Data return method and device based on user value layering, equipment and medium

Abstract

The invention relates to a data return method and device based on user value layering, equipment and a medium, and the method comprises the steps: obtaining a user set completing a target conversion event in a preset time window and corresponding user attribute data, and sequentially applying a preset layering judgment rule, based on the user attribute data, each user is divided to a unique corresponding value layer, the value layer at least comprises a core realization layer, a behavior depth layer and a user feature layer, according to preset return quota configuration, the sample extraction number of each value layer of the core realization layer, the behavior depth layer and the user feature layer is determined, and according to the sample extraction number, the user attribute data is obtained; and respectively extracting a corresponding number of users from each value layer, merging to generate a return user list, and returning the return user list to the media platform. According to the method, the diversity of user sample distribution can be ensured in the total data return amount, the data return efficiency and value are improved, and a more accurate and more stable advertisement putting model can be constructed.

Description

Technical Field

[0001] This application relates to the field of computer advertising data processing, and in particular to a data feedback method, apparatus, device, and medium based on user value stratification. Background Technology

[0002] In mobile advertising performance optimization, advertisers need to send in-app user conversion data back to the media platform so that it can use this data to train and optimize ad delivery algorithms. Currently, there are two main data transmission methods commonly used in the industry: full transmission and simple random sampling transmission.

[0003] While full data transmission ensures data integrity, it has significant technical drawbacks: it generates a large amount of redundant data transmission, consuming valuable network bandwidth and server computing resources, resulting in high processing costs. More importantly, the large number of low-value samples mixed in with converted users can interfere with the media platform's model training, making it difficult for the model to accurately capture the true characteristics of high-value users, thus affecting the long-term return on investment for advertising.

[0004] To reduce backhaul costs, a simple random sampling backhaul scheme was proposed. While this scheme alleviates system load, it lacks accurate identification of user value, making it easy for some high-value samples to be missed during the sampling process. This leads to biases in the trained advertising model, resulting in poor performance in new user exploration and cross-group generalization.

[0005] Therefore, existing technologies cannot intelligently filter out the most valuable and diverse subset of users within a limited data feedback quota. This technological bottleneck makes it difficult to achieve an optimal balance between the data feedback efficiency of advertisers and the model learning effect of media platforms, thus hindering further improvements in advertising effectiveness. Summary of the Invention

[0006] Therefore, it is necessary to provide a data feedback method, device, equipment, and medium based on user value stratification to address the aforementioned technical issues. This method can identify and prioritize feedback of users with high value for model training, while ensuring the diversity of user sample distribution within the total amount of data feedback. This improves the efficiency and value of data feedback and is beneficial for media platforms to build more accurate and robust advertising delivery models.

[0007] In a first aspect, embodiments of the present invention provide a data backhaul method based on user value stratification, the method comprising:

[0008] Get the set of users who complete the target conversion event within a preset time window and the corresponding user attribute data;

[0009] The preset stratification rules are applied sequentially to classify each user into a unique corresponding value stratification based on the user attribute data; the value stratification includes at least a core monetization stratification, a behavior depth stratification, and a user feature stratification.

[0010] Based on the preset feedback quota configuration, determine the number of samples to be extracted for each value layer of the core monetization layer, behavior depth layer, and user feature layer;

[0011] Based on the number of samples extracted from each value layer, a corresponding number of users are extracted from the core monetization layer, the behavior depth layer, and the user feature layer, and then merged to generate a user list to be returned.

[0012] The returned user list will be sent back to the media platform.

[0013] Furthermore, by sequentially applying preset stratification rules, each user is categorized into a unique corresponding value stratification based on the user attribute data, including:

[0014] Applying the first judgment rule, based on the monetization data in the user attribute data, users whose recovered amount is higher than the first threshold or whose recovered amount ranks in the top N percentiles are classified into the core monetization layer, where N is a real number greater than 0;

[0015] For users not included in the core monetization layer, the second judgment rule is applied, and users who have completed at least one preset key deep behavior are included in the behavior depth layer based on the behavior sequence data in the user attribute data.

[0016] Users who are not classified into the core monetization layer and the behavior depth layer are classified into the user feature layer.

