Customer asset intelligent management system based on hierarchical dynamic driving model

Abstract

The invention discloses a customer asset intelligent management system based on a hierarchical dynamic driving model, which relates to the technical field of business data processing and comprises a unified data middle table, a hierarchical dynamic driving engine, a strategy driving unit, a feedback optimization unit, a loss early warning and intervention unit and a quantitative stripping decision unit. According to the invention, dynamic and automatic customer asset management is realized, and the system can automatically execute a precise strategy for intervention, upgrading or stripping, so that the overall value and operation efficiency of customer assets are remarkably improved, and the management efficiency of the customer assets is improved.

Description

Technical Field

[0001] This invention relates to the technical field of business data processing, specifically to a customer asset intelligent management system based on a hierarchical dynamic driving model. Background Technology

[0002] In the digital economy era, customer relationships have become one of the most core strategic assets for enterprises. The traditional concept of "customer relationship management" is evolving into "customer asset management," which emphasizes the quantitative assessment, proactive operation, and dynamic optimization of customer relationships from an investment return perspective. With the maturity of big data, machine learning, and cloud computing technologies, enterprises are now able to collect and store massive amounts of multi-dimensional customer interaction data, laying the technological foundation for building intelligent and automated customer asset management systems. How to transform this data into precise insights into the future value of customers, and thereby drive efficient resource allocation and personalized interactions, has become a key challenge for enterprises to enhance their core competitiveness and achieve sustainable growth.

[0003] Early customer asset management primarily relied on simple statistical analysis models based on historical transaction data. The most representative example is the RFM model, which segments customers using three static dimensions: most recent purchase, purchase frequency, and purchase amount. Additionally, the basic discounted cash flow (CLV) model was widely used, typically assuming stable customer behavior and that future revenue could be simply predicted based on past averages. These methods depended on regular manual analysis, with results presented in static reports. Marketing decisions heavily relied on managers' personal experience, resulting in long lead times, delayed actions, and difficulty in achieving large-scale personalization.

[0004] With technological advancements, existing technologies have evolved from traditional methods towards more complex models and partial automation. Academics and commercial software developers are introducing more sophisticated statistical models to predict Customer Churn Value (CLV), such as the Pareto / NBD model and its derivatives, like the BG / NBD model, to predict future customer transaction behavior, and Markov chain models to describe customer state transitions. These models improve the accuracy of value prediction, especially when dealing with non-contractual customer relationships. Some advanced CRM or customer data platforms are beginning to utilize clustering algorithms, such as K-Means, for dynamic customer segmentation and are attempting to use machine learning classification models to predict customer churn.

[0005] However, existing technical solutions typically operate at fixed points in time, generating snapshot-style analysis results. Within the interval between two analyses, customers' value status, loyalty, and behavioral intentions may have undergone significant changes, but the system cannot perceive and respond in real time, leading to severely delayed interventions and missed optimal operational opportunities. Existing technologies mostly focus on predicting single indicators, lacking the ability to predict the dynamic transition paths of customers throughout the entire asset value hierarchy, resulting in a lack of accurate data support for the formulation of preventative maintenance and proactive upgrade strategies. In CLV measurement, existing solutions often use a single model for all customers or require manual selection of models for different groups, failing to flexibly and accurately handle complex mixed behavioral patterns such as intermittent and regular purchases in reality. Existing models have significant prediction errors, leading to inaccurate asset valuations and affecting the fundamental reliability of stratification and resource allocation.

[0006] The shortcomings of existing technologies have resulted in customer asset management remaining in a relatively static, passive, and manual intervention-based stage, leading to problems such as misallocation of marketing resources and increased risk of losing high-value customers, thus resulting in low efficiency in customer asset management. Summary of the Invention

[0007] The purpose of this invention is to provide a customer asset intelligent management system based on a hierarchical dynamic driving model to solve the problems mentioned in the background art.

[0008] This invention provides a customer asset intelligent management system based on a hierarchical dynamic driving model, and achieves its objective through the following technical solution:

[0009] A customer asset intelligent management system based on a hierarchical dynamic driving model includes the following modules:

[0010] Unified Data Platform: Used to access and integrate raw customer data from transaction systems, interaction platforms and external data sources in real time. By performing data cleaning, entity association and feature engineering, it outputs dynamic feature vectors with unique customer identifiers as keys, containing time-series behavioral tags and statistical indicators.

[0011] Hierarchical Dynamic Drive Engine: Used to dynamically assess and classify customers and predict future state changes, including collaborative computing unit, dynamic mapping unit and state inference unit;

[0012] Collaborative computing unit: used to perform customer lifetime value prediction and customer loyalty measurement in parallel based on the dynamic feature vector, wherein the customer lifetime value prediction calls different prediction models for calculation according to customer behavior patterns, and the customer loyalty measurement is performed by fitting the preset latent variable observation indicators through structural equation model.

[0013] Dynamic mapping unit: It is used to take the customer lifetime value prediction result and the customer loyalty measurement result as inputs to the first coordinate axis and the second coordinate axis respectively, and map the customer to a specified level in the hierarchical architecture composed of platinum layer, gold layer, steel layer and heavy lead layer in real time through a preset two-dimensional threshold grid.

[0014] State deduction unit: used to deduce the state transition probability distribution vector of customers from the current level to other levels in the future preset time period based on the customer's historical level sequence and the behavioral trend indicators in the dynamic feature vector, through a time series classification model;

[0015] Strategy-driven unit: Based on the customer's current level, predicted state transition probability, and real-time triggered behavioral events, it matches the corresponding automated operation strategy from the strategy knowledge base and converts it into executable instructions to be distributed to downstream business systems.

[0016] By adopting the above technical solutions, the unified data platform transforms raw data from heterogeneous sources into standardized dynamic feature vectors through real-time data cleaning, correlation, and feature engineering. This provides a highly timely and dimensionally unified input data foundation for upper-level analysis, overcoming the modeling difficulties and analysis delays caused by data dispersion and inconsistent formats in existing technologies. The collaborative computing unit executes customer lifetime value prediction and loyalty measurement in parallel. CLV prediction does not rely on a single fixed model but can adaptively call the most suitable prediction model based on customer behavior patterns, significantly improving prediction accuracy and system applicability. Especially for customer groups with complex and variable transaction patterns, the system achieves accurate fitting of different behavioral characteristics (such as intermittent and regular patterns) through an intelligent model selection mechanism, solving the problem of large prediction bias caused by model-behavior mismatch in traditional methods. Simultaneously, customer loyalty measurement uses structural equation modeling to fit multi-dimensional latent variable observation indicators, achieving a scientific and quantitative measurement of the abstract concept of customer loyalty. This transforms non-monetary values ​​such as reputation and trust, which are difficult to observe directly, into calculable indicators, providing key input for a comprehensive assessment of customer assets. The dynamic mapping unit utilizes the two quantitative results mentioned above to perform two-dimensional hierarchical classification, establishing a more scientific and stable customer asset classification system. The classification results simultaneously reflect the customer's current profitability and future relationship stability, providing more guidance than static clustering based on a single RFM indicator. The state inference unit predicts state transition probabilities through time series models, enabling the system to possess forward-looking insight capabilities and providing key decision-making basis for proactive intervention. The strategy-driven unit automatically matches and executes strategies based on the current level, predicted probabilities, and real-time events. It associates complex analytical conclusions with specific business actions and achieves automated execution through instruction distribution. This solves the inefficiency and missed opportunities caused by the traditional chain from analysis reports to manual decision-making and manual configuration. It realizes integrated operation of dynamic assessment of customer asset status, risk warning and automation, and personalized intervention. Fluctuations in customer value can be perceived in real time and transformed into specific actions through pre-set strategy logic, thereby significantly improving the efficiency and accuracy of marketing resource allocation and effectively stabilizing and increasing the total value of customer assets. This setup realizes dynamic and automated customer asset management. The system can not only accurately measure and classify customer assets, but also proactively predict asset change trends and automatically execute precise strategies for intervention, upgrading, or divestiture, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0017] Optionally, the collaborative computing unit is configured with a first prediction model, which is a fusion algorithm that combines the BG / NBD model and the Markov chain model. Its specific execution process is as follows: using the BG / NBD model to process the customer's historical transaction time and number of transactions in the dynamic feature vector, predicting the customer's transaction probability in a specific future period, dynamically constructing the transaction probability sequence into a non-homogeneous state transition matrix in the Markov chain model, and outputting the customer's total remaining lifetime value by solving the steady-state expected return of the Markov chain.

[0018] By adopting the above technical solution, the BG / NBD model is a typical application of the negative binomial distribution model in customer transaction prediction. By using the BG / NBD model for initial prediction and the Markov chain for subsequent valuation, the specific technical challenge of long-term value prediction for intermittently purchasing customers is effectively solved. The BG / NBD model is good at predicting the probability of future transactions of individuals based on transaction history, but it does not directly output monetary value itself. This solution transforms the probability sequence output by BG / NBD, rather than single-point probabilities, into non-homogeneous state transition probabilities that change over time in the Markov chain model. This constructs a refined probability model that can characterize the dynamic changes in customers' future transaction behavior. Compared with homogeneous Markov chains that use fixed transition probabilities, non-homogeneous chains are more in line with the reality that customer activity may decay or fluctuate over time, thus making the model's simulation of the customer lifecycle more realistic. By defining state-related returns, such as single-transaction profits, on this non-homogeneous Markov chain and solving for their steady-state or long-term expected returns, it calculates not the expected value over a specific time period, but the mathematical expectation of the total remaining lifetime value over possible future time spans. This solves the problem that traditional CLV calculations often require subjective setting of prediction periods or cannot handle infinite time horizons. For business decision-making, this value is a more comprehensive and stable long-term indicator, providing a solid quantitative basis for determining whether to make long-term investments in a customer. Therefore, this specific fusion algorithm is not a simple stacking of two models, but rather forms a complementary predictive value chain. For the common and challenging customer type of intermittent purchases, it provides an end-to-end automated calculation method that automatically learns behavioral patterns from historical data, accurately predicts future transaction times, and scientifically converts them into lifetime monetary value. This method greatly improves the accuracy and objectivity of CLV measurement for this customer group, provides reliable input for dynamic stratification and strategy-driven approaches, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0019] Optionally, the collaborative computing unit is specifically used to execute the following model selection logic: based on the customer dynamic feature vector obtained from the unified data platform, especially the coefficient of variation of historical transaction intervals and the most recent transaction time, to determine the customer's purchase behavior pattern; if it is determined to be intermittent purchase behavior and the most recent transaction time is less than or equal to a preset activity threshold, then the first prediction model is called to use the fusion algorithm of the negative binomial distribution model and the Markov chain model to predict the customer's lifetime value; if it is determined to be intermittent purchase behavior and the most recent transaction time is greater than the preset activity threshold, then the improved Pareto / NBD model is used; if it is determined to be regular purchase behavior, then the prediction is made using the Shifted-Beta-Geometric model; for new customers, the estimation is made using the mean initialization model based on similar customer groups.

