Big data-based customer information updating and management method and system

Abstract

This invention discloses a method and system for updating and managing customer information based on big data, relating to the fields of big data and customer relationship management technology. The method includes: firstly, collecting multi-source customer data and constructing a multi-dimensional time-series feature panel; then, simultaneously predicting the dynamic lifetime value and continuous churn risk probability of customers through a shared underlying integrated model; and quantifying the network influence weight of customers using an influence penetration model and Monte Carlo simulation; furthermore, integrating the above three dimensions and generating resource allocation coefficients through a multi-objective optimization function that dynamically configures weights based on customer type; finally, automatically updating customer tags based on these coefficients and combining atomic operations through a strategy orchestration engine to form a personalized strategy workflow execution and closed-loop optimization. This invention solves the problems of static isolation and policy lag in traditional customer information management, achieving accurate insight into customer value and dynamic intelligent allocation of management resources.

Description

Technical Field

[0001] This invention belongs to the field of big data and customer relationship management technology, specifically relating to a method and system for updating and managing customer information based on big data. Background Technology

[0002] As enterprises continue to improve their IT capabilities, customer data is becoming increasingly massive, multi-source, and dynamic.

[0003] Traditional customer information management systems rely primarily on static, fixed customer attributes and historical transaction snapshots, resulting in significant shortcomings: First, customer value assessment is largely based on the sum of past transactions, failing to dynamically and continuously predict future changes in customer value; second, customer churn warnings are typically binary or based on simple rules, unable to provide continuous risk probabilities at specific future points in time, leading to untimely or inaccurate warnings; third, existing methods generally treat customers as isolated individuals, ignoring their influence within social or business networks and failing to tap into the network value-added effects they bring; finally, customer management strategy formulation often depends on human experience, resulting in a single decision-making dimension and a long cycle from information updates to strategy execution, making it difficult to achieve accurate and efficient personalized management.

[0004] Therefore, there is an urgent need for a customer information update and management method and system that can integrate multi-dimensional dynamic information, accurately predict future states, quantify network effects, and automatically drive personalized strategies. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a customer information updating and management method and system based on big data, to solve the following technical problems:

[0006] First, customer value assessment is mostly based on the sum of past transactions, making it impossible to dynamically and continuously predict changes in customer value in the future. Second, customer churn warnings are usually binary classifications or based on simple rules, failing to provide continuous risk probabilities at specific future points in time, resulting in untimely or inaccurate warnings. Third, existing methods generally treat customers as isolated individuals, ignoring their influence in social or business networks and failing to tap into the network value-added effects they bring. Finally, the formulation of customer management strategies often relies on human experience, resulting in a single decision-making dimension and a long cycle from information updates to strategy execution, making it difficult to achieve accurate and efficient personalized management.

[0007] To address the aforementioned problems, the first aspect of this invention provides a method for updating and managing customer information based on big data, comprising the following steps:

[0008] S1: Collect customer internal transaction data, external public behavior data and customer relationship data, preprocess them, and construct a multi-dimensional time series feature panel;

[0009] S2: Input the multi-dimensional temporal feature panel into the ensemble model, which performs two tasks by sharing the underlying feature encoding layer;

[0010] Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value.

[0011] Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future;

[0012] S3: Using the customer-interconnected data, construct a directed graph with dynamically updated edge weights. Utilize the influence penetration model and Monte Carlo simulation method to calculate the expected value of the entire network that a single customer node can activate, and quantify it as the customer's network influence weight.

[0013] S4: Integrate the dynamic lifecycle value, continuous risk probability, and network influence weight to construct and solve a multi-objective optimization function to generate resource allocation coefficients;

[0014] S5: Automatically updates customer tags based on resource allocation coefficients, and combines atomic operations through a strategy orchestration engine to form a personalized customer management strategy workflow, track strategy effects, and provide feedback.

[0015] Preferably, step S1 includes the following steps:

[0016] The preprocessing includes cleaning, aligning, and privacy protection of the data;

[0017] Based on the preprocessed customer time-series data, features are extracted for each customer entity, including rolling window statistical features, behavioral sequence trend features, and interactive periodic fluctuation features;

[0018] The feature group generated from the three features is horizontally concatenated with the customer's static attribute features to form a multi-dimensional time-series feature panel with fixed dimensions that is unique to each customer.

[0019] Preferably, the feature group generated from the three features includes the following steps:

[0020] The rolling window statistical feature is defined by defining rolling windows of different time lengths on the transaction and behavior data stream. Within each window, the statistics of transaction frequency and amount, the aggregation intensity of behavioral events, and the slope of the change of the statistics and aggregation intensity within the window are calculated in parallel, which together constitute a multi-scale window feature vector.

[0021] Behavioral sequence trend features: The ordered behavioral sequence of customers is transformed into an embedded vector. A lightweight time series decomposition model is used to decompose the sequence into trend terms, periodic terms and residual terms. The fitting coefficient and significance level of the trend terms are extracted as long-term intention features, and the main period amplitude of the periodic terms is calculated as behavioral regularity features.

[0022] Interactive periodic fluctuation features identify the correlation between transaction behavior and external time context, calculate the coupling strength coefficient between the customer's personal behavior cycle and the external public cycle, and encode the coefficient and the customer's behavioral fluctuation amplitude at key cycle phases into context-aware periodic features.