[0017] Furthermore, when classifying users into the user feature layer, sampling is performed based on the static attribute data in the user attribute data so that the sampled users cover different regions and / or different device models.

[0018] Furthermore, the method also includes:

[0019] After applying the second judgment rule, the third judgment rule is applied, and based on the user attribute data, users who are visiting for the first time and complete the target conversion or key deep behavior in the first session are classified into a new user layer.

[0020] The new user layer will be treated as a separate value layer.

[0021] Furthermore, based on the sample extraction quantity for each value layer, the corresponding number of users extracted from the core monetization layer, behavior depth layer, and user feature layer respectively include:

[0022] Based on the number of samples extracted from each value layer and the weight configuration set for the key fields in the user attribute data, the sampling priority score of each user in their respective value layer is calculated from the core monetization layer, behavior depth layer and user feature layer.

[0023] Based on the sampling priority scores, users in the core monetization layer, behavior depth layer, and user feature layer are sorted respectively.

[0024] Based on the ranking results, users with higher sampling priority scores are selected first.

[0025] Furthermore, in the core monetization layer, the key field used to calculate the sampling priority score is the amount recovered;

[0026] In the behavior depth layer, the key field used to calculate the sampling priority score is the behavior completion depth.

[0027] Furthermore, the backhaul quota configuration is dynamically optimized through an automated testing framework, and the specific steps include:

[0028] Configure multiple test groups, each test group corresponding to a different return quota configuration;

[0029] During the testing period, different advertising campaigns used different test group return quota configurations for data return;

[0030] The advertising performance metrics of each test group are collected, and the optimal backhaul quota configuration is selected based on the advertising performance metrics to update the global backhaul quota configuration. The advertising performance metrics include long-term return on investment and / or the diversity of new user characteristic distribution.

[0031] Secondly, embodiments of the present invention provide a data backhaul device based on user value stratification, the device comprising:

[0032] The acquisition module is used to acquire the set of users who have completed the target conversion event within a preset time window and the corresponding user attribute data;

[0033] The segmentation module is used to sequentially apply preset hierarchical judgment rules to segment each user into a unique corresponding value layer based on the user attribute data; the value layer includes at least a core monetization layer, a behavior depth layer, and a user feature layer;

[0034] The determination module is used to determine the number of samples to be extracted for each of the core monetization layer, the behavior depth layer, and the user feature layer according to the preset return quota configuration.

[0035] The extraction module is used to extract a corresponding number of users from the core monetization layer, the behavior depth layer, and the user feature layer respectively, based on the number of samples extracted from each value layer, and merge them to generate a user list to be returned.

[0036] The return module is used to return the user list to the media platform.

[0037] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any one of the first aspects.

[0038] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in any one of the first aspects.

[0039] The aforementioned data feedback method, apparatus, equipment, and medium based on user value stratification first acquires user data and their attributes that have completed the target conversion. Then, it sequentially applies preset rules to precisely categorize each user into a single, prioritized layer within a value system: the core monetization layer, the behavioral depth layer, and the user characteristic layer. Next, a preset feedback quota is configured to independently allocate sample extraction quantities for each of the three value layers, thereby achieving coordinated control of the total feedback volume and sample structure. Finally, the system samples according to the independent quotas for each layer and merges the samples to generate the final feedback list. Therefore, this invention, through stratified sampling, reduces the total amount of data returned while ensuring the value density of the returned samples, overcoming the resource consumption problem of full data return and the low sample value problem of random sampling. By establishing a stratified system of core monetization layer, behavior depth layer, and user feature layer, and coordinating with intra-layer priority ranking, it can identify high-value users from multiple dimensions, ensuring the maximization of the comprehensive value of the returned list. Through an automated testing framework, the returned quota configuration is continuously verified and updated, enabling the returned strategy to adapt to business changes and achieve stable improvement in long-term advertising effectiveness. By performing sampling covering different regions and device models in the user feature layer, it provides training data with a wide feature distribution for the advertising model, which helps to improve the model's generalization ability, achieves accurate identification of high-value users and effective control of sample diversity, significantly improves the value utility of data returned while reducing the amount of returned data, and provides a high-quality data foundation for media platforms to build accurate and robust advertising delivery models. Attached Figure Description

[0040] Figure 1 These are internal structural diagrams of the computer device in some embodiments;

