Member marketing automatic recommendation method and device based on user life cycle

Abstract

The invention relates to a member marketing automatic recommendation method and device based on a user life cycle, and belongs to the technical field of information push, a user behavior data set is collected, a user activity curve of a latest time period is generated, and a time period width threshold is obtained by presetting; identifying a user activeness curve, judging that the current period belongs to any period of an introduction period, a growth period, a stable period and a loss early warning period in a user life cycle, and regarding the current period as a current period; based on the member ID channel, inputting the data content mapped in the current period into a preset prediction model, and predicting the evolution of the next period under the non-interference condition; when it is predicted that the evolution of the next period belongs to a loss early warning period, interference information matched with the user behavior data set is generated, and recommendation content is adjusted based on the interference information; the loss risk pre-judgment accuracy is obviously improved; specific interference measures such as member exclusive coupons, gift rights and interests and customer service retention channels are automatically generated for early warning users.

Description

Technical Field

[0001] The present invention relates to the field of information push technology, and in particular to a member marketing automation recommendation method and device based on user life cycle. Background Art

[0002] Current membership marketing generally relies on static lifecycle segmentation rules (such as fixed time segments or RFM models), which are difficult to adapt to the dynamic changes in consumer behavior, resulting in significant lag in marketing strategies. In particular, in the user churn warning process, existing technologies rely primarily on manually set thresholds to trigger standardized retention plans. This is not only slow to respond but also unable to predict the path of natural user churn. Furthermore, cross-channel recommendation resource allocation lacks deep integration with lifecycle stages, resulting in insufficient services for high-value users and inefficient resource allocation. Summary of the Invention

[0003] To achieve the above-mentioned purpose, the present invention provides a member marketing automated recommendation method based on user life cycle, comprising the following steps:

[0004] Collect user behavior data sets, including consumption intervals, consumption frequencies, and the timestamp of the most recent interaction;

[0005] Generate a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, where the period width threshold is preset;

[0006] Identify the user activity curve and determine whether the current period belongs to any of the introduction period, growth period, stabilization period, and churn warning period in the user life cycle, and regard it as the current period;

[0007] Based on the member ID channel, the data content of the current period mapping is input into a preset prediction model to predict the evolution of the next period without interference;

[0008] When it is predicted that the next cycle evolution belongs to the churn warning period, interference information matching the user behavior dataset is generated to adjust the recommended content based on the interference information.

[0009] Furthermore, the step of collecting the user behavior data set includes:

[0010] The user interaction event log is captured by the tracking program, and the raw data of consumption interval, consumption frequency and the latest interaction timestamp are extracted from the log. The raw data is then time-zone calibrated and outlier filtered to form a structured user row data set.

[0011] Furthermore, the step of generating a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp includes:

[0012] The consumption interval, consumption frequency and the most recent interaction timestamp are decomposed into time series, and the trend item component is extracted as the core activity indicator. The trend item sequence in the most recent period is intercepted by a preset time period width threshold, and the discrete points of the sequence are interpolated with cubic spline to generate a continuous and differentiable activity curve.

[0013] Furthermore, the step of identifying the user activity curve and determining whether the current period belongs to any of the introduction period, growth period, stabilization period, and churn warning period in the user life cycle includes:

[0014] Perform wavelet transform on the user activity curve to extract local feature vectors, input the local feature vectors into a preset life cycle classifier for pattern matching, and the classifier outputs a stage identifier of the introduction period, growth period, stable period, or churn warning period based on a gradient amplitude threshold judgment rule, where:

[0015] When the first-order derivative of the curve is continuously positive and the number of mutations of the second-order derivative exceeds the threshold, it is determined to be in the growth stage;

[0016] When the standard deviation of the curve amplitude is lower than the stable interval and the first-order derivative value approaches zero, it is determined to be in the stable period;

[0017] When the curve shows a continuous negative gradient within the preset monitoring window and the amplitude exceeds the loss critical line, the loss warning period is triggered.

[0018] Furthermore, based on the member ID channel, the data content of the current period mapping is input into a preset prediction model to predict the evolution of the next period without interference, including:

[0019] Recombining the current life cycle stage identifier and the corresponding user activity curve feature vector into time series input data through a distributed data pipeline associated with the member identifier;

[0020] Load the preset hidden Markov state transition probability matrix model, use the time series input data as the observation sequence, calculate the maximum probability state path under the condition of no external interference, and output the prediction result of the next cycle stage, wherein the state transition probability matrix is ​​obtained by training the historical life cycle stage migration data.