[0020] By adopting the above technical solution, this logic intelligently determines purchasing behavior patterns based on customer dynamic feature vectors, particularly the coefficient of variation of historical transaction intervals and the time of the most recent transaction. It then assigns different prediction models or initialization methods to different patterns, introducing adaptability and context awareness into the customer lifetime value prediction module of this invention. This results in a significant improvement in the overall prediction accuracy, robustness, and practicality of the system. The core of this model selection logic lies in automating and objectively diagnosing complex customer behavior patterns through quantifiable statistical indicators, such as the coefficient of variation and the time of the most recent transaction. Using the coefficient of variation of historical transaction intervals to distinguish between intermittent and regular purchases is a scientific discrimination method based on the inherent fluctuation characteristics of data, and its effect is far superior to the crude classification relying on human experience or fixed rules. Combining the time of the most recent transaction further confirms the customer's current activity level, which helps maintain the best match between the subsequent value prediction model and the customer's actual behavioral characteristics, reducing the model's systematic prediction bias. Differentiated prediction strategies are adopted for different diagnostic results. For intermittent purchasing behavior, a fusion algorithm or an improved Pareto / NBD model is used based on the most recent transaction time. As mentioned earlier, this approach solves the prediction challenges for such customers. For regular purchasing behavior, a Shifted-Beta-Geometric model is used for prediction. This model typically has higher computational efficiency and goodness of fit for customers with frequent transactions and stable patterns. This choice optimizes the allocation of computational resources while ensuring prediction accuracy and avoids using unnecessary complex models for regular customers. For new customers, an initialization model based on the mean of similar customer groups is used. This provides a reasonable initial value estimate based on group characteristics, allowing the system to incorporate new customers into management from the earliest stage of their lifecycle, rather than waiting for them to accumulate sufficient data before analysis. This setup allows the system to automatically call the most appropriate prediction tool based on the specific customer's data characteristics, enhancing the system's generalization ability and prediction accuracy when facing heterogeneous customer groups. This ensures that different types of assets receive scientifically appropriate evaluations, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0021] Optionally, in the state inference unit, a trained recurrent neural network is used as the time series classification model. The input layer of the recurrent neural network receives the customer level encoding sequence arranged in chronological order provided by the dynamic mapping unit, as well as the customer lifetime value change rate and behavioral activity slope derived from the dynamic feature vector. The output layer of the recurrent neural network provides the state transition probability distribution vector, where each element represents the probability of a customer transitioning to the corresponding level.

[0022] By employing the aforementioned technical solution and selecting Recurrent Neural Networks (RNNs), particularly their variants such as Long Short-Term Memory (LSTM) networks, to handle customer state deduction problems, the inherent architectural advantages of RNNs—designed specifically for processing sequential data—are leveraged. The recurrent connections of RNNs enable them to retain historical information, making them suitable for analyzing time-dependent customer hierarchy encoding sequences. They can automatically capture and learn long-term dependencies and complex patterns in customer hierarchy transitions. For example, they might learn nonlinear patterns such as a high probability of quickly returning to the platinum level after a brief fall from the platinum level to the gold level if the CLV change rate is positive. This is something traditional Markov chains or logistic regression models struggle to discover automatically. The design of input features, such as hierarchy sequences, CLV change rates, and behavioral activity slopes, provides the RNN model with high-information-density learning materials. Inputting these heterogeneous but related temporal features into the RNN allows the model to perform multi-dimensional, multi-signal fusion analysis, thereby making more comprehensive and sensitive inferences about state transitions, significantly enhancing the timeliness and foresight of predictions. The output is a state transition probability distribution vector, where each element corresponds to the probability of transitioning to a specific level. This output format provides a quantitative and complete measure of uncertainty, offering rich decision-making support for the strategy-driven unit. For example, the system can trigger different strategy combinations based on two different probability distribution patterns: high probability of upgrading and high probability of churn. It can even calculate expected value changes for strategy benefit-cost assessment, achieving high-precision, fine-grained probability prediction of the complex time-series process of customer asset dynamics. This significantly improves the overall value and operational efficiency of customer assets, thereby enhancing customer asset management efficiency.

[0023] Optionally, it also includes a feedback optimization unit: used to collect customer feedback and behavior change data returned by the downstream business system after executing the automated operation strategy, as effect data, convert the effect data into reward signals, and iteratively optimize the parameters of the time series classification model in the state inference unit and the strategy matching logic of the strategy knowledge base in the strategy driving unit through reinforcement learning algorithms.

[0024] By adopting the above technical solutions, although traditional automated systems can execute actions according to rules, their rules and models are static and cannot learn from the execution results. The feedback optimization unit continuously collects customer feedback and behavioral change data, such as whether they respond to offers, whether they complete upgrade purchases, and whether the average order value increases, and systematically transforms this data into quantifiable performance data. This provides the system with an objective basis for self-evaluation, enabling the system to have the ability to self-diagnose based on actual business results and objectively measure the short-term and long-term impact of each automated decision. By applying reinforcement learning algorithms to handle feedback optimization, the entire customer asset management process is modeled as a sequential decision problem. The system selects strategy actions based on the current customer state, and the environment provides feedback. The system updates its decision model accordingly to maximize long-term cumulative rewards. This allows the computational model to be dynamically calibrated based on changes in actual customer behavior after strategy intervention, thereby increasingly reflecting the patterns of customer behavior under the influence of existing operational strategies and reducing model errors. The system can automatically identify and strengthen strategy rules or action combinations that bring high rewards in specific customer states, while weakening or eliminating ineffective strategies. This solves the rigidity and lag problems caused by traditional systems that rely on fixed rules and periodic manual tuning, enabling the system to maintain high adaptability and high operational efficiency, achieving sustainable growth in asset returns, and thus significantly improving the overall value and operational efficiency of customer assets, thereby enhancing the efficiency of customer asset management.

[0025] Optionally, the unified data platform also includes a customer asset quality evaluation unit: used to model the evolution of the hierarchical distribution of newly acquired customer groups in the hierarchical architecture as a three-state warehouse model, the three states corresponding to the steel heavy lead layer, the gold layer and the platinum layer, based on the real-time output of the collaborative computing unit and the dynamic mapping unit, to fit the time-varying transition rate parameters of customers between states, and to predict the long-term steady-state hierarchical composition ratio of the new customer group by solving the differential equation system corresponding to the three-state warehouse model, and to generate a quantitative score for the quality of new customer channels or marketing activities based on this ratio.

[0026] By adopting the above technical solutions, traditional evaluation methods typically rely on short-term, static indicators, such as the number of new customers, first-purchase conversion rate, or initial average order value. These indicators fail to reflect the long-term value potential and structural health of new customer groups, leading to short-sighted marketing budget allocation decisions. By introducing a warehouse model from system dynamics into the customer asset quality evaluation unit, dynamic simulation and forward-looking evaluation of the lifecycle value evolution of new customer groups are achieved. This unit enables dynamic and process-oriented tracking and modeling of the quality of new customer groups. It does not evaluate a static snapshot at a certain point in time, but continuously utilizes the real-time output of the collaborative computing unit and the dynamic mapping unit to fit time-varying transfer rate parameters between three representative asset quality states: steel heavy lead layer, gold layer, and platinum layer. For example, the rate of upgrading from steel heavy lead layer to gold layer, the rate of upgrading from gold layer to platinum layer, and the rate of reversal or churn. It aggregates discrete, individual hierarchical jump events into a continuous, group flow process for analysis, which can accurately depict the dynamic evolution trajectory and internal flow speed of the overall value structure of new customer groups. By solving the differential equations corresponding to the three-state warehouse model, the long-term steady-state hierarchical composition ratio when the new customer base reaches dynamic equilibrium is predicted. This extends the evaluation perspective from short-term fluctuations to long-term equilibrium, providing a stable quality assessment benchmark based on long-term value orientation. Compared to traditional indicators that only focus on initial purchase amount or short-term retention rate, steady-state ratio prediction is better able to identify high-quality customer acquisition channels or marketing activities that may have high initial conversion costs but huge potential for total customer lifetime value and good prospects for loyalty cultivation. Based on the predicted long-term steady-state hierarchical composition ratio, a quantitative score for the quality of new customer channels or marketing activities is generated. This score integrates the quantity and quality of customer assets, providing objective and long-term meaningful data support for marketing decisions. It constructs a feedback loop from front-end customer acquisition to back-end asset value assessment, enabling enterprises not only to manage existing assets but also to scientifically evaluate and optimize the quality of incremental assets. This guides future marketing budgets and resources, allowing for more precise allocation to channels or activities that can generate high-quality customer assets, thus achieving continuous optimization of marketing ROI. This setup applies insights from dynamic stratification and state transitions to new customer assessment. Through a system dynamics model, it enables a long-term, dynamic, and value-based scientific evaluation of customer acquisition effectiveness, significantly enhancing the foresight and scientific nature of corporate marketing strategic planning. This, in turn, significantly improves the overall value and operational efficiency of customer assets, and enhances the management efficiency of customer assets.

[0027] Optionally, it also includes a churn warning and intervention unit: for customers in the gold and platinum layers, continuously extract key covariates related to churn risk from the dynamic feature vector, input the key covariates into a pre-trained Cox proportional hazards regression model, calculate and update the individual churn risk index of the customer, and when the individual churn risk index exceeds the threshold dynamically calculated based on historical data, generate a high-risk warning signal and send it to the strategy driving unit.

[0028] The strategy-driven unit is also used to prioritize matching and activating customer retention strategies when the high-risk warning signal is received.

[0029] By adopting the above technical solutions, traditional methods often rely on batch scoring based on static classification models such as logistic regression, or simple rules such as the number of consecutive days of inactivity. These methods cannot handle censored data that has not churned during the customer observation period, and the response after an alert depends on manual intervention, which is inefficient and easily leads to missed opportunities for optimal retention. By using the Cox proportional hazards regression model as the core risk identification tool in the churn warning and intervention unit, the scientific rigor and timeliness of the warnings are significantly improved. The Cox model is a semi-parametric survival analysis model. Its core advantage lies in its ability to effectively utilize a complete sample of customer data, including censored data (i.e., customer data that has not yet resulted in churn at the end of the observation period), for training, thereby making fuller use of data information and obtaining a more robust risk estimate. The risk index output by the model, i.e., the risk function value, not only quantifies the relative probability of churn but also includes time-dimensional information, reflecting the dynamic evolution of risk as customer characteristics change. By continuously extracting key covariates highly correlated with churn from the latest dynamic feature vectors—such as a sharp drop in login frequency, a continuous decline in average order value, a surge in negative emotions in customer service interactions, and a spike in competitor app activity—this unit can update the individual churn risk index for each high-value customer in near real-time, achieving dynamic, continuous monitoring and quantification of churn risk. By setting thresholds dynamically calculated based on historical data—for example, using a high quantile of the historical churn customer risk index distribution as the threshold—and automatically generating a high-risk warning signal when this threshold is exceeded, this unit achieves automatic triggering and dynamic calibration of warnings. This ensures a reasonable balance between the sensitivity and specificity of warnings, enabling timely capture of genuine risk signals while effectively avoiding false alarms caused by noise or normal fluctuations, thus improving the overall reliability and usability of the warning system. When the strategy-driven unit receives a high-risk warning signal, it prioritizes and activates customer retention strategies. This allows intervention actions targeting high-risk, high-value customers to be automatically triggered and executed by the system with high priority and low latency. This significantly shortens the critical time window from risk identification to retention action, compressing the manual judgment and execution process, which may traditionally take days or even weeks, to minutes or even seconds. This provides an opportunity to implement effective intervention at the customer's decision-making critical point and successfully recover core assets. It effectively addresses the core objectives of reducing core customer churn and protecting the foundation of corporate profits, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0030] Optionally, it also includes a quantitative stripping decision unit: used to construct a discrete state space based on the time interval between the customer's last transaction and the transaction frequency within a preset time period, and use a Markov decision process model, combined with a preset state transition probability matrix and a revenue matrix, to solve for the optimal strategy that maximizes the customer's long-term expected net profit. The optimal strategy specifies the action of "continue investment" or "stop investment" for each state. For the customer group corresponding to the state indicated by the optimal strategy as "stop investment", the strategy driving unit is controlled to no longer match and distribute any resource investment-type operation strategies to them.

