Intelligent marketing frequency control method based on user behavior prediction

Abstract

The invention discloses an intelligent marketing frequency control method based on user behavior prediction. According to the method, a marketing request is received and a user identifier is analyzed, user historical behavior data is acquired to construct an interest portrait, a machine learning model is utilized to predict a user response probability and a complaint risk, and a frequency control period is dynamically adjusted according to a prediction result: high-value users are shortened to 3 days, high-risk users are prolonged to 30 days, and common users are kept for 7 days. The system adopts a micro-service architecture, encapsulates an AI model into an independent service, and provides a prediction capability to the outside through an RPC framework. A cache, database and message queue three-writing mechanism is innovatively realized to guarantee data consistency, and a natural period alignment and sliding window mechanism is introduced to improve cache efficiency. And meanwhile, the functions of gray release strategy linkage, Mock time test, automatic recovery of anomaly detection and the like are supported. According to the method, the marketing precision is remarkably improved, the marketing conversion rate is improved by 30%, the user complaint rate is reduced by 60%, and personalized intelligent control over the marketing frequency is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed system and digital marketing, and particularly relates to an intelligent marketing frequency control method based on user behavior prediction. BACKGROUND

[0002] In today's digital marketing environment, it has become common for enterprises to push marketing information to users through various channels. However, how to control the marketing frequency to balance business value and user experience has become a key technical challenge. The core goal of the marketing frequency control mechanism is to prevent the same user from repeatedly receiving marketing information in a short period of time, and to avoid causing user dissatisfaction and brand trust decline. With the continuous accumulation of user behavior data and the rapid development of artificial intelligence technology, the traditional fixed rule-based frequency control method has been difficult to meet the needs of fine operation, and enterprises urgently need an intelligent solution that can dynamically adjust the marketing frequency according to user characteristics and behavior patterns. This demand has driven the marketing frequency control technology to evolve from simple rule control to intelligent and personalized direction.

[0003] Existing marketing frequency control technologies mainly adopt fixed period control strategies, and common implementation methods include N-day clearing, week clearing and month clearing mechanisms. These traditional solutions usually rely on Redis cache in combination with databases to record the number of user marketing, and limit the marketing frequency by setting fixed time windows and number thresholds. For example, the system will maintain a counter for each user in Redis to record the number of marketing information received by the user within a certain period, and stop pushing marketing information to the user when the preset threshold is reached. However, this fixed period method has obvious limitations: it cannot differentiate between users based on their actual behavior and value, and uses the same frequency control strategy for all users. High-value users may miss important marketing opportunities due to strict frequency control, while users who are not interested in marketing may feel disturbed even within the threshold. In addition, in a distributed system environment, the data consistency problem between Redis cache and database often plagues developers, and when the cache is invalid or the system fails, it may cause frequency control failure or counting errors.

[0004] Based on the above analysis, there are five major technical problems in the prior art that need to be solved. First, the lack of dynamic adjustment capability, unable to automatically adjust the frequency control period according to user behavior characteristics, resulting in low marketing efficiency; second, not making full use of artificial intelligence technology, unable to predict user response probability and complaint risk, difficult to achieve precision marketing; third, the data consistency guarantee mechanism is not perfect, and in high concurrency scenarios, the problem of cache and database data inconsistency may occur; fourth, lack of linkage mechanism with gray release strategy, unable to effectively control the risk when new strategy is online; fifth, no exception detection and automatic recovery capability, unable to respond in time when malicious user behavior or system failure occurs. The existence of these problems seriously restricts the intelligent level and operation efficiency of the marketing system, therefore the present application aims to provide an intelligent marketing frequency control method based on user behavior prediction, which comprehensively solves the above technical problems by introducing machine learning model, multi-level frequency control mechanism, cache optimization strategy and other technical means. SUMMARY

[0005] The present application provides an intelligent marketing frequency control method based on user behavior prediction, which realizes intelligent and personalized control of marketing frequency by deeply integrating artificial intelligence technology and distributed system architecture. The core idea of the whole technical solution is to take user behavior data as input, predict user response probability and complaint risk through machine learning model, and dynamically adjust the frequency control period of each user accordingly, so as to maximize marketing effect while protecting user experience.