[0021] Furthermore, the steps of loading a preset hidden Markov state transition probability matrix model, taking the time series input data as the observation sequence, calculating the maximum probability state path under the condition of no external interference, and outputting the prediction result of the next cycle stage include:

[0022] Loading pre-built state transition probability matrices and observation probability matrices from distributed storage nodes, where the matrices are generated by training with historical lifecycle stage migration data;

[0023] Encoding the time series input data into a discrete observation sequence, wherein the life cycle stage identifier is mapped to a set of hidden states, and the activity curve feature vector is quantized to a set of observation symbols;

[0024] Calculating a maximum probability state path corresponding to the observation sequence based on a Viterbi algorithm, wherein the state transition depends only on the state transition probability matrix;

[0025] The terminal state of the state path is extracted as a prediction result of the next cycle stage under the condition of no interference.

[0026] Furthermore, the pre-training method of the prediction model includes:

[0027] Obtain a historical user lifecycle stage migration dataset, the dataset comprising annotated stage transition sequences and corresponding activity feature vectors;

[0028] Initialize the random parameters of the state transition probability matrix and the observation probability matrix;

[0029] The matrix parameters are iteratively optimized by the Baum-Welch algorithm until the convergence condition is met:

[0030] -E step calculates the expected statistics of the state sequence given the current parameters

[0031] -M steps to maximize the expected statistics to update the transition probability and emission probability;

[0032] Use the confusion matrix to verify the stage prediction accuracy of the optimized model;

[0033] The verified model parameters are persistently stored in the distributed storage nodes.

[0034] Furthermore, when it is predicted that the next cycle evolution belongs to the churn warning period, the step of generating interference information matching the user behavior dataset includes:

[0035] When the prediction result of the next cycle stage is the churn warning period, execute:

[0036] Extracting historical consumption preference characteristics and churn sensitivity factors from the user behavior dataset;

[0037] According to the characteristics, a preset marketing resource library is matched to generate a combination plan including at least one of the following intervention measures:

[0038] - Targeted distribution of member-exclusive coupons with limited-time discounts;

[0039] -Activate gift rights related to user consumption records;

[0040] -Assign high-priority customer service retention service channels;

[0041] The combination plan is bound to the corresponding member identifier to generate an executable intervention instruction.

[0042] Furthermore, the step of adjusting the recommended content based on the interference information includes:

[0043] Analyze the member-exclusive coupons, gift rights and customer service retention channels in the intervention instructions;

[0044] Generate a new product sequence with weighted priority based on the product range applicable to the coupon covering the corresponding categories in the original recommendation queue;

[0045] Convert the gift rights into a visual resource identifier and insert it into the recommended content display template;

[0046] Open a real-time interaction channel entrance for users who match customer service retention services;

[0047] The new product sequence, resource identifier and interactive channel entry are integrated to generate a final recommended content data package.

[0048] The present invention also proposes a member marketing automation recommendation device based on user life cycle, comprising:

[0049] A collection unit, configured to collect a user behavior data set, wherein the behavior data includes consumption interval, consumption frequency, and the most recent interaction timestamp;

[0050] An analysis unit, configured to generate a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, wherein the period width threshold is preset;

[0051] an identification unit, configured to identify the user activity curve and determine whether the current period belongs to any of the following periods: an introduction period, a growth period, a stable period, and a churn warning period in the user life cycle, and regard the current period as the current period;

[0052] A prediction unit, configured to input the data content mapped in the current period into a preset prediction model based on the member ID channel, and predict the evolution of the next period without interference;

[0053] The recommendation unit is configured to generate interference information matching the user behavior dataset when predicting that the next cycle evolution belongs to the churn warning period, so as to adjust the recommended content based on the interference information.

[0054] The member marketing automated recommendation method and device based on user life cycle provided by the present invention have the following beneficial effects:

[0055] By accurately dividing the life cycle stages through dynamic time period activity curves and combining them with the hidden Markov model to predict the stage transition path under the non-interference state, the accuracy of churn risk prediction is significantly improved; concrete intervention measures such as member-exclusive coupons, gift rights and customer service retention channels are automatically generated for warned users, realizing the "prediction-decision-execution" closed loop and improving the response speed of retention strategies; through the real-time recommendation reconstruction mechanism bound to the member identifier, it is ensured that the intervention measures accurately reach the target users. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flowchart of an automated member marketing recommendation method based on user life cycle in one embodiment of the present invention;