[0031] By adopting the above technical solution, the quantitative separation decision-making unit discretizes the two key behavioral dimensions of customer transaction intervals and transaction frequencies, constructing a discrete state space with clear business meaning. This simplifies complex historical customer behavior patterns into computable and analyzable system states, providing a stable modeling foundation for applying Markov decision processes, a stochastic dynamic programming tool, allowing customer behavior dynamics to be formally described and analyzed. This unit utilizes a Markov decision process model for optimization. The core inputs of the model include a state transition probability matrix estimated based on historical data, describing the probability of a customer transitioning from the current state to the next state, and a payoff matrix, describing the immediate net profit gained by the firm in each state by taking actions to continue or cease investment. By solving the MDP model, the optimal strategy that maximizes the long-term expected discounted net benefit is obtained, providing an objective and quantitative economic basis for divestiture decisions. For the entire customer group corresponding to the state of ceasing investment as determined by the optimal strategy, this unit directly controls the strategy-driven unit, so that it systematically ignores these customers in the subsequent strategy matching and distribution process, and no longer triggers any resource-consuming operational actions for them, such as sending promotional text messages, pushing personalized advertisements, or allocating customer service resources. This realizes the automated and large-scale execution of negative asset divestiture. Enterprises do not need manual screening and operation. The system can automatically identify and silently process these customers, thereby releasing marketing resources and reallocating them to customer groups with higher value potential. This improves the overall efficiency of marketing resource allocation, thereby significantly enhancing the overall value and operational efficiency of customer assets and improving the management efficiency of customer assets.

[0032] Optionally, the customer loyalty measurement in the collaborative computing unit specifically involves: pre-setting a structural equation model containing multiple latent variables such as customer satisfaction, customer trust, relationship commitment, and future loyalty; each latent variable is associated with multiple observation indicators from questionnaire scales or behavioral logs; periodically inputting the observation indicator data generated by the unified data platform into the measurement model; using a partial least squares algorithm for model fitting and validation; outputting the latent variable factor score for each customer; and using the factor score of future loyalty as the current loyalty measurement result.

[0033] By adopting the above-mentioned technical solutions, traditional methods of measuring loyalty either rely on single survey questions, such as Net Promoter Score (NPS), or use simple behavioral substitutes, such as repeat purchase rate. These methods fail to reveal the multidimensional composition of loyalty and its underlying formation mechanism, and the measurement results are often unstable and have limited explanatory power. This new solution establishes a solid scientific theoretical framework for loyalty measurement by pre-setting a multi-latent variable structural equation model. The latent variables included in the model systematically cover multiple progressive levels of loyalty, such as the cognitive evaluation basis, emotional dependence core, behavioral intention intensity, and long-term relationship orientation. Furthermore, the variables are pre-set with causal paths consistent with consumer behavior theory. Compared to single indicators, this multi-dimensional model can more comprehensively and profoundly diagnose the true quality and stability of customer relationships, providing rich insights for subsequent differentiation strategy development. It achieves multi-source data fusion measurement, improves the reliability and effectiveness of loyalty measurement, reduces the potential bias of single data sources, and makes the measurement results more robust, objective, and closer to reality. Future loyalty reflects customers' fundamental willingness to maintain long-term relationships, make incremental or cross-purchases, and actively recommend products. Using it as a core dimension of segmentation ensures that loyalty measurement and customer lifetime value prediction are highly aligned in terms of objectives. Together, they constitute a complete evaluation system for the current value and future potential of customer assets, providing this system with a scientific and model-based tool for measuring relationship quality. This provides direct data support and action direction for a deeper understanding of customers and the development of precise relationship strengthening and loyalty enhancement strategies, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0034] Optionally, the policy-driven unit specifically includes:

[0035] Strategy knowledge base: It contains pre-stored relationships, which are associated with hierarchical transition scenarios, triggering conditions and at least one automated operation action;

[0036] Strategy Matcher: Used to query the strategy knowledge base based on the customer's current level, target jump level, and real-time triggered behavioral events, and match the corresponding target automated operation actions;

[0037] Instruction Distributor: Used to convert the target automated operational actions into executable instructions and distribute them to downstream business execution systems.

[0038] By adopting the above technical solutions, the strategy knowledge base stores strategies in a structured, relational format, achieving modularity, scenario-based approach, and manageability. Operations personnel can design strategy packages around clear business objectives and set sophisticated, multi-condition triggering logic within them. For example, a strategy can be defined as follows: when the scenario is "upgrade from Steel Layer to Gold Layer," the triggering condition is "predicted upgrade probability > 40%" and "real-time event is adding a high-priced item to the favorites list," then the strategy will execute two actions: "pushing a coupon exclusively for that item" and "generating a notification in the customer service workbench." This greatly enhances the complexity and business relevance of strategy expression, supporting precise personalized intervention. The strategy matcher continuously monitors input streams from various parts of the system, including the customer's current static level, the predicted dynamic target level or transition probability, and real-time behavioral event streams captured by the unified data platform. The strategy matcher integrates these multi-source, heterogeneous input information and queries the strategy knowledge base in real time, achieving precise real-time strategy triggering based on multi-dimensional context awareness. This decision-making capability, which integrates long-term trends and instantaneous intentions, is something traditional rule engines lack. The instruction dispatcher transforms the abstract automated operational actions output by the strategy matcher into standardized instructions or API calls that can be recognized and processed by various downstream business execution systems, and ensures their reliable distribution. This makes the system highly scalable. When new marketing channels or business systems need to be integrated, only the corresponding adapter needs to be added to the instruction dispatcher. There is no need to change the core strategy management and matching logic, thus building an intelligent and easy-to-integrate strategy decision-making and execution engine.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] 1. The unified data platform transforms raw data from heterogeneous sources into standardized dynamic feature vectors through real-time data cleaning, correlation, and feature engineering. This provides a highly timely and dimensionally consistent input data foundation for upper-level analysis, overcoming the modeling difficulties and analysis delays caused by data dispersion and inconsistent formats in existing technologies. The collaborative computing unit performs customer lifetime value prediction and loyalty measurement in parallel. CLV prediction does not rely on a single fixed model but can adaptively call the most suitable prediction model for calculation based on customer behavior patterns, thereby significantly improving the accuracy of prediction and the applicability of the system. Especially for customer groups with complex and ever-changing transaction patterns, the system achieves accurate fitting of different behavioral characteristics (such as intermittent and regularity) through an intelligent model selection mechanism, solving the problem of large prediction deviations caused by model-behavior mismatch in traditional methods. Meanwhile, customer loyalty quantification uses structural equation modeling to fit multi-dimensional latent variable observation indicators, achieving a scientific and quantitative measurement of the abstract concept of customer loyalty. It transforms non-monetary values ​​such as reputation and trust, which are difficult to directly observe, into calculable indicators, providing key input for a comprehensive assessment of customer assets. The dynamic mapping unit uses the above two quantitative results to perform two-dimensional hierarchical classification. The classification results reflect the customer's current profitability and future relationship stability. The state inference unit predicts state transition probabilities using a time series model. The strategy-driven unit automatically matches and executes strategies based on the current level, predicted probabilities, and real-time events, associating complex analytical conclusions with specific business actions and achieving automated execution through instruction distribution. This solves the problems inherent in traditional methods. The system addresses the inefficiencies and missed opportunities inherent in the traditional chain of analysis reports, human decision-making, and manual configuration. It achieves integrated operation of dynamic assessment of customer asset status, risk warning and automation, and personalized intervention. Fluctuations in customer value can be perceived in real time and translated into specific actions through pre-set strategy logic, significantly improving the efficiency and accuracy of marketing resource allocation. This effectively stabilizes and enhances the total value of customer assets. This setup enables dynamic and automated customer asset management. The system not only accurately measures and classifies customer assets but also proactively predicts asset change trends and automatically executes precise strategies for intervention, upgrading, or divestiture, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0041] 2. By using the BG / NBD model for initial prediction and the Markov chain for subsequent valuation, this approach effectively solves the specific technical challenge of long-term value prediction for intermittently purchasing customers. This solution transforms the probability sequence output by BG / NBD, rather than single-point probabilities, into non-homogeneous state transition probabilities that vary over time within the Markov chain model. Compared to homogeneous Markov chains using fixed transition probabilities, non-homogeneous chains better reflect the reality that customer activity may decline or fluctuate over time, thus making the model's simulation of the customer lifecycle more realistic. By defining state-related returns, such as single-transaction profits, on this non-homogeneous Markov chain and solving for their steady-state or long-term expected returns, the problem of traditional CLV calculations often requiring subjective setting of the prediction period or being unable to handle infinite time horizons is resolved. This specific fusion algorithm is not a simple stacking of two models, but rather forms a predictive value chain with complementary functions. For the common and challenging customer type of intermittent purchases, it provides an end-to-end automated calculation method that automatically learns behavioral patterns from historical data, accurately predicts future transaction times, and scientifically converts them into lifetime monetary value. This method greatly improves the accuracy and objectivity of CLV measurement for this type of customer group, provides reliable input for dynamic stratification and strategy-driven approaches, and thus significantly improves the overall value and operational efficiency of customer assets, thereby enhancing the management efficiency of customer assets.

[0042] 3. By intelligently determining purchasing behavior patterns based on customer dynamic feature vectors and assigning different prediction models or initialization methods to different patterns, the customer lifetime value prediction module of this invention introduces adaptability and context awareness, resulting in a significant improvement in the overall prediction accuracy, robustness, and practicality of the system. Through quantifiable statistical indicators, such as the coefficient of variation and the most recent transaction time, complex customer behavior patterns are diagnosed automatically and objectively. The coefficient of variation of historical transaction intervals distinguishes between intermittent and regular purchases, with results far superior to coarse classifications relying on human experience or fixed rules. Combining the most recent transaction time further confirms the customer's current activity level, which helps maintain the best match between the subsequent value prediction model and the customer's actual behavioral characteristics, reduces the model's systematic prediction bias, optimizes the allocation of computing resources, and avoids using unnecessary complex models for regular customers. With this setup, the system can automatically call the most suitable prediction tool based on the specific customer's data characteristics, enhancing the system's generalization ability and prediction accuracy when facing heterogeneous customer groups. This allows different types of assets to receive scientifically appropriate evaluations, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0043] 4. The feedback optimization unit continuously collects customer feedback and behavior change data, systematically transforming it into quantifiable performance data. This provides the system with objective evidence for self-evaluation, enabling it to self-diagnose based on actual business results and objectively measure the short-term and long-term impact of each automated decision. Reinforcement learning algorithms are applied to handle feedback optimization, modeling the entire customer asset management process as a sequential decision problem. The system selects strategy actions based on the current customer state, receives feedback from the environment, and updates its decision model accordingly to maximize long-term cumulative rewards. This allows the computational model to dynamically calibrate based on changes in actual customer behavior after strategy intervention, increasingly accurately reflecting customer behavior patterns under the influence of existing operational strategies, reducing model errors. The system can automatically identify and strengthen strategy rules or action combinations that bring high rewards in specific customer states, while weakening or eliminating ineffective strategies. This solves the rigidity and lag problems caused by traditional systems relying on fixed rules and periodic manual tuning, ensuring the system maintains high adaptability and operational efficiency.

