Marketing red envelope digital regulation and control method and system based on user life cycle value

Abstract

The invention discloses a marketing red envelope digital regulation and control method and system based on a user life cycle value, and relates to the field of marketing red envelope digital regulation and control, and the method comprises the steps: constructing time sequence features through multi-dimensional behavior data, precisely predicting the user life cycle value through employing Temporal Convolutional Network and a time attention mechanism, and carrying out the digital regulation and control of the marketing red envelope. Reasonable allocation of the red envelope quota is realized in combination with a nonlinear excitation function, the marketing resource utilization efficiency and the customer maintenance effect are remarkably improved in cooperation with a value layering and directional putting strategy, and intelligent, precise and dynamic development of marketing is promoted.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of marketing red envelope digital regulation, in particular to a marketing red envelope digital regulation method and system based on user life cycle value. BACKGROUND

[0002] In the prior art, the distribution of marketing red envelopes usually relies on the static labels or simple historical consumption data of users, and customer stratification and incentive amount setting are performed through rule-based or empirical methods. Most systems focus on user value evaluation based on single-dimensional data, such as consumption amount or recent purchase time, and combine fixed threshold strategies to divide user levels, thereby determining the amount and frequency of red envelope issuance. At the same time, some technologies begin to introduce time series analysis or simple prediction models to make trend judgments on user behavior, but overall, they still mainly use fixed and unified incentive schemes, lacking dynamic correlation analysis of multi-channel user behavior.

[0003] However, these methods have obvious deficiencies in practical application, mainly in the insufficient mining of dynamic changes and multi-dimensional interaction characteristics of user behavior, which makes it difficult for incentive strategies to accurately reflect the time-varying characteristics of user value. In addition, marketing activities often lack fine-grained time dimension weight allocation, ignoring the difference in user behavior contribution to future value at different time periods, and cannot achieve key focus on key behavior time points, resulting in low efficiency of incentive resource allocation and the risk of over-concentration or waste. At the same time, existing systems mostly use unified incentive rules, lack flexible stratification mechanisms and personalized red envelope structures, and are difficult to meet the diverse user needs and complex marketing scenario requirements, affecting the long-term activity and loyalty of customers. SUMMARY

[0004] To solve the problems in the prior art, the purpose of the present application is to provide a marketing red envelope digital regulation method and system based on user life cycle value, which aims to accurately predict user future value through multi-dimensional behavior data deep mining and time series feature modeling, combine dynamic weighting and nonlinear incentive mechanism, realize personalized and scientific allocation of red envelope incentives, and improve marketing efficiency and customer retention through value stratification and targeted delivery, and improve overall marketing effectiveness and resource utilization.

[0005] To achieve the above purpose, the technical solution adopted by the present application is as follows: The marketing red envelope digital regulation method based on user life cycle value comprises the following steps: S1. Synchronously collect user behavior data from transaction dimension, interaction dimension and propagation dimension respectively, and construct user behavior feature sequence by cleaning, standardizing and time window dividing the user behavior data; S2. input the user behavior feature sequence as the model input, build and train the life cycle value prediction model, and predict the short-term change trend of the user value by introducing a dynamic weighting structure into the life cycle value prediction model; S3. input the new user behavior feature to be predicted into the life cycle value prediction model, output the user life cycle value prediction value at the current time, and substitute the output prediction value into the nonlinear incentive function to calculate the red envelope quota that the user should obtain; S4. embed the calculated red envelope quota into the activity mechanism, and perform targeted delivery to the corresponding user.

[0006] Specifically, the method proposed by the present application takes the user life cycle value as the core index, uses the time series behavior data collected in multiple dimensions (transaction, interaction, and propagation) synchronously, builds a rich and dynamic user portrait, and breaks through the bottleneck that the traditional static label and single-point data cannot reflect the dynamic evolution of the long-term value of the user; by introducing the TemporalConvolutional Network, the long-term and short-term dependencies and complex time sequence relationships of the behavior sequence are deeply mined, the capturing ability of the model to the internal law of the user behavior is improved, and the problems of gradient disappearance and insufficient calculation efficiency of the traditional sequence model such as RNN in long sequence processing are overcome; the time dimension attention mechanism is combined to dynamically weight, the key behavior time slice is focused on, the model expression deficiency caused by the imbalance of behavior contribution in time is solved, and the sensitive capture of the short-term value fluctuation of the user is further enhanced; multi-task joint training is adopted, the life cycle value regression and the short-term behavior trend judgment are optimized synchronously, the robustness and generalization ability of the model are improved, and the overfitting risk of single target training is reduced; a nonlinear incentive function is designed based on the accurate prediction of LTV, the relationship between the user value and the red envelope quota is mapped, the marginal effect of the incentive is reflected, the influence of the user activity is reflected, resource waste is avoided, and the incentive effect is maximized; combined with the user value stratification and the red envelope structure strategy, the precise targeted delivery of the red envelope is realized, not only the utilization rate of marketing resources is improved, but also the long-term participation and loyalty of the user are enhanced. The The scheme effectively solves the technical problems of inaccurate customer value evaluation, extensive incentive strategy and low response rate in traditional marketing , realizes the intelligent, precise and dynamic transformation of digital marketing.