[0057] Figure 2 This is a structural block diagram of a member marketing automation recommendation device based on user life cycle in one embodiment of the present invention;

[0058] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0059] To make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] Reference Figure 1 , which is a flow chart of a member marketing automation recommendation method based on user life cycle proposed by the present invention, comprising the following steps:

[0061] S1, collect user behavior data sets, the behavior data includes consumption interval, consumption frequency and the latest interaction timestamp;

[0062] S2, generating a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, wherein the period width threshold is preset;

[0063] S3, identifying the user activity curve, determining whether the current period belongs to any of the introduction period, growth period, stabilization period, and churn warning period in the user life cycle, and treating it as the current period;

[0064] S4, based on the member ID channel, inputting the data content of the current period mapping into a preset prediction model to predict the evolution of the next period without interference;

[0065] S5. When it is predicted that the next cycle evolution belongs to the churn warning period, interference information matching the user behavior dataset is generated to adjust the recommended content based on the interference information.

[0066] Specifically, by dynamically monitoring user behavior data, automatically identifying their lifecycle stages, and proactively triggering precision marketing strategies when churn risks are predicted, user retention rates can be improved. The specific implementation process is as follows:

[0067] First, the user's behavioral data such as consumption intervals, consumption frequency and the latest interaction timestamp on the platform are collected in real time, and a structured data set is formed after cleaning and standardization.

[0068] Subsequently, a user activity curve for the most recent preset time period is generated based on this data. This curve, calculated using a dynamic time window, captures short-term trends in user behavior, avoiding the lag inherent in traditional static periodization. By analyzing the curve's morphological characteristics (such as fluctuation amplitude and slope), the user's current lifecycle stage can be accurately determined, including introduction, growth, stability, or churn warning.

[0069] After determining the current cycle, historical behavioral data associated with the member ID is fed into the predictive model to simulate the user's natural evolution path without external intervention. If the prediction results indicate that the user may enter the churn warning period, targeted intervention information is automatically generated, such as issuing limited-time coupons, activating gift benefits, or allocating dedicated customer service channels. These measures are not random or universal, but are tailored to the user's specific behavioral characteristics (such as decreased consumption frequency and longer interaction intervals), ensuring the accuracy and effectiveness of the intervention.

[0070] Finally, intervention measures can be dynamically integrated into user recommendations. For example, coupon-eligible products can be prioritized in the product recommendation queue, or the gift redemption link can be highlighted in the user profile. Through seamless adjustments, retention actions can be taken before users are aware of churn, significantly improving user stickiness and repurchase rates.

[0071] The implementation process of the present invention is fully automated, does not require human intervention, and can flexibly adjust parameters according to the characteristics of different industries and user groups, and has wide applicability and scalability.

[0072] In one embodiment, the step of collecting a user behavior dataset includes:

[0073] The user interaction event log is captured by the tracking program, and the raw data of consumption interval, consumption frequency and the latest interaction timestamp are extracted from the log. The raw data is then time-zone calibrated and outlier filtered to form a structured user row data set.

[0074] During implementation, a lightweight tracking program is deployed on the user client to capture all user interaction events on the platform in real time, including but not limited to product browsing, adding to cart, order payment, and membership service usage. After the raw interaction event logs are securely transmitted to the server, the data processing module extracts three core metrics: consumption interval data is obtained by calculating the time difference between adjacent order completion times; consumption frequency data is calculated by counting the number of transactions within a preset period; and the last interaction timestamp records the specific time of the user's last active action. To ensure data accuracy and consistency, the raw data is normalized, including converting all timestamps to Coordinated Universal Time (UTC) to eliminate time zone differences. Pre-defined outlier filtering rules are also applied to eliminate data records with obvious illogicality (such as repeated actions with short intervals). After processing, the normalized data is linked and consolidated by member ID, ultimately generating a structured behavioral dataset for subsequent analysis. This dataset not only contains basic numerical information but also preserves complete time series features, providing a reliable data foundation for lifecycle analysis. The entire collection process ensures comprehensive data acquisition and efficient processing while protecting user privacy.

[0075] The above-mentioned step of generating a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp includes:

[0076] The consumption interval, consumption frequency and the most recent interaction timestamp are decomposed into time series, and the trend item component is extracted as the core activity indicator. The trend item sequence in the most recent period is intercepted by a preset time period width threshold, and the discrete points of the sequence are interpolated with cubic spline to generate a continuous and differentiable activity curve.