[0023] Preferably, in step S2, the shared underlying feature encoding layer of the integrated model is constructed through a multilayer perceptron, and its output is used as the common input of Task 1 and Task 2. The output of the temporal convolutional network in Task 1 is connected to a multi-head self-attention mechanism layer to perform importance weighting on the features at different time steps in the multi-dimensional temporal feature panel.

[0024] Preferably, in Task 1, the multi-branch time-series value prediction network includes the following steps:

[0025] The main sequence modeling branch uses a temporal convolutional network to process the multi-dimensional temporal feature panel to extract basic temporal patterns;

[0026] The mutation point detection branch uses a lightweight temporal anomaly detector to identify positive and negative mutation points in customer behavior sequences in real time, and outputs the mutation point confidence and feature vector.

[0027] The periodic memory branch utilizes a gated recurrent unit network to capture long-term periodic patterns of customer behavior and generate periodic context vectors.

[0028] The multi-head self-attention mechanism layer performs cross-branch attention fusion on the output features of the three branches, calculates the contribution of different branch features to the final prediction, and aggregates them to generate the final value prediction curve.

[0029] Preferably, Task 2 is implemented using a deep survival network based on a temporal model memory, and includes the following steps:

[0030] A temporal pattern memory is introduced to learn from the historical behavior sequences of all customers in an unsupervised manner and store K typical customer behavior evolution prototype patterns.

[0031] For any customer, calculate the attention weights between its current temporal features and all K prototype patterns in the memory bank to obtain a K-dimensional pattern matching weight vector a. k ;

[0032] The baseline risk function of the survival analysis model is deconstructed into a weighted sum of the baseline risk sub-functions corresponding to the K prototype models, specifically as follows:

[0033]

[0034] Where h0(t|X) is the baseline risk function, t is the time point, and h 0k (t) is the baseline risk subfunction specifically learned for the k-th prototype pattern;

[0035] The probability of continuous risk for an individual customer is ultimately determined by the benchmark risk function and their individual characteristics.

[0036] Preferably, step S3 includes the following steps:

[0037] Construct a customer association directed graph with dynamically updated edge weights, where nodes represent customers and the weights of directed edges are determined by association strength, association direction, and association freshness.

[0038] We utilize a value penetration model based on multi-threshold coupling and quantify the network influence weight of a single customer node through a parallel Monte Carlo simulation method.

[0039] The value penetration model sets a dynamic value activation threshold function, so that the probability of a neighboring node being activated is non-linearly positively correlated with its own static value and edge weight.

[0040] The network influence weight is calculated by using the current customer node as the initial source of infection in multiple Monte Carlo simulations, propagating the virus on the network according to the value penetration model, and statistically analyzing the total value of the activated neighbor nodes in each simulation. Finally, the average value of all simulation results is quantified as the network influence weight of the customer node.

[0041] Preferably, in step S4, constructing and solving the multi-objective optimization function includes the following steps:

[0042] The construction of the multi-objective optimization function specifically involves: constructing an objective vector for each customer, wherein the objective vector consists of three sub-objective functions, including:

[0043] ; ;

[0044] Where CLV(c) is the dynamic lifecycle value, P churn (c) represents the continuous risk probability, and L(c) represents the network influence weight;

[0045] Based on dynamic lifecycle value and continuous risk probability, customers are divided into four core types through a dual-threshold network, and a specific network influence weight coefficient w3(c) is configured for each type.

[0046] Calculate the resource allocation coefficient for each customer, specifically as follows:

[0047]

[0048] Where w1 and w2 are the fixed weight coefficients of the corresponding sub-functions, w3(c) is the network influence weight coefficient, and N[·] function is used to normalize the target value within the parentheses.

[0049] Preferably, step S5 includes the following steps:

[0050] Based on resource allocation coefficients, customer level tags in the customer information management system are automatically updated using preset threshold rules.

[0051] The strategy orchestration engine matches and invokes multiple atomic operations from a predefined strategy atomic operation library based on the updated customer level label and the continuous risk probability.

[0052] The strategy orchestration engine, based on the logic of a directed acyclic graph, combines multiple atomic operations into an ordered, personalized customer management strategy workflow and drives its execution.

[0053] The system continuously tracks the execution process of the strategy workflow and the subsequent behavior data of the customer after completion. The change in the resource allocation coefficient is used as the core feedback signal to evaluate the effect of the strategy and trigger the retraining of the model parameters.

[0054] A second aspect of the present invention provides a customer information update and management system based on big data, comprising the following modules:

[0055] Multi-source data fusion and feature extraction module: Collects internal customer transaction data, external public behavior data and customer relationship data, performs preprocessing, and constructs a multi-dimensional time series feature panel;

[0056] Customer Value and Risk Integrated Prediction Module: Inputs a multi-dimensional time-series feature panel into the integrated model, which performs two tasks by sharing a bottom-level feature encoding layer;

[0057] Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value.

[0058] Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future;

[0059] Customer network influence quantification module: Based on the customer relationship data, a directed graph with dynamically updated edge weights is constructed. Using the influence penetration model and Monte Carlo simulation method, the expected value of the entire network that a single customer node can activate is calculated and quantified as the customer's network influence weight.