[0006] To solve the above problems, the present application provides an intelligent marketing frequency control method based on user behavior prediction, comprising the following steps: Receiving a marketing request, parsing the user identifier and strategy code in the marketing request; Obtaining user historical behavior data, including click record, conversion record and complaint record, and constructing user interest portrait based on historical behavior data; Inputting historical behavior data into a trained machine learning model, which includes response prediction model and complaint risk prediction model, respectively outputting user response probability value and complaint risk value; Dynamically determining the frequency control period length according to the response probability value and the complaint risk value, setting the frequency control period to 3 days when the response probability value is greater than 0.8 and the complaint risk value is less than 0.1, setting the frequency control period to 30 days when the complaint risk value is greater than 0.5, and setting the frequency control period to 7 days in other cases; Querying the number of marketing times of the user in the current frequency control period, if the number of marketing times is less than or equal to the preset threshold, executing the marketing action and updating the number of marketing times, if the number of marketing times is greater than the preset threshold, skipping this marketing; The updating marketing times includes writing update data to Redis cache and MySQL database at the same time, and sending operation logs asynchronously through Kafka message queue.

[0007] Further, the constructing user interest portrait includes: inputting the user behavior sequence data into a deep learning model including Embedding layer, LSTM layer and Dense layer; mapping discrete behavior ID to continuous vector space through Embedding layer; capturing time sequence features of behavior sequence through LSTM layer; outputting 128-dimensional user interest vector through Dense layer, and the user interest vector is used as user interest portrait for subsequent strategy matching.

[0008] Further, the above method further includes an experimental group hit prediction step: extracting user portrait features and experimental group configuration features, the user portrait features including gender, age and activity, and the experimental group configuration features including exposure times and historical conversion rate; inputting the user portrait features and experimental group configuration features into XGBoost or LightGBM prediction model to output response probability of the user to the experimental group; sorting multiple experimental groups according to the response probability, and preferentially selecting the experimental group with the highest response probability to perform marketing.

[0009] Further, the Redis cache adopts natural cycle alignment mechanism: setting the expiration time of Redis cache to align with the natural cycle, and the natural cycle is 2:00 am every day; adopting sliding window mechanism to refresh cache data every hour; when the cache is expired or missed, reading data from MySQL database and reloading to the cache.

[0010] Further, the above method further includes a gray release control step: dividing users into test group and control group, the test group applying new version marketing strategy, and the control group keeping original marketing strategy; monitoring performance indicators of the test group and the control group, the performance indicators including conversion rate, complaint rate and system response time; when the performance indicators of the test group are better than those of the control group and meet preset conditions, promoting the new version marketing strategy to all users.

[0011] Further, the above method further includes a Mock time test mechanism: receive a test request containing a mockTime parameter, which specifies a simulated future date; replace the system time with the time specified by the mockTime parameter; perform frequency control logic judgment based on the replaced time to verify the marketing execution results under different periods; restore the system real time after the test is completed.

[0012] Further, the machine learning model is deployed using a microservice architecture: encapsulate the trained machine learning model as an independent microservice; expose the prediction service interface to the outside through an RPC framework, including SpringFeign, Dubbo or RESTAPI; support model hot loading and version switching to update model parameters without downtime.

