate, the activity decay rate, and the consumption growth rate change rate are weighted and summed, and the weighted sum is mapped by the sigmoid function to obtain the basic risk score; Based on the number of days since the last interaction, a time decay adjustment is performed on the basic risk score, and the time decay adjustment is implemented according to an exponential decay function; The churn risk score is generated by multiplying the base risk score (adjusted for time decay) by the customer type weighting coefficient.

[0009] In one possible implementation, determining the probability of the customer's membership level advancement based on the original data includes: Based on the raw data, determine the customer's current level code, consumption change trend index, consumption accumulation speed, activity change trend index, and historical upgrade count; Construct a feature vector X, which includes at least the current level code, the consumption change trend index, the consumption accumulation speed, the activity change trend index, and the historical upgrade count, respectively serving as the first feature X1, the second feature X2, the third feature X3, the fourth feature X4, and the fifth feature X5. Input the feature vector X into the logistic regression model, and calculate the upgrade probability P_u according to the following formula: P_u = 1 / (1 + exp(-(β0 + β1×X1 + β2×X2 + β3×X3 + β4×X4 + β5×X5 +... + βn×Xn))), where n ≥ 5, X1 to X5 are the first to fifth features, X6 to Xn are the other features in the feature vector X besides the first to fifth features, and β0, β1, β2, ..., βn are the corresponding model parameters; Input the feature vector X into the logistic regression model, and calculate the downgrade probability P_d according to the following formula: P_d = 1 / (1 + exp(-(γ0 + γ1×X1 + γ2×X2 + γ3×X3 + γ4×X4 + γ5×X5 + ... + γn×Xn))), where γ0, γ1, γ2, ..., γn are the corresponding model parameters.

[0010] In one possible implementation, determining the customer's transition time window based on the original data includes: Based on the raw data, determine the customer's current behavioral indicators, the level threshold corresponding to the membership level jump, and the customer's historical average rate of change. The rate of change is calculated by dividing the absolute value of the difference between the current behavior indicator and the level threshold by the historical average rate of change. The transition time window is predicted by multiplying the rate of change by a preset adjustment coefficient.

[0011] In one possible implementation, the intervention strategy identifier includes at least a first type of intervention strategy identifier, a second type of intervention strategy identifier, and a third type of intervention strategy identifier; the membership level jump probability includes an upgrade probability and a downgrade probability; the step of generating the customer's intervention strategy identifier using a decision tree algorithm based on the stratification result, the churn risk score, the membership level jump probability, and the jump time window includes: The root node of the decision tree is constructed based on the customer's current level in the hierarchical results; Based on the customer's current level represented by the root node, the upgrade probability, the downgrade probability, and the transition time window, the branching conditions of the decision tree are determined; When the upgrade probability is greater than or equal to the first probability threshold and the transition time window is less than or equal to the first time threshold, a first type of intervention strategy identifier is generated based on the churn risk score. When the upgrade probability is between the second probability threshold and the first probability threshold and the transition time window is between the second time threshold and the first time threshold, a second type of intervention strategy identifier is generated based on the churn risk score; When the downgrade probability is greater than or equal to the third probability threshold, a third type of intervention strategy identifier is generated based on the churn risk score.

[0012] In one possible implementation, after determining the customer's lifetime value and before determining the customer's segmentation result based on the lifetime value, the method further includes: Calculate the correlation between the prediction results of the XGBoost model and the LSTM model, and use the absolute value of the correlation as the model consistency score; A historical accuracy score is obtained based on the historical prediction accuracy of a customer group similar to the current customer. A data integrity score is obtained based on the current level of customer data integrity. The model consistency score, the historical accuracy score, and the data integrity score are multiplied by preset weights and then summed to obtain the overall confidence score, wherein the sum of each preset weight is 1.

[0013] In one possible implementation, the step of constructing the comprehensive feature vector includes: Based on the customer's group type, the weight coefficients of each dimension feature are dynamically adjusted using a multilayer perceptron. The historical behavior features and the temporal features are weighted and fused based on the adjusted weights to obtain the comprehensive feature vector.

[0014] Secondly, the present invention also provides a customer value prediction and tiered management system, comprising: A construction unit is used to construct a comprehensive feature vector based on the customer's original data, wherein the comprehensive feature vector includes historical behavioral features and time-series features; The prediction unit is used to input the comprehensive feature vector into the customer value prediction engine to perform a hybrid prediction using the XGBoost model and the LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, while the LSTM model processes the time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor. The first determining unit is used to determine the customer segmentation result based on the customer's lifetime value; The second determining unit is used to determine the customer's churn risk score, membership level upgrade probability, and upgrade time window based on the original data. The generation unit is used to generate the customer's intervention strategy identifier based on the hierarchical results, the churn risk score, the membership level jump probability and the jump time window, using a decision tree algorithm.

[0015] Thirdly, the present invention also provides a computer-readable storage medium for storing a computer-readable program or instruction, which, when executed by a processor, can implement the steps of the customer value prediction and hierarchical management method described in any of the above implementations.

[0016] The beneficial effects of this invention are: The customer value prediction and tiered management method provided by this invention constructs a comprehensive feature vector based on the customer's original data. This comprehensive feature vector includes historical behavioral features and time-series features, enhancing the information representation capability of the feature vector. This allows the constructed feature vector to more comprehensively reflect customer behavior patterns, ensuring that subsequent prediction models can perform value assessments based on richer information, thereby improving the accuracy of customer lifetime value prediction. The comprehensive feature vector is input into a customer value prediction engine for hybrid prediction using an XGBoost model and an LSTM model. The XGBoost model processes historical behavioral features and outputs a base value score, while the LSTM model processes time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor, improving the utilization efficiency of feature information and increasing... The prediction accuracy of Customer Lifetime Value (CMW) is improved, and the responsiveness of the prediction results to dynamic changes in customer behavior is enhanced. Customer segmentation based on CMW facilitates subsequent rule-based processing. Based on raw data, customer churn risk scores, membership level transition probabilities, and transition time windows are determined, ensuring that subsequent multi-dimensional integrated decisions are based on more comprehensive and refined input information, thereby improving the targeting and effectiveness of subsequent segmentation management strategies. Based on segmentation results, churn risk scores, membership level transition probabilities, and transition time windows, a decision tree algorithm is used to generate customer intervention strategy identifiers. This enables comprehensive judgment of customer status and differentiated strategy generation, allowing the final generated strategy identifiers to more accurately match the current multi-dimensional status of customers. This achieves refined processing and intelligent analysis of customer data, improving the targeting and effectiveness of customer segmentation management. Attached Figure Description

[0017] Figure 1 A schematic flowchart of an embodiment of the customer value prediction and tiered management method provided by the present invention; Figure 2 Provided by the present invention Figure 1 A schematic diagram of an embodiment of S104; Figure 3 Provided by the present invention Figure 1 Another embodiment of the process diagram of S104; Figure 4 This is a schematic diagram of the customer value prediction and hierarchical management system provided by the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] In the description of the embodiments of the present invention, unless otherwise stated, "a plurality of" means two or more.

[0020] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] This invention provides a customer value prediction and tiered management method, system, and storage medium, which are described below.