[0060] Multi-objective resource allocation decision module: Integrates the dynamic life cycle value, continuous risk probability and network influence weight, constructs and solves a multi-objective optimization function to generate resource allocation coefficients;

[0061] The strategy automation execution and feedback optimization module automatically updates customer tags based on resource allocation coefficients, combines atomic operations through the strategy orchestration engine to form a personalized customer management strategy workflow, tracks strategy effectiveness, and provides feedback.

[0062] The beneficial effects of this invention are:

[0063] This invention extracts multi-scale dynamic features of rolling window statistics, behavioral sequence trends, and interactive periodic fluctuations, and uses temporal convolutional networks, attention mechanisms, and multi-branch networks to predict value curves. Combined with a survival analysis model based on interval learning and temporal pattern memory to predict the probability of continuous risks, this invention achieves a refined and forward-looking assessment of customer lifetime value and churn risk, providing an unprecedented time window and decision-making basis for precise intervention.

[0064] This invention constructs a directed graph with dynamically updated edge weights and uses a value penetration model based on multi-threshold coupling and Monte Carlo simulation to calculate the expected value of the entire network that a single customer can activate. This transforms the abstract concept of network influence into a concrete and quantifiable weight indicator, enabling enterprises to identify and prioritize services for key nodes that can bring about customer growth and maximize the network effect of marketing investment.

[0065] This invention constructs value, risk, and influence into a multi-objective vector, and assigns dynamic weights to network influence objectives based on customer type, ultimately generating a unified resource allocation coefficient. This breaks down multi-dimensional information silos, encodes complex business logic into a computable optimization problem, and generates a comprehensive and clearly oriented decision indicator, enabling limited customer management resources to be directed towards the customer group with the greatest overall benefits. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0067] Figure 2 This is a schematic diagram of the module flow of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1 As shown, this invention is a customer information update and management method based on big data, including the following steps:

[0070] S1: Collect customer internal transaction data, external public behavior data and customer relationship data, preprocess them, and construct a multi-dimensional time series feature panel;

[0071] S2: Input the multi-dimensional temporal feature panel into the ensemble model, which performs two tasks by sharing the underlying feature encoding layer;

[0072] Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value.

[0073] Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future;

[0074] S3: Using the customer-interconnected data, construct a directed graph with dynamically updated edge weights. Utilize the influence penetration model and Monte Carlo simulation method to calculate the expected value of the entire network that a single customer node can activate, and quantify it as the customer's network influence weight.

[0075] S4: Integrate the dynamic lifecycle value, continuous risk probability, and network influence weight to construct and solve a multi-objective optimization function to generate resource allocation coefficients;

[0076] S5: Automatically updates customer tags based on resource allocation coefficients, and combines atomic operations through a strategy orchestration engine to form a personalized customer management strategy workflow, track strategy effects, and provide feedback.

[0077] In one embodiment of the present invention, step S1 includes the following steps:

[0078] The preprocessing includes cleaning, aligning, and privacy protection of the data;

[0079] Based on the preprocessed customer time-series data, features are extracted for each customer entity, including rolling window statistical features, behavioral sequence trend features, and interactive periodic fluctuation features;

[0080] The feature group generated from the three features is horizontally concatenated with the customer's static attribute features to form a multi-dimensional time-series feature panel with fixed dimensions that is unique to each customer.

[0081] In one embodiment of the present invention, the feature group generated from the three features includes the following steps:

[0082] The rolling window statistical feature is defined by defining rolling windows of different time lengths on the transaction and behavior data stream. Within each window, the statistics of transaction frequency and amount, the aggregation intensity of behavioral events, and the slope of the change of the statistics and aggregation intensity within the window are calculated in parallel, which together constitute a multi-scale window feature vector.

[0083] Behavioral sequence trend features: The ordered behavioral sequence of customers is transformed into an embedded vector. A lightweight time series decomposition model is used to decompose the sequence into trend terms, periodic terms and residual terms. The fitting coefficient and significance level of the trend terms are extracted as long-term intention features, and the main period amplitude of the periodic terms is calculated as behavioral regularity features.

[0084] Interactive periodic fluctuation features identify the correlation between transaction behavior and external time context, calculate the coupling strength coefficient between the customer's personal behavior cycle and the external public cycle, and encode the coefficient and the customer's behavioral fluctuation amplitude at key cycle phases into context-aware periodic features.

[0085] Specifically, the scrolling window statistical features are as follows: For each customer, at least two different time-length scrolling windows are defined on their transaction and behavior data streams. For example, a 7-day window (short-term) and a 30-day window (medium-term) are defined. Within each window, three indicators are calculated in parallel: transaction statistics (total transaction frequency, average order value, standard deviation of average order value, and cumulative transaction amount within the window), behavior aggregation intensity (counting or weighting predefined key behavioral events to obtain the total behavioral intensity within the window; key behavioral events include page views, favorited items, and customer service inquiries), and change slope (based on the above transaction statistics and behavior aggregation intensity, the linear change slope within the window time is calculated; for example, the slope of average order value within the 7-day window is calculated to determine the short-term trend of their spending power). All indicators calculated within all the above windows are concatenated to form a comprehensive multi-scale window feature vector.