[0013] Further, the above method further includes an exception detection and recovery mechanism: collect system performance indicators in real time through monitoring tools, including API response time, error rate, database connection number and cache hit rate; when the performance indicators deviate from the preset threshold range, trigger an exception alarm; when the Redis cache service is abnormal, automatically switch to the MySQL database direct query mode; when the microservice instance fails, automatically switch the traffic to the healthy instance The technical solution of the present application includes a complete execution process. The system first receives and parses the user identifier and policy code in the marketing request, then obtains the historical behavior data of the user from multiple data sources, including click, browse, purchase, complaint and other multi-dimensional information. These data are input into the pre-trained deep learning model, which uses an LSTM network structure to capture the time sequence characteristics of user behavior and generate a 128-dimensional user interest vector representation. Based on this vector representation, the system calls the response prediction model and the complaint risk prediction model to calculate the response probability of the user to the marketing activity and the risk value of possible complaints. According to these two key indicators, the system dynamically determines the user's frequency control period through a nonlinear decision algorithm: when the response probability is higher than 0.8 and the complaint risk is lower than 0.1, the system determines that the user is a high-value low-risk user, and shortens the frequency control period to 3 days; when the complaint risk is higher than 0.5, the period is extended to 30 days to reduce the disturbance; in other cases, the standard period of 7 days is maintained. After determining the period, the system queries the number of marketing times of the user within the current period, and only when it does not exceed the threshold value, the marketing action is executed.

[0014] The present application realizes breakthrough innovation in multiple technical aspects. First, in terms of intelligent decision-making, the present application first deeply integrates AI prediction models into the marketing frequency control process, uses machine learning algorithms (including XGBoost, LightGBM, etc.) to build user response prediction and complaint risk assessment models, and realizes the transition from rule-driven to data-driven. This way can capture the complex patterns of user behavior and provide more accurate personalized frequency control strategies.

[0015] Second, in terms of system architecture design, the present application innovatively proposes a cache, database, and message queue triple-write mechanism. When the marketing frequency needs to be updated, the system simultaneously writes real-time counts to Redis cache, writes persistent records to MySQL database, and asynchronously sends operation logs to data warehouse through Kafka message queue. This triple guarantee mechanism completely solves the data consistency problem in distributed environment. Even if a component fails, the system can still recover accurate count data from other channels.

[0016] Third, in terms of cache optimization strategy, the present application designs a natural cycle alignment combined with sliding window cache update mechanism. The expiration time of Redis cache is uniformly set to 2:00 am every day, aligned with the natural date cycle, avoiding the count deviation caused by cache misalignment. At the same time, the system performs sliding window refresh every hour to ensure the timeliness and accuracy of cache data, significantly improving the cache hit rate.

[0017] Fourth, in terms of risk control, the present application realizes the deep linkage of gray release strategy and frequency control mechanism. When a new marketing strategy is online, the system first selects a small part of users as the test group, monitors key indicators such as conversion rate and complaint rate in real time, and evaluates the strategy effect. Only when all indicators meet the expectations, gradually expand the scope of deployment, this progressive release method effectively reduces the strategy risk.

[0018] Finally, the present application also provides perfect test and operation support mechanism, including Mock time test function and operation manual intervention interface. Mock time mechanism allows testers to simulate future dates to verify the frequency control logic under different period settings, greatly improving test coverage. The operation intervention interface allows operation personnel to quickly suspend, resume or adjust marketing tasks in emergency situations, enhancing the controllability of the system.

[0019] The beneficial effects of the present application are: The application achieves remarkable technical effects and practical application value through the implementation of the above technical solutions. In terms of marketing efficiency improvement, through AI-driven dynamic period adjustment, the system can identify high-value users and increase their reach frequency, while reducing disturbance to low-response users. Practice shows that the marketing conversion rate can be increased by more than 30%. In terms of user experience improvement, the personalized frequency control strategy effectively reduces user complaint rate. According to the actual application data, the user complaint rate is reduced by 60%, significantly improving brand image and user satisfaction.

[0020] In terms of system performance optimization, the cache mechanism of the application controls the average response time of frequency control judgment within 200 milliseconds, meeting the performance requirements in high-concurrency scenarios. The three-write mechanism ensures 99.99% data consistency, so that the system can maintain accurate marketing counting even in the case of network partition or partial component failure. In addition, the design of micro-service architecture enables AI models to be updated and expanded independently, supporting hot loading and version switching, greatly improving the flexibility and maintainability of the system.