[0044] 5. By introducing the compartmental model from system dynamics through the customer asset quality evaluation unit, dynamic simulation and forward-looking assessment of the lifecycle value evolution of new customer groups are realized. This unit enables dynamic and process-oriented tracking and modeling of the quality of new customer groups. It does not evaluate a static snapshot at a certain point in time, but continuously uses the real-time output of the collaborative computing unit and the dynamic mapping unit to fit the time-varying transfer rate parameters of customers between three representative asset quality states: steel heavy lead layer, gold layer and platinum layer. It can accurately depict the dynamic evolution trajectory and internal flow speed of the overall value structure of new customer groups. By solving the differential equations corresponding to the three-state warehouse model, the long-term steady-state hierarchical composition ratio when the new customer base reaches dynamic equilibrium is predicted. This extends the evaluation perspective from short-term fluctuations to long-term equilibrium. Based on the predicted long-term steady-state hierarchical composition ratio, a quantitative score for the quality of new customer channels or marketing activities is generated. This score integrates the quantity and quality of customer assets, providing objective and long-term data support for marketing decisions. It constructs a feedback loop from front-end customer acquisition to back-end asset value assessment, enabling enterprises not only to manage existing assets but also to scientifically evaluate and optimize the quality of incremental assets. This guides future marketing budgets and resources, allowing for more precise allocation to channels or activities that can generate high-quality customer assets, thereby achieving continuous optimization of marketing ROI. Attached Figure Description

[0045] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort. In the drawings:

[0046] Figure 1 This is a block diagram of a customer asset intelligent management system based on a hierarchical dynamic driving model, according to an embodiment of the present invention.

[0047] Figure 2 This is a schematic diagram of the data flow of a customer asset intelligent management system based on a hierarchical dynamic driving model, according to an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the model selection logic of a customer asset intelligent management system based on a hierarchical dynamic driving model according to an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram of a two-dimensional value mapping matrix for a customer asset intelligent management system based on a hierarchical dynamic driving model, according to an embodiment of the present invention. Detailed Implementation

[0050] The following will be based on embodiments of the present invention. Figures 1-4 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0051] This embodiment discloses a customer asset intelligent management system based on a hierarchical dynamic driving model, referring to... Figures 1-4 It includes a unified data platform, a layered dynamic driving engine, a strategy-driven unit, a feedback optimization unit, a churn warning and intervention unit, and a quantitative stripping decision-making unit.

[0052] Unified Data Platform: Used to access and integrate raw customer data from transaction systems, interaction platforms and external data sources in real time. Through data cleaning, entity association and feature engineering, it outputs dynamic feature vectors with unique customer identifiers as keys, containing time-series behavioral tags and statistical indicators. It includes user behavior data collection modules, transaction data collection modules, loyalty data collection modules and customer asset quality evaluation units.

[0053] User behavior data collection module: used to capture the sequence of customers' online browsing, clicking, searching and interaction behaviors, from front-end tracking logs, customer service system and social media, including user ID, event time, event type, page / product ID, dwell time, etc.

[0054] Transaction data acquisition module: used to record customers' purchase time, frequency, amount, category and payment information, which comes from the order database.

[0055] Loyalty data collection module: Used to collect data on customer satisfaction, willingness to recommend, and commitment levels through questionnaires, ratings, or interactive feedback channels.

[0056] Customer Asset Quality Evaluation Unit: This unit models the evolution of the hierarchical distribution of newly acquired customers within the hierarchical architecture as a three-state warehouse model. The three states correspond to the combined steel and heavy lead layer, the gold layer, and the platinum layer. Based on the real-time output of the collaborative computing unit and the dynamic mapping unit, it fits the time-varying transition rate parameters of customers between states. By solving the differential equations corresponding to the three-state warehouse model, it predicts the long-term steady-state hierarchical composition ratio of the new customer group and generates a quantitative score for the quality of new customer channels or marketing activities based on this ratio.

[0057] The customer asset quality evaluation unit models the flow of new customers as a three-state compartment model, whose differential equations are expressed as follows:

[0058] set up , , These represent the number of customers in the steel-lead layer, gold layer, and platinum layer at time t, respectively. , These represent the upgrade rate constants from the steel-lead layer to the gold layer, and from the gold layer to the platinum layer, respectively. , This is the corresponding degradation or loss rate constant. Let be the rate at which new customers flow into the steel heavy lead layer. Then the kinetic equation of the system can be expressed as:

[0059] ;

[0060] ;

[0061] ;

[0062] After obtaining the rate parameters by fitting actual data, setting the left side of the equation system to zero allows us to solve for the steady-state solution. Then, the ratio of platinum to gold layers under steady state is calculated as the basis for quality scoring.

[0063] The core processing flow of the unified data platform is illustrated as follows: Raw data is deduplicated, outliers are handled, and formats are standardized. Using the user ID as the core, all records of the same user across different data sources and timestamps are linked to form a raw behavior log at the user granularity. Based on this behavior log, two types of features are continuously calculated and updated.

[0064] Statistical indicators: such as total transaction amount in the past 30 days, number of logins in the past 7 days, historical average order value, and number of days since the last purchase.

[0065] Time-series behavior tags: Using a sliding window, generate Boolean or enumerated event tags such as "Browsed high-end products but did not place an order in the past 24 hours" or "Average order value has decreased by more than 20% in the past week".

[0066] Finally, a dynamic feature vector is generated for each active user. The vector can be in JSON format or a record from a specific vector database. This vector is pushed to a cache such as Redis and a feature library such as Feast, for downstream engines to subscribe to and query in real time.

[0067] When constructing dynamic feature vectors, the unified data platform employs a multi-frequency data alignment and asynchronous update mechanism for data sources with different update frequencies (such as second-level transaction data, daily log data, and monthly questionnaire data). This mechanism uses high-frequency timestamps as a benchmark and generates synchronous feature values ​​for low-frequency data using methods such as forward padding or linear interpolation. This ensures that the feature vector input to the engine is complete and time-consistent at any given real-time calculation moment. Simultaneously, the system possesses data loss tolerance capabilities: when low-frequency feature data, such as loyalty questionnaire data, is temporarily missing, the system automatically calls the historical average of that indicator for that customer or a proxy indicator calculated based on their recent transaction behavior (such as repeat purchase rate) as a temporary substitute, maintaining the uninterrupted operation of the strategy-driven unit.

[0068] An example of the application of the Customer Asset Quality Evaluation Unit is as follows: After a large-scale promotional event, this unit evaluates the quality of 100,000 newly acquired customers, modeling customer tier flow as a three-state warehouse model: X1 (steel / heavy lead), X2 (gold), and X3 (platinum). Over the next 90 days, the unit uses the outputs of the collaborative computing unit and the dynamic mapping unit to statistically analyze the flow of people between each state, fitting transfer rate parameters such as k12 (steel -> gold) and k23 (gold -> platinum). A system of differential equations is solved to predict the steady-state proportion of this batch of customers in the long term (e.g., after 2 years). If the predicted proportion of platinum + gold tiers is higher than the historical average, the quality of new customers acquired during the promotional event is deemed "excellent," and a quantitative score (e.g., 85 points) is generated. This score is used to guide the allocation of future marketing budgets.

[0069] Hierarchical Dynamic Drive Engine: Used to dynamically assess and stratify customers and predict future state changes, including collaborative computing unit, dynamic mapping unit and state inference unit.

[0070] Collaborative computing unit: used to perform customer lifetime value (CLV) prediction and customer loyalty measurement in parallel based on the dynamic feature vector. The customer lifetime value prediction calls different prediction models for calculation according to customer behavior patterns. The customer loyalty measurement is performed by fitting the preset latent variable observation indicators through structural equation model. Based on the customer dynamic feature vector obtained from the unified data platform, especially the coefficient of variation of historical transaction intervals and the most recent transaction time, the customer purchase behavior pattern is determined.

[0071] The collaborative computing unit is equipped with a first prediction model, which is a fusion algorithm that combines the BG / NBD model and the Markov chain model. Its specific execution process is as follows: using the BG / NBD model to process the customer's historical transaction time and number of transactions in the dynamic feature vector, predicting the customer's transaction probability in a specific future period, dynamically constructing the transaction probability sequence into a non-homogeneous state transition matrix in the Markov chain model, and outputting the customer's total remaining lifetime value by solving the steady-state expected return of the Markov chain.

[0072] If the purchase behavior is determined to be intermittent and the recent transaction time is less than or equal to the preset activity threshold, the first prediction model is invoked to use the fusion algorithm of the BG / NBD model and the Markov chain model to predict the customer's lifetime value.

[0073] If the purchase behavior is determined to be intermittent and the most recent transaction time is greater than the preset activity threshold, the improved Pareto / NBD model will be used.

[0074] If the purchase behavior is determined to be regular, a prediction based on the Shifted-Beta-Geometric (sBG) model is used. This type of model is specifically designed to describe customer retention and renewal behavior in contractual situations.

[0075] For new customers with sparse data, a mean-based initialization model based on similar customer groups is used for estimation. The similarity is determined by calculating the Euclidean distance or cosine similarity between the new customer's channel source, geographic attributes, and first-order feature vector and the existing customer group. Then, a clustering algorithm (such as K-Means) is used to match the new customer to the closest existing customer cluster, and the historical average CLV of the customers in that cluster is used as its initial estimate.

[0076] The customer loyalty measurement specifically involves: pre-setting a structural equation model containing multiple latent variables such as customer satisfaction, customer trust, relationship commitment, and future loyalty; each latent variable is associated with multiple observation indicators from questionnaires or behavioral logs; periodically inputting the observation indicator data generated by the unified data platform into the measurement model; using a partial least squares algorithm for model fitting and validation; outputting the latent variable factor score for each customer; and using the factor score of future loyalty as the current loyalty measurement result.

[0077] The core of the fusion algorithm is to transform the future transaction probability sequence predicted by the BG / NBD model into a non-homogeneous Markov chain model, and then calculate the expected discounted return of this Markov chain to solve for the total remaining lifetime value of the customer. The specific steps are as follows:

[0078] First, using the BG / NBD model, input the customer's historical transaction count x and the most recent transaction time t. x And a preset activity threshold (observation period length) T is used to predict the conditional probability p(τ) of the customer making a transaction in the τth time unit (e.g., a month). The preset activity threshold is used to distinguish between "active intermittent users" and "potentially churned users" (e.g., order cycles twice the industry average), thereby deciding whether to initiate a Markov chain for in-depth valuation.

[0079] Secondly, a two-state nonhomogeneous Markov chain is constructed, with a state space S = {0 (dormant), 1 (active)}. The probability of transitioning from state 'dormant' to state 'active' at time τ is defined as the transaction probability p(τ) predicted by BG / NBD, while the probability of transitioning from 'active' to 'dormant' is a fixed churn rate q (which can be estimated from historical data or used as a model parameter). The nonhomogeneous one-step state transition matrix P(τ) corresponding to time τ is: P(τ) = [[1-p(τ),p(τ)];[q,1-q]], where the rows and columns of the matrix correspond to states 0 and 1, respectively. Then, the revenue vector R = [0,m] is defined, where m represents the average gross profit brought to the enterprise by a single transaction by a customer in the 'active' state.

[0080] Finally, the total remaining lifetime value (CLV) of the customer can be obtained by calculating the expected discounted returns of the non-homogeneous Markov chain over an infinite time span, expressed as: CLV = ∑{τ=1}^{∞}[(1+d)^{-τ}*(initial state distribution vector)*(∏{t x =1}^{τ-1}P(t x ))*R].