[0086] The behavioral sequence trend feature transforms the customer's ordered behavioral sequence (e.g., logging into the APP -> searching for the keyword 'laptop' -> browsing the product A details page -> adding to the shopping cart) into a numerical embedding vector through an embedding model to capture the semantic relationships between behaviors. The embedding vector is regarded as a time series and decomposed into three components using a lightweight time series decomposition model, including a trend term, a periodic term, and a residual term. The trend term is linearly fitted, and its fitting coefficient and significance level are extracted as long-term intention features. The periodic term is subjected to spectral analysis, and its main period amplitude is extracted as a behavioral regularity feature. The larger the amplitude, the stronger the regularity of the customer's behavior.

[0087] Interactive periodic fluctuation features are achieved by analyzing the correlation between the timestamps of customers' identified transaction behaviors and external time contexts (e.g., weekdays, public holidays, etc.). The Pearson coefficient between the customer's personal behavior cycle (derived from the above cycle terms) and the external common cycle is calculated as the coupling strength coefficient. The amplitude of customer behavior fluctuations in key cycle phases (e.g., the increase in spending during this period compared to the average daily value) is recorded. The coupling strength coefficient and the amplitude of customer behavior fluctuations in key cycle phases are jointly encoded to form a context-aware periodic feature.

[0088] The three types of dynamic feature groups are horizontally combined with the customer's static attribute features (e.g., registration duration, region, membership level) to form a unique, fixed-dimensional multi-dimensional time-series feature panel for each customer, providing comprehensive and in-depth input for subsequent intelligent prediction models.

[0089] In one embodiment of the present invention, in step S2, the shared underlying feature encoding layer of the integrated model is constructed through a multilayer perceptron, and its output is used as the common input of Task 1 and Task 2. The output of the temporal convolutional network in Task 1 is connected to a multi-head self-attention mechanism layer to perform importance weighting on the features at different time steps in the multi-dimensional temporal feature panel.

[0090] Specifically, the input to the ensemble model is a multi-dimensional temporal feature panel, which is a two-dimensional matrix where rows represent clients and columns represent various features at different time steps. The feature panel is organized sequentially by time step to form a sequence input, preserving temporal dependencies. The shared underlying feature encoding layer is constructed using a multilayer perceptron (MLP), which consists of two fully connected layers using the ReLU activation function and introducing a Dropout layer to prevent overfitting. The feature vector of each time step in the multi-dimensional temporal feature panel is independently input into the MLP. The MLP's role is to perform non-linear transformations and abstractions on the original features, mapping them to a higher layer. A more semantically rich shared feature space: After processing all time steps, the MLP outputs a deeply encoded temporal feature representation, which serves as the common input for the subsequent two independent task branches (Task 1 and Task 2). The ensemble model is trained alternately or jointly, with the loss function being a weighted sum of Task 1 (regression problem, e.g., using mean squared error) and Task 2 (survival analysis problem, e.g., using negative log-likelihood loss). The parameters of the shared underlying MLP, TCN, attention layer, and task-specific networks are simultaneously optimized through the backpropagation algorithm, ultimately resulting in a unified model that can accurately and efficiently output dynamic lifetime value and continuous risk probability simultaneously.

[0091] In one embodiment of the present invention, the multi-branch temporal value prediction network in Task 1 includes the following steps:

[0092] The main sequence modeling branch uses a temporal convolutional network to process the multi-dimensional temporal feature panel to extract basic temporal patterns;

[0093] The mutation point detection branch uses a lightweight temporal anomaly detector to identify positive and negative mutation points in customer behavior sequences in real time, and outputs the mutation point confidence and feature vector.

[0094] The periodic memory branch utilizes a gated recurrent unit network to capture long-term periodic patterns of customer behavior and generate periodic context vectors.

[0095] The multi-head self-attention mechanism layer performs cross-branch attention fusion on the output features of the three branches, calculates the contribution of different branch features to the final prediction, and aggregates them to generate the final value prediction curve.

[0096] Specifically, the main sequence modeling branch: Input: Receives temporal features encoded from a shared underlying MLP; Processing: Processed using a Temporal Convolutional Network (TCN), which utilizes dilated causal convolutional layers to effectively capture long-term dependencies in the sequence and extract the underlying temporal patterns of customer behavior; Output: Outputs a feature tensor containing deep temporal information.

[0097] Multi-head self-attention mechanism weighting: Input: Receives a feature tensor from the TCN; Processing: Inputs this feature tensor into a multi-head self-attention mechanism layer, which assigns an importance weight to each time step by calculating the correlation between features at different time steps within the sequence. For example, it may identify time steps near the prediction start point or time steps where specific behavioral changes occur as having higher importance; Output: Outputs an importance-weighted temporal feature representation that focuses more on the key time points that contribute the most to the value prediction.

[0098] Multi-branch feature fusion and curve generation: Input: The attention-weighted features are concatenated with the outputs of the mutation point detection branch and the periodic memory branch; Processing: The concatenated fused features are dimensionality-reduced through a fully connected layer and finally input into a time-distributed fully connected layer; Output: This layer simultaneously outputs the predicted value for multiple consecutive time points in the future (e.g., the next 12 months), thus forming a complete value prediction curve. The net present value of this curve is calculated as the dynamic life cycle value.