[0021] The application also has good scalability and universality. The frequency control idea based on user behavior prediction is not only applicable to marketing scenarios, but also can be extended to push notifications, content recommendations and other business scenarios that require frequency control. The interface design reserved by the system makes it easy to access more advanced AI algorithms such as reinforcement learning in the future, continuously optimizing the frequency control strategy. Overall, the application provides a complete, efficient and reliable technical solution for the field of intelligent marketing frequency control, with important theoretical value and wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0023] Figure 1 The system overall architecture diagram of the application: shows the overall architecture of the intelligent marketing frequency control system; Figure 2 The intelligent frequency control main flowchart of the application: describes in detail the complete process from receiving marketing requests to executing or skipping marketing; Figure 3 The dynamic period adjustment algorithm flowchart of the application: shows the decision logic of dynamically adjusting the frequency control period based on AI prediction results; Figure 4 The three-write mechanism data flowchart of the application: explains the concurrent write process of the cache, database and message queue three-write mechanism; Figure 5 Modeling architecture diagram for user interest: shows a four-layer architecture of a deep learning model. DETAILED DESCRIPTION

[0024] The application will be described in detail below with reference to the drawings and specific embodiments. It should be noted here that in order to make the embodiments more detailed, the following embodiments are the best, preferred embodiments, and other alternative ways can also be used by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the application.

[0025] It should be noted that in the specification, "one embodiment", "embodiment", "exemplary embodiment", "some embodiments" and the like indicate that the described embodiment can include a specific feature, structure or property, but not necessarily every embodiment includes this specific feature, structure or property. In addition, when a specific feature, structure or property is described in combination with an embodiment, it should be within the knowledge of those skilled in the related art to implement such a feature, structure or property in combination with other embodiments (whether or not explicitly described).

[0026] Generally, the terms can be understood at least in part from the context of their use. For example, depending on the context, the term "one or more" as used herein can be used to describe any feature, structure, or characteristic in the singular or can be used to describe combinations of features, structures, or characteristics that are combinable into one or more instances. In addition, the term "based on" can be understood as not necessarily intended to convey a set of exclusive factors, but can instead, depending on the context, allow for the existence of other factors that are not necessarily explicitly described.

[0027] Reference Figures 1 to 5 shown The application provides an intelligent marketing frequency control method based on user behavior prediction. By introducing AI model prediction, multi-level frequency control mechanism, cache optimization strategy and other technical means, the intelligent control of marketing frequency is realized. This method mainly solves the technical problems of fixed cycle frequency control in traditional marketing system, such as lack of flexibility, difficulty in ensuring the consistency of cache and database data, lack of gray release linkage, etc. The whole system adopts micro-service architecture design, encapsulates the AI model as an independent service, exposes the service interface to the outside through the RPC framework, including for example SpringFeign, Dubbo or RESTAPI, supports hot loading and version switching of the model, and ensures the flexibility and scalability of the system.

[0028] System architecture and core process implementation method The intelligent marketing frequency control system of the application is implemented according to the following steps. First, the system receives a marketing request and parses the user identification and policy code information therein. After receiving the request, the system immediately obtains the historical behavior data of the user, including the user's click record, conversion record, complaint record and other multi-dimensional information, and constructs a user interest portrait based on these data. Subsequently, the system calls the AI model service and uses the trained machine learning model to predict the user's response probability and complaint risk. According to the prediction result, the system dynamically adjusts the frequency control period length, and the high-value low-complaint user can shorten the frequency control period to 3 days, while the high-complaint risk user is extended to 30 days, and the ordinary user maintains the default period of 7 days. After determining the frequency control period, the system queries the number of marketing times of the user in the current period, and if it is less than or equal to the preset threshold, the marketing action is executed and the count is updated, otherwise the current marketing is skipped. The entire process also supports a mock time test function, allowing test personnel to simulate future dates to verify the execution logic under different periods. At the same time, the system realizes the linkage of the gray release strategy and the frequency control mechanism, and constructs a double risk control system. In order to ensure data consistency, a cache, database, message queue three-write mechanism is adopted to ensure the accuracy of data in a distributed environment. In addition, the system also provides an interface for the operator to manually intervene in the marketing batch state, and supports a cache invalidation mechanism that aligns with the sliding window and natural cycle.