[0023] The entity executing the customer value prediction and tiered management method in this application embodiment can be the customer value prediction and tiered management system provided in this application embodiment, or different types of electronic devices such as server equipment, physical host, or user equipment (UE) that integrate the customer value prediction and tiered management system. The customer value prediction and tiered management system can be implemented in hardware or software. The UE can specifically be a terminal device such as a smartphone, tablet computer, laptop computer, handheld computer, desktop computer, or personal digital assistant (PDA).

[0024] The customer value prediction and tiered management method provided in this application can be applied to data processing systems in industries with membership systems, such as aviation, hotels, retail, and finance. Specifically, it can be used for multi-dimensional feature extraction from customer behavior data, value prediction model construction, risk score calculation, tier migration probability analysis, and personalized service strategy generation. By fusing historical and time-series customer behavior data, and combining the predictive capabilities of a hybrid XGBoost and LSTM model, it can output multi-dimensional analysis results such as customer lifetime value, churn risk score, tier migration probability, and time window, providing data support for subsequent customer tiered management and differentiated services.

[0025] Figure 1 This is a schematic flowchart of an embodiment of the customer value prediction and tiered management method provided by the present invention, as shown below. Figure 1 As shown, customer value forecasting and tiered management methods include: S101. Based on the customer's original data, construct a comprehensive feature vector, wherein the comprehensive feature vector includes historical behavioral features and time-series features.

[0026] The raw data may include, but is not limited to, basic customer attribute information, historical behavior records, consumption records, and time-series behavior data. This data may originate from the database or data warehouse of the enterprise customer relationship management system.

[0027] Historical behavioral characteristics reflect customers' long-term accumulated behavioral patterns. Taking air passengers as an example, historical behavioral characteristics include static statistics such as historical cumulative spending amount, average spending frequency, and accumulated points or miles.

[0028] Temporal features reflect the dynamic patterns of customer behavior over time, such as the rate of change in monthly spending and the fluctuation trend of quarterly active days. A time window sliding mechanism can be introduced to capture these temporal patterns. For example, extract the customer's monthly spending change rate sequence for the past 6 months [+2%, +3%, +5%, +4%, -1%, -2%], and the monthly active days trend for the past 6 months [8 days, 9 days, 10 days, 9 days, 7 days, 6 days]. Multiple temporal features constitute the input for characterizing the dynamic evolution of customer behavior.

[0029] Specifically, the raw data can be preprocessed, including data cleaning, missing value imputation, and outlier removal. Then, two types of features are extracted from the preprocessed data: historical behavioral features and time-series features. These features are combined to construct a multi-dimensional comprehensive feature vector. This comprehensive feature vector contains both long-term accumulated information and recent behavioral trend information about the customer, providing rich input data for subsequent predictive models.

[0030] Understandably, by constructing a comprehensive feature vector that includes temporal features, the problem of single feature dimensions and inability to represent dynamic changes in customer behavior in existing technologies is solved. This enhances the information representation capability of the feature vector, enabling the constructed feature vector to more comprehensively reflect customer behavior patterns. This ensures that subsequent prediction models can conduct value assessments based on richer information, thereby improving the accuracy of customer lifetime value prediction.

[0031] S102. Input the comprehensive feature vector into the customer value prediction engine to perform hybrid prediction using the XGBoost model and LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, while the LSTM model processes the time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor.

[0032] The customer value prediction engine performs in-depth analysis and modeling on the constructed comprehensive feature vector, outputting the predicted customer lifetime value. It consists of two different types of prediction models: the XGBoost model and the LSTM model. These two models are processed in parallel, each handling different features within the comprehensive feature vector.

[0033] XGBoost (eXtreme Gradient Boosting) is an ensemble learning algorithm based on gradient boosting decision trees. It excels at handling static features and can effectively capture nonlinear relationships and interactions between features.

[0034] LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN). By introducing three gating mechanisms—input gate, forget gate, and output gate—LSTM can selectively retain or forget historical information, effectively capturing long-term dependencies in time-series data. The LSTM model models these temporal features and outputs a time-series adjustment factor α, which quantifies the direction and extent of the impact of recent customer behavior trends on future value. A positive α value indicates a positive upward trend in customer behavior; a negative α value indicates a negative downward trend.

[0035] Specifically, the customer value prediction engine employs a hybrid prediction model combining XGBoost and LSTM. The XGBoost model processes historical behavioral features from the comprehensive feature vector, outputting a base value score V_base, which represents the value foundation assessed based on the customer's historical accumulation. The LSTM model processes temporal features from the comprehensive feature vector, outputting a temporal adjustment factor α, which quantifies the direction and extent of the impact of recent customer behavioral trends on their future value.

[0036] Understandably, this embodiment uses the XGBoost model and the LSTM model to differentiate the processing of different types of features. The XGBoost model fully leverages its advantages in processing tabular static features, while the LSTM model utilizes its ability to capture temporal dependencies, thereby improving the efficiency of feature information utilization, increasing the prediction accuracy of customer lifetime value, and enhancing the responsiveness of the prediction results to dynamic changes in customer behavior.

[0037] S103. Determine the customer segmentation results based on the stated lifetime value.

[0038] The stratification results can be either continuous value scores or discrete stratification labels, such as high-value customers, medium-value customers, and low-value customers.

[0039] Specifically, customers are segmented based on their lifetime value and corresponding segmentation thresholds. These thresholds can be dynamically set according to business needs and data distribution. The segmentation results serve as a quantitative output of customer value, providing a basis for subsequent differentiated processing. This embodiment transforms continuous predicted value into actionable segmentation labels, facilitating subsequent rule-based processing.

[0040] S104. Based on the original data, determine the customer's churn risk score, membership level upgrade probability, and upgrade time window.

[0041] The churn risk score characterizes the degree to which a customer's current behavior deviates from normal behavior patterns, reflecting the likelihood of customer churn. This score is calculated through real-time monitoring and anomaly detection of recent behavioral data in the raw data. The score typically ranges from 0 to 1, with higher scores indicating a greater risk of churn.

[0042] The probability of a customer moving up or down a membership level is used to characterize the likelihood of a customer migrating from their current membership level to a higher or lower level in the future. This indicator includes the probability of upgrading and the probability of downgrading, with values ​​ranging from 0 to 1. A higher score indicates a greater likelihood of a change in the corresponding level.

[0043] The transition time window is used to predict the approximate length of time required for a customer to reach a new level from their current state.

[0044] Specifically, multiple basic indicators can be extracted from the raw data. Anomalies are calculated based on these indicators, then weighted and summed. The weighted sum is then mapped to a range of 0 to 1 using a sigmoid function to obtain a basic risk score. Subsequently, an exponential decay function is used to adjust the basic risk score over time, generating the final churn risk score.

[0045] Extract relevant feature data from the original data, construct feature vectors, and input the feature vectors into two logistic regression models to calculate the upgrade probability and downgrade probability, respectively.

[0046] The system extracts current customer behavior metrics, corresponding level thresholds for level transitions, and historical average rates of change from the raw data. Then, it calculates the rate of change required to reach the level threshold. Finally, it multiplies the calculated rate of change by a preset adjustment factor to predict the transition time window.