[0041] Figure 2This is one of the flowcharts illustrating a data backhaul method based on user value stratification in some embodiments;

[0042] Figure 3 This is a schematic diagram of the process in some embodiments to classify each user into a unique corresponding value tier;

[0043] Figure 4 This is a schematic diagram of the process for extracting a corresponding number of users from each value stratum in some embodiments;

[0044] Figure 5 This is a flowchart illustrating the process of dynamically optimizing the backhaul quota configuration in some embodiments;

[0045] Figure 6 This is a schematic diagram of the framework of a data backhaul device based on user value stratification in some embodiments. Detailed Implementation

[0046] In the embodiments of this application, the term "and / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0047] In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar.

[0048] In the embodiments of this application, the term "at least one" means one or more. For example, at least one of A, B and C can represent six situations: A exists alone, B exists alone, C exists alone, A and B exist simultaneously, A and C exist simultaneously, B and C exist simultaneously, and A, B and C exist simultaneously.

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0050] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0051] In some embodiments, the data feedback method based on user value stratification provided in this application can be implemented by a server-side system of the advertiser or data manager. This server-side system, as the core execution unit for data processing and distribution, is responsible for completing key tasks such as user value analysis, stratified sampling, and data feedback.

[0052] In practical implementation, the system is deployed on server equipment with data computing and network communication capabilities, including but not limited to cloud servers, physical servers, or server clusters. This server equipment comprises the following hardware components: a processor as the computing core, responsible for performing calculations such as user value stratification, sampling priority ranking, and quota optimization; a memory including non-volatile storage media and internal memory, storing the operating system, user attribute database, and computer programs implementing this method; a communication interface supporting data connection with media platforms via network for encrypted transmission of user lists; and a database system for storing user attribute data, stratified configuration parameters, and historical transmission records.

[0053] In actual operation, the system achieves efficient processing of massive amounts of user data through the collaborative work of its various hardware components. The system stratifies users who have completed target conversion events based on preset stratification rules, determines the sampling quantity for each stratum according to the backhaul quota configuration, and completes the final user list selection based on a sampling priority algorithm. Centralized processing on the server side ensures both the consistency of the stratification logic and unified management of the backhaul strategy.

[0054] In terms of system architecture, a distributed processing model can be adopted: user data collection and preprocessing tasks are deployed on front-end data collection nodes, while core computing tasks such as value stratification and quota allocation are centrally executed by back-end servers. This architecture ensures both the efficiency of data processing and the consistency of the stratification strategy and the security of the returned data. Those skilled in the art will understand that specific server equipment may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0055] In some embodiments, such as Figure 2 As shown, a data feedback method based on user value stratification is provided, which is then applied to... Figure 1 Taking a computer device as an example, the explanation includes the following steps:

[0056] S201, obtain the set of users who completed the target conversion event within the preset time window and the corresponding user attribute data.

[0057] This step aims to accurately filter out candidate user groups and their related characteristic data with feedback value from massive amounts of user behavior data. The specific implementation process is as follows:

[0058] First, the system triggers the data acquisition process through a preset scheduled task scheduler (such as Apache Airflow or Crontab). The preset time window is a configurable traceability period, usually set to the last 7 days or the last 30 days, to ensure the timeliness of the acquired data and avoid the performance overhead of processing full historical data.

[0059] During the data acquisition process, the system initiates query requests to the data warehouse or big data platform. The query conditions include two key dimensions: first, the time dimension, which is limited to the preset time window; second, the event dimension, which is limited to users who have completed the target conversion event. The target conversion event includes, but is not limited to, key behaviors with clear commercial value such as application installation, user registration, completion of purchase, and submission of form.

[0060] To achieve efficient querying, the system employs partitioned tables and columnar storage techniques to optimize the underlying data. Query statements utilize the partitioned index on the event time field to quickly locate data blocks within the target time range, while the inverted index on the event type field precisely filters out user records that have completed the conversion event. For tables with hundreds of millions of records, this optimization can reduce query time from hours to minutes.

[0061] While acquiring the user set, the system retrieves corresponding user attribute data through multi-table join queries. This data includes three dimensions: 1) monetization data dimension, including historical cumulative payment amount, most recent payment amount, payment frequency, and other data that directly reflect the user's economic value; 2) behavioral sequence data dimension, including session duration, access frequency, and completion status of key deep behaviors, reflecting user activity; 3) static attribute data dimension, including basic characteristic data such as user region, device model, and operating system version.