[0077] Specifically, the method constructs a user activity curve using advanced time series analysis methods. First, the collected time series data—consumption interval, consumption frequency, and the timestamp of the most recent interaction—is decomposed and processed. Statistical methods are used to isolate the core components representing long-term trends, filtering out interfering factors such as seasonal fluctuations and random noise. The trend component truly reflects the underlying dynamics of user activity. Based on a preset time period width threshold (such as 7 or 30 days), a data segment representing the most recent analysis period is extracted from the complete trend series. To ensure the continuity and smoothness of the curve, the extracted discrete trend values ​​are processed using a cubic spline interpolation algorithm. This method preserves the dynamic patterns of the original data while generating a smooth, continuous curve with excellent mathematical properties. The resulting activity curve not only accurately depicts the trajectory of user activity over the recent period, but its differentiability also provides a mathematical foundation for identifying key lifecycle turning points by analyzing the curve's slope. Compared to traditional simple weighted calculations, this curve generation method better captures the underlying patterns of user behavior, avoids the subjectivity of manually set weights, and ensures more objective and accurate analysis results. The entire process is fully automated without manual intervention, ensuring consistency and repeatability of analysis efficiency.

[0078] In one embodiment, the step of identifying the user activity curve and determining whether the current period belongs to any of the introduction period, growth period, stable period, and churn warning period in the user life cycle includes:

[0079] Perform wavelet transform on the user activity curve to extract local feature vectors, input the local feature vectors into a preset life cycle classifier for pattern matching, and the classifier outputs a stage identifier of the introduction period, growth period, stable period, or churn warning period based on a gradient amplitude threshold judgment rule, where:

[0080] When the first-order derivative of the curve is continuously positive and the number of mutations of the second-order derivative exceeds the threshold, it is determined to be in the growth stage;

[0081] When the standard deviation of the curve amplitude is lower than the stable interval and the first-order derivative value approaches zero, it is determined to be in the stable period;

[0082] When the curve shows a continuous negative gradient within the preset monitoring window and the amplitude exceeds the loss critical line, the loss warning period is triggered.

[0083] During implementation, a multi-level curve feature analysis method is used to accurately identify the user's lifecycle stage. Wavelet transform technology is first used to perform a multi-scale decomposition of the activity curve, extracting key parameters that characterize the curve's local morphological features. These parameters form a feature vector that describes the curve's dynamic characteristics. This feature vector is then input into a pre-trained lifecycle classification model, which has built-in intelligent judgment rules based on the curve's differential characteristics:

[0084] To identify the growth phase, we check whether the first derivative of the curve (reflecting the rate of change) remains positive, and analyze the frequency of sudden changes in the second derivative (reflecting the acceleration of change). If a sustained upward trend with multiple changes in growth rate is detected, it can be confirmed that the user is in a growth phase with rapidly increasing demand.

[0085] The determination of the stable period focuses on the smoothness of the curve. By calculating the curve fluctuation amplitude and comparing it with the standard deviation of the preset stable interval, and combining the characteristic of the first-order derivative approaching zero, a comprehensive judgment is made as to whether the user has entered a mature stage with stable consumption behavior.

[0086] The most critical churn warning period detection uses a dynamic threshold mechanism, continuously tracking the slope of the curve within a preset monitoring window. If the curve continues to decline and the decline exceeds the industry's empirical value, the warning mechanism is immediately triggered. This mathematical feature-based judgment method can identify potential churn risks earlier and more accurately than traditional fixed threshold rules.

[0087] In one embodiment, based on the member ID channel, the data content of the current period mapping is input into a preset prediction model to predict the evolution of the next period without interference, including:

[0088] Recombining the current life cycle stage identifier and the corresponding user activity curve feature vector into time series input data through a distributed data pipeline associated with the member identifier;

[0089] Load the preset hidden Markov state transition probability matrix model, use the time series input data as the observation sequence, calculate the maximum probability state path under the condition of no external interference, and output the prediction result of the next cycle stage, wherein the state transition probability matrix is ​​obtained by training the historical life cycle stage migration data.

[0090] During implementation, a complete data flow and analysis chain is constructed using the member ID as a key identifier, enabling accurate prediction of user lifecycle evolution. First, a distributed data pipeline (such as a Kafka-based message queue) is used to obtain the current user's lifecycle stage identifier and its corresponding activity curve feature vector in real time. After standardization and reorganization, the data is constructed into structured input data with time series characteristics, including a record of the user's stage transitions over the last N cycles and the corresponding activity feature change trends.