[0047] Understandably, churn risk scoring enables early detection and quantification of abnormal customer behavior, membership level hopping probability transforms level management from passive response to proactive prediction, and the hopping time window provides a quantitative basis for the urgency of level changes, facilitating the matching of differentiated processing methods based on the degree of time urgency. This embodiment improves the real-time performance and sensitivity of abnormal customer behavior monitoring and enhances the quantitative prediction capability of customer level change trends by calculating churn risk score, membership level hopping probability, and hopping time window. These three indicators comprehensively characterize the dynamic behavioral features of customers from three dimensions: risk level, probability of change, and urgency of change, ensuring that subsequent multi-dimensional integrated decision-making can be based on more comprehensive and refined input information, thereby improving the pertinence and effectiveness of subsequent tiered management strategies.

[0048] S105. Based on the stratification results, the churn risk score, the membership level jump probability and jump time window, the intervention strategy identifier for the customer is generated using a decision tree algorithm.

[0049] Among them, the decision tree algorithm is a machine learning algorithm based on a tree structure. It recursively splits the data through a series of feature judgment rules and finally reaches the leaf node to output the decision result.

[0050] Specifically, a decision tree algorithm is used to integrate customer segmentation results, churn risk scores, upgrade probabilities, downgrade probabilities, and transition time windows to generate intervention strategy identifiers that match the customer's current status. This embodiment achieves comprehensive judgment of customer status and differentiated strategy generation by integrating multi-dimensional information such as segmentation results, churn risk scores, membership level transition probabilities, and transition time windows using a decision tree algorithm. This allows for more accurate matching of the final strategy identifiers to the customer's current multi-dimensional status, enabling refined processing and intelligent analysis of customer data, and improving the targeting and effectiveness of customer segmentation management.

[0051] Understandably, this application embodiment constructs and dynamically weights customer raw data using multi-dimensional features, employs a hybrid XGBoost and LSTM model to predict customer lifetime value, and calculates churn risk scores, membership level transition probabilities, and transition time windows in parallel. Finally, it uses a decision tree algorithm to fuse multi-dimensional information to generate differentiated intervention strategy identifiers. This method effectively improves the representational ability of customer behavioral characteristics, the accuracy of value prediction, the sensitivity of risk identification, and the fusionability of multi-dimensional decision results, thereby achieving refined processing and intelligent analysis of customer data and improving the targeting and effectiveness of customer segmentation management.

[0052] In summary, the customer value prediction and hierarchical management method provided by this invention constructs a comprehensive feature vector based on the customer's original data. This comprehensive feature vector includes historical behavioral features and time-series features, enhancing the information representation capability of the feature vector. This allows the constructed feature vector to more comprehensively reflect customer behavior patterns, ensuring that subsequent prediction models can perform value assessments based on richer information, thereby improving the accuracy of customer lifetime value prediction. The comprehensive feature vector is then input into the customer value prediction engine for hybrid prediction using the XGBoost and LSTM models. The XGBoost model processes historical behavioral features and outputs a base value score, while the LSTM model processes time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor, improving the efficiency of feature information utilization. This approach improves the accuracy of customer lifetime value (CWD) predictions and enhances the responsiveness of prediction results to dynamic changes in customer behavior. Customer segmentation based on CWD facilitates subsequent rule-based processing. Based on raw data, it determines customer churn risk scores, membership level transition probabilities, and transition time windows, ensuring that subsequent multi-dimensional integrated decisions are based on more comprehensive and refined input information, thereby improving the targeting and effectiveness of subsequent segmentation management strategies. Based on segmentation results, churn risk scores, membership level transition probabilities, and transition time windows, a decision tree algorithm is used to generate customer intervention strategy identifiers. This enables comprehensive judgment of customer status and differentiated strategy generation, allowing the final generated strategy identifiers to more accurately match the current multi-dimensional status of customers. This achieves refined processing and intelligent analysis of customer data, improving the targeting and effectiveness of customer segmentation management.

[0053] In some embodiments of the present invention, determining the customer's lifetime value based on the base value score and the time-series adjustment factor includes: calculating the lifetime value CLV according to the following formula: CLV = V_base×(1+α)×β where V_base is the base value score, α is the time-series adjustment factor, β is the time decay factor, and β= exp(-λt), λ is the time decay parameter, and t is the predicted time span.

[0054] Among them, the base value score V_base represents the value base assessed based on the customer's long-term accumulated information.

[0055] The time decay factor β is a preset exponential decay function used to weight and adjust the results for different prediction time spans.

[0056] Specifically, in this embodiment, the customer lifetime value is calculated using the following formula: CLV = V_base × (1 + α) × β; Where β = exp(-λt), λ is the time decay parameter, and t is the prediction time span.

[0057] The time-series adjustment factor α typically ranges from -1 to 1. When α is positive, it indicates that the customer's recent behavior is showing a positive upward trend, such as increased purchase frequency and increased activity. (1+α) is greater than 1, and a positive bonus is applied to the base value score. When α is negative, it indicates that the customer's recent behavior is showing a negative downward trend, such as decreased purchases and decreased activity. (1+α) is less than 1, and a negative discount is applied to the base value score. When α is close to 0, it indicates that the customer's behavior trend is stable, and the lifetime value is basically consistent with the base value score.

[0058] This embodiment integrates static accumulated information with dynamic trend information by multiplying the base value score by a time-series adjustment factor. This allows the lifetime value prediction results to simultaneously reflect both the customer's long-term historical performance and recent behavioral changes, enhancing the prediction results' dynamic responsiveness to changes in customer behavior. The introduction of the time-series adjustment factor enables the model to promptly capture rising or falling customer trends and make forward-looking adjustments before changes in customer value occur. The introduction of the time decay factor differentiates the weighting of results across different prediction time spans, ensuring the rationality and stability of long-term prediction results and avoiding over-extrapolation of long-term trends. Through the above calculation methods, the accuracy of lifetime value is improved, and the prediction is dynamic and interpretable.

[0059] In some embodiments of the present invention, such as Figure 2 As shown, step S104 includes: S201. Based on the original data, determine the current monthly consumption, historical average consumption, current monthly active days, historical average active days, current quarterly consumption growth rate, historical average growth rate, and number of days since the last interaction; S202. Determine the consumption decline rate based on the current monthly consumption and the historical average consumption. The consumption decline rate is calculated by dividing the difference between the current monthly consumption and the historical average consumption by the historical average consumption. S203. Based on the current monthly active days and the historical average active days, determine the activity decay rate, which is calculated by dividing the difference between the current monthly active days and the historical average active days by the historical average active days. S204. Based on the current quarterly consumption growth rate and the historical average growth rate, determine the consumption growth rate change rate, which is calculated by dividing the difference between the current quarterly consumption growth rate and the historical average growth rate by the historical average growth rate. S205. The consumption decline rate, the activity decay rate, and the consumption growth rate change rate are weighted and summed, and the weighted sum is mapped by the sigmoid function to obtain the basic risk score; S206. Based on the number of days since the last interaction, perform a time decay adjustment on the basic risk score, wherein the time decay adjustment is implemented according to an exponential decay function; S207. Multiply the base risk score adjusted for time decay by the customer type weighting coefficient to generate the churn risk score.