[0029] User behavior modeling module implementation method User behavior modeling is the foundation of the entire system, and this module uses machine learning algorithms to model user behavior data. In specific implementation, the system uses algorithms such as decision tree, random forest, XGBoost, etc. for model training. First, a user behavior feature library is constructed, which contains user basic information (such as age, gender, region, etc.), historical behavior data (browsing, clicking, purchasing, complaining, etc.), and user preference information. Through feature engineering techniques, key features such as user activity, purchase frequency, complaint rate, etc. are extracted from the original data, which will be used for subsequent prediction and recommendation. After model training, the system encapsulates it as a microservice and provides prediction services through standardized interfaces.

[0030] Dynamic period adjustment module implementation method Dynamic cycle adjustment is one of the core innovations of the invention. This module dynamically adjusts the cycle of marketing activities based on user behavior data and market trends. When implemented, time series analysis methods such as ARIMA, Prophet, etc. are used to predict the trend of user behavior. According to the prediction results, the system dynamically adjusts the frequency and intensity of marketing activities. The specific adjustment algorithm is implemented through the DynamicFrequencyAdjuster class, which contains a calculateDynamicCycle method. The input is the user ID. First, the fetchUserBehavior method is called to obtain user behavior data from Redis or the database. Then, the AI model service is called to predict the response rate and complaint rate of the user. When the response rate is greater than 0.8 and the complaint rate is less than 0.1, the system determines that it is a high-value low-complaint user, and the frequency control cycle is shortened to 3 days. When the complaint rate is greater than 0.5, it is determined to be a high-complaint user, and the cycle is extended to 30 days. In other cases, the default cycle of 7 days is maintained. This non-linear decision-making mechanism can adjust the cycle according to the cross-judgment of response rate and complaint rate, and has strong scalability. In the future, reinforcement learning models can be connected for automatic optimization.

[0031] Task scheduling module implementation method The task scheduling module uses a distributed task scheduling framework to realize the automatic scheduling of tasks. Mature frameworks such as Apache Airflow or Azkaban can be selected. The system designs a task priority mechanism to ensure that high-priority marketing tasks are executed first. At the same time, a task retry and fault tolerance mechanism is implemented. When a task fails to execute, the system will automatically re-execute according to the configured retry strategy, improving the reliability of task execution. The scheduler will calculate the next executable marketing time point according to the user's frequency control cycle and the current time, and add the task to the scheduling queue.

[0032] Gray release module implementation method The gray release module uses blue-green deployment or canary release strategy to gradually apply new versions of marketing strategies to user groups. When implemented, the system first selects a small number of users as the test group and applies the new marketing strategy. At the same time, it monitors key performance indicators such as conversion rate, complaint rate, and system response time. Through real-time monitoring, the system can quickly roll back to the old version when problems are found. According to the results of gray release and data analysis, the operation personnel decide whether to fully promote the new version. This gradual release method greatly reduces the risk brought by new strategies.

[0033] Abnormality detection and recovery module implementation method The module uses monitoring tools such as Prometheus, Grafana, etc. to monitor system performance indicators in real time, including API response time, error rate, database connection number, cache hit rate, etc. The system designs an abnormality detection algorithm based on statistics, which triggers an alarm when a certain indicator deviates from the normal range. At the same time, an automatic abnormality recovery mechanism is implemented, for example, when the cache service is detected to be abnormal, the system automatically switches to the database direct query mode; when a microservice instance fails, the system automatically switches the traffic to a healthy instance, ensuring that the system can quickly recover to normal service under abnormal conditions.

[0034] Distributed storage module implementation method Distributed storage uses HBase or Cassandra and other distributed databases to store user behavior data and marketing strategy configurations. The system implements sharded storage of data, hashes user data according to user ID, and distributes it to different storage nodes to improve the scalability and query performance of the system. At the same time, a master-slave replication data backup mechanism is designed, with at least two copies of each data to ensure data security. When the master node fails, the system automatically switches to the slave node to ensure service continuity.