[0091] A pre-trained hidden Markov model is then loaded from the model repository. The core of this model is a state transition probability matrix that reflects the user's natural behavior patterns. This matrix is ​​trained by analyzing the real lifecycle migration data of a large number of historical users (at least 100,000+ samples), accurately capturing the natural transition patterns between different stages. During the prediction phase, the Viterbi algorithm is applied, using the current user's time series data as the observation sequence, to calculate the most likely state transition path without any external intervention.

[0092] For example, when a user is identified as currently in a "stable phase" and their activity curve shows a slight downward trend, the predictive model, based on the subsequent behavior of users with similar patterns in historical data, calculates that the probability of this user naturally entering the "churn warning phase" is 78%, while the probability of remaining stable is only 22%. This prediction considers the user's current state and the evolution of their historical behavior patterns, making the results more accurate and reliable. The entire prediction process is completed within a distributed computing framework, ensuring real-time processing of large-scale user data and providing data support for subsequent precision marketing interventions.

[0093] In one embodiment, the steps of loading a preset hidden Markov state transition probability matrix model, taking the time series input data as an observation sequence, calculating the maximum probability state path under the condition of no external interference, and outputting the prediction result of the next cycle stage include:

[0094] Loading pre-built state transition probability matrices and observation probability matrices from distributed storage nodes, where the matrices are generated by training with historical lifecycle stage migration data;

[0095] Encoding the time series input data into a discrete observation sequence, wherein the life cycle stage identifier is mapped to a set of hidden states, and the activity curve feature vector is quantized to a set of observation symbols;

[0096] Calculating a maximum probability state path corresponding to the observation sequence based on a Viterbi algorithm, wherein the state transition depends only on the state transition probability matrix;

[0097] The terminal state of the state path is extracted as a prediction result of the next cycle stage under the condition of no interference.

[0098] In a specific embodiment, accurate prediction of user lifecycle evolution is achieved through the following steps: First, pre-trained hidden Markov model parameters are retrieved from a distributed storage cluster. These parameters, including the state transition probability matrix and the observation probability matrix, are trained by analyzing massive amounts of historical user behavior data and accurately reflect the natural transition patterns between different lifecycle stages. Real-time user data collected is converted into a format that the model can process, mapping the current lifecycle stage to a hidden state in the model. Key features of the activity curve are quantized into specific observation symbols to form a complete observation sequence. Based on this pre-processed data, a Viterbi algorithm is used for in-depth analysis. The algorithm traverses all possible state transition paths and calculates the state evolution trajectory with the highest probability without any external intervention. This calculation process is entirely based on the state transition probability matrix in the model, ensuring that the prediction results objectively reflect the user's natural behavior patterns. Finally, the last state node of this optimal path is extracted as the prediction conclusion for the user's most likely lifecycle stage in the next cycle.

[0099] In one embodiment, a method for pre-training a prediction model includes:

[0100] Obtain a historical user lifecycle stage migration dataset, the dataset comprising annotated stage transition sequences and corresponding activity feature vectors;

[0101] Initialize the random parameters of the state transition probability matrix and the observation probability matrix;

[0102] The matrix parameters are iteratively optimized by the Baum-Welch algorithm until the convergence condition is met:

[0103] -E step calculates the expected statistics of the state sequence given the current parameters

[0104] -M steps to maximize the expected statistics to update the transition probability and emission probability;

[0105] Use the confusion matrix to verify the stage prediction accuracy of the optimized model;

[0106] The verified model parameters are persistently stored in the distributed storage nodes.

[0107] In a specific embodiment, the prediction model is trained using the following method: First, historical user data is collected and organized. This data includes a large number of fully annotated lifecycle stage transition records and corresponding activity characteristic change curves. This raw data is cleaned and standardized to ensure that the data quality meets modeling requirements. Subsequently, the model parameters are initialized, assigning reasonable random initial values ​​to the state transition probability matrix and the observation probability matrix, laying the foundation for subsequent iterative optimization.

[0108] The core of model training involves parameter optimization using the Baum-Welch algorithm, a classic expectation-maximization (EM) algorithm. In the expectation step (E-step), the algorithm calculates expected statistics for all possible state sequences based on the current parameters. In the maximization step (M-step), these statistics are used to update the model's transition and emission probability parameters. This process is repeated iteratively, with each iteration improving the model's log-likelihood function until the parameter change falls below a preset threshold, indicating that the model has converged to an optimal state.