[0060] The sigmoid function maps any real number to the interval between 0 and 1, and its mathematical expression is f(x) = 1 / (1 + e^(-x)). Through the sigmoid function mapping, the weighted summation result is transformed into a basic risk score with probabilistic significance.

[0061] Specifically, basic indicator data such as current monthly consumption, historical average consumption, current monthly active days, historical average active days, current quarterly consumption growth rate, historical average growth rate, and days since the last interaction are extracted from the raw data. Based on the basic indicator data, the following abnormal indicators are calculated using the following formula: Consumption decline rate = (Current monthly consumption - Historical average consumption) / Historical average consumption; Activity decay rate = (Current monthly active days - Historical average active days) / Historical average active days; The rate of change in consumption growth = (current quarterly consumption growth rate - historical average growth rate) / historical average growth rate; The base risk score is calculated using the following formula: R_base = sigmoid(w1×consumption decline rate + w2×activity decay rate + w3×consumption growth rate change rate); The time decay adjustment is performed according to the following formula: R_time = R_base × exp(-λ×number of days since the last interaction); Adjust customer type according to the following formula: R_final = R_time × customer type weight coefficient, and obtain the churn risk score.

[0062] It should be noted that, considering the close correlation between the degree of customer behavioral abnormality and the time of their most recent interaction, the more recent the interaction, the greater its contribution to the risk score. Therefore, a time decay adjustment is applied to the base risk score based on the number of days since the last interaction. This time decay adjustment uses an exponential decay function, which allows the risk score to decrease smoothly as the number of days since the last interaction increases, consistent with the objective law that the impact of customer behavior gradually weakens over time.

[0063] It's important to further clarify that different customer groups may perceive the same degree of behavioral abnormality with different risk implications. For example, a slight decrease in activity from a high-value customer may be more concerning than a significant decrease from a low-value customer. Therefore, the final churn risk score is generated by multiplying the time-decay adjusted base risk score by a customer type weighting coefficient. This customer type weighting coefficient can be pre-set based on customer segmentation or clustering results, ensuring the risk score reflects the differentiated characteristics of different customer groups.

[0064] Understandably, this embodiment achieves comprehensive monitoring of customer spending amount, interaction frequency, and growth trend across three dimensions by calculating three abnormal indicators: consumption decline rate, activity decay rate, and consumption growth rate change rate. This enhances the multidimensionality and comprehensiveness of risk identification. Mapping the weighted summation result to a basic risk score between 0 and 1 using the sigmoid function ensures the probabilistic interpretability of the risk score and its comparability across different customers. Exponential decay adjustment based on the number of days since the last interaction improves the timeliness of the risk score, making recent behavioral changes contribute more significantly to the risk score. Introducing a customer type weighting coefficient enables dynamic adaptation of the risk score to customer group characteristics, ensuring differentiated sensitivity of risk scores for different value customer groups. This improves the rationality and accuracy of churn risk score determination.

[0065] In some embodiments of the present invention, step S104 includes: determining the customer's current level code, consumption change trend index, consumption accumulation speed, activity change trend index, and historical upgrade count based on the original data; constructing a feature vector X, wherein the feature vector X includes at least the current level code, the consumption change trend index, the consumption accumulation speed, the activity change trend index, and the historical upgrade count, respectively serving as the first feature X1, the second feature X2, the third feature X3, the fourth feature X4, and the fifth feature X5; inputting the feature vector X into a logistic regression model, and calculating the upgrade probability P_u according to the following formula: P_u = 1 / (1 + exp(-(β0 + β1×X1 + β2×X2 + β3×X3 + β4×X4 + β5×X5 + ... + βn×Xn))), where n ≥ 5, X1 to X5 are the first to fifth features, X6 to Xn are the other features in the feature vector X besides the first to fifth features, and β0, β1, β2, ..., ... ..., βn are the corresponding model parameters; input the feature vector X into the logistic regression model, and calculate the downgrade probability P_d according to the following formula: P_d= 1 / (1 + exp(-(γ0 + γ1×X1 + γ2×X2 + γ3×X3 + γ4×X4 + γ5×X5 + ... + γn×Xn))), where γ0, γ1, γ2, ..., γn are the corresponding model parameters.

[0066] Specifically, a logistic regression model can be used for probabilistic prediction. First, feature parameters related to changes in membership levels are extracted from the raw data. These features include the customer's current level code, which characterizes the customer's current membership level; a consumption trend indicator, quantified by the consumption growth rate over the past three months, reflecting the direction of recent changes in customer spending behavior; a consumption accumulation rate, measuring the rate at which a customer accumulates mileage or points over a certain period; an activity trend indicator, describing the increase or decrease in the number of active days; and historical upgrade frequency, recording the frequency with which a customer has previously jumped from lower to higher levels. These features characterize customer behavior patterns and their correlation with level changes from different dimensions.

[0067] Based on feature extraction, a feature vector X is constructed. This vector contains at least the current level code, consumption change trend index, consumption accumulation rate, activity change trend index, and historical upgrade count, which are respectively designated as the first feature X1, the second feature X2, the third feature X3, the fourth feature X4, and the fifth feature X5. In addition to the above five core features, the feature vector X can also incorporate other behavioral features related to level transitions as extended features X6 to Xn, depending on the needs of the actual application scenario, to further improve the prediction accuracy of the model.

[0068] The constructed feature vector X is input into two independent logistic regression models to calculate the upgrade probability and downgrade probability, respectively. Logistic regression is a machine learning algorithm widely used in binary classification problems. Its core lies in mapping the result of linear regression to a probability value between 0 and 1 through the sigmoid function, thereby achieving a quantitative prediction of the likelihood of an event occurring. In this embodiment, the formula for calculating the upgrade probability P_u is: P_u = 1 / (1 + exp(-(β0 + β1×X1 + β2×X2 + β3×X3 + β4×X4 + β5×X5 + ... + βn×Xn))), where β0, β1, β2, ..., βn are parameters learned by the model through training data, and these parameters reflect the contribution weight of each feature to the upgrade probability. Similarly, the formula for calculating the downgrade probability P_d is: P_d = 1 / (1 + exp(-(γ0 + γ1×X1 + γ2×X2 + γ3×X3 + γ4×X4 + γ5×X5 + ... + γn×Xn))), where γ0, γ1, γ2, ...,γn are the parameters of the corresponding downgrade probability model.

[0069] This embodiment transforms the traditional fixed-rule-based membership level determination into dynamic probability prediction based on real-time customer behavior data. This allows for more sensitive capture of the potential impact of subtle changes in customer behavior on membership level status. By outputting specific probability values, it provides quantitative input for subsequent decision tree algorithms, enabling the generation of intervention strategies to no longer rely on a single rule but to comprehensively consider the customer's current status, behavioral trends, and historical performance. This improves the precision and responsiveness of customer segmentation management.

[0070] In some embodiments of the present invention, such as Figure 3 As shown, step S104 includes: S301. Based on the raw data, determine the customer's current behavior indicators, the level threshold corresponding to the membership level jump, and the customer's historical average rate of change. S302. The rate of change is calculated by dividing the absolute value of the difference between the current behavior indicator and the level threshold by the historical average rate of change. S303. Based on the change rate multiplied by a preset adjustment coefficient, the transition time window is predicted.