[0081] Where d is the discount rate, and the initial state is usually set to [1,0] (i.e., currently in a 'dormant' state). In actual calculations, approximate values ​​can be obtained within a finite time horizon using dynamic programming or iterative methods.

[0082] The core processing flow of the collaborative computing unit is exemplified by the following: the unit performs two core calculations in parallel, namely customer lifetime value (CLV) prediction and customer loyalty measurement.

[0083] Customer Lifetime Value Prediction: The unit first analyzes the user's dynamic feature vector. It calculates the coefficient of variation (CV) of historical transaction intervals. If CV > 0.5 and the most recent transaction is recent, it is determined to be intermittent purchasing behavior; if CV ≤ 0.5 and the transactions are regular, it is determined to be regular purchasing behavior; if the customer is a new customer (less than 2 transaction records), it enters the new customer processing flow. Here, CV (coefficient of variation) is the ratio of the standard deviation to the mean of historical order intervals, used to quantify the heterogeneity of transaction frequency.

[0084] Intermittent Purchases: For customers identified as exhibiting intermittent purchase behavior, if the most recent transaction time is less than or equal to a preset activity threshold, a fusion algorithm combining a negative binomial distribution model and a Markov chain model is used to predict customer lifetime value. If the most recent transaction time exceeds the preset activity threshold, an improved Pareto / NBD model is employed. The improved Pareto / NBD model includes four key parameters: transaction rate parameter λ, churn rate parameter μ, shape parameter r, and scale parameter α. These parameters are derived from historical customer transaction data using maximum likelihood estimation. In practice, the values ​​of parameters λ and μ are typically related to industry and business characteristics. For example, in a retail e-commerce scenario, λ might range from [0.1, 10] (representing the average transaction rate), and μ might range from [0.01, 2] (representing the average churn rate). The shape parameter r and scale parameter α are positive numbers, and their values ​​determine the heterogeneity of customer activity. They are usually obtained by fitting historical data, for example, r ∈ [0.5, 5] and α ∈ [1, 20]. The system periodically (e.g., monthly) refits the model parameters using the latest data to ensure predictive adaptability.

[0085] Regular purchases: Prediction is made using a Shifted-Beta-Geometric model, a common process in this field.

[0086] New Customers: Find customers in the feature database who are similar to the new customer in terms of demographic attributes and source channels, and use their average CLV as the initial estimate.

[0087] Customer loyalty measurement: A structural equation model (SEM) is pre-set, which includes four latent variables: customer satisfaction (observed indicators: overall satisfaction score, gap with expectations score), customer trust (observed indicators: trust in product quality, trust in after-sales service), relationship commitment (observed indicators: willingness to continue using, willingness to recommend), and future loyalty (observed indicators: willingness to pay a premium, willingness to cross-purchase). Some of these observed indicator data come from periodic NPS questionnaires, and the other part is inferred from behavioral data (e.g., "willingness to recommend" can be replaced by sharing behavior).

[0088] At the beginning of each month, questionnaire data and behavioral inference indicators from the past month are input into the model. Partial Least Squares-SEM (PLS-SEM) is used for model fitting and path coefficient estimation to calculate the factor scores for each user on each latent variable. The factor scores of the future loyalty latent variable are standardized from 0 to 100 and used as the current loyalty metric for that user.

[0089] Dynamic mapping unit: Used to take the customer lifetime value prediction result and the customer loyalty measurement result as inputs to the first coordinate axis and the second coordinate axis respectively, and map the customer to a specified level in a hierarchical architecture consisting of a platinum layer, a gold layer, a steel layer and a heavy lead layer in real time through a preset two-dimensional threshold grid.

[0090] The dynamic mapping unit receives the output of the collaborative computing unit. During initialization, the system administrator can set thresholds through the management interface. The mapping unit then maps users to levels in real time according to the following rules:

[0091] Platinum layer: CLV ≥ high threshold and loyalty ≥ high threshold;

[0092] Gold Layer: CLV ≥ High Threshold and Loyalty < High Threshold; or Low Threshold ≤ CLV < High Threshold and Loyalty ≥ High Threshold;

[0093] Steel layer: CLV < low threshold and loyalty ≥ low threshold; or low threshold ≤ CLV < high threshold and loyalty < high threshold;

[0094] Heavy lead layer: CLV < low threshold and loyalty < low threshold.

[0095] Each mapping result will be updated in the user profile and the history of hierarchical changes will be recorded.

[0096] The criteria for dividing the two-dimensional threshold grid are jointly set by business experts based on historical data distribution and strategic objectives, and support dynamic adjustment. (Refer to...) Figure 4 A specific example of this division is as follows:

[0097] Customer lifetime value (CLV) predictions were standardized to a score of 0-100, and customer loyalty measurement results were also standardized to a score of 0-100.

[0098] Platinum layer threshold: CLV ≥ 80 points and loyalty ≥ 80 points.

[0099] Gold layer threshold: CLV ≥ 80 points and loyalty < 80 points; or CLV ∈ [60, 80) points and loyalty ≥ 70 points.

[0100] Steel layer threshold: CLV < 60 points and loyalty ≥ 50 points; or CLV ∈ [60, 80) points and loyalty < 70 points.

[0101] Heavy lead layer threshold: CLV < 60 points and loyalty < 50 points.

[0102] In actual deployment, the above thresholds can be dynamically calibrated by analyzing the joint distribution of historical customer value and loyalty (e.g., using quantiles) and combining it with business objectives (e.g., aiming for a 10% platinum customer share).

[0103] Reference Figure 4 This is a schematic diagram of the two-dimensional value mapping matrix used by the dynamic mapping unit in this embodiment, which is used to demonstrate the hierarchical division logic based on the two dimensions of customer lifetime value (CLV) and customer loyalty.

[0104] Specifically, the system uses a two-dimensional value mapping matrix to transform the quantified results output by the collaborative computing unit into business-executable customer tiers. The construction and mapping logic of this matrix are as follows:

[0105] Coordinate axis definition:

[0106] The horizontal axis (X-axis) represents the Customer Lifetime Value (CLV) score, which is output by the collaborative computing unit and standardized from 0 to 100, reflecting the customer's potential for long-term economic contribution to the enterprise.

[0107] Vertical axis (Y-axis): Represents the customer loyalty score based on structural equation model (SEM) factor scores and standardized from 0 to 100 by the collaborative computing unit, reflecting the stability of the relationship between customers and enterprises and the strength of psychological contract.

[0108] Four-quadrant hierarchical division:

[0109] Platinum Layer (Core Value Zone): Located in the upper right corner of the matrix. The criteria for selection are CLV ≥ high threshold (e.g., 80 points) and loyalty ≥ high threshold (e.g., 80 points). These customers are the core assets of the company, possessing extremely high economic value and brand recognition.

[0110] Gold Tier (Potential Enhancement Zone): The criteria for selection are: 1) CLV ≥ High Threshold and Loyalty < High Threshold; 2) Low Threshold (e.g., 60 points) ≤ CLV < High Threshold and Loyalty ≥ Medium Threshold (e.g., 70 points). These customers exhibit "high value, low to medium loyalty" or "medium to high value, high loyalty" and are the focus of enhancement and maintenance efforts.

[0111] Steel Layer (Basic Maintenance Zone): The following criteria must be met: 1) CLV < low threshold and loyalty ≥ low threshold (e.g., 50 points); 2) low threshold ≤ CLV < high threshold and loyalty < medium threshold. These customers are primarily composed of mass-market consumers, exhibiting significant value fluctuations but also a degree of stability.

[0112] Heavy Lead Layer (Stripping / Observation Area): Located in the lower left corner of the matrix. The criteria are CLV < low threshold and loyalty < low threshold. These customers not only have low economic contribution but also lack emotional connection, making them the primary target for the system's quantitative stripping decisions.

[0113] Threshold dynamic calibration mechanism: Figure 4 The thresholds shown (e.g., 80, 70, 60, 50) are not fixed. System administrators can dynamically calibrate these thresholds through the management interface based on the joint distribution of historical customer value (e.g., using quantiles) and in conjunction with business strategic objectives, thereby achieving flexible control over the proportion of customers at each level.

[0114] State deduction unit: used to deduce the state transition probability distribution vector of customers from the current level to other levels in the future preset time period based on the customer's historical level sequence and the behavioral trend indicators in the dynamic feature vector, through a time series classification model;

[0115] The recurrent neural network is used as the time series classification model. The input layer of the recurrent neural network receives the customer level encoding sequence arranged in chronological order provided by the dynamic mapping unit, as well as the customer lifetime value change rate and behavioral activity slope derived from the dynamic feature vector. The output layer of the recurrent neural network provides the state transition probability distribution vector, where each element represents the probability of a customer transitioning to the corresponding level.

[0116] The recurrent neural network (such as LSTM) used in the state extrapolation unit is trained with the goal of minimizing prediction error. Specifically, the network uses a cross-entropy loss function, and the training objective is to minimize the difference between the network's output state transition probability distribution vector and the one-hot encoded vector of the customer's actual hierarchical transition in the next observation period. Training with a large amount of historical customer hierarchical sequence data (such as monthly hierarchical snapshots from the past 24 months) enables the network to learn complex temporal patterns of hierarchical transitions.

[0117] Application Example: The recurrent neural network used in the state deduction unit is a Long Short-Term Memory (LSTM) network, with the following specific network structure parameters: The network consists of one input layer, two LSTM hidden layers, and a fully connected output layer. The input layer dimension corresponds to the number of input features, such as customer hierarchy encoding (one-hot, 4D), CLV change rate (1D), behavioral activity slope (1D), etc., totaling 6 dimensions. The first LSTM hidden layer contains 128 neurons, and the second LSTM hidden layer contains 64 neurons, both using the tanh activation function. The output layer has 4 neurons, corresponding to the transition probabilities of the four levels, and uses the Softmax activation function to generate the probability distribution. The network is trained using the Adam optimizer with an initial learning rate of 0.001, and Dropout (dropout rate 0.2) is used to prevent overfitting.

[0118] Strategy-driven unit: Based on the customer's current level, predicted state transition probability, and real-time triggered behavioral events, it matches the corresponding automated operation strategy from the strategy knowledge base and converts it into executable instructions to be distributed to downstream business systems. It includes a strategy knowledge base, a strategy matcher, and an instruction distributor.

[0119] When the high-risk warning signal is received, customer retention strategies will be prioritized and activated.

[0120] Strategy knowledge base: Pre-stored relationships, which are associated with hierarchical transition scenarios, triggering conditions and at least one automated operation action.

[0121] Strategy Matcher: Used to query the strategy knowledge base based on the customer's current level, target jump level, and real-time triggered behavioral events, and match the corresponding target automated operation actions.

[0122] Instruction Distributor: Used to convert the target automated operational actions into executable instructions and distribute them to downstream business execution systems.

[0123] An example of the internal operation of the strategy-driven unit is as follows:

[0124] Strategy Knowledge Base: Stores a large number of condition-action rules. For example, when a customer triggers the condition: the predicted upgrade probability is >40% and the real-time event is "high-priced items added to the shopping cart have not been paid for for more than 24 hours", the customer is upgraded from the Steel Layer to the Gold Layer. The triggering strategy is to send an APP message containing an exclusive coupon for the product, generate a prompt in the customer service workbench, and suggest that a dedicated consultant communicate with the customer online.

[0125] As a real-time event processor, the policy matcher monitors two information streams: one is the periodic prediction results from the state deduction unit (event stream A), and the other is real-time user behavior events from the unified data platform (event stream B). When, for a given user, both the high-probability transition warning in event stream A and the specific triggering condition in event stream B are met simultaneously, the matcher immediately activates the corresponding policy rule.