[0007] Further, the present application also proposes a marketing red envelope digital regulation and control system based on user life cycle value, which is realized based on the marketing red envelope digital regulation and control method based on user life cycle value according to any one of the above, comprising: A data acquisition module is used to synchronously acquire user behavior data from transaction dimension, interaction dimension and propagation dimension respectively, and to build a user behavior feature sequence by cleaning, standardizing and time window dividing the user behavior data; The model construction module is configured to input the user behavior feature sequence as a model input, construct and train a life cycle value prediction model, and capture a short-term change trend of user value by introducing a dynamic weighting structure into the life cycle value prediction model. The model prediction module is configured to input new user behavior features to be predicted into the life cycle value prediction model, output a user life cycle value prediction value at a current time, and substitute the output prediction value into a nonlinear incentive function to calculate a red envelope amount that the user should obtain. The marketing delivery module is configured to embed the calculated red envelope amount into an activity mechanism and perform targeted delivery to the corresponding user. The model construction module specifically includes: The model construction unit is configured to construct a life cycle value prediction model with a TemporalConvolutional Network as a backbone structure according to the constructed user behavior feature sequence. The dynamic weighting unit is configured to introduce a time attention mechanism as a dynamic weighting structure according to an influence degree of each time slice in the user historical behavior on the life cycle value, score and weight fuse hidden feature representations of each time slice in the life cycle value prediction model, generate trend-aware global user features, and obtain user trend-aware features. The model training unit is configured to set a main task as a regression prediction of the user life cycle value according to the weighted fused user trend-aware features, construct an auxiliary task for judging a short-term behavior trend of the user, train the life cycle value prediction model by constructing a joint loss function, and output a life cycle value prediction value at a current time.

[0008] Specifically, the system proposed in the application comprehensively reflects dynamic changes of user behavior through multi-dimensional time sequence data, deeply mines long-term and short-term dependencies in the behavior sequence by using a Temporal Convolutional Network, and ensures reasonable transmission and effective modeling of historical behavior on future value. The time attention mechanism gives different weights to behaviors of each time slice, highlights the influence of key time periods on user value, and solves the information dilution problem caused by equal time processing. Joint multi-task training integrates life cycle value regression and short-term behavior trend judgment, and improves the stability and generalization ability of the model. The nonlinear incentive function calculates the red envelope amount according to the prediction value, reasonably allocates incentive resources through a decreasing function, avoids excessive concentration, and improves the efficiency of marketing investment. The system as a whole realizes deep fusion of behavior data and value prediction, dynamically adjusts the incentive strategy, and realizes precise and efficient customer maintenance and marketing resource allocation.

[0009] The application has the following beneficial effects: This invention, through the fusion of multi-dimensional and multi-time-segmented user behavior data, can more comprehensively and dynamically reflect the changing trends of user value, significantly improving the accuracy and timeliness of lifecycle value prediction; and by employing TemporalConvolutional Network combined with a time attention mechanism, it effectively captures the long-term and short-term dependencies of user behavior and key time nodes, enhancing the model's ability to perceive short-term fluctuations in user value and overcoming the shortcomings of traditional time series models in long sequence learning. This invention effectively integrates lifetime value regression and short-term behavioral trend judgment through a multi-task joint training strategy, improving the model's adaptability to the complexity of user behavior and the stability of prediction, while reducing overfitting. By designing a nonlinear red envelope incentive function, it achieves a scientific mapping between red envelope amount and user future value, reasonably controls the incentive marginal effect, and avoids the ineffective waste of marketing resources. This invention combines user value tiering tags and a pre-set red envelope structure strategy to achieve precise targeted distribution of red envelopes, thereby increasing red envelope usage and user activity, and promoting long-term user retention and repeat purchases. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the digital control method for marketing red envelopes based on user lifetime value, as proposed in an embodiment of the present invention. Detailed Implementation

[0011] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0012] Example 1 like Figure 1 The digital control method for marketing coupons based on user lifetime value includes the following steps: S1. Collect user behavior data simultaneously from the transaction, interaction, and dissemination dimensions. Construct a user behavior feature sequence by cleaning, standardizing, and dividing the user behavior data into time windows. S2. Using user behavior feature sequences as model input, construct and train a lifetime value prediction model, and predict the short-term trend of user value changes by introducing a dynamic weighting structure into the lifetime value prediction model. S3. Input new user behavior features to be predicted into the lifetime value prediction model, output the predicted value of the user's lifetime value at the current moment, and substitute the output predicted value into the nonlinear activation function to calculate the amount of red envelope that the user should receive; S4. Embed the calculated red envelope amount into the activity mechanism and target it to the corresponding users.

[0013] Specifically, for step S1, user behavior data is collected simultaneously from the transaction, interaction, and dissemination dimensions to comprehensively capture diverse user behaviors during consumption, usage, and social dissemination, overcoming the limitations of traditional single-data collection that leads to information gaps. After data collection, standardization is used to eliminate dimensional differences, and behavioral data is divided into fixed time windows (daily or weekly) to ensure the timeliness and continuity of the time series. The three-dimensional features within each time window are statistically summarized and concatenated to form a unified multi-dimensional behavioral feature vector, constituting a high-dimensional multi-channel time series. This process fully utilizes time alignment technology to keep different dimensions of behavioral features synchronized over time, providing a standard format input for the time series model and enabling a fine depiction of dynamic changes in user behavior.