[0071] Among them, current behavior indicators can be quantitative values ​​that reflect a customer's recent activity level or spending level, such as the average spending amount, cumulative mileage, or number of active days in the past three months. These indicators can represent the customer's current position in the membership level system.

[0072] The threshold for membership level advancement can be the indicator data for advancing to the corresponding membership level, such as the spending threshold required to upgrade from a regular member to a silver member, or the minimum activity level required to downgrade from a gold member to a silver member.

[0073] The difference represents the incremental action a customer needs to take to reach the required level for a change. The rate of change is obtained by dividing the absolute value of this difference by the historical average rate of change.

[0074] The rate of change refers to the number of time units required to reach the upgrade threshold based on the customer's historical rate of change. For example, if a customer's current spending is 1,000 yuan short of the upgrade threshold, and their historical average monthly spending growth rate is 500 yuan, then the rate of change is 2 months.

[0075] A preset adjustment coefficient is used to calibrate the initially calculated rate of change. In practical applications, customer behavior may be affected by seasonal factors, market activities, or individual random fluctuations, leading to some bias in extrapolation based solely on historical rates of change. Therefore, a preset adjustment coefficient is introduced to correct the rate of change. This coefficient can be set based on business validation results, model training feedback, or industry experience. For example, in scenarios with relatively stable customer behavior, the preset adjustment coefficient can be set to 1.0, maintaining the original rate of change; in scenarios with large fluctuations in customer behavior, the preset adjustment coefficient can be set to 1.5 to 2.0, appropriately extending the prediction time window to reduce the risk of misjudgment; in scenarios where customer behavior shows an accelerating trend, the preset adjustment coefficient can be set to 0.8 to 1.0, shortening the prediction time window to improve responsiveness. By introducing the preset adjustment coefficient, the prediction of the transition time window is no longer a fixed mathematical transformation result, but rather possesses flexible calibration capabilities, adapting to changes in different customer group characteristics and business environments, thereby improving the accuracy and practicality of time window prediction.

[0076] Specifically, based on the customer's original data, the current behavioral indicators related to level advancement are determined. Then, the level threshold corresponding to the membership level advancement is determined. Next, the absolute value of the difference between the current behavioral indicator and the level threshold is calculated. Multiplying the rate of change by a preset adjustment coefficient, the advancement time window can be predicted. This achieves personalized quantification of customer level change time.

[0077] Understandably, this embodiment predicts based on the rate of change of an individual customer's historical behavior. Compared to prediction methods using fixed-time rules, this approach more accurately reflects the behavioral differences among different customers and improves the accuracy of time window predictions. By introducing a preset adjustment coefficient, the prediction results can be flexibly calibrated according to different business scenarios or model feedback, enhancing the adaptability and controllability of the method.

[0078] In some embodiments of the present invention, the intervention strategy identifier includes at least a first type of intervention strategy identifier, a second type of intervention strategy identifier, and a third type of intervention strategy identifier, and the membership level jump probability includes an upgrade probability and a downgrade probability; step S105 includes: constructing the root node of a decision tree based on the customer's current level in the stratification result; determining the branching conditions of the decision tree based on the customer's current level represented by the root node, the upgrade probability, the downgrade probability, and the jump time window; when the upgrade probability is greater than or equal to a first probability threshold and the jump time window is less than or equal to a first time threshold, generating a first type of intervention strategy identifier based on the churn risk score; when the upgrade probability is between a second probability threshold and a first probability threshold and the jump time window is between a second time threshold and a first time threshold, generating a second type of intervention strategy identifier based on the churn risk score; when the downgrade probability is greater than or equal to a third probability threshold, generating a third type of intervention strategy identifier based on the churn risk score.

[0079] The root node serves as the initial classification of customers, providing a starting point for subsequent branch decisions. Taking an airline's membership system as an example, if the stratification result indicates the customer is currently a Silver member, the root node of the decision tree is "Silver"; if the customer is a Gold member, the root node is "Gold"; and if the customer is a Platinum member, the root node is "Platinum". Different root nodes correspond to different subsequent branch paths, thus achieving differentiated processing for customers of different levels.

[0080] Branch conditions are rules that judge the values ​​of input features and determine which child node to continue traversing from the current node.

[0081] Specifically, based on the customer segmentation results, the customer's current level information is obtained. This level information represents the customer's basic level in the value segmentation system, such as Silver, Gold, or Platinum. In the decision tree algorithm, this level information is used as the root node. Then, based on the customer's current level represented by the root node, combined with the upgrade probability, downgrade probability, and transition time window, the branching conditions of the decision tree are constructed. Specifically, for each root node (e.g., the "Silver" node), its branching conditions can be set as follows: First branch condition: Determine if the upgrade probability is greater than or equal to the first probability threshold (e.g., 0.8). If this condition is met, proceed to the first branch path, where the leaf nodes correspond to the first type of intervention strategy identifier (e.g., the points acceleration strategy).

[0082] Second branch condition: If the first branch condition is not met, determine whether the upgrade probability is between the second probability threshold and the first probability threshold (e.g., between 0.6 and 0.8), and simultaneously determine whether the transition time window is between the second time threshold and the first time threshold (e.g., between 1 and 2 months). If both conditions are met, proceed to the second branch path, where the leaf nodes correspond to the second type of intervention strategy identifier (e.g., exclusive discount strategy).

[0083] Third branch condition: If none of the aforementioned conditions are met, determine whether the downgrade probability is greater than or equal to the third probability threshold (e.g., 0.7). If this condition is met, proceed to the third branch path, where the leaf nodes correspond to the third type of intervention strategy identifier (e.g., retention service strategy).

[0084] Default branch: If none of the above conditions are met, the default branch path will be entered. The leaf nodes under this path can correspond to the conventional processing strategy identifier (such as maintaining the existing service method).

[0085] For different root nodes (such as the "Gold Card" node and the "Platinum Card" node), the specific parameters of their branching conditions (such as probability thresholds and time window thresholds) can be set differently according to the different customer levels. For example, for higher-value Platinum Card customers, the downgrade probability threshold can be set more sensitively (such as 0.5) to identify potential downgrade risks earlier; while for Silver Card customers, the upgrade probability threshold can be set relatively leniently (such as 0.7) to more actively incentivize them to jump to higher levels.

[0086] In this way, the root node and branching conditions of the decision tree together constitute a complete tree-like classification structure. The root node performs initial classification based on the customer's current level, while the branching conditions perform refined classification based on upgrade probability, downgrade probability, and transition time window. Finally, the leaf nodes output intervention strategy identifiers that match the customer's status. The root node of the decision tree is constructed based on the customer's current level in the stratified results. The customer's current level reflects the customer's basic positioning in the value dimension, such as a high-value customer, a medium-value customer, or a low-value customer. The setting of the root node ensures that subsequent branching decisions are made within a specific value level of the customer group, thereby guaranteeing the targeting of strategy recommendations.

[0087] It should be noted that the decision tree algorithm described above is a machine learning method for classification and regression. It uses a tree structure to progressively evaluate input features, ultimately outputting the corresponding category or value at the leaf nodes. In this embodiment, the root node of the decision tree represents the customer's current level, the branching conditions are the probability of upgrading, the probability of downgrading, and the transition time window, and the leaf nodes represent the corresponding intervention strategy identifiers. Specific implementations of the decision tree algorithm include, but are not limited to, Classification and Regression Tree (CART), the C4.5 algorithm, or decision tree construction methods based on information gain.