[0126] The instruction dispatcher converts action sequences into instructions that can be executed by downstream systems. For example, it converts sending an app message into calling an API call to the company's internal message push service, and converts generating customer service prompts into mentioning relevant customer service personnel in the WeChat Work robot.

[0127] Feedback Optimization Unit: Used to collect customer feedback and behavior change data returned by downstream business systems after executing the automated operation strategy, as effect data, and convert the effect data into reward signals. Iterative optimization is performed on the parameters of the time series classification model in the state inference unit and the strategy matching logic of the strategy knowledge base in the strategy driving unit through reinforcement learning algorithm.

[0128] The feedback optimization unit employs a reinforcement learning algorithm, the core of which is defining a reward function directly linked to the long-term value growth of customer assets. The entire system is viewed as a Markov Decision Process (MDP), and the operational flow at each time step (e.g., day or week) is as follows: the system observes the customer's current state. (Composed of dynamic feature vectors and hierarchies, etc.), the policy-driven unit selects and executes an operational action based on policy π. (For example, sending coupons), and then the customer generates new behavioral data, and the system transitions to a new state. And receive an instant reward. .

[0129] The instant reward The design goal is to quantify actions. The immediate impact on long-term customer value. Specifically defined as:

[0130] ,in, This refers to the change in the customer's future lifetime value (CLV) after the action is executed at time step t, as predicted by the collaborative computing unit, compared to the predicted value before the action was executed. This directly measures the expected impact of the action on the monetary value of the customer's assets. It is the resource cost consumed in performing the action (such as marketing costs, customer service manpower costs). β is a scaling factor used to balance the dimension of value change and cost, and can also be regarded as a discount factor for long-term value changes.

[0131] The goal of a reinforcement learning agent is to learn an optimal policy π* to maximize the expected cumulative discounted reward from the current state to the future: E[∑{k=0}^{∞}γ^k*r{t+k}], where γ is the discount factor. This is achieved by continuously collecting interaction data ( , , , By using algorithms such as Q-learning and policy gradient to update the policy, the system can automatically tend to execute operations that bring long-term net growth in customer assets.

[0132] An example of the feedback optimization unit's processing flow is as follows: Track each executed policy; for example, record whether the user used the coupon and completed an order within 48 hours after the rule was executed, and the order amount. The entire system is modeled as a Markov Decision Process (MDP), where the state is the user's dynamic feature vector and current level, the actions are various policies in the policy knowledge base, and the reward is the change in CLV (Cost Per Volume) after policy execution (which can be replaced by conversion profit in the short term). Using Q-Learning or policy gradient methods, a neural network-driven Q-value function or policy function is periodically updated with collected empirical data. The output of this function can be used to adjust the policy matching logic, prioritizing policies with high Q-values, and indirectly guiding feature weights, influencing hierarchical structure and prediction.

[0133] Churn warning and intervention unit: For customers in the gold and platinum layers, continuously extract key covariates related to churn risk from the dynamic feature vector, input the key covariates into the pre-trained Cox proportional hazards regression model, calculate and update the individual churn risk index of the customer, and when the individual churn risk index exceeds the threshold dynamically calculated based on historical data, generate a high-risk warning signal and send it to the strategy driving unit.

[0134] The key covariates do not include data such as the real-time activity of competitor apps that may involve compliance risks, but instead use industry trend data, publicly available data, or alternative indicators built based on the company's own user surveys obtained through legally authorized third-party data platforms.

[0135] The list of key covariates includes, but is not limited to, the following quantifiable metrics, all of which are extracted or derived from the dynamic feature vector of the unified data platform:

[0136] Trading activity metrics: Recency (number of days since the last transaction), Frequency (number of transactions in the last 30 days), and Monetary amount (transaction amount in the last 30 days) compared to the previous period.

[0137] Interaction engagement metrics: Number of logins in the past 7 days, average dwell time in the past 7 days, and number of customer service interactions (including complaints) in the past 30 days.

[0138] Value change metrics: recent monthly change rate of Customer Lifetime Value (CLV) and slope of average order value trend.

[0139] Competitive product attention index (based on legal authorization or survey data): The competitor product usage tendency index of the customer's segment group as shown in the industry report obtained through a third-party data platform, or the customer's competitor product usage frequency score (1-5 points) based on internal surveys.

[0140] Service contact metrics: most recent customer service satisfaction rating (if any), response status of the most recent marketing campaign (whether or not a response was received).

[0141] Lifecycle stage indicator: Duration of customer-business relationship (in months).

[0142] These covariates need to be standardized before model training, and a subset that is significantly related to churn risk should be selected for use in the final model using feature selection methods (such as the Cox model based on LASSO).

[0143] An example of the churn warning and intervention unit's application is as follows: A Cox model is run daily only for Gold and Platinum level users. Key covariates in the model include: login frequency decline rate, month-on-month change in average order value, recent customer service complaint count, and competitor app activity. The model is trained based on churn data from the past 24 months and can output the risk ratio for each covariate.

[0144] The system calculates the individual churn risk index for each user, which is the cumulative risk function value based on the current covariate value. The threshold is set as the 80th percentile of the risk index of historical churned users. When the user's risk index exceeds the threshold, the system immediately generates an early warning. After receiving this early warning signal, the strategy-driven unit will trigger a high-priority "churn recovery" scenario matching, which may execute powerful retention strategies such as "emergency call back from customer manager" or "gifting high-value benefits".

[0145] The quantitative stripping decision unit is used to construct a discrete state space based on the time interval between a customer's last transaction and the transaction frequency within a preset time period. Utilizing a Markov decision process model and combining a preset state transition probability matrix and payoff matrix, it solves for the optimal strategy that maximizes the customer's long-term expected net profit. The optimal strategy specifies either "continue investment" or "stop investment" for each state. For customer groups corresponding to states where the optimal strategy indicates "stop investment," the strategy-driven unit is controlled to no longer match or distribute any resource-investment-based operational strategies to them. The decision results of this unit serve as a management aid tool, and the system is designed with interfaces for manual review and intervention to ensure the transparency and controllability of automated decision-making.

[0146] In the quantitative divestiture decision unit, the immediate revenue value in the revenue matrix is ​​generated by real-time netting of the "customer remaining lifetime value" output by the collaborative computing unit and the "resource input cost" returned by the downstream business system. When constructing the Markov Decision Process (MDP), each element R(s,a) of its revenue matrix R represents the immediate net revenue obtained by taking action a (continue investing / stop investing) in state s (defined by the transaction interval level R and the transaction frequency level F). This net revenue is not subjectively set, but is automatically calculated by the technology closed loop: R(s,a='continue investing')=λ*CLV_remaining(s)-Cost(s). Wherein, CLV_remaining(s) is the expected value of the remaining lifetime value estimated in real time by the collaborative computing unit based on the customer characteristics of this state; Cost(s) is the average cost feedback of the downstream business system performing one "investment" action (such as sending marketing information); λ is an adjustment coefficient between 0 and 1, used to discount long-term value into immediate revenue. By solving this MDP using strategy iteration or value iteration algorithms, a stripping strategy based on long-term economic optimality can be obtained.

[0147] Specific application examples are as follows:

[0148] User Zhang's behavior on the e-commerce platform triggered the system's full-chain operation. The data platform recorded that Zhang browsed high-end laptops multiple times but did not place an order (behavioral tag), and his average order value has recently decreased (statistical indicator).

[0149] Engine calculation: The collaborative computing unit determined that the purchase was intermittent, with a predicted CLV of 3000 yuan and a loyalty quantification score of 72 (slightly lower due to recent complaints). The dynamic mapping unit classified it into the Gold layer, while the state inference unit predicted a 25% probability of downgrading it to the Steel layer next month.

[0150] The churn warning and intervention unit calculates a high churn risk index and triggers an alert based on the decline in average order value and complaint records. Simultaneously, the real-time event of Mr. Zhang "adding a high-end computer to his cart more than 48 hours ago" is captured. The strategy-driven unit receives both the "churn warning" and the "adding a cart but not paying" event, matching them with a composite strategy: immediately sending a "dedicated manager service" text message informing him that feedback has been received and will be followed up; issuing a "dedicated concern-removal coupon" for the laptop to his account (which can be redeemed for an in-depth video sales guide or an additional extended warranty).

[0151] Feedback optimization: Zhang used the video shopping guide service and finally placed an order. This positive feedback was recorded and used to strengthen the Q value of the "human service + exclusive benefits" combination strategy in the scenario of "gold layer + churn risk + high-priced items added to cart".

[0152] The implementation principle of the customer asset intelligent management system based on a hierarchical dynamic driving model in this embodiment is as follows:

[0153] This system addresses the real-time challenge of high- and low-frequency data fusion through a multi-frequency data alignment mechanism; it enhances the universality and accuracy of CLV prediction through precise model selection logic that aligns with academic consensus; and by linking the revenue matrix with CLV prediction in real time, it decouples quantitative stripping decisions from subjective business rules, thus forming a purely technical closed-loop optimization system.

[0154] The unified data platform transforms raw data from heterogeneous sources into standardized dynamic feature vectors by performing real-time data cleaning, correlation, and feature engineering. This provides a highly timely and dimensionally consistent input data foundation for upper-level analysis, overcoming the modeling difficulties and analysis delays caused by data dispersion and inconsistent formats in existing technologies.

[0155] Traditional evaluation methods typically rely on short-term, static indicators such as the number of new customers, first-purchase conversion rate, or initial average order value. These methods fail to reflect the long-term value potential and structural health of new customer groups, leading to short-sighted marketing budget allocation decisions. By introducing a warehouse model from system dynamics into the customer asset quality evaluation unit, we can achieve dynamic simulation and forward-looking evaluation of the lifecycle value evolution of new customer groups. This unit enables dynamic and process-oriented tracking and modeling of the quality of new customer groups. It does not evaluate a static snapshot at a certain point in time, but continuously uses the real-time output of the collaborative computing unit and the dynamic mapping unit to fit the time-varying transfer rate parameters of customers between three representative asset quality states: steel heavy lead layer, gold layer, and platinum layer.

[0156] For example, the rate of upgrading from the steel-lead layer to the gold layer, the rate of upgrading from the gold layer to the platinum layer, and the rate of reversal or loss, aggregate discrete, individual hierarchical leap events into a continuous, collective flow process for analysis. This can accurately depict the dynamic evolution trajectory and internal flow speed of the overall value structure of the new customer group.

[0157] By solving the differential equations corresponding to the three-state warehouse model, the long-term steady-state hierarchical composition ratio when the new customer group reaches dynamic equilibrium is predicted. This extends the evaluation perspective from short-term fluctuations to long-term equilibrium, providing a stable quality assessment benchmark based on long-term value orientation. Compared with traditional indicators that only focus on the initial purchase amount or short-term retention rate, steady-state ratio prediction is better able to identify high-quality customer acquisition channels or marketing activities that may have high initial conversion costs but huge potential for total customer lifetime value and good prospects for loyalty cultivation.

[0158] Based on the predicted long-term steady-state hierarchical composition ratio, a quantitative score is generated for the quality of new customer channels or marketing activities. This score integrates the quantity and quality of customer assets, providing objective and long-term meaningful data support for marketing decisions. It constructs a feedback loop from front-end customer acquisition to back-end asset value assessment, enabling enterprises not only to manage existing assets but also to scientifically evaluate and optimize the quality of incremental assets. This guides future marketing budgets and resources to be more accurately targeted at channels or activities that can bring high-quality customer assets, achieving continuous optimization of marketing ROI.