[0109] After training is complete, an independent test dataset is used to evaluate the model's performance, and a confusion matrix is ​​constructed to quantify the model's prediction accuracy at each lifecycle stage. Only models that meet the preset accuracy standards (such as an overall accuracy rate exceeding 85%) will be ultimately adopted. Qualified model parameters will be serialized and stored in distributed files for online prediction services to call at any time. The entire training process is usually performed regularly in an offline environment (such as once a week) to ensure that the model can continuously learn the latest user behavior patterns. The training mechanism not only ensures the model's prediction accuracy, but also adapts to dynamic changes in business data.

[0110] In one embodiment, when it is predicted that the next cycle evolution falls within the churn warning period, the step of generating interference information matching the user behavior dataset includes:

[0111] When the prediction result of the next cycle stage is the churn warning period, execute:

[0112] Extracting historical consumption preference characteristics and churn sensitivity factors from the user behavior dataset;

[0113] According to the characteristics, a preset marketing resource library is matched to generate a combination plan including at least one of the following intervention measures:

[0114] - Targeted distribution of member-exclusive coupons with limited-time discounts;

[0115] -Activate gift rights related to user consumption records;

[0116] -Assign high-priority customer service retention service channels;

[0117] The combination plan is bound to the corresponding member identifier to generate an executable intervention instruction.

[0118] In practice, precise intervention for churned users is achieved through the following mechanisms: When the predictive model determines that a user is about to enter the churn warning period, the intervention plan generation process is immediately initiated. First, an in-depth analysis of the user's historical behavioral data is conducted to extract key characteristic indicators, including but not limited to core dimensions such as consumer category preferences (e.g., frequent purchases of maternal and infant products), price sensitivity (e.g., response rate to promotions), and service usage habits (e.g., frequent customer service consultations). A comprehensive assessment is also conducted, incorporating churn risk characteristic factors (e.g., recent decrease in visit frequency, shopping cart abandonment rate, etc.).

[0119] Based on these profile analysis results, the company intelligently matches available interventions from its marketing resource library. For example, for users who frequently purchase high-end cosmetics but whose engagement has recently decreased, a package offering a limited-edition sample plus exclusive beauty consultant service might be generated. For price-sensitive consumers of daily necessities, an incentive package offering a ¥50 discount coupon and free shipping on purchases over ¥199 might be offered. The company then selects 3-5 intervention options that best match the user's profile from the resource library, and automatically adjusts the offer based on the predicted churn risk level (e.g., offering higher discounts to high-risk users).

[0120] The intervention instructions finally generated will contain complete execution parameters: the specific denomination and usage conditions of the coupon, the method of collecting gifts and the validity period, the priority level of the customer service channel, etc. This information will be strictly bound to the user's membership account. These instructions will be encapsulated into standardized format messages and pushed to various businesses (such as coupon issuance, customer service call center, APP message push, etc.) in real time to ensure that the intervention measures can be accurately implemented. The entire process is usually completed within 5 minutes after the prediction is generated, achieving a seamless connection from risk identification to intervention execution.

[0121] In one embodiment, the step of adjusting the recommended content based on the interference information includes:

[0122] Analyze the member-exclusive coupons, gift rights and customer service retention channels in the intervention instructions;

[0123] Generate a new product sequence with weighted priority based on the product range applicable to the coupon covering the corresponding categories in the original recommendation queue;

[0124] Convert the gift rights into a visual resource identifier and insert it into the recommended content display template;

[0125] Open a real-time interaction channel entrance for users who match customer service retention services;

[0126] The new product sequence, resource identifier and interactive channel entry are integrated to generate a final recommended content data package.