[0014] Furthermore, the content related to red envelope marketing mentioned in the above embodiments refers to the subcategories of electronic incentive resources / data and virtual incentive resources / data. The solutions in the embodiments are not only applicable to the field of marketing red envelope incentive strategies, but also to other technical solutions related to electronic incentive resources / data and virtual incentive resources / data. For example, the digital control method of electronic incentive resources based on user lifetime value and the digital control method of virtual incentive resources based on user lifetime value can both be implemented using the technical solutions in steps S1-S4. Specifically, the digital control method of electronic incentive resources based on user lifetime value includes the following steps: A1. Collect user behavior data simultaneously from the transaction, interaction, and dissemination dimensions. Construct a user behavior feature sequence by cleaning, standardizing, and dividing the user behavior data into time windows. A2. Using user behavior feature sequences as model input, construct and train a lifetime value prediction model, and capture the short-term trend of user value changes by introducing a dynamic weighting structure into the lifetime value prediction model. A3. Input new user behavior features to be predicted into the lifetime value prediction model, output the predicted value of the user's lifetime value at the current moment, and substitute the output predicted value into the nonlinear activation function to calculate the amount of red envelope that the user should receive; A4. Embed the calculated electronic incentive resources into the activity mechanism and target them to the corresponding users.

[0015] Furthermore, in step S1, the transaction dimension includes: user's historical spending amount and repurchase frequency; the interaction dimension includes: page clicks, page user dwell time, user browsing path, number of times products are added to the shopping cart, and bounce rate; the dissemination dimension includes: number of times users share, number of times they recommend to friends, number of interactive comments, and user's social influence.

[0016] Specifically, the transaction dimension mainly includes data directly related to purchasing behavior, such as users' historical spending amount and repurchase frequency. This type of data usually comes from e-commerce platforms, payment systems, or order management systems. It involves capturing user transaction flows, order details, and payment records in real-time or periodically through interfaces to achieve precise monitoring of user purchasing behavior. The interaction dimension covers user behavior and interaction data on the platform, such as page clicks, dwell time, browsing paths, number of times items are added to the shopping cart, and bounce rate. This data is mostly collected by the website or app's behavior tracking system, capturing user operation trajectories and dwell states in real-time through front-end logs or third-party analysis tools, reflecting user interests, preferences, and usage habits. The dissemination dimension focuses on users' sharing behavior on social networks, including the number of shares, the number of friend recommendations, the number of interactive comments, and user social influence metrics. This data is usually collected through social platform API interfaces or the company's own social marketing system, combined with users' social relationship chains and content dissemination paths to achieve quantitative analysis of users' word-of-mouth dissemination capabilities and social activity.

[0017] Furthermore, in step S1, the specific process of constructing a user behavior feature sequence by cleaning, standardizing, and dividing user behavior data into time windows includes: S101. Clean and standardize user behavior data, and segment the processed user behavior data according to a fixed time window, where the fixed time window is a day or a week; S102. Extract statistical features for each time slice from the transaction dimension, interaction dimension, and propagation dimension, and concatenate the extracted statistical features to form the behavioral feature vector of that time slice; S103. Arrange the behavioral feature vectors of each user in each time window in chronological order to form... Transaction sequence, interaction sequence and propagation sequence are three time sequences To reveal the stage characteristics and trend changes of user behavior evolution over time The three time series were aligned according to time segments, and features were spliced ​​together in each time segment to construct a multidimensional user behavior feature sequence in a unified format.

[0018] Specifically, time window segmentation divides continuously and asynchronously generated user behavior data into discrete and standardized time segments by dividing the data according to a preset fixed time period (such as day or week). The core principle of this process is to simplify complex dynamic behavior sequences into measurable, phased units. ​ ​This effectively captures the patterns of user activity, purchase intention, and social interaction fluctuations over time. Statistical features are extracted from the transaction, interaction, and dissemination dimensions. High-dimensional, heterogeneous, and complex behavioral data are abstracted and transformed using aggregation and quantification methods. By calculating key indicators such as frequency, duration, and number of occurrences, massive amounts of raw behavior are condensed into easily understandable and information-rich numerical features, enhancing the data's expressive power and the model's interpretability. Strict alignment and feature stitching of time series from various dimensions on a unified time slice achieves spatiotemporal fusion of cross-dimensional data. This ensures the model can simultaneously capture the comprehensive performance of different types of behavior at the same time point, promoting interaction and information sharing among multi-dimensional features and greatly improving the accuracy of characterizing and predicting the overall picture of user behavior.