[0088] According to the preset rules, the decision tree generates policy identifiers according to the following branching logic: When the upgrade probability is greater than or equal to the first probability threshold and the transition time window is less than or equal to the first time threshold, it indicates that the customer is highly likely to upgrade and time is of the essence. At this point, a first-type intervention strategy identifier is generated based on the churn risk score. The first-type intervention strategy identifier can correspond to an immediate incentive strategy, such as a points acceleration strategy, which aims to encourage customers to complete the upgrade conversion as quickly as possible.

[0089] When the upgrade probability falls between the second probability threshold and the first probability threshold, and the transition time window falls between the second time threshold and the first time threshold, it indicates that the customer is in a medium upgrade probability and medium urgency range. At this point, a second type of intervention strategy identifier is generated based on the churn risk score. The second type of intervention strategy identifier can correspond to a mid-term nurturing strategy, such as an exclusive discount strategy, which aims to continuously stimulate customer activity and steadily promote upgrades.

[0090] When the downgrade probability is greater than or equal to the third probability threshold, it indicates a high risk of customer downgrade. At this point, a third type of intervention strategy identifier is generated based on the churn risk score. This third type of intervention strategy identifier can correspond to retention strategies, such as retention service strategies, which aim to reduce the risk of churn.

[0091] In the aforementioned branches, the first probability threshold, second probability threshold, third probability threshold, first time threshold, and second time threshold are all configurable parameters that can be dynamically adjusted based on business needs and data distribution. Meanwhile, the churn risk score serves as an auxiliary input, used to further refine the strategy strength when generating strategy identifiers. For example, for customers who also meet upgrade conditions, a higher churn risk score can correspond to a more intense incentive measure in the generated strategy identifier. Furthermore, the branching conditions and leaf node strategy identifiers of the decision tree can also be iteratively optimized based on feedback from historical data. For instance, the success probability of the currently generated strategy can be estimated based on the response rates of similar customers to different strategies in the past, thereby achieving dynamic strategy optimization.

[0092] It is worth noting that the success probability of the current strategy can also be estimated based on the strategy response rate of similar customers in the past.

[0093] In some embodiments of the present invention, after determining the customer's lifetime value and before determining the customer's stratification result based on the lifetime value, the method further includes: calculating the correlation between the prediction results of the XGBoost model and the LSTM model, and using the absolute value of the correlation as a model consistency score; obtaining a historical accuracy score based on the historical prediction accuracy of a customer group similar to the current customer; obtaining a data integrity score based on the data completeness of the current customer; and multiplying the model consistency score, the historical accuracy score, and the data integrity score by preset weights and then summing them to obtain a comprehensive confidence score, wherein the sum of each preset weight is 1.

[0094] The correlation degree is used to measure the statistical association between the prediction results of two models. It can be calculated using the Pearson correlation coefficient, which ranges from -1 to 1. The closer the absolute value is to 1, the stronger the correlation.

[0095] Similar customer groups refer to sets of customers with similar behavioral patterns to current customers. They can be obtained by dividing historical customer data using clustering algorithms. Clustering algorithms include K-means clustering and hierarchical clustering, among others.

[0096] Data completeness refers to the ratio of the number of features that the customer can actually use during the feature construction process to the total number of features that should theoretically be available. The closer this ratio is to 1, the more complete the data is and the more sufficient the information is on which the prediction is based.

[0097] The overall confidence score is a quantitative indicator obtained by weighting and fusing scores from multiple dimensions. The weighting coefficients can be adjusted according to different application scenarios. For example, higher weights can be given to the model consistency score and historical accuracy score to reflect the importance of model performance and historical experience.

[0098] The sum of the preset weights is 1. The specific weight values ​​can be adjusted according to the actual application scenario. For example, higher weights can be given to the model consistency score and historical accuracy score to reflect the importance of model performance and historical experience.

[0099] Specifically, the steps for determining the prediction confidence level are as follows: Calculate the correlation coefficient r between the prediction results of the XGBoost model and the LSTM model to obtain the model consistency score C1 = |r|; Based on the historical prediction accuracy P_hist of similar customer groups, the historical accuracy score C2 = P_hist is obtained; Based on the customer's data completeness D_complete, the data completeness score C3 = D_complete is obtained; The overall confidence level C_final is calculated using the following formula: C_final = w1×C1 + w2×C2 + w3×C3, where w1+w2+w3=1.

[0100] Understandably, this embodiment introduces a multi-verification mechanism to comprehensively evaluate the reliability of the prediction results from three dimensions: model consistency, historical accuracy, and data integrity. This enables the system to output not only the predicted value of customer lifetime value but also the confidence level of that prediction. This provides important reference for subsequent customer segmentation and strategy generation: when the overall confidence level is high, the prediction results can be trusted more and automated processing can be performed accordingly; when the overall confidence level is low, it prompts for manual review or data supplementation, thereby improving the reliability and usability of the entire prediction system.

[0101] In some embodiments of the present invention, the step of constructing a comprehensive feature vector includes: dynamically adjusting the weight coefficients of each dimension feature according to the group type to which the customer belongs through a multilayer perceptron; and performing weighted fusion of the historical behavior feature and the time-series feature based on the adjusted weights to obtain the comprehensive feature vector.

[0102] Specifically, customer group types can be pre-determined using clustering algorithms, such as K-means clustering or hierarchical clustering. Based on customers' historical behavioral data, customers are divided into different group categories, with each group category corresponding to a set of customers with similar behavioral characteristics. The customer group type can be represented by an embedding vector, which is obtained by mapping the group category to a continuous vector space and can effectively characterize the intrinsic attributes of the group category.

[0103] After determining the customer's group type, the group type embedding vector is input into a Multi-Layer Perceptron (MLP) for processing. A MLP is a feedforward neural network structure containing an input layer, hidden layers, and an output layer, capable of learning a non-linear mapping from input data to output data. In this embodiment, the input to the MLP is the group type embedding vector, and the output is the weight coefficients corresponding to each dimension feature. To ensure that the sum of all weight coefficients is 1 and that each weight coefficient is between 0 and 1, the output layer of the MLP is activated using the softmax function, i.e., the weight coefficients are calculated according to w_i = softmax(MLP(customer_type_embedding)). Here, MLP represents the mapping function of the MLP, customer_type_embedding is the embedding vector corresponding to the customer group type, and w_i is the weight coefficient corresponding to the i-th dimension feature.

[0104] The aforementioned dynamic weight adjustment process can be performed periodically, for example, by recalculating the weight coefficients monthly based on changes in customer behavior, to ensure that the weights can reflect the evolution of customer group behavior patterns in a timely manner.

[0105] After obtaining the dynamic weight coefficients of each dimension's features, historical behavioral features and time-series features are weighted and fused based on the adjusted weights to obtain a comprehensive feature vector. The weighted fusion can be achieved using linear weighted summation, where the value of each dimension in the comprehensive feature vector is the original feature value of that dimension multiplied by its corresponding weight coefficient. The comprehensive feature vector constructed in this way retains the information content of the original features while incorporating the differentiated influence of customer group characteristics on feature importance.