[0127] In practice, multi-dimensional collaborative processing enables intelligent adjustment of recommended content. Upon receiving an intervention command, the content engine first analyzes the intervention measures contained in the command: extracting the applicable product categories, usage conditions, and expiration date of the coupon; identifying the specific content and display requirements of the gift benefits; and confirming the access method and priority settings for the customer service channel. Based on these parameters, a real-time recommendation content reconstruction process is initiated. A two-level mechanism is used to adjust the product recommendation queue: first, eligible candidate products are filtered from the original product pool based on the coupon's applicability. Then, these products are re-prioritized and weighted for display based on historical user preference data. For example, for users who receive large appliance coupons, the recommendation weight for high-priced items such as refrigerators and washing machines will be increased by 300%, and a coupon usage prompt will be highlighted on the product card. Gift benefits are then transformed into visual UI elements, such as a prominent label saying "Get a free gift worth XX yuan." These elements automatically adapt to the display templates of different devices according to platform design specifications. The integration of customer service channels uses dynamic entry technology, inserting exclusive customer service floating buttons in key paths after the user logs in (such as the top of the personal center, the shopping cart page, etc.), which will jump directly to the priority queue channel after clicking. After strict verification, all adjustment elements will be packaged into a structured data package, including: the updated product list and its sorting weight, the style and position parameters of the gift display elements, the interaction logic of the customer service entrance, etc. The data package is quickly distributed to each terminal through the CDN network to ensure that users can see the personalized adjusted recommendation content in the shortest time (usually within 30 seconds), thereby maximizing the conversion rate of retention intervention. The entire process ensures stability while achieving the real-time update capability of recommended content at the millisecond level.

[0128] Reference Attachment Figure 2 This is a structural block diagram of a member marketing automation recommendation device based on user life cycle proposed by the present invention, including:

[0129] A collection unit, configured to collect a user behavior data set, wherein the behavior data includes consumption interval, consumption frequency, and the most recent interaction timestamp;

[0130] An analysis unit, configured to generate a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, wherein the period width threshold is preset;

[0131] an identification unit, configured to identify the user activity curve and determine whether the current period belongs to any of the following periods: an introduction period, a growth period, a stable period, and a churn warning period in the user life cycle, and regard the current period as the current period;

[0132] A prediction unit, configured to input the data content mapped in the current period into a preset prediction model based on the member ID channel, and predict the evolution of the next period without interference;

[0133] The recommendation unit is configured to generate interference information matching the user behavior dataset when predicting that the next cycle evolution belongs to the churn warning period, so as to adjust the recommended content based on the interference information.

[0134] To summarize, by collecting user behavior data sets, the behavior data includes consumption intervals, consumption frequencies and the most recent interaction timestamps; based on the consumption intervals, consumption frequencies and the most recent interaction timestamps, a user activity curve for the most recent time period is generated, and the time period width threshold is preset; the user activity curve is identified, and it is determined that the current time period belongs to any period of the introduction period, growth period, stabilization period and churn warning period in the user life cycle, which is regarded as the current period; based on the member ID channel, the data content mapped by the current period is input into the preset prediction model to predict the evolution of the next period without interference; when it is predicted that the evolution of the next period belongs to the churn warning period, interference information matching the user behavior data set is generated to adjust the recommended content based on the interference information; so as to significantly improve the accuracy of churn risk prediction; concrete intervention measures such as member-exclusive coupons, gift rights and customer service retention channels are automatically generated for warned users.

[0135] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A membership marketing automation recommendation method based on user life cycle, characterized in that: The following steps are involved: Collect user behavior data sets, including consumption intervals, consumption frequencies, and the timestamp of the most recent interaction; Generate a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, where the period width threshold is preset; Identify the user activity curve and determine whether the current period belongs to any of the introduction period, growth period, stabilization period, and churn warning period in the user life cycle, and regard it as the current period; Based on the member ID channel, the data content of the current period mapping is input into a preset prediction model to predict the evolution of the next period without interference; When it is predicted that the next cycle evolution belongs to the churn warning period, interference information matching the user behavior dataset is generated to adjust the recommended content based on the interference information.

2. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: The steps for collecting user behavior datasets include: The user interaction event log is captured by the tracking program, and the raw data of consumption interval, consumption frequency and the latest interaction timestamp are extracted from the log. The raw data is then time-zone calibrated and outlier filtered to form a structured user row data set.

3. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: The step of generating a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp includes: The consumption interval, consumption frequency and the most recent interaction timestamp are decomposed into time series, and the trend item component is extracted as the core activity indicator. The trend item sequence in the most recent period is intercepted by a preset time period width threshold, and the discrete points of the sequence are interpolated with cubic spline to generate a continuous and differentiable activity curve.

4. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: The step of identifying the user activity curve and determining whether the current period belongs to any of the introduction period, growth period, stabilization period, and churn warning period in the user life cycle includes: Perform wavelet transform on the user activity curve to extract local feature vectors, input the local feature vectors into a preset life cycle classifier for pattern matching, and the classifier outputs a stage identifier of the introduction period, growth period, stable period, or churn warning period based on a gradient amplitude threshold judgment rule, where: When the first-order derivative of the curve is continuously positive and the number of mutations of the second-order derivative exceeds the threshold, it is determined to be in the growth stage; When the standard deviation of the curve amplitude is lower than the stable interval and the first-order derivative value approaches zero, it is determined to be in the stable period; When the curve shows a continuous negative gradient within the preset monitoring window and the amplitude exceeds the loss critical line, the loss warning period is triggered.

5. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: Based on the member ID channel, the data content of the current period mapping is input into a preset prediction model to predict the evolution of the next period without interference, including: Recombining the current life cycle stage identifier and the corresponding user activity curve feature vector into time series input data through a distributed data pipeline associated with the member identifier; Load the preset hidden Markov state transition probability matrix model, use the time series input data as the observation sequence, calculate the maximum probability state path under the condition of no external interference, and output the prediction result of the next cycle stage, wherein the state transition probability matrix is ​​obtained by training the historical life cycle stage migration data.

6. The member marketing automation recommendation method based on user life cycle according to claim 5 is characterized in that: The steps of loading a preset hidden Markov state transition probability matrix model, taking the time series input data as the observation sequence, calculating the maximum probability state path under the condition of no external interference, and outputting the prediction result of the next cycle stage include: Loading pre-built state transition probability matrices and observation probability matrices from distributed storage nodes, where the matrices are generated by training with historical lifecycle stage migration data; Encoding the time series input data into a discrete observation sequence, wherein the life cycle stage identifier is mapped to a set of hidden states, and the activity curve feature vector is quantized to a set of observation symbols; Calculating a maximum probability state path corresponding to the observation sequence based on a Viterbi algorithm, wherein the state transition depends only on the state transition probability matrix; The terminal state of the state path is extracted as a prediction result of the next cycle stage under the condition of no interference.

7. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: Pre-training methods for prediction models, including: Obtain a historical user lifecycle stage migration dataset, the dataset comprising annotated stage transition sequences and corresponding activity feature vectors; Initialize the random parameters of the state transition probability matrix and the observation probability matrix; The matrix parameters are iteratively optimized by the Baum-Welch algorithm until the convergence condition is met: -E step calculates the expected statistics of the state sequence given the current parameters -M steps to maximize the expected statistics to update the transition probability and emission probability; Use the confusion matrix to verify the stage prediction accuracy of the optimized model; The verified model parameters are persistently stored in the distributed storage nodes.

8. The member marketing automation recommendation method based on user life cycle according to claim 1 is characterized in that: When it is predicted that the next cycle evolution falls within the churn warning period, the step of generating interference information matching the user behavior dataset includes: When the prediction result of the next cycle stage is the churn warning period, execute: Extracting historical consumption preference characteristics and churn sensitivity factors from the user behavior dataset; According to the characteristics, a preset marketing resource library is matched to generate a combination plan including at least one of the following intervention measures: - Targeted distribution of member-exclusive coupons with limited-time discounts; -Activate gift rights related to user consumption records; -Assign high-priority customer service retention service channels; The combination plan is bound to the corresponding member identifier to generate an executable intervention instruction.

9. The member marketing automation recommendation method based on user life cycle according to claim 1, characterized in that: The step of adjusting the recommended content based on the interference information includes: Analyze the member-exclusive coupons, gift rights and customer service retention channels in the intervention instructions; Generate a new product sequence with weighted priority based on the product range applicable to the coupon covering the corresponding categories in the original recommendation queue; Convert the gift rights into a visual resource identifier and insert it into the recommended content display template; Open a real-time interaction channel entrance for users who match customer service retention services; The new product sequence, resource identifier and interactive channel entry are integrated to generate a final recommended content data package.

10. A member marketing automation recommendation device based on user life cycle, characterized in that: include: A collection unit, configured to collect a user behavior data set, wherein the behavior data includes consumption interval, consumption frequency, and the most recent interaction timestamp; An analysis unit, configured to generate a user activity curve for a recent period based on the consumption interval, consumption frequency, and the latest interaction timestamp, wherein the period width threshold is preset; an identification unit, configured to identify the user activity curve and determine whether the current period belongs to any of the following periods: an introduction period, a growth period, a stable period, and a churn warning period in the user life cycle, and regard the current period as the current period; A prediction unit, configured to input the data content mapped in the current period into a preset prediction model based on the member ID channel, and predict the evolution of the next period without interference; The recommendation unit is configured to generate interference information matching the user behavior dataset when predicting that the next cycle evolution belongs to the churn warning period, so as to adjust the recommended content based on the interference information.

Images