[0019] Specifically, for step S2, the lifecycle value prediction model uses Temporal Convolutional Network (TCN) as the backbone network for temporal feature extraction. Leveraging its causal convolution and dilated convolution properties, it efficiently captures long-term and short-term dependencies in user behavior sequences. To address the time-varying nature of user value, a temporal attention mechanism is designed as a dynamic weighting structure. This mechanism learns the importance weights of hidden features in time slices to dynamically adjust for different time slices. Specifically, attention weights are generated based on the hidden representations of each time slice, and then all time slice information is weighted and fused to form trend-aware global user features. This design fundamentally solves the problem that traditional equal-weighting or fixed-weighting methods cannot flexibly reflect the temporal differences in user behavior, improving the model's sensitivity and expressive power to short-term fluctuations in user value. Furthermore, to finely evaluate the response potential of user behavior to the red envelope incentive strategy at different times, the scheme constructs a behavior incentive sensitivity factor. Based on the time slice features output by TCN, a parameterized Sigmoid network is used for mapping to quantify the behavior incentive responsiveness. Simultaneously, a lifecycle guiding vector is introduced to capture the overall trend of user value through global feature mapping. The two methods are integrated using element-wise multiplication and a non-linear activation function to achieve deep interaction, ultimately converting the projection vector into a scalar attention score. This mechanism effectively couples the sensitivity of local behavioral incentives with the global lifecycle trend, ensuring that the time attention weight not only reflects the importance of the behavior itself but also its position in the user's value development trajectory, thus providing more precise guidance for adjusting the amount of red envelopes.

[0020] Furthermore, step S2 specifically includes the following sub-steps: S201. Based on the constructed user behavior feature sequence, construct a lifecycle value prediction model with Temporal Convolutional Network as the backbone structure; S202. Based on the degree of influence of each time slice on the lifetime value in the user's historical behavior, a time attention mechanism is introduced as a dynamic weighting structure, and the hidden feature representations of each time slice in the Temporal Convolutional Network are scored and weighted and fused to generate trend-aware global user features, thus obtaining user trend-aware features. S203. Based on the weighted and fused user trend perception features, set the main task as the regression prediction of user lifetime value, and construct an auxiliary task to judge the short-term behavior trend of users. Train the lifetime value prediction model by constructing a joint loss function and output the lifetime value prediction value at the current time.

[0021] Specifically, the Temporal Convolutional Network (TCN), as the core component of the lifecycle value prediction model, undertakes the task of temporal modeling of multidimensional user behavioral feature sequences. Specifically, after cleaning, standardization, and time window segmentation, user behavior data forms transaction sequences, interaction sequences, and propagation sequences, which are then input into the TCN. The TCN utilizes its causal and dilated convolutional structures to effectively capture long-term and short-term dependencies in the user behavior sequences, encoding hidden feature representations for each time segment, reflecting the user's behavioral characteristics and potential value changes over different time periods. Subsequently, the model introduces a dynamic weighting structure—a time attention mechanism—to perform importance scoring and weighted fusion of the hidden features of each time segment output by the TCN, forming a trend-aware global user feature. This feature not only considers the temporal information of the behavioral sequence but also dynamically captures the short-term fluctuation trend of user value. By combining the joint training of the main task (lifecycle value regression prediction) and the auxiliary task (short-term behavioral trend discrimination), the entire lifecycle value prediction model, using the TCN as the temporal feature, improves the ability to perceive dynamic changes in user value and the prediction accuracy, supporting subsequent personalized calculation and precise distribution of bonus amounts based on predicted values. Furthermore, step S202 specifically includes the following sub-steps: S2021. Based on the feature representation of each time segment output by the Temporal Convolutional Network, construct a temporal-dimensional attention scoring network, and score the importance of the behavioral features of each time segment through the attention scoring network. The attention scoring network includes a multilayer perceptron structure. S2022. By normalizing the attention scores of each time segment, a time attention weight vector is obtained. Based on the time attention weight vector, the features of all time segments output by the Temporal Convolutional Network are weighted and fused to obtain the overall trend-aware feature representation of the user, i.e., the user trend-aware feature.

[0022] Specifically, the TCN is responsible for deep temporal feature extraction of user behavior time series in the lifecycle value prediction model. The generated hidden features carry behavioral information and potential value signals of users in each time segment. The lifecycle value prediction model relies on these high-dimensional temporal representations to capture the patterns of user value changes over time. However, simply using the TCN output may not fully reflect the differentiated contribution of behavior at different times to the user's future value. In the lifecycle value prediction model, by applying a time attention mechanism to the TCN output, the influence of hidden features in each time segment is dynamically adjusted to form trend-aware global user features. This allows the lifecycle value prediction model to not only rely on the temporal dependencies captured by the TCN, but also focus on behaviors at key time nodes, enhancing the model's ability to perceive and predict short-term fluctuations in user value, thereby improving the accuracy and application effectiveness of the entire lifecycle value prediction. Furthermore, in step S2021, the specific calculation process for assigning importance scores to the behavioral features of each time slice using the attention scoring network includes the following sub-steps: S20211. Set the Temporal Convolutional Network output to a time series vector set for the user within the time window. ,in, The behavior characteristics of the t-th time slice are represented by the following. These represent the behavioral characteristics of time slices from 1 to T, where T represents the length of the time window, i.e., the total number of time slices contained in the user behavior sequence, a positive integer, in units of days or weeks; S20212. Define behavioral motivation sensitivity factors And construct a lifecycle guidance vector related to the user's historical value tags. , wherein Let represent the behavioral incentive sensitivity factor for the t-th time slice, used to quantify the response potential of the behavioral characteristics of that time slice to the incentive strategy. This represents the Sigmoid activation function, the... The weight matrix represents the behavioral sensitivity scoring network. This represents the user behavior feature vector at time slice t, encoded by a Temporal Convolutional Network. The bias term representing the behavioral sensitivity scoring network, the Represents the lifecycle guidance vector, the Represents the lifecycle-guided vector constructor; S20213. Integrate behavioral sensitivity factors with lifecycle guidance vectors, and introduce a guided attention mechanism to score the importance of behavioral features in each time slice: ; Among them, the This represents the attention score for the t-th time slice. The transposed attention projection vector is used to assign scalar weights to the fused representation. This represents the hyperbolic tangent activation function, used for nonlinear mapping. This indicates element-wise multiplication.