[0099] In one embodiment of the present invention, task two is implemented using a deep survival network based on temporal model memory, and includes the following steps:

[0100] A temporal pattern memory is introduced to learn from the historical behavior sequences of all customers in an unsupervised manner and store K typical customer behavior evolution prototype patterns.

[0101] For any customer, calculate the attention weights between its current temporal features and all K prototype patterns in the memory bank to obtain a K-dimensional pattern matching weight vector a. k ;

[0102] The baseline risk function of the survival analysis model is deconstructed into a weighted sum of the baseline risk sub-functions corresponding to the K prototype models, specifically as follows:

[0103]

[0104] Where h0(t|X) is the baseline risk function, t is the time point, and h 0k (t) is the baseline risk subfunction specifically learned for the k-th prototype pattern;

[0105] The probability of continuous risk for an individual customer is ultimately determined by the benchmark risk function and their individual characteristics.

[0106] Specifically, we collect complete historical behavioral sequence data of all customers (i.e., multi-dimensional time-series feature panels of historical versions), ensuring that these customers have a clear churn or retention outcome. We then use a time-series K-Means clustering algorithm or an autoencoder-based sequence clustering method to perform unsupervised learning on the aforementioned historical sequences. The algorithm will automatically identify K most representative customer behavior evolution patterns from the data, such as: Pattern 1: long-term activity followed by sudden inactivity; Pattern 2: slow decline; Pattern 3: high-frequency interaction and stability; and Pattern 4: periodic fluctuations. Each pattern represents a typical customer journey. These K learned patterns are then used to... The prototype patterns of customer behavior evolution (i.e., the sequence vectors of each cluster center) are stored to form a fixed temporal pattern memory for matching during subsequent online predictions. For any current customer to be predicted, their behavior sequence is matched with the patterns in the memory. Specifically, the customer's current multi-dimensional temporal feature sequence is obtained, and an attention network (usually a shallow neural network) is used to calculate the similarity score between the current customer's temporal features and each prototype pattern k (k from 1 to K) in the memory. This similarity score is normalized using a Softmax function to obtain a K-dimensional pattern matching weight vector, where the weights are a. k This represents the degree of matching between the current customer's behavioral sequence and the k-th prototype pattern, with all weights summing to 1. During the model training phase, for each prototype pattern k in the memory, a corresponding baseline risk sub-function h is independently learned. 0k (t), this sub-function describes the baseline pattern of churn risk over time t if a customer fully follows the prototype pattern. These sub-functions can be learned using parametric models (Weiber distribution) or non-parametric methods (piecewise constants). For the current customer, their personalized baseline risk function h0(t|X) is no longer a fixed function, but is determined by the above k baseline risk sub-functions according to their pattern matching weights a. k The risk probability is obtained by weighted summation; by combining personalized benchmark risk with the individual static characteristics of the client, the final risk probability is derived. Specifically, a proportional risk model is constructed using the framework of the Cox proportional risk model, and the client's instantaneous risk rate is calculated.

[0107] ,

[0108] Where h(t|X) is the instantaneous risk rate at time t, given customer characteristics X, i.e., the instantaneous probability density of a customer churning immediately at time t, z=(z1,z2,…,z…). p ) represents the individual feature vector of this customer, including the sum of static and dynamic attributes, β=(β1,β2,…,β2). p ) represents the coefficient vector obtained from model training, where each β i For the corresponding feature zi The degree to which it contributes to the risk of attrition;

[0109] The customer's continuous risk probability is derived through a survival function, specifically:

[0110] The survival function S(t|X) represents the probability that a customer remains active (not churned) from the time of initial observation to time t. The relationship between the survival function and the hazard ratio is as follows:

[0111] ,

[0112] Among them, the integral term Let P be the cumulative risk function, and P be the continuous risk probability. churn (t) represents the probability that a customer will churn before time t, specifically: ;

[0113] By calculating the risk probability from the current time to multiple future time points t (e.g., 30 days, 60 days, 90 days, etc.), a continuous risk probability curve for the customer is formed.

[0114] In one embodiment of the present invention, step S3 includes the following steps:

[0115] Construct a customer association directed graph with dynamically updated edge weights, where nodes represent customers and the weights of directed edges are determined by association strength, association direction, and association freshness.

[0116] We utilize a value penetration model based on multi-threshold coupling and quantify the network influence weight of a single customer node through a parallel Monte Carlo simulation method.

[0117] The value penetration model sets a dynamic value activation threshold function, so that the probability of a neighboring node being activated is non-linearly positively correlated with its own static value and edge weight.

[0118] The network influence weight is calculated by using the current customer node as the initial source of infection in multiple Monte Carlo simulations, propagating the virus on the network according to the value penetration model, and statistically analyzing the total value of the activated neighbor nodes in each simulation. Finally, the average value of all simulation results is quantified as the network influence weight of the customer node.