[0035] User interest Embedding representation model implementation method The user interest Embedding model is constructed using deep learning technology, specifically implemented using the TensorFlow framework. The model input is the user's behavior sequence, such as the time series data of click, browse, purchase, and other behaviors. The model structure includes Embedding layer, LSTM layer, and Dense layer. The Embedding layer maps discrete behavior IDs to continuous vector space, the LSTM layer captures the time sequence features of the behavior sequence, and finally the Dense layer outputs a 128-dimensional user interest vector. The model code is implemented through the build_user_interest_model function, which receives input dimensions, embedding dimensions, and hidden layer dimensions as parameters to build and return the complete model. The trained model outputs the Embedding vector for strategy matching and experimental group selection, and the system supports incremental training, which can update the model in real time according to new user behavior data, continuously improving the prediction effect.

[0036] Experimental group hit prediction model implementation method The hit prediction of the experimental group is implemented by using the XGBoost or LightGBM algorithm. The input features of the model include user portraits (gender, age, activity, etc.), experimental group configurations (exposure times, historical conversion rates), and current time information. The model outputs the response probability of a user to a certain experimental group. The specific implementation is completed by the ExperimentHitPredictor class, which contains the predictHitProbability method. First, the extractFeatures method is called to extract the features of the user and the experimental group, and the features are converted into JSON format. Then, the Python script is executed by calling the PythonScriptExecutor. Finally, the probability value is obtained by parsing the returned JSON result. This data-driven approach trains the model based on real data, avoiding the bias of manual rules, and supports A / B test optimization. The system will preferentially display the experimental group with high hit rate, improving marketing efficiency.

[0037] Method for implementing data consistency guarantee mechanism The application innovatively proposes a three-write mechanism of cache, database, and message queue to guarantee data consistency. When the marketing times need to be updated, the system first updates the counter in the Redis cache, uses the atomic operation of Redis to ensure concurrent safety, writes the update operation into the MySQL database as persistent storage, and sends the operation log to the data warehouse for archiving through the Kafka message queue asynchronously. This three-write mechanism ensures that even if a component fails, the system can recover data from other components, greatly improving the reliability of the system. In addition, the system also implements a natural cycle alignment cache invalidation mechanism, the invalidation time of the Redis cache is aligned with the natural cycle (such as 2:00 every day), avoiding the counting deviation caused by cache misalignment. At the same time, the sliding window mechanism is introduced, and the cache is refreshed every hour, further improving the cache hit rate.

[0038] Embodiment one The application process of the present application is illustrated below through a specific example. Assume that an e-commerce platform needs to push a promotion activity to user Zhang San, after the system receives the marketing request, it first parses the user ID as "user_001" and the strategy code as "promo_2025". The system obtains Zhang San's behavior data in the last 30 days from the database and cache, including browsing 50 goods, clicking 10 times on the promotion link, completing 3 purchases, and no complaint record. Based on these data, the system calls the user interest Embedding model to generate a 128-dimensional interest vector of Zhang San. Then the response prediction model is called to calculate the response probability of Zhang San as 0.85 and the complaint risk as 0.05. According to the dynamic frequency control algorithm, since the response rate is greater than 0.8 and the complaint rate is less than 0.1, the system sets the frequency control period of Zhang San to 3 days. It is found that Zhang San has only received 1 marketing message in the last 3 days, which is less than the threshold of 3 times, so the system performs marketing push. After the push is completed, the system updates the Redis cache (increases the counter by 1), the MySQL database (inserts the marketing record), and sends the marketing log through Kafka. The whole process is completed within 200 milliseconds, ensuring the smoothness of user experience. Through this intelligent frequency control mechanism, the marketing conversion rate of the platform is improved by 30%, and the user complaint rate is reduced by 60%, fully verifying the effectiveness and practicality of the present application.

[0039] The present application encompasses any alternatives, modifications, equivalent methods and solutions made to the essence and scope of the present application. In order to fully understand the present application by the public, specific details are described in the following preferred embodiments of the present application, and the present application can also be fully understood without the description of these details to those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits, etc. are not described in detail.