[0023] Specifically, in e-commerce promotional activities, if a user browses products but doesn't make a purchase, the model may misjudge this short-term activity as a high-value signal, resulting in the distribution of larger rewards and causing marketing resources to be tilted towards users with low conversion potential. The reason for setting incentive sensitivity factors is to accurately distinguish the true potential of behavior in different time slots for incentive responses, dynamically adjust incentive weights, and avoid mistaking occasional, non-purchasing browsing behavior—"superficial activity"—as high-value signals. Incentive sensitivity factors, through in-depth quantification of users' historical behavior and behavioral characteristics, capture the correlation between behavior and incentive effects, effectively filtering out inefficient incentive recipients. This ensures that rewards are distributed more to users with genuine purchasing potential and sensitive to incentive responses, achieving precise allocation and maximum utilization of marketing resources, improving overall promotional effectiveness and ROI.

[0024] Furthermore, the lifecycle guidance vector, related to users' historical value tags, is designed to compensate for the shortcomings of purely behavioral data in comprehensively reflecting changes in users' long-term value. In e-commerce promotional scenarios, users' purchasing power and activity levels change significantly with different lifecycle stages (such as the new user development period, active period, and churn warning period). Models struggle to distinguish the value stage a user is in when a particular behavior occurs, easily leading to the application of the same incentive strategies to high-potential growth users and users nearing churn, resulting in wasted resources or increased churn risk. By using a lifecycle guidance vector based on users' historical value tags, long-term value trends are embedded into the model. The model is then guided to adjust the weight allocation of different behavioral characteristics, allowing incentive strategies to dynamically adapt based on accurate identification of the user's current stage. This strengthens the incentive development of high-potential users while promptly retaining users at risk of churn, thereby improving overall customer lifetime value and promotional effectiveness.

[0025] Specifically, for step S3, the user's red envelope amount is dynamically calculated using a non-linear incentive function based on the lifecycle value prediction results, replacing the traditional experience-based fixed amount and improving the individual adaptability of the incentive strategy. Combined with a multi-dimensional value stratification tagging system, user group management is achieved, and corresponding red envelope parameter configuration schemes are invoked for different user levels, including minimum spending thresholds, usage time limits, applicable categories, and stacking rules. The red envelope amount is decomposed into specific distribution templates under the guidance of preset rules, and each template is bound to a target user for precise distribution through a targeted push mechanism. The core principle of this process is to combine value prediction results with a rule engine to achieve structured, controllable, and personalized red envelope distribution, fully improving the utilization rate of red envelope resources and the user incentive effect, overcoming the resource waste and incentive failure problems caused by universal red envelopes in traditional marketing.

[0026] Furthermore, step S3 specifically includes the following sub-steps: S301. Input a new user behavior feature sequence and output the corresponding lifecycle value prediction value through the trained lifecycle value prediction model; S302. Using the predicted lifetime value as an input variable for the incentive strategy, substitute it into the bonus amount incentive function to calculate the personalized bonus amount that the user can obtain in the current period. , wherein Indicates the amount of the red envelope, the stated This represents the incentive function for the red envelope amount. This represents the predicted lifecycle value, the This represents the preset maximum upper limit coefficient for the red envelope amount. The base of the natural exponential function is denoted as . This represents the sensitivity coefficient of the increasing red envelope amount. It is a positive real number used to control the speed and slope of the growth of the red envelope amount with its lifetime value. The larger the value, the more sensitive the red envelope amount is to the lifetime value.

[0027] Specifically, in step S4, based on the red envelope amount obtained from lifecycle value prediction and combined with pre-set value stratification tags, users are precisely divided into categories such as high potential, high activity, churn warning, and inactive users, achieving fine-grained management of user value levels. By calling the red envelope structure strategy library, parameters such as minimum spending thresholds, usage time limits, applicable categories, and stacking rules are pre-configured for different value levels to ensure the targeting and rationality of red envelope incentives. Based on the red envelope amount and strategy library parameters, a structured decomposition algorithm refines the red envelope amount into specific distribution templates, including elements such as red envelope value, validity period, and applicable scope. The templates are precisely bound to target users, and red envelope distribution is implemented using an intelligent targeted push mechanism to ensure that red envelope resources are allocated on demand.