[0062] To ensure data quality, the system performs a data cleaning process during data extraction, including: removing duplicate records, filling in missing fields, standardizing data formats, and verifying data logical consistency. The cleaned data will be temporarily stored in a distributed caching system (such as a Redis cluster) or a temporary data table, providing a complete and accurate data foundation for subsequent value stratification processing.

[0063] The output of step S201 is a structured dataset containing the mapping relationship between user unique identifiers and multi-dimensional attribute features, forming the original data pool for subsequent value stratification and sample extraction. Through the implementation of this step, the system establishes a high-quality and timely data input source for the entire data feedback process.

[0064] S202, apply the preset hierarchical judgment rules in sequence, and divide each user into a unique corresponding value layer based on the user attribute data; the value layer includes at least the core monetization layer, the behavior depth layer and the user feature layer.

[0065] This step is the core of this invention. Its main task is to accurately classify candidate users into different value strata based on multi-dimensional user attribute data and through preset stratification rules. The key terms and implementation process involved in this step are described in detail below:

[0066] Hierarchical judgment rules refer to a series of conditional judgment logics for value assessment based on user attribute data. These rules are executed sequentially according to a preset priority order to ensure that each user can be assigned to a unique value hierarchy. The rules can employ a tree-like decision logic, exhibiting clear exclusivity and sequential characteristics in value hierarchy.

[0067] Value stratification is a classification system based on user value characteristics, which includes at least the following three levels: the core monetization layer, which consists of users with direct economic value and is mainly identified through monetization data indicators; the behavior depth layer, which consists of users with high participation and potential value and is mainly identified through behavior sequence data; and the user feature layer, which consists of users with specific static characteristics and is mainly used to ensure the diversity of the sample.

[0068] This step can be implemented based on a hierarchical decision engine, which executes the following processing flow according to a preset rule priority order:

[0069] First, the system loads the user attribute dataset obtained in step S201 and initializes the hierarchical decision engine. The engine applies the judgment rules sequentially according to the priority order from the core monetization layer to the behavioral depth layer and then to the user feature layer.

[0070] For the division of the core monetization layer, the system applies the first judgment rule, and the specific execution process is as follows: extract monetization data dimensions from user attribute data, including indicators such as historical cumulative payment amount, highest single payment amount, and payment frequency; set dynamic threshold judgment conditions, such as marking users whose cumulative payment amount in the last 30 days exceeds 100 yuan, or whose payment amount ranks in the top 10%, as candidates; perform multi-dimensional cross-validation to ensure that the economic value of users reaches the preset standard; classify users who meet the conditions into the core monetization layer, and mark the corresponding layer identifier in the original dataset.

[0071] For users not classified into the core monetization layer, the system continues to apply the second judgment rule to classify the behavior depth layer: parsing the user's behavior sequence data, including session duration, page access depth, key behavior completion path, etc.; calculating the behavior value score of each user based on the preset behavior weight configuration; identifying users who have completed at least one key deep behavior, including but not limited to: completing the entire new user tutorial process, sharing content more than 3 times, being active for 7 consecutive days, etc.; classifying users who meet the behavior depth criteria into the behavior depth layer and updating the layer identifier.

[0072] For users remaining after the first two layers of screening, the system automatically categorizes them into the user feature layer: extracting users' static attribute data, including basic features such as geographical distribution, device model, and network environment; establishing user feature profiles to provide a data foundation for subsequent diversity sampling; and recording the stratification results to ensure that each user has one and only one value stratification identifier.

[0073] Through this step, the system completes the transformation from raw user data to structured value stratification, laying a solid foundation for subsequent sample extraction. This stratification method ensures both the priority identification of high-value users and comprehensive coverage of sample features, maximizing the value of data feedback.

[0074] S203, based on the preset backhaul quota configuration, determine the number of samples to be extracted for each value layer of the core monetization layer, behavior depth layer, and user feature layer.