[0119] Specifically, the weight w of the directed edge ij (From node i to node j), the following three factors are jointly determined and dynamically updated over time, including: association strength (assigning a basic weight based on the relationship type, for example, a direct recommendation relationship has a weight of 0.8, a joint purchase relationship has a weight of 0.5, and a weak social association has a weight of 0.3), association direction (clarifying the directionality of the relationship, w) ij With wji Independent calculations are performed to differentiate between different impacts, and association freshness is assessed (introducing a time decay factor; the more recent the association occurred, the higher the weight, implemented through an exponential decay function, specifically, freshness factor = exp(-λ*Δt), where Δt is the number of days since today, and λ is the decay coefficient, adjusted according to the specific business scenario and the time density of the data; the final weight w) ij = Association Strength * Freshness Factor;

[0120] Among them, the dynamic value activation threshold function is the activation threshold d for any neighbor node j. j With its own static value v j and the received edge weight w ij Related: d j =(v j ,w ij This function is designed to be a non-linear positive correlation, specifically defined as: , where d base With the base threshold and γ as the adjustment coefficient, this function means that the higher the value of a neighboring customer and the closer the relationship with the source of the dissemination, the lower the threshold for activation, that is, the easier it is to be influenced. This aligns with the business intuition that it is easier to persuade a high-value and closely related customer.

[0121] In this method, the total number of Monte Carlo simulations is set to M. At the beginning of each simulation, only the target customer node i is marked as activated (the source of infection), while all other nodes in the network are marked as inactive. In each round of propagation, the outgoing neighbors of all activated nodes are traversed, and the above steps are repeated until no new nodes are activated. The set of all activated nodes in this simulation is recorded, and the sum of the static values ​​of all activated nodes in this simulation is calculated. The distributed computing framework is used to execute M simulations, and the average of the results of all M Monte Carlo simulations is taken to obtain the network influence weight of the customer node.

[0122] In one embodiment of the present invention, step S4, which involves constructing and solving a multi-objective optimization function, includes the following steps:

[0123] The construction of the multi-objective optimization function specifically involves: constructing an objective vector for each customer, wherein the objective vector consists of three sub-objective functions, including:

[0124] ; ; ;

[0125] Where CLV(c) is the dynamic lifecycle value, P churn (c) represents the continuous risk probability, and L(c) represents the network influence weight;

[0126] Based on dynamic lifecycle value and continuous risk probability, customers are divided into four core types through a dual-threshold network, and a specific network influence weight coefficient w3(c) is configured for each type.

[0127] Calculate the resource allocation coefficient for each customer, specifically as follows:

[0128] ;

[0129] Where w1 and w2 are the fixed weight coefficients of the corresponding sub-functions, w3(c) is the network influence weight coefficient, and N[·] function is used to normalize the target value within the parentheses.

[0130] Specifically, a high threshold C_high for dynamic lifetime value and a low threshold C_low for continuous risk probability are set. These thresholds are obtained based on the statistical quantiles of dynamic lifetime value and continuous risk probability. In this embodiment, C_high is set to the 70th quantile of CLV(c), and C_low is set to the P... churn (c) 30th percentile;

[0131] Based on dynamic lifecycle value and continuous risk probability, customers are divided into four core types through a dual-threshold network, and a specific network influence weight coefficient w3(c) is assigned to each type, as follows:

[0132] Core customers: CLV(c) ≥ C_high and P churn (c)≤C_low, which is configured with a high weight coefficient, and the strategy lies in the diffusion effect of the exciter network;

[0133] Risky clients: CLV(c) ≥ C_high and P churn (c) > C_low, and configure it with the first weight coefficient. The strategy is to balance value maintenance and risk control.

[0134] Developing customers, CLV(c) < C_high and P churn (c)≤C_low, and configure it with the second weighting coefficient. The strategy is to encourage it to convert into core customers.

[0135] Observe the customer, CLV(c) < C_high and P churn (c) > C_low, which is configured with a low weight coefficient. The strategy is to reduce unnecessary resource investment.

[0136] In this embodiment, the high weight coefficient is 0.8, the first medium weight coefficient and the second medium weight coefficient are 0.5, and the low weight coefficient is 0.3.

[0137] The fixed weight coefficients w1 and w2 were determined using a grid search method based on historical data, with w1 having a value of 0.6 and w2 having a value of 0.3.

[0138] In one embodiment of the present invention, step S5 includes the following steps:

[0139] Based on resource allocation coefficients, customer level tags in the customer information management system are automatically updated using preset threshold rules.

[0140] The strategy orchestration engine matches and invokes multiple atomic operations from a predefined strategy atomic operation library based on the updated customer level label and the continuous risk probability.

[0141] The strategy orchestration engine, based on the logic of a directed acyclic graph, combines multiple atomic operations into an ordered, personalized customer management strategy workflow and drives its execution.

[0142] The system continuously tracks the execution process of the strategy workflow and the subsequent behavior data of the customer after completion. The change in the resource allocation coefficient is used as the core feedback signal to evaluate the effect of the strategy and trigger the retraining of the model parameters.

[0143] Specifically, the system predefines the threshold range of resource allocation coefficients and maps them to action level labels. These labels directly determine the priority and intensity of resource investment, including: Strategic level: R(c)≥0.8; Priority level: 0.6≤R(c)<0.8; Routine level: 0.3≤R(c)<0.6; Observation level: R(c)<0.3.

[0144] A library of predefined atomic operations, where each atomic operation is an independent, executable function or service call, such as: A0001: issue a coupon for 20 off purchases over 100; A0002: assign a dedicated account manager, etc.