[0028] Furthermore, step S4 specifically includes the following sub-steps: S401. Based on the output red envelope amount, obtain the current user's value level through preset value stratification tags; S402. Based on the user's value stratification, call the preset red envelope structure strategy library to determine the red envelope parameter configuration scheme for that type of user; S403. Based on the red envelope amount and red envelope structure strategy, and combined with the preset matching rules, the red envelope amount is decomposed in a structured way to generate a specific and executable red envelope distribution template; S404. Bind the generated red envelope distribution template to the target user, and push the red envelope to the user through the targeted push mechanism.

[0029] Furthermore, the preset value stratification tags include at least: high-potential users, highly active users, churn warning users, and inactive users; The red envelope structure strategy library includes at least the following red envelope parameter configuration schemes: red envelope minimum discount threshold, usage time limit, applicable categories, and stacking rules.

[0030] Example 2 Furthermore, for the preset value hierarchy labels, an exemplary preset principle process is given: Based on historical user behavior data and predicted user lifetime value, a user profile database is constructed, including total historical spending, repurchase frequency, activity metrics (such as login frequency and interaction count in the past 30 days), churn risk metrics (such as the time since last active activity and the downward trend of page dwell time), and social influence metrics (such as the number of shares and recommended friends). These metrics are then normalized to eliminate dimensional differences, and different weights are assigned based on business needs and historical data analysis experience to generate a weighted comprehensive score reflecting the user's overall value and activity status. Based on the comprehensive score, and considering the actual business situation and the statistical distribution of the metrics, corresponding threshold rules are set. High-potential users are defined as those with a predicted lifetime value higher than the set threshold and medium-to-high activity levels, exhibiting strong spending potential and loyalty. Highly active users are those with significantly higher-than-average activity levels and frequent recent interactions, but with a moderate lifetime value. Churn warning users are those with significantly declining activity and churn risk metrics exceeding the warning threshold, requiring close monitoring. Silent users are those who have been inactive for a long time, have no significant recent interactions, and have low lifetime value or sparse historical behavior.

[0031] Example 3 Furthermore, for the preset red envelope structure strategy library, an exemplary preset strategy rule is provided: Lower spending thresholds are set for high-potential users, such as a ¥20 discount for purchases over ¥100, aiming to lower the spending barrier and incentivize frequent repeat purchases. The coupons are also valid for 7 days to ensure a short-term boost in user activity. For applicable product categories, the coupons are limited to the three categories with the highest historical purchase frequency to improve usage and accuracy. Regarding stacking rules, the coupons can be used in conjunction with platform discount coupons, but cannot be combined with store-specific offers to prevent excessive discounting that could lead to profit loss.

[0032] Example 4 Furthermore, as a preferred embodiment of the above embodiments, a marketing red envelope digital control system based on user lifetime value is proposed, wherein the system is implemented based on the marketing red envelope digital control method based on user lifetime value as described in any of the above embodiments, including: The data acquisition module is used to simultaneously collect user behavior data from the transaction, interaction and dissemination dimensions. By cleaning, standardizing and dividing the user behavior data into time windows, a user behavior feature sequence is constructed. The model building module is used to take user behavior feature sequences as model input, build and train a lifecycle value prediction model, and capture the short-term trend of user value changes by introducing a dynamic weighting structure into the lifecycle value prediction model. The model prediction module is used to input new user behavior features to be predicted into the lifetime value prediction model, output the user's lifetime value prediction value at the current moment, and substitute the output prediction value into the nonlinear activation function to calculate the amount of red envelope that the user should receive. The marketing delivery module is used to embed the calculated red envelope amount into the activity mechanism and target the corresponding users. Specifically, the model building module includes: The model building unit is used to construct a lifecycle value prediction model with TemporalConvolutional Network as the backbone structure based on the constructed user behavior feature sequence. The dynamic weighting unit is used to introduce a time attention mechanism as a dynamic weighting structure based on the degree of influence of each time slice on the lifetime value in the user's historical behavior. It also scores and weights the hidden feature representations of each time slice in the lifetime value prediction model to generate trend-aware global user features and obtain user trend-aware features. The model training unit is used to set the main task as regression prediction of user lifetime value based on the weighted and fused user trend perception features, and at the same time construct an auxiliary task to judge the short-term behavior trend of users. The lifetime value prediction model is trained by constructing a joint loss function and outputs the lifetime value prediction value at the current time.

[0033] Furthermore, the dynamic weighting unit specifically includes: The attention scoring subunit is used to construct a time-dimensional attention scoring network based on the feature representation of each time slice output by the lifecycle value prediction model. The attention scoring network scores the importance of the behavioral features of each time slice. The attention scoring network includes a multilayer perceptron structure. The weighted fusion subunit is used to normalize the attention scores of each time segment to obtain a time attention weight vector. Based on the time attention weight vector, the features of all time segments output by the life cycle value prediction model are weighted and fused to obtain the overall trend-aware feature representation of the user, i.e., the user trend-aware feature.