[0075] The core task of this step is to determine the number of samples to be extracted from each value layer—the core monetization layer, the behavioral depth layer, and the user feature layer—based on the preset backhaul quota configuration, so as to achieve the optimal balance between the volume and quality of data backhaul.

[0076] The data return quota configuration refers to a pre-defined scheme for allocating the number of samples for each value stratum. This configuration is determined based on the media platform's model training needs, data return cost constraints, and historical performance data. In practice, the system first loads the data return quota configuration parameters, which include the overall upper limit of the return scale, the basic allocation ratio for each value stratum, the minimum guaranteed sample size for each stratum, and the dynamically adjusted weighting coefficients.

[0077] The implementation process employs a tiered quota calculation algorithm. First, the basic sample size for each tier is calculated according to a preset allocation ratio, completing the basic quota allocation. Next, a sample size compliance check is performed, where the system checks whether the existing user count for each tier meets the basic quota requirements. When a tier has insufficient users, the system initiates a quota reallocation mechanism, using the actual available users for that tier as the final quota, and redistributes the surplus quota to other tiers according to their weights. Simultaneously, a quota adjustment log is recorded for performance analysis.

[0078] During the dynamic optimization phase, the system can perform intelligent tuning based on real-time data characteristics. When the quality of core monetization layer users significantly improves, the system will automatically increase their quota ratio; for newly discovered high-value user characteristics, the system will temporarily increase the sampling quota for the corresponding layer; at the same time, based on feedback data from media platforms, the system will dynamically optimize the quota weight of each layer.

[0079] At the technical implementation level, the system establishes a configuration-based management system, allowing business personnel to intuitively adjust quota parameters through a graphical configuration interface, and enabling configuration version management, supporting rapid rollback and testing verification. The system can also establish a real-time monitoring and early warning mechanism, displaying the sampling progress of each tier in real time through a quota execution monitoring dashboard, and setting up a quota utilization rate early warning mechanism to proactively issue alerts when a tier's samples are about to be exhausted. Furthermore, the system implements an automatic scaling mechanism, intelligently adjusting the total backhaul scale in specific scenarios.

[0080] Through a feedback loop, the system can collect data on the transmission performance under various quota configurations, evaluate the optimal quota ratio using analytical models, and establish an automatic configuration optimization mechanism to continuously improve transmission efficiency. This refined quota allocation mechanism ensures the maximization of the value of limited transmission resources, meeting the media platform's demand for training data while guaranteeing the quality and diversity of transmission samples, thus laying a solid foundation for subsequent accurate sampling.

[0081] S204. Based on the number of samples extracted from each value layer, extract the corresponding number of users from the core monetization layer, behavior depth layer, and user feature layer respectively, and merge them to generate a user list to be returned.

[0082] The core task of this step is to extract the corresponding number of users from the core monetization layer, the behavioral depth layer, and the user feature layer according to the sample extraction quantity determined by each value layer, and then merge the extraction results to generate the final user list to be returned.

[0083] The sample size refers to the specific number of samples to be drawn from each stratum, calculated using the quota allocation algorithm in step S203. This number specifies the number of users to be selected from each value stratum. The returned user list refers to the final set of user identifiers formed after stratified sampling and merging. This list will be the direct object of the data return.

[0084] The system first reads the sample size for each stratum determined in step S203. For the core monetization layer, samples are sorted in descending order based on monetization data such as user recovery amount, prioritizing users with the highest value. For the behavior depth layer, samples are sorted based on the number and depth of key behaviors completed by users, selecting users with the highest engagement. For the user characteristic layer, sampling is conducted while ensuring diversity in geographical location and device model.

[0085] During the sampling process, the system employs a priority queue to manage the user ordering across different strata, ensuring efficient and stable sampling. For user feature strata, a weighted rotation algorithm guarantees uniform coverage of different feature groups. The system also implements a deduplication mechanism to prevent the same user from appearing repeatedly in multiple strata.

[0086] After each stratified sampling is completed, the system merges the sampling results to generate a unified list of users to be returned. The system standardizes the format of the list and performs data quality checks to ensure that the sample size is accurate and the user identifiers are valid, and finally outputs a list of users to be returned that meets the requirements of the media platform interface.

[0087] S205, the returned user list is returned to the media platform.