[0145] The strategy orchestration engine matches the list of atomic operations to be executed from the preset strategy rule table based on the customer's latest level label and continuous risk probability. For example, if customer level = strategic customer ANDP_churn(90) < 0.1THEN call [A0002, A0005];

[0146] The engine combines multiple matched atomic operations according to the logic of a directed acyclic graph (DAG) to form an ordered and personalized policy workflow. The policy orchestration engine submits the assembled DAG workflow to the workflow executor. The executor calls the microservices or APIs corresponding to each atomic operation in sequence according to the order and dependencies defined in the DAG to complete the automated execution of the entire policy and records detailed logs during the execution process.

[0147] The system continuously tracks the execution process of the strategy workflow and collects subsequent customer behavior data after the strategy is completed (e.g., after waiting 7 days or 30 days), including whether new transactions have been generated, whether the frequency of interaction has changed, and whether the resource allocation coefficient has changed. The system uses the change in the resource allocation coefficient as the core quantitative indicator to measure the effectiveness of the strategy. A positive change indicates that the strategy is effective and improves the overall value of the customer.

[0148] Please see Figure 2 As shown, this invention is a customer information update and management system based on big data, including the following modules:

[0149] Multi-source data fusion and feature extraction module: Collects internal customer transaction data, external public behavior data and customer relationship data, performs preprocessing, and constructs a multi-dimensional time series feature panel;

[0150] Customer Value and Risk Integrated Prediction Module: Inputs a multi-dimensional time-series feature panel into the integrated model, which performs two tasks by sharing a bottom-level feature encoding layer;

[0151] Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value.

[0152] Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future;

[0153] Customer network influence quantification module: Based on the customer relationship data, a directed graph with dynamically updated edge weights is constructed. Using the influence penetration model and Monte Carlo simulation method, the expected value of the entire network that a single customer node can activate is calculated and quantified as the customer's network influence weight.

[0154] Multi-objective resource allocation decision module: Integrates the dynamic life cycle value, continuous risk probability and network influence weight, constructs and solves a multi-objective optimization function to generate resource allocation coefficients;

[0155] The strategy automation execution and feedback optimization module automatically updates customer tags based on resource allocation coefficients, combines atomic operations through the strategy orchestration engine to form a personalized customer management strategy workflow, tracks strategy effectiveness, and provides feedback.

[0156] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A customer information update and management method based on big data, characterized in that, Includes the following steps: S1: Collect customer internal transaction data, external public behavior data and customer relationship data, preprocess them, and construct a multi-dimensional time series feature panel; S2: Input the multi-dimensional temporal feature panel into the ensemble model, which performs two tasks by sharing the underlying feature encoding layer; Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value. Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future; S3: Using the customer-interconnected data, construct a directed graph with dynamically updated edge weights. Utilize the influence penetration model and Monte Carlo simulation method to calculate the expected value of the entire network that a single customer node can activate, and quantify it as the customer's network influence weight. S4: Integrate the dynamic lifecycle value, continuous risk probability, and network influence weight to construct and solve a multi-objective optimization function to generate resource allocation coefficients; S5: Automatically updates customer tags based on resource allocation coefficients, and combines atomic operations through a strategy orchestration engine to form a personalized customer management strategy workflow, track strategy effectiveness, and provide feedback; Task 2 is implemented using a deep survival network based on a temporal model memory, and includes the following steps: A temporal pattern memory is introduced to learn from the historical behavior sequences of all customers in an unsupervised manner and store K typical customer behavior evolution prototype patterns. For any customer, calculate the attention weights between its current temporal features and all K prototype patterns in the memory bank to obtain a K-dimensional pattern matching weight vector a. k ; The baseline risk function of the survival analysis model is deconstructed into a weighted sum of the baseline risk sub-functions corresponding to the K prototype models, specifically as follows: ; Where h0(t|X) is the baseline risk function, t is the time point, and h 0k (t) is the baseline risk subfunction specifically learned for the k-th prototype pattern; The probability of continuous risk for an individual customer is ultimately determined by the benchmark risk function and their individual characteristics.

2. The customer information update and management method based on big data according to claim 1, characterized in that, Step S1 includes the following steps: The preprocessing includes cleaning, aligning, and privacy protection of the data; Based on the preprocessed customer time-series data, features are extracted for each customer entity, including rolling window statistical features, behavioral sequence trend features, and interactive periodic fluctuation features; The feature group generated from the three features is horizontally concatenated with the customer's static attribute features to form a multi-dimensional time-series feature panel with fixed dimensions that is unique to each customer.

3. The customer information update and management method based on big data according to claim 2, characterized in that, The feature group generated from the three features includes the following steps: The rolling window statistical feature is defined by defining rolling windows of different time lengths on the transaction and behavior data stream. Within each window, the statistics of transaction frequency and amount, the aggregation intensity of behavioral events, and the slope of the change of the statistics and aggregation intensity within the window are calculated in parallel, which together constitute a multi-scale window feature vector. Behavioral sequence trend features: The ordered behavioral sequence of customers is transformed into an embedded vector. A lightweight time series decomposition model is used to decompose the sequence into trend terms, periodic terms and residual terms. The fitting coefficient and significance level of the trend terms are extracted as long-term intention features, and the main period amplitude of the periodic terms is calculated as behavioral regularity features. Interactive periodic fluctuation features identify the correlation between transaction behavior and external time context, calculate the coupling strength coefficient between the customer's personal behavior cycle and the external public cycle, and encode the coefficient and the customer's behavioral fluctuation amplitude at key cycle phases into context-aware periodic features.