[0034] Specifically, the implementation principle of the above embodiments is as follows: First, the data acquisition module simultaneously collects user behavior data from transaction, interaction, and dissemination dimensions. The collected data is then cleaned and standardized, and segmented into data fragments according to preset time windows (daily or weekly). Based on this, a unified format user behavior feature sequence is constructed to ensure that the multi-dimensional behavior data accurately reflects the user's time-series characteristics. Next, the model building module receives this user behavior feature sequence and uses the model building unit to construct a lifecycle value prediction model with a Temporal Convolutional Network (TCN) as its backbone. The TCN is responsible for extracting rich temporal features from the multi-dimensional time-series data. Subsequently, the dynamic weighting unit processes the hidden features output by the TCN in each time segment. An attention scoring network is constructed in the time dimension through the attention scoring subunit, and the features of each time segment are scored using a multi-layer perceptron structure to form a quantitative representation of the importance of the behavioral features in each time segment. Then, the weighted fusion subunit normalizes the scores of each time segment, generating a time attention weight vector. Based on this weight, the features of all time segments are weighted and fused to obtain the global user features for trend perception, i.e., the user trend perception features. Then, the model... The training unit sets the primary task of lifecycle value regression prediction based on this trend-aware feature, and the auxiliary task of short-term behavioral trend judgment. It constructs a joint loss function to jointly optimize the model parameters and completes the training of the lifecycle value prediction model. After training, the model prediction module accepts new user behavior feature sequence inputs, outputs the user's lifecycle value prediction value at the current moment through the trained lifecycle value prediction model, and then substitutes the prediction value into a nonlinear activation function to calculate the personalized red envelope amount. Finally, the marketing delivery module embeds a preset activity mechanism according to the calculated red envelope amount and distributes the red envelope to the corresponding users through a targeted delivery strategy, thereby achieving precise incentives for high-value users and long-term customer maintenance.

[0035] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A digital control method for marketing coupons based on user lifetime value, characterized in that: Includes the following steps: S1. Collect user behavior data simultaneously from the transaction, interaction, and dissemination dimensions. Construct a user behavior feature sequence by cleaning, standardizing, and dividing the user behavior data into time windows. S2. Using user behavior feature sequences as model input, construct and train a lifetime value prediction model, and predict the short-term trend of user value changes by introducing a dynamic weighting structure into the lifetime value prediction model. S3. Input new user behavior features to be predicted into the lifetime value prediction model, output the predicted value of the user's lifetime value at the current moment, and substitute the output predicted value into the nonlinear activation function to calculate the amount of red envelope that the user should receive; S4. Embed the calculated red envelope amount into the activity mechanism and target it to the corresponding users.

2. The marketing red envelope digital control method based on user lifetime value as described in claim 1, characterized in that, In step S1, the transaction dimension includes: user's historical spending amount and repurchase frequency; the interaction dimension includes: page clicks, page user dwell time, user browsing path, number of times products are added to the shopping cart, and bounce rate; the dissemination dimension includes: number of times users share, number of times they recommend to friends, number of interactive comments, and user's social influence.

3. The marketing red envelope digital control method based on user lifetime value as described in claim 1, characterized in that, In step S1, the specific process of constructing a user behavior feature sequence by cleaning, standardizing, and dividing user behavior data into time windows includes: S101. Clean and standardize user behavior data, and segment the processed user behavior data according to a fixed time window, where the fixed time window is a day or a week; S102. Extract statistical features for each time slice from the transaction dimension, interaction dimension, and propagation dimension, and concatenate the extracted statistical features to form the behavioral feature vector of that time slice; S103. Arrange the behavioral feature vectors of each user in each time window in chronological order to form three time series: transaction sequence, interaction sequence, and propagation sequence. Align the three time series according to time slices and splice features in each time slice to construct a multidimensional user behavior feature sequence in a unified format.

4. The marketing red envelope digital control method based on user lifetime value as described in claim 1, characterized in that, Step S2 specifically includes the following sub-steps: S201. Based on the constructed user behavior feature sequence, construct a lifecycle value prediction model with Temporal Convolutional Network as the backbone structure; S202. Based on the degree of influence of each time segment in the user's historical behavior on the lifetime value, a time attention mechanism is introduced as a dynamic weighting structure, and the hidden feature representations of each time segment in the Temporal Convolutional Network are scored and weighted and fused to generate trend-aware global user features, thus obtaining user trend-aware features. S203. Based on the weighted and fused user trend perception features, set the main task for regression prediction of user lifetime value, and construct an auxiliary task for judging short-term user behavior trends. Train the lifetime value prediction model by constructing a joint loss function and output the lifetime value prediction value at the current moment.

5. The marketing red envelope digital control method based on user lifetime value as described in claim 4, characterized in that, Step S202 specifically includes the following sub-steps: S2021. Based on the feature representation of each time segment output by the Temporal Convolutional Network, construct a temporal-dimensional attention scoring network, and score the importance of the behavioral features of each time segment through the attention scoring network. The attention scoring network includes a multilayer perceptron structure. S2022. By normalizing the attention scores of each time segment, a time attention weight vector is obtained. Based on the time attention weight vector, the features of all time segments output by the Temporal Convolutional Network are weighted and fused to obtain the overall trend-aware feature representation of the user, i.e., the user trend-aware feature.