[0088] This step involves transmitting the generated list of returned users to the media platform, completing the final stage of data return. The media platform refers to the advertising service platform that receives the returned data, including but not limited to various digital advertising platforms and performance monitoring platforms.

[0089] In its implementation, the system first encapsulates the returned user list, converting it into the standard data format required by the media platform. This encapsulation process includes data field mapping, format standardization, and metadata addition. The metadata includes basic information such as the batch identifier, data generation timestamp, and data version number. Simultaneously, the system can perform data encryption, employing industry-standard encryption algorithms to encrypt sensitive fields, ensuring user privacy and data security.

[0090] After data encapsulation, the system can establish a connection with the media platform's application programming interface (API) via a secure network protocol. During the connection process, the system uses pre-configured authentication credentials for authentication, including platform access tokens, API keys, and other authentication elements. In the data transmission phase, the system employs a chunked transmission mechanism, dividing the large dataset into appropriately sized data blocks for batch transmission, and performing integrity checks after each data block is transmitted.

[0091] Meanwhile, the system establishes a robust retry and fault tolerance mechanism. When a network anomaly or transmission failure is detected, the system automatically records the failed node and retryes according to an exponential backoff strategy. For data blocks that continuously fail to transmit, the system marks them as abnormal data and transfers them to a pending queue, while simultaneously generating alarm information to notify the system administrator. After the transmission task is completed, the system collects the reception confirmation information returned by the media platform, updates the return transmission task status, and records detailed transmission logs for subsequent auditing and analysis.

[0092] Through this step, the system achieved secure and reliable transmission of the user list, ensuring that the media platform could obtain high-quality training data in a timely manner, providing necessary data support for subsequent advertising optimization.

[0093] In some embodiments, such as Figure 3 As shown, by applying preset hierarchical judgment rules sequentially, each user can be divided into a unique corresponding value hierarchy based on the user attribute data through the following steps:

[0094] S2021, applying the first judgment rule, based on the monetization data in the user attribute data, users whose recovered amount is higher than the first threshold or whose recovered amount ranks in the top N percentile are included in the core monetization layer, where N is a real number greater than 0.

[0095] Among them, monetization data is a numerical indicator reflecting the user's economic contribution, mainly including data that directly reflects the user's monetization ability, such as historical recovery amount, payment amount, and transaction frequency. Recovery amount refers to the total actual economic value generated by the user through various transaction behaviors within a specific time period, and is the core indicator for measuring the user's monetization ability.

[0096] This step applies the first judgment rule to classify high-value users into the core monetization layer based on user monetization data.

[0097] The first judgment rule can adopt a parallel screening mechanism, with the criteria being: the user's recovered amount is higher than a preset first threshold, or their recovered amount ranks in the top N percentiles (N is a real number greater than 0). Meeting either of these two conditions is sufficient, ensuring the coverage and flexibility of the screening logic. Users who meet either condition are categorized into the core monetization layer.

[0098] In its implementation, the system first extracts monetization metrics such as recovered amount from user attribute data. For threshold determination, the system compares each user's recovered amount with a first threshold dynamically set according to business objectives. For percentile determination, the system calculates the distribution of recovered amounts for all users and identifies the top N percentile group. These two determination processes can be executed in parallel, ultimately identifying users who meet either condition and incorporating them into the core monetization layer, laying the foundation for subsequent sampling.

[0099] S2022, For users not included in the core monetization layer, the second judgment rule is applied, and users who have completed at least one preset key deep behavior are included in the behavior depth layer based on the behavior sequence data in the user attribute data.

[0100] This step executes the second judgment rule, filtering out potentially valuable users from those not yet categorized into the core monetization layer and placing them into the behavioral depth layer. Behavioral sequence data records the user's complete operational path within the application, including access duration, page views, and feature usage logs. Key deep behaviors are pre-defined, high-value behavioral indicators, such as completing new user tasks, sharing content, and multiple return visits—core behaviors reflecting user engagement.

[0101] The system first receives user data from S2021 that is not categorized into the core monetization layer. Then, the system extracts behavioral sequence data from the user attribute data and analyzes the user's operation path and behavioral characteristics using a behavioral analysis engine. The system matches user behavior with a pre-defined list of key deep behaviors. These key deep behaviors are set according to business objectives and typically include completing specific tasks, reaching a certain usage frequency, or achieving important interactions. Users who complete at least one key deep behavior are categorized into the behavioral depth layer. The system adds a corresponding layer identifier to them and records the specific matching key behavior type, thus completing the user segmentation for that value layer.