[0040] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can also be made, which should be considered as the protection scope of the present application.

Claims

1. A smart marketing frequency control method based on user behavior prediction, characterized in that, Includes the following steps: Receive marketing requests and parse the user identifier and policy code in the marketing requests; Acquire users’ historical behavior data, including click records, conversion records, and complaint records, and construct user interest profiles based on the historical behavior data; The historical behavior data is input into a trained machine learning model, which includes a response prediction model and a complaint risk prediction model, and outputs the user's response probability value and complaint risk value, respectively. The frequency control period is dynamically determined based on the response probability value and the complaint risk value. When the response probability value is greater than 0.8 and the complaint risk value is less than 0.1, the frequency control period is set to 3 days; when the complaint risk value is greater than 0.5, the frequency control period is set to 30 days; otherwise, the frequency control period is set to 7 days. The system queries the number of marketing campaigns a user has conducted within the current frequency control period. If the number of marketing campaigns is less than or equal to a preset threshold, a marketing action is executed and the number of marketing campaigns is updated. If the number of marketing campaigns is greater than the preset threshold, the current marketing campaign is skipped. The number of marketing updates includes simultaneously writing update data to the Redis cache and the MySQL database, and asynchronously sending operation logs through the Kafka message queue.

2. The method according to claim 1, characterized in that, The construction of user interest profiles includes: Input the user's behavioral sequence data into a deep learning model that includes an Embedding layer, an LSTM layer, and a Dense layer; The Embedding layer maps discrete behavior IDs to a continuous vector space. The temporal features of behavioral sequences are captured through LSTM layers; The Dense layer outputs a 128-dimensional user interest vector, which serves as a user interest profile for subsequent strategy matching.

3. The method according to claim 1, characterized in that, It also includes the experimental group hit prediction step: Extract user profile features and experimental group configuration features. The user profile features include gender, age, and activity level. The experimental group configuration features include exposure count and historical conversion rate. Input the user profile features and experimental group configuration features into the XGBoost or LightGBM prediction model, and output the probability of the user's response to the experimental group. Based on the response probabilities, multiple experimental groups are ranked, and the experimental group with the highest response probability is selected first for marketing.

4. The method according to claim 1, characterized in that, The Redis cache uses a natural lifecycle alignment mechanism: Set the expiration time of the Redis cache to align with the natural cycle, which is 2:00 AM every day; A sliding window mechanism is used to refresh the cached data every hour; When the cache expires or misses, data is read from the MySQL database and reloaded into the cache.

5. The method according to claim 1, characterized in that, It also includes the canary release control steps: Users were divided into a test group and a control group. The test group applied the new version of the marketing strategy, while the control group maintained the original marketing strategy. Monitor the performance metrics of the experimental and control groups, including conversion rate, complaint rate, and system response time; When the performance indicators of the experimental group are better than those of the control group and meet the preset conditions, the new version of the marketing strategy will be promoted to all users.

6. The method according to claim 1, characterized in that, It also includes a mock time testing mechanism: Receive a test request containing a mockTime parameter, which specifies a simulated future date; Replace the system time with the time specified by the mockTime parameter; Based on the replaced time, the frequency control logic is used to determine and verify the marketing execution results under different periods; The system time will be restored to its real time after the test is completed.

7. The method according to claim 1, characterized in that, The machine learning model is deployed using a microservices architecture. The trained machine learning model is packaged into an independent microservice; The prediction service interface is exposed to the outside world through an RPC framework, including SpringFeign, Dubbo, or RESTAPI. It supports hot reloading and version switching of models, updating model parameters without downtime.

8. The method according to any one of claims 1 to 7, characterized in that, It also includes anomaly detection and recovery mechanisms: The system performance metrics are collected in real time through monitoring tools. These metrics include API response time, error rate, database connection count, and cache hit rate. When a performance metric is detected to deviate from a preset threshold range, an anomaly alarm is triggered. When the Redis caching service fails, it automatically switches to the direct query mode of the MySQL database. When a microservice instance fails, traffic is automatically switched to a healthy instance.

Images