6. The marketing red envelope digital control method based on user lifetime value as described in claim 5, characterized in that, In step S2021, the specific calculation process for scoring the importance of behavioral features in each time slice using the attention scoring network includes the following sub-steps: S20211. Set the Temporal Convolutional Network output to a time series vector set for the user within the time window. ,in, The behavior characteristics of the t-th time slice are represented by the following. These represent the behavioral characteristics of time slices from 1 to T, where T represents the length of the time window, i.e., the total number of time slices contained in the user behavior sequence, a positive integer, in units of days or weeks; S20212. Define behavioral motivation sensitivity factors And construct a lifecycle guidance vector related to the user's historical value tags. , wherein Let represent the behavioral incentive sensitivity factor for the t-th time slice, used to quantify the response potential of the behavioral characteristics of that time slice to the incentive strategy. This represents the Sigmoid activation function, the... The weight matrix represents the behavioral sensitivity scoring network. This represents the user behavior feature vector at time slice t, encoded by a Temporal Convolutional Network. The bias term representing the behavioral sensitivity scoring network, the Represents the lifecycle guidance vector, the Represents the lifecycle-guided vector constructor; S20213. Integrate behavioral sensitivity factors with lifecycle guidance vectors, and introduce a guided attention mechanism to score the importance of behavioral features in each time slice: ； Among them, the This represents the attention score for the t-th time slice. The transposed attention projection vector is described. This represents the hyperbolic tangent activation function, used for nonlinear mapping. This indicates element-wise multiplication.

7. The marketing red envelope digital control method based on user lifetime value as described in claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S301. Input a new user behavior feature sequence and output the corresponding lifecycle value prediction value through the trained lifecycle value prediction model; S302. Using the predicted lifetime value as an input variable for the incentive strategy, substitute it into the bonus amount incentive function to calculate the personalized bonus amount that the user can obtain in the current period. , wherein Indicates the amount of the red envelope, the stated This represents the incentive function for the red envelope amount. This represents the predicted lifecycle value, the This represents the preset maximum upper limit coefficient for the red envelope amount. The base of the natural exponential function is denoted as . This represents the sensitivity coefficient of the increasing red envelope amount. It is a positive real number used to control the speed and slope of the growth of the red envelope amount with its lifetime value. The larger the value, the more sensitive the red envelope amount is to the lifetime value.

8. The marketing red envelope digital control method based on user lifetime value as described in claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S401. Based on the output red envelope amount, obtain the current user's value level through preset value stratification tags; S402. Based on the user's value stratification, call the preset red envelope structure strategy library to determine the red envelope parameter configuration scheme for that type of user; S403. Based on the red envelope amount and red envelope structure strategy, and combined with the preset matching rules, the red envelope amount is decomposed in a structured way to generate a specific and executable red envelope distribution template; S404. Bind the generated red envelope distribution template to the target user, and push the red envelope to the user through the targeted push mechanism.

9. The marketing red envelope digital control method based on user lifetime value as described in claim 8, characterized in that, The preset value stratification tags include at least: high-potential users, highly active users, churn warning users, and inactive users; The red envelope structure strategy library includes at least the following red envelope parameter configuration schemes: red envelope minimum discount threshold, usage time limit, applicable categories, and stacking rules.

10. A digital control system for marketing red envelopes based on user lifetime value, wherein, This system is implemented based on any one of the marketing red envelope digital control methods based on user lifetime value according to claims 1-9, characterized in that it includes: The data acquisition module is used to simultaneously collect user behavior data from the transaction, interaction and dissemination dimensions. By cleaning, standardizing and dividing the user behavior data into time windows, a user behavior feature sequence is constructed. The model building module is used to take user behavior feature sequences as model input, build and train a lifecycle value prediction model, and capture the short-term trend of user value changes by introducing a dynamic weighting structure into the lifecycle value prediction model. The model prediction module is used to input new user behavior features to be predicted into the lifetime value prediction model, output the user's lifetime value prediction value at the current moment, and substitute the output prediction value into the nonlinear activation function to calculate the amount of red envelope that the user should receive. The marketing delivery module is used to embed the calculated red envelope amount into the activity mechanism and target the corresponding users. Specifically, the model building module includes: The model building unit is used to construct a lifecycle value prediction model with Temporal Convolutional Network as the backbone based on the constructed user behavior feature sequence. The dynamic weighting unit is used to introduce a time attention mechanism as a dynamic weighting structure based on the degree of influence of each time slice on the lifetime value in the user's historical behavior. It also scores and weights the hidden feature representations of each time slice in the lifetime value prediction model to generate trend-aware global user features and obtain user trend-aware features. The model training unit is used to set the main task as regression prediction of user lifetime value based on the weighted and fused user trend perception features, and at the same time construct an auxiliary task to judge the short-term behavior trend of users. The lifetime value prediction model is trained by constructing a joint loss function and outputs the lifetime value prediction value at the current time.

11. The marketing red envelope digital control system based on user lifetime value as described in claim 10, characterized in that, The dynamic weighting unit specifically includes: The attention scoring subunit is used to construct a time-dimensional attention scoring network based on the feature representation of each time slice output by the lifecycle value prediction model. The attention scoring network scores the importance of the behavioral features of each time slice. The attention scoring network includes a multilayer perceptron structure. The weighted fusion subunit is used to normalize the attention scores of each time segment to obtain a time attention weight vector. Based on the time attention weight vector, the features of all time segments output by the life cycle value prediction model are weighted and fused to obtain the overall trend-aware feature representation of the user, i.e., the user trend-aware feature.

Images