[0106] It should be noted that the above-mentioned multilayer perceptron can be implemented in various ways. For example, a shallow network structure with 2-3 hidden layers can be used, with each hidden layer containing 64-128 neurons and the ReLU activation function. Alternatively, a deeper or wider network structure can be chosen depending on the actual data scale. The dimension of the group type embedding vector can be set according to the number of groups. For example, when the number of groups is K, the dimension of the embedding vector can be set to between 1 / 2 and 1 / 4 of K. The update frequency of the weight coefficients can be flexibly adjusted according to business needs. In addition to monthly updates, quarterly updates or updates triggered by changes in behavioral data can also be used.

[0107] In some embodiments of the present invention, after generating the intervention strategy identifier, the strategy execution results can be tracked, and the model and parameters can be continuously optimized based on the feedback data to form a closed-loop processing flow.

[0108] Specifically, once the intervention strategy is identified and executed, the system records relevant information about the strategy implementation, including the implementation time, customer response behavior, and the final strategy execution result. Customer response behavior can include whether the customer accepted the strategy and any changes in behavior after acceptance; the strategy execution result can include indicators such as whether the customer's level changed and whether their lifetime value increased. This information constitutes feedback data, which is used for subsequent model optimization and parameter adjustment.

[0109] Based on the feedback data, the system continuously optimizes the prediction model. Specifically, incremental learning can be used to incorporate new feedback data into the training set, fine-tuning the parameters of the XGBoost and LSTM models to adapt them to changes in customer behavior patterns. For the logistic regression model, the probability calculation parameters are adjusted based on the feedback results, such as updating the model parameters β0, β1, ..., βn and γ0, γ1, ..., γn to improve the accuracy of predicting grade transition probabilities. The update frequency can be set according to actual needs, such as weekly incremental updates and monthly full training.

[0110] Furthermore, based on the effectiveness data of strategy execution, the strategy recommendation rules can be optimized. Specifically, the response rate and success rate of different strategies in different customer groups can be statistically analyzed. For strategies with low response rates, the branching conditions or threshold settings of the decision tree can be adjusted; for strategy combinations with high success rates, their weight in the strategy library can be increased. By continuously collecting and analyzing the effectiveness of strategy execution, the strategy recommendation rules can be continuously optimized, gradually improving the matching degree between strategies and actual customer conditions.

[0111] In a further embodiment, the system can also dynamically adjust various threshold parameters. For example, the churn risk scoring threshold, first probability threshold, second probability threshold, third probability threshold, first time threshold, and second time threshold can be dynamically adjusted based on the statistical analysis results of historical data. When a model performance indicator (such as prediction accuracy) is detected to have decreased beyond a preset range, the system automatically triggers parameter readjustment or model retraining processes to ensure that the system can adapt to changes in data distribution.

[0112] This embodiment introduces a feedback optimization mechanism, enabling the system to continuously improve the prediction model and recommendation rules based on the policy execution results. The collection of feedback data allows model parameters to dynamically adapt to changes in customer behavior patterns, while the optimization of policy rules improves the match between intervention strategies and actual needs. This closed-loop process gives the system self-optimization capabilities, maintaining prediction accuracy and policy effectiveness over long-term operation and avoiding performance degradation caused by model aging.

[0113] The customer value prediction and hierarchical management method of this application embodiment is executed in a distributed microservice architecture, including: accessing real-time and batch data through message queues; performing feature engineering processing based on a distributed computing framework; deploying the XGBoost model and the LSTM model in containers and managing them through container orchestration tools; and providing API interfaces for integration with external systems.

[0114] In one specific implementation, a distributed microservice architecture may include the following core components: Data access layer: Supports real-time and batch data access, and uses Kafka message queues to ensure reliable data transmission; Feature engineering services: Feature extraction module: Processes large-scale customer data based on the Spark distributed computing framework; Feature storage: Redis is used to cache hot features, and HBase is used to store historical feature data; Feature update: Supports both incremental update and full update modes; Model service layer: XGBoost model service: Deployed in Docker containers, supporting horizontal scaling; LSTM Model Service: Provides high-performance inference based on TensorFlow Serving; Model version management: Supports A / B testing and canary releases; Prediction Engine: Real-time prediction: Response time <100ms, supports 10000+ QPS concurrency; Batch forecasting: Supports batch processing of tens of millions of customer data; Result caching: Prediction results are cached for 24 hours to reduce redundant calculations; API Gateway: Provides RESTful API interfaces and supports integration with existing CRM and marketing systems; The model training and update mechanism is as follows: Offline training: Training data: Customer behavior data from the past 24 months; Training frequency: Incremental training every week, full training every month; Model validation: Time series cross-validation is used to ensure the model's generalization ability; Online learning: Real-time feedback: Collect actual customer behavior data and update model parameters; Concept drift detection: Monitor model performance metrics and automatically trigger model retraining; Parameter tuning: Automatically adjust hyperparameters based on Bayesian optimization algorithm; Deployment and operation are specifically as follows: Containerized deployment: All services are deployed based on Docker containers and support Kubernetes orchestration; Monitoring Alerts: System monitoring: Monitoring of basic metrics such as CPU, memory, and network; Business monitoring: Monitoring business metrics such as prediction accuracy, response time, and throughput; Alarm mechanism: A multi-level alarm system based on thresholds and anomaly detection; Data security: Data encryption: Both transmitted and stored data are encrypted using AES-256. Access control: Fine-grained permission control based on RBAC; Audit logs: Complete records of data access and model call logs; To better implement the customer value prediction and tiered management method in the embodiments of the present invention, based on the customer value prediction and tiered management method, correspondingly, as follows: Figure 4 As shown, this embodiment of the invention also provides a customer value prediction and tiering management system. The customer value prediction and tiering management system 400 includes: Construction unit 401 is used to construct a comprehensive feature vector based on the customer's original data, wherein the comprehensive feature vector includes historical behavioral features and time-series features; Prediction unit 402 is used to input the comprehensive feature vector into the customer value prediction engine to perform hybrid prediction through the XGBoost model and LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, the LSTM model processes the time series features and outputs a time series adjustment factor, and the customer's lifetime value is determined based on the base value score and the time series adjustment factor. The first determining unit 403 is used to determine the customer segmentation result based on the lifecycle value; The second determining unit 404 is used to determine the customer's churn risk score, membership level upgrade probability, and upgrade time window based on the original data. The generation unit 405 is used to generate the customer's intervention strategy identifier based on the hierarchical results, the churn risk score, the membership level jump probability and the jump time window, using a decision tree algorithm.

[0115] The customer value prediction and tiering management system 400 provided in the above embodiments can realize the technical solutions described in the above customer value prediction and tiering management method embodiments. The specific implementation principles of each module or unit can be found in the corresponding content in the above customer value prediction and tiering management method embodiments, and will not be repeated here.

[0116] Accordingly, embodiments of this application also provide a computer-readable storage medium for storing computer-readable programs or instructions. When the programs or instructions are executed by a processor, they can implement the steps or functions of the customer value prediction and hierarchical management methods provided in the above-described method embodiments.