4. The customer information update and management method based on big data according to claim 1, characterized in that, In step S2, the shared underlying feature encoding layer of the integrated model is constructed through a multilayer perceptron, and its output is used as the common input of Task 1 and Task 2. The output of the temporal convolutional network in Task 1 is connected to a multi-head self-attention mechanism layer, which is used to weight the importance of features at different time steps in the multi-dimensional temporal feature panel.

5. The customer information update and management method based on big data according to claim 4, characterized in that, In Task 1, the multi-branch time-series value prediction network includes the following steps: The main sequence modeling branch uses a temporal convolutional network to process the multi-dimensional temporal feature panel to extract basic temporal patterns; The mutation point detection branch uses a lightweight temporal anomaly detector to identify positive and negative mutation points in customer behavior sequences in real time, and outputs the mutation point confidence and feature vector. The periodic memory branch utilizes a gated recurrent unit network to capture long-term periodic patterns of customer behavior and generate periodic context vectors. The multi-head self-attention mechanism layer performs cross-branch attention fusion on the output features of the three branches, calculates the contribution of different branch features to the final prediction, and aggregates them to generate the final value prediction curve.

6. The customer information update and management method based on big data according to claim 1, characterized in that, Step S3 includes the following steps: Construct a customer association directed graph with dynamically updated edge weights, where nodes represent customers and the weights of directed edges are determined by association strength, association direction, and association freshness. We utilize a value penetration model based on multi-threshold coupling and quantify the network influence weight of a single customer node through a parallel Monte Carlo simulation method. The value penetration model sets a dynamic value activation threshold function, so that the probability of a neighboring node being activated is non-linearly positively correlated with its own static value and edge weight. The network influence weight is calculated by using the current customer node as the initial source of infection in multiple Monte Carlo simulations, propagating the virus on the network according to the value penetration model, and statistically analyzing the total value of the activated neighbor nodes in each simulation. Finally, the average value of all simulation results is quantified as the network influence weight of the customer node.

7. The customer information update and management method based on big data according to claim 1, characterized in that, In step S4, constructing and solving the multi-objective optimization function includes the following steps: The construction of the multi-objective optimization function specifically involves: constructing an objective vector for each customer, wherein the objective vector consists of three sub-objective functions, including: ； ； ; Where CLV(c) is the dynamic lifecycle value, P churn (c) represents the continuous risk probability, and L(c) represents the network influence weight; Based on dynamic lifecycle value and continuous risk probability, customers are divided into four core types through a dual-threshold network, and a specific network influence weight coefficient w3(c) is configured for each type. Calculate the resource allocation coefficient for each customer, specifically as follows: ; Where w1 and w2 are the fixed weight coefficients of the corresponding sub-functions, w3(c) is the network influence weight coefficient, and N[·] function is used to normalize the target value within the parentheses.

8. The customer information update and management method based on big data according to claim 1, characterized in that, Step S5 includes the following steps: Based on resource allocation coefficients, customer level tags in the customer information management system are automatically updated using preset threshold rules. The strategy orchestration engine matches and invokes multiple atomic operations from a predefined strategy atomic operation library based on the updated customer level label and the continuous risk probability. The strategy orchestration engine, based on the logic of a directed acyclic graph, combines multiple atomic operations into an ordered, personalized customer management strategy workflow and drives its execution. The system continuously tracks the execution process of the strategy workflow and the subsequent behavior data of the customer after completion. The change in the resource allocation coefficient is used as the core feedback signal to evaluate the effect of the strategy and trigger the retraining of the model parameters.

9. A customer information update and management system based on big data, used to implement the customer information update and management method based on big data as described in any one of claims 1-8, characterized in that, Includes the following modules: Multi-source data fusion and feature extraction module: Collects internal customer transaction data, external public behavior data and customer relationship data, performs preprocessing, and constructs a multi-dimensional time series feature panel; Customer Value and Risk Integrated Prediction Module: Inputs a multi-dimensional time-series feature panel into the integrated model, which performs two tasks by sharing a bottom-level feature encoding layer; Task 1: Use a multi-branch time series value prediction network combined with an attention mechanism to predict the value prediction curve for future time intervals, and use the net present value of the curve as the dynamic life cycle value. Task 2: Use an interval learning-based survival analysis model to predict the probability of continuous risk of customer churn at a specific point in the future; Customer network influence quantification module: Based on the customer relationship data, a directed graph with dynamically updated edge weights is constructed. Using the influence penetration model and Monte Carlo simulation method, the expected value of the entire network that a single customer node can activate is calculated and quantified as the customer's network influence weight. Multi-objective resource allocation decision module: Integrates the dynamic life cycle value, continuous risk probability and network influence weight, constructs and solves a multi-objective optimization function to generate resource allocation coefficients; The strategy automation execution and feedback optimization module automatically updates customer tags based on resource allocation coefficients, combines atomic operations through the strategy orchestration engine to form a personalized customer management strategy workflow, tracks strategy effectiveness, and provides feedback.

Images