[0159] This setup applies insights from dynamic stratification and state transitions to new customer assessment. Through a system dynamics model, it enables a long-term, dynamic, and value-based scientific evaluation of customer acquisition effectiveness, significantly enhancing the foresight and scientific nature of corporate marketing strategic planning. This, in turn, significantly improves the overall value and operational efficiency of customer assets, and enhances the management efficiency of customer assets.

[0160] The collaborative computing unit performs customer lifetime value prediction and loyalty measurement in parallel. The CLV prediction does not rely on a single fixed model, but can adaptively call the most suitable prediction model for calculation based on customer behavior patterns, thereby significantly improving the prediction accuracy and the applicability of the system. Especially for customer groups with complex and ever-changing transaction patterns, the system achieves accurate fitting of different behavioral characteristics (such as intermittent and regularity) through an intelligent model selection mechanism, solving the problem of large prediction deviation caused by model and behavior mismatch in traditional methods.

[0161] Meanwhile, customer loyalty measurement uses structural equation modeling to fit multi-dimensional latent variable observation indicators, realizing a scientific and quantitative measurement of the abstract concept of customer loyalty. It transforms non-monetary values ​​such as reputation and trust, which are difficult to observe directly, into calculable indicators, providing key inputs for a comprehensive assessment of customer assets.

[0162] By using the BG / NBD model for initial prediction and the Markov chain for subsequent valuation, this approach effectively solves the specific technical challenge of long-term value prediction for intermittently purchasing customers. While the BG / NBD model excels at predicting the probability of future transactions based on transaction history, it does not directly output monetary value. This solution transforms the probability sequence output by BG / NBD, rather than single-point probabilities, into non-homogeneous state transition probabilities that change over time in the Markov chain model. This constructs a sophisticated probability model that can characterize the dynamic changes in customers' future transaction behavior. Compared to homogeneous Markov chains that use fixed transition probabilities, non-homogeneous chains better reflect the reality that customer activity may decline or fluctuate over time, thus making the model's simulation of the customer lifecycle more realistic.

[0163] By defining state-associated returns, such as single-transaction profits, on this non-homogeneous Markov chain and solving for its steady-state or long-term expected returns, it calculates not the expected value over a certain period of time, but the mathematical expectation of all remaining lifetime value over possible future time spans. This solves the problem in traditional CLV calculations that often require subjective setting of prediction years or cannot handle infinite time horizons.

[0164] For corporate decision-making, this value is a more comprehensive and stable long-term indicator, providing a solid quantitative basis for determining whether to make long-term investments in a customer. Therefore, this specific fusion algorithm is not a simple stacking of two models, but rather forms a complementary predictive value chain. For the common and challenging customer type of intermittent purchases, it provides an end-to-end automated calculation method that automatically learns behavioral patterns from historical data, accurately predicts future transaction times, and scientifically converts them into lifetime monetary value. This method greatly improves the accuracy and objectivity of CLV measurement for this customer group, providing reliable input for dynamic stratification and strategy-driven approaches, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0165] Reference Figure 3 In the collaborative computing unit, the model selection logic intelligently judges the purchase behavior pattern based on the customer's dynamic feature vector, especially the coefficient of variation of historical transaction intervals and the most recent transaction time, and assigns different prediction models or initialization methods to different patterns accordingly. This introduces adaptability and context awareness into the customer lifetime value prediction module of the present invention, resulting in a significant improvement in the overall prediction accuracy, robustness and practicality of the system.

[0166] Reference Figure 3 In the collaborative computing unit, the core of the model selection logic lies in automating and objectively diagnosing complex customer behavior patterns through quantifiable statistical indicators, such as the coefficient of variation and the most recent transaction time. Utilizing the coefficient of variation of historical transaction intervals to distinguish between intermittent and regular purchases is a scientific judgment method based on the inherent fluctuation characteristics of data, and its effectiveness far surpasses that of coarse classification relying on human experience or fixed rules. Combining the most recent transaction time further confirms the customer's current activity level, which helps maintain the best match between the subsequent value prediction model and the customer's actual behavioral characteristics, reducing the model's systematic prediction bias.

[0167] Differentiated prediction strategies are adopted for different diagnostic results. For intermittent purchasing behavior, a fusion algorithm or an improved Pareto / NBD model is used based on the most recent transaction time. As mentioned above, the effect is to solve the prediction problem of such customers. For regular purchasing behavior, a Shifted-Beta-Geometric (sBG) model is adopted. This model usually has higher computational efficiency and goodness of fit for customers with frequent transactions and stable patterns. The effect of this choice is to optimize the allocation of computational resources while ensuring prediction accuracy and avoiding the use of unnecessary complex models for regular customers.

[0168] For new customers, an initialization model based on the mean of similar customer groups is used to provide a reasonable initial value estimate based on group characteristics, enabling the system to incorporate them into management from the very early stages of the customer lifecycle, rather than waiting for them to accumulate enough data before analysis.

[0169] With this setup, the system can automatically call the most suitable prediction tool based on the specific customer's data characteristics, enhancing the system's generalization ability and prediction accuracy when facing heterogeneous customer groups. This allows different types of assets to receive scientific evaluations that match their characteristics, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing the management efficiency of customer assets.

[0170] Traditional methods of measuring loyalty either rely on single survey questions, such as Net Promoter Score (NPS), or use simple behavioral substitutes, such as repeat purchase rate. These methods fail to reveal the multidimensional composition of loyalty and its underlying formation mechanisms, and the measurement results are often unstable and have limited explanatory power. This proposed solution establishes a solid scientific theoretical framework for loyalty measurement by presupposing a multi-latent variable structural equation model. The latent variables included in the model systematically cover multiple progressive levels of loyalty, such as the cognitive evaluation basis, emotional dependence core, behavioral intention intensity, and long-term relationship orientation, and the variables are presupposed to follow causal paths consistent with consumer behavior theory.

[0171] Compared to a single indicator, this multidimensional model can more comprehensively and profoundly diagnose the true quality and stability of customer relationships, providing rich insights for the formulation of subsequent differentiation strategies. It achieves multi-source data fusion measurement, improves the reliability and effectiveness of loyalty measurement, reduces the possible bias of a single data source, and makes the measurement results more robust, objective and closer to reality.

[0172] Future loyalty reflects customers' fundamental willingness to maintain long-term relationships, make incremental or cross-purchases, and actively recommend products. Using it as a core dimension of segmentation ensures that loyalty measurement and customer lifetime value prediction are highly aligned in terms of objectives. Together, they constitute a complete evaluation system for the current value and future potential of customer assets, providing this system with a scientific and model-based tool for measuring relationship quality. This provides direct data support and action direction for a deeper understanding of customers and the development of precise relationship strengthening and loyalty enhancement strategies, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0173] The dynamic mapping unit uses the two quantitative results mentioned above to perform two-dimensional hierarchical classification, establishing a more scientific and stable customer asset classification system. The classification results simultaneously reflect the customer's current profitability and future relationship stability, which is more instructive than static clustering based on a single RFM indicator.

[0174] The state inference unit predicts state transition probabilities through time series models, enabling the system to have forward-looking insight capabilities and providing key decision-making basis for proactive intervention.

[0175] The state deduction unit uses recurrent neural networks (RNNs), especially their variants such as long short-term memory networks (LSTMs), to handle customer state deduction problems. It leverages the inherent architectural advantages of RNNs, which are specifically designed for processing sequential data. The recurrent connections of RNNs enable them to retain the memory of historical information, making them suitable for analyzing time-dependent customer hierarchy coding sequences. They can automatically capture and learn long-term dependencies and complex patterns in customer hierarchy changes.

[0176] For example, it might learn a nonlinear pattern: after briefly falling from the platinum layer to the gold layer, if the CLV change rate is positive, there's a higher probability of quickly returning to the platinum layer. This is something traditional Markov chain or logistic regression models struggle to discover automatically. The design of input features, such as hierarchical sequences, CLV change rates, and behavioral activity slopes, provides the RNN model with high-information-density learning material. Inputting these heterogeneous but related temporal features into the RNN enables the model to perform multi-dimensional, multi-signal fusion analysis, thereby making more comprehensive and sensitive inferences about state transitions, significantly enhancing the timeliness and foresight of predictions.

[0177] The output is a state transition probability distribution vector, where each element corresponds to the probability of transitioning to a specific level. This output format provides a quantified and complete measure of uncertainty, offering rich decision-making support for the policy-driven unit.

[0178] For example, the system can trigger different strategy combinations for two different probability distribution patterns: high probability of upgrading and high probability of churn. It can even calculate the expected value change for the benefit and cost assessment of the strategy, achieving high-precision, fine-grained probability prediction of the complex time-series process of customer asset dynamics. This significantly improves the overall value and operational efficiency of customer assets, and enhances the management efficiency of customer assets.

[0179] The strategy-driven unit automatically matches and executes strategies based on the current level, predicted probabilities, and real-time events. It associates complex analytical conclusions with specific business actions and achieves automated execution through instruction distribution. This solves the inefficiency and missed opportunities caused by the traditional chain from analysis reports to manual decision-making and manual configuration. It realizes integrated operation of dynamic assessment of customer asset status, risk warning and automation, and personalized intervention. Fluctuations in customer value can be perceived in real time and transformed into specific actions through pre-set strategy logic, thereby significantly improving the efficiency and accuracy of marketing resource allocation and effectively stabilizing and increasing the total value of customer assets.

[0180] This setup enables dynamic and automated customer asset management. The system can not only accurately measure and classify customer assets, but also proactively predict asset change trends and automatically execute precise strategies for intervention, upgrading, or divestiture, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0181] The strategy knowledge base stores strategies in a structured, relational format, enabling modularity, contextualization, and manageability of strategies. Operations personnel can design strategy packages around clear business objectives and set up detailed, multi-condition triggering logic within them.

[0182] For example, a strategy can be defined as follows: when the scenario is "upgrade from Steel Layer to Gold Layer", the triggering condition is "predicted upgrade probability > 40%" and "real-time event is adding a high-priced product to the favorites", then execute two actions: "push an exclusive coupon for the product" and "generate a prompt in the customer service workbench". This greatly enhances the complexity of strategy expression and business fit, and supports precise personalized intervention.

[0183] The strategy matcher continuously monitors input streams from various parts of the system, including the customer's current static level, the predicted dynamic target level or transition probability, and real-time behavioral event streams captured by the unified data platform. The strategy matcher integrates these multi-source heterogeneous input information and queries the strategy knowledge base in real time, realizing accurate real-time strategy triggering based on multi-dimensional context awareness. This decision-making capability, which integrates long-term trends and instantaneous intentions, is something that traditional rule engines do not possess.

[0184] The instruction dispatcher transforms the abstract automated operational actions output by the strategy matcher into standardized instructions or API calls that can be recognized and processed by various downstream business execution systems, and ensures their reliable distribution. This makes the system highly scalable. When new marketing channels or business systems need to be integrated, only the corresponding adapter needs to be added to the instruction dispatcher. There is no need to change the core strategy management and matching logic, thus building an intelligent and easy-to-integrate strategy decision-making and execution engine.

[0185] While traditional automated systems can execute actions according to rules, their rules and models are static and cannot learn from the execution results. The feedback optimization unit continuously collects customer feedback and behavioral change data, such as whether customers respond to offers, complete upgrade purchases, and whether the average order value increases, and systematically transforms this data into quantifiable performance data. This provides the system with an objective basis for self-evaluation, enabling the system to perform self-diagnosis based on actual business results and objectively measure the short-term and long-term impact of each automated decision.