[0117] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0118] The above provides a detailed description of the customer value prediction and tiered management method, system, and storage medium provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A customer value forecasting and tiered management method, characterized in that, include: Based on the customer's original data, a comprehensive feature vector is constructed, which includes historical behavioral features and time-series features; The comprehensive feature vector is input into the customer value prediction engine to perform hybrid prediction using the XGBoost model and LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, while the LSTM model processes the time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor. The customer segmentation results are determined based on the stated lifetime value; Based on the raw data, the customer's churn risk score, membership level upgrade probability, and upgrade time window are determined. Based on the stratification results, the churn risk score, the membership level transition probability and transition time window, the intervention strategy identifier for the customer is generated using a decision tree algorithm.

2. The customer value prediction and tiered management method according to claim 1, characterized in that, The process of determining a customer's lifetime value based on the base value score and the time-series adjustment factor includes: The Lifetime Value (CLV) is calculated using the following formula: CLV = V_base×(1+α)×β Where V_base is the base value score, α is the time series adjustment factor, β is the time decay factor, and β=exp(-λt), where λ is the time decay parameter and t is the prediction time span.

3. The customer value forecasting and tiered management method according to claim 1, characterized in that, The process of determining the customer's churn risk score based on the raw data includes: Based on the original data, determine the current monthly consumption, historical average consumption, current monthly active days, historical average active days, current quarterly consumption growth rate, historical average growth rate, and number of days since the last interaction; The consumption decline rate is determined based on the current monthly consumption and the historical average consumption. The consumption decline rate is calculated by dividing the difference between the current monthly consumption and the historical average consumption by the historical average consumption. The activity decay rate is determined based on the current monthly active days and the historical average active days. The activity decay rate is calculated by dividing the difference between the current monthly active days and the historical average active days by the historical average active days. The consumption growth rate is determined based on the current quarterly consumption growth rate and the historical average growth rate. The consumption growth rate is calculated by dividing the difference between the current quarterly consumption growth rate and the historical average growth rate by the historical average growth rate. The consumption decline rate, the activity decay rate, and the consumption growth rate change rate are weighted and summed, and the weighted sum is mapped by the sigmoid function to obtain the basic risk score; Based on the number of days since the last interaction, a time decay adjustment is performed on the basic risk score, and the time decay adjustment is implemented according to an exponential decay function; The churn risk score is generated by multiplying the base risk score (adjusted for time decay) by the customer type weighting coefficient.

4. The customer value prediction and tiered management method according to claim 1, characterized in that, The step of determining the probability of a customer advancing to a higher membership level based on the original data includes: Based on the raw data, determine the customer's current level code, consumption change trend index, consumption accumulation speed, activity change trend index, and historical upgrade count; Construct a feature vector X, which includes at least the current level code, the consumption change trend index, the consumption accumulation speed, the activity change trend index, and the historical upgrade count, respectively serving as the first feature X1, the second feature X2, the third feature X3, the fourth feature X4, and the fifth feature X5. Input the feature vector X into the logistic regression model, and calculate the upgrade probability P_u according to the following formula: P_u = 1 / (1 + exp(-(β0 + β1×X1 + β2×X2 + β3×X3 + β4×X4 + β5×X5 + ... + βn×Xn))), where n ≥ 5, X1 to X5 are the first to fifth features, X6 to Xn are the other features in the feature vector X besides the first to fifth features, and β0, β1, β2, ..., βn are the corresponding model parameters; Input the feature vector X into the logistic regression model, and calculate the downgrade probability P_d according to the following formula: P_d = 1 / (1 + exp(-(γ0 + γ1×X1 + γ2×X2 + γ3×X3 + γ4×X4 + γ5×X5+ ... + γn×Xn))), where γ0, γ1, γ2, ..., γn are the corresponding model parameters.

5. The customer value forecasting and tiered management method according to claim 1, characterized in that, Determining the customer's transition time window based on the original data includes: Based on the raw data, determine the customer's current behavioral indicators, the level threshold corresponding to the membership level jump, and the customer's historical average rate of change. The rate of change is calculated by dividing the absolute value of the difference between the current behavior indicator and the level threshold by the historical average rate of change. The transition time window is predicted by multiplying the rate of change by a preset adjustment coefficient.

6. The customer value prediction and tiered management method according to claim 1, characterized in that, The intervention strategy identifier includes at least a first type of intervention strategy identifier, a second type of intervention strategy identifier, and a third type of intervention strategy identifier; the membership level jump probability includes an upgrade probability and a downgrade probability; the step of generating the customer's intervention strategy identifier using a decision tree algorithm based on the stratification results, the churn risk score, the membership level jump probability, and the jump time window includes: The root node of the decision tree is constructed based on the customer's current level in the hierarchical results; Based on the customer's current level represented by the root node, the upgrade probability, the downgrade probability, and the transition time window, the branching conditions of the decision tree are determined; When the upgrade probability is greater than or equal to the first probability threshold and the transition time window is less than or equal to the first time threshold, a first type of intervention strategy identifier is generated based on the churn risk score. When the upgrade probability is between the second probability threshold and the first probability threshold and the transition time window is between the second time threshold and the first time threshold, a second type of intervention strategy identifier is generated based on the churn risk score; When the downgrade probability is greater than or equal to the third probability threshold, a third type of intervention strategy identifier is generated based on the churn risk score.

7. The customer value prediction and tiered management method according to claim 1, characterized in that, After determining the customer's lifetime value, and before determining the customer segmentation result based on the lifetime value, the process also includes: Calculate the correlation between the prediction results of the XGBoost model and the LSTM model, and use the absolute value of the correlation as the model consistency score; A historical accuracy score is obtained based on the historical prediction accuracy of a customer group similar to the current customer. A data integrity score is obtained based on the current level of customer data integrity. The model consistency score, the historical accuracy score, and the data integrity score are multiplied by preset weights and then summed to obtain the overall confidence score, wherein the sum of each preset weight is 1.

8. The customer value prediction and tiered management method according to claim 1, characterized in that, The steps for constructing the comprehensive feature vector include: Based on the customer's group type, the weight coefficients of each dimension feature are dynamically adjusted using a multilayer perceptron. The historical behavior features and the temporal features are weighted and fused based on the adjusted weights to obtain the comprehensive feature vector.

9. A customer value forecasting and tiered management system, characterized in that, include: A construction unit is used to construct a comprehensive feature vector based on the customer's original data, wherein the comprehensive feature vector includes historical behavioral features and time-series features; The prediction unit is used to input the comprehensive feature vector into the customer value prediction engine to perform a hybrid prediction using the XGBoost model and the LSTM model in the customer value prediction engine. The XGBoost model processes the historical behavioral features and outputs a base value score, while the LSTM model processes the time-series features and outputs a time-series adjustment factor. The customer's lifetime value is determined based on the base value score and the time-series adjustment factor. The first determining unit is used to determine the customer segmentation result based on the customer's lifetime value; The second determining unit is used to determine the customer's churn risk score, membership level upgrade probability, and upgrade time window based on the original data. The generation unit is used to generate the customer's intervention strategy identifier based on the hierarchical results, the churn risk score, the membership level jump probability and the jump time window, using a decision tree algorithm.

10. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the customer value prediction and tiered management method according to any one of claims 1 to 8.

Images