[0186] By applying reinforcement learning algorithms to handle feedback optimization, the entire customer asset management process is modeled as a sequential decision problem. The system selects strategic actions based on the current customer status, and the environment provides feedback. The system updates its decision model accordingly to pursue the maximization of long-term cumulative rewards. This allows the computational model to be dynamically calibrated based on changes in actual customer behavior after strategic intervention, thereby increasingly reflecting the patterns of customer behavior under the influence of existing operational strategies and reducing model errors.

[0187] The system can automatically identify and strengthen strategies, rules, or action combinations that bring high rewards in specific customer states, while weakening or eliminating ineffective strategies. This solves the rigidity and lag problems caused by traditional systems that rely on fixed rules and periodic manual adjustments, enabling the system to maintain high adaptability and high operational efficiency, achieving sustainable growth in return on assets, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing the management efficiency of customer assets.

[0188] Traditional methods often rely on batch scoring based on static classification models such as logistic regression, or simple rules such as the number of consecutive days without logging in. These methods cannot handle censored data that did not churn during the customer observation period, and the response after an alert depends on manual intervention, which is inefficient and can easily lead to missing the best retention opportunity.

[0189] By employing the Cox proportional hazards regression model as the core risk identification tool in the churn early warning and intervention unit, the scientific rigor and timeliness of early warnings have been significantly improved. The Cox model is a semi-parametric survival analysis model whose core advantage lies in its ability to effectively utilize a complete sample of customer data, including censored data (i.e., customer data that has not yet resulted in churn at the end of the observation period), for training. This allows for more comprehensive use of data information and the acquisition of more robust risk estimates.

[0190] The risk index output by the model, i.e., the risk function value, not only quantifies the relative probability of churn but also includes time-dimensional information, reflecting the dynamic evolution of risk as customer characteristics change. This is achieved by continuously extracting key covariates highly correlated with churn from the latest dynamic feature vectors.

[0191] For example, this unit can update the individual churn risk index of each high-value customer in near real-time, which can detect phenomena such as a sharp drop in login frequency, a continuous decline in average order value, a surge in negative emotions in customer service interactions, and a surge in the activity of competitor apps. This enables dynamic and continuous monitoring and quantification of churn risk.

[0192] By setting a threshold dynamically calculated based on historical data, such as taking a high quantile of the historical churn customer risk index distribution as the threshold, and automatically generating a high-risk warning signal when this threshold is exceeded, this unit realizes automatic triggering and dynamic calibration of the warning, ensuring that the sensitivity and specificity of the warning are in a reasonable balance. It can capture real risk signals in a timely manner, while effectively avoiding the proliferation of false alarms caused by noise or normal fluctuations, thereby improving the overall reliability and availability of the warning system.

[0193] When the strategy-driven unit receives a high-risk warning signal, it prioritizes and activates customer retention strategies. This allows intervention actions targeting high-risk, high-value customers to be automatically triggered and executed by the system with high priority and low latency. This significantly shortens the critical time window from risk identification to retention action, compressing the manual judgment and execution process, which may traditionally take days or even weeks, to minutes or even seconds. This provides an opportunity to implement effective intervention at the customer's decision-making critical point and successfully recover core assets. It effectively addresses the core objectives of reducing core customer churn and protecting the foundation of corporate profits, thereby significantly improving the overall value and operational efficiency of customer assets and enhancing customer asset management efficiency.

[0194] The quantitative separation of decision units discretizes the two key behavioral dimensions of customer transaction intervals and transaction frequencies, constructing a discrete state space with clear business meaning. This simplifies complex historical customer behavior patterns into computable and analyzable system states, providing a stable modeling foundation for applying Markov decision processes, a stochastic dynamic programming tool, and enabling the dynamics of customer behavior to be formally described and analyzed.

[0195] This unit utilizes a Markov decision process model for optimization. The core inputs of the model include a state transition probability matrix estimated based on historical data, which describes the probability of a customer transitioning from the current state to the next state, and a payoff matrix, which describes the immediate net payoff obtained by the firm in each state by taking the action of continuing investment or ceasing investment.

[0196] By solving the MDP model, the optimal strategy that maximizes the long-term expected discounted net benefit is obtained, providing an objective and quantitative economic basis for divestiture decisions. For the entire customer group corresponding to the state of ceasing investment as determined by the optimal strategy, this unit directly controls the strategy-driven unit, so that it systematically ignores these customers in the subsequent strategy matching and distribution process, and no longer triggers any resource-consuming operational actions for them, such as sending promotional text messages, pushing personalized advertisements, or allocating customer service resources. This realizes the automated and large-scale execution of negative asset divestiture. Enterprises do not need manual screening and operation. The system can automatically identify and silently process these customers, thereby releasing marketing resources and reallocating them to customer groups with higher value potential. This improves the overall efficiency of marketing resource allocation, thereby significantly enhancing the overall value and operational efficiency of customer assets and improving the management efficiency of customer assets.

[0197] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A customer asset intelligent management system based on a hierarchical dynamic driving model, characterized in that, Includes the following modules: Unified Data Platform: Used to access and integrate raw customer data from transaction systems, interaction platforms and external data sources in real time. By performing data cleaning, entity association and feature engineering, it outputs dynamic feature vectors with unique customer identifiers as keys, containing time-series behavioral tags and statistical indicators. Hierarchical Dynamic Drive Engine: Used to dynamically assess and classify customers and predict future state changes, including collaborative computing unit, dynamic mapping unit and state inference unit; Collaborative computing unit: used to perform customer lifetime value prediction and customer loyalty measurement in parallel based on the dynamic feature vector, wherein the customer lifetime value prediction calls different prediction models for calculation according to customer behavior patterns, and the customer loyalty measurement is performed by fitting the preset latent variable observation indicators through structural equation model. Dynamic mapping unit: It is used to take the customer lifetime value prediction result and the customer loyalty measurement result as inputs to the first coordinate axis and the second coordinate axis respectively, and map the customer to a specified level in the hierarchical architecture composed of platinum layer, gold layer, steel layer and heavy lead layer in real time through a preset two-dimensional threshold grid. State deduction unit: used to deduce the state transition probability distribution vector of customers from the current level to other levels in the future preset time period based on the customer's historical level sequence and the behavioral trend indicators in the dynamic feature vector, through a time series classification model; Strategy-driven unit: Based on the customer's current level, predicted state transition probability, and real-time triggered behavioral events, it matches the corresponding automated operation strategy from the strategy knowledge base and converts it into executable instructions to be distributed to downstream business systems.

2. The customer asset intelligent management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: The collaborative computing unit is equipped with a first prediction model, which is a fusion algorithm that combines the BG / NBD model and the Markov chain model. Its specific execution process is as follows: using the BG / NBD model to process the customer's historical transaction time and number of transactions in the dynamic feature vector, predicting the customer's transaction probability in a specific future period, dynamically constructing the transaction probability sequence into a non-homogeneous state transition matrix in the Markov chain model, and outputting the customer's total remaining lifetime value by solving the steady-state expected return of the Markov chain.

3. The intelligent customer asset management system based on a hierarchical dynamic driving model according to claim 2, characterized in that: The collaborative computing unit is specifically used to execute the following model selection logic: based on the customer dynamic feature vector obtained from the unified data platform, especially the coefficient of variation of historical transaction intervals and the most recent transaction time, it determines the customer's purchase behavior pattern. If it is determined to be intermittent purchase behavior and the most recent transaction time is less than or equal to a preset activity threshold, the first prediction model is called to use the fusion algorithm of the negative binomial distribution model and the Markov chain model to predict the customer's lifetime value. If it is determined to be intermittent purchase behavior and the most recent transaction time is greater than the preset activity threshold, the improved Pareto / NBD model is used. If it is determined to be regular purchase behavior, the prediction is made using the Shifted-Beta-Geometric model. For new customers, the estimation is made using the mean initialization model based on similar customer groups.

4. The customer asset intelligent management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: In the state deduction unit, a trained recurrent neural network is used as the time series classification model. The input layer of the recurrent neural network receives the customer level encoding sequence arranged in chronological order provided by the dynamic mapping unit, as well as the customer lifetime value change rate and behavioral activity slope derived from the dynamic feature vector. The output layer of the recurrent neural network provides the state transition probability distribution vector, where each element represents the probability of a customer transitioning to the corresponding level.

5. The intelligent customer asset management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: It also includes a feedback optimization unit: used to collect customer feedback and behavior change data returned by the downstream business system after executing the automated operation strategy, as effect data, convert the effect data into reward signals, and iteratively optimize the parameters of the time series classification model in the state inference unit and the strategy matching logic of the strategy knowledge base in the strategy driving unit through reinforcement learning algorithms.

6. The customer asset intelligent management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: The unified data platform also includes a customer asset quality evaluation unit: used to model the evolution of the hierarchical distribution of newly acquired customer groups in the hierarchical architecture as a three-state warehouse model, the three states corresponding to the steel heavy lead layer, the gold layer, and the platinum layer. Based on the real-time output of the collaborative computing unit and the dynamic mapping unit, the time-varying transition rate parameters of customers between states are fitted. By solving the differential equation system corresponding to the three-state warehouse model, the long-term steady-state hierarchical composition ratio of the new customer group is predicted, and a quantitative score for the quality of new customer channels or marketing activities is generated based on this ratio.

7. The intelligent customer asset management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: It also includes a churn warning and intervention unit: for customers in the gold and platinum layers, it continuously extracts key covariates related to churn risk from the dynamic feature vector, inputs the key covariates into a pre-trained Cox proportional hazards regression model, calculates and updates the individual churn risk index of the customer, and when the individual churn risk index exceeds the threshold dynamically calculated based on historical data, it generates a high-risk warning signal and sends it to the strategy driving unit. The strategy-driven unit is also used to prioritize matching and activating customer retention strategies when the high-risk warning signal is received.

8. The intelligent customer asset management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: It also includes a quantitative stripping decision unit: used to construct a discrete state space based on the time interval between the customer's last transaction and the transaction frequency within a preset time period, and use a Markov decision process model, combined with a preset state transition probability matrix and a revenue matrix, to solve for the optimal strategy that maximizes the customer's long-term expected net profit. The optimal strategy specifies the action of "continue investment" or "stop investment" for each state. For the customer group corresponding to the state indicated by the optimal strategy as "stop investment", the strategy driving unit is controlled to no longer match and distribute any resource investment-type operation strategies to them.

9. A customer asset intelligent management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: The customer loyalty measurement in the collaborative computing unit specifically involves: pre-setting a structural equation model containing multiple latent variables such as customer satisfaction, customer trust, relationship commitment, and future loyalty; each latent variable is associated with multiple observation indicators from questionnaire scales or behavioral logs; periodically inputting the observation indicator data generated by the unified data platform into the measurement model; using a partial least squares algorithm for model fitting and validation; outputting the latent variable factor score for each customer; and using the factor score of future loyalty as the current loyalty measurement result.

10. A customer asset intelligent management system based on a hierarchical dynamic driving model according to claim 1, characterized in that: The strategy-driven unit specifically includes: Strategy knowledge base: It contains pre-stored relationships, which are associated with hierarchical transition scenarios, triggering conditions and at least one automated operation action; Strategy Matcher: Used to query the strategy knowledge base based on the customer's current level, target jump level, and real-time triggered behavioral events, and match the corresponding target automated operation actions; Instruction Distributor: Used to convert the target automated operational actions into executable instructions and distribute them to downstream business execution systems.

Images