[0102] S2023, For users who are not classified into the core monetization layer and the behavior depth layer, they are classified into the user feature layer.

[0103] This step is responsible for the final processing of user segmentation, categorizing users remaining after the first two layers of screening into the user feature layer. The user feature layer comprises users who do not meet the value criteria of the first two layers, and its subsequent processing is primarily based on static attribute data. Static attribute data refers to relatively stable basic characteristic data of users, including information that remains unchanged or changes infrequently, such as region, device model, and registration channel.

[0104] The system receives the remaining user group after the screening in S2021 and S2022. These users are no longer subjected to in-depth value assessment; instead, they are directly categorized into a unified user characteristic layer. During this process, the system extracts users' static attribute data and establishes feature profiles, providing a foundation for subsequent diversity-based sampling. The system also adds a unified identifier to each user in this layer, ensuring that each user is fully categorized into the layered system, achieving complete coverage of user segmentation.

[0105] In some embodiments, when classifying users into the user feature layer, sampling is performed based on static attribute data in the user attribute data so that the sampled users cover different regions and / or different device models.

[0106] In some embodiments, when samples need to be extracted from the user feature layer, the system performs targeted sampling based on the user's static attribute data. Specifically, the system identifies key attributes such as the user's geographic location and device model, and configures the sampling algorithm to ensure that the final selected samples can broadly cover different geographical regions and device types. This method effectively ensures the diversity of the returned user list in terms of basic feature dimensions, providing a data foundation for the generalization capability of the advertising model.

[0107] In some embodiments, after step S2023, the present invention may further include the following steps:

[0108] S2024, applying the third judgment rule, based on the user attribute data, users who are visiting for the first time and complete the target conversion or key deep behavior in the first session are classified into a new user layer; the new user layer is processed as an independent value layer.

[0109] In this step, the initial session refers to the continuous usage period from the user's first access to the application until they exit it. The new user tier is an independent value stratification composed of new users who meet specific criteria, and it has a unique sample processing strategy.

[0110] After completing the behavioral depth segmentation, this step performs a third judgment rule screening on the remaining user group. The system first filters out first-time users based on their access records in the user attribute data. Building on this, the system further examines these users' behavioral records during their initial session to determine whether they have completed a target conversion event (such as registration or purchase) or at least one preset key deep behavior (such as completing a tutorial or sharing content).

[0111] For users who simultaneously meet the dual criteria of first-time visit and completion of key actions in their first session, the system categorizes them into a newly established user tier. This tier will participate as an independent value level in subsequent sample quota allocation and sampling processes, ensuring that these high-potential new users receive appropriate feedback priority and providing important new user characteristic data for media platform model optimization.

[0112] In some embodiments, such as Figure 4 As shown, extracting a corresponding number of users from each value tier can be achieved through the following steps:

[0113] S2041, based on the number of samples extracted from each value layer and the weight configuration set for the key fields in the user attribute data, calculate the sampling priority score for each user in their respective value layer from the core monetization layer, behavior depth layer and user feature layer.

[0114] Weighting configuration refers to the importance coefficients assigned to different user attribute fields, used to quantify the contribution of each field to sampling priority. Sampling priority score refers to the numerical evaluation result obtained through weighted calculation, used to measure the relative importance of a user within its respective value stratum.

[0115] This step is responsible for establishing a user sampling priority sequence within each value stratum. The system first reads the preset weight configuration, which specifies the weight coefficients of key fields (such as the recovered amount in the core monetization layer and the behavioral completion depth in the behavioral depth layer) in each value stratum. Then, the system performs weighted calculations on the attribute data of each user based on the weight configuration, generating a corresponding sampling priority score. In the core monetization layer, the calculation mainly relies on monetization data fields; in the behavioral depth layer, the calculation mainly relies on behavioral sequence data; and in the user characteristic layer, the focus is on the diverse characteristics of static attribute data. Through this differentiated weighted calculation, the system establishes a quantitative ranking basis for user priority within each value stratum.

[0116] In the core monetization layer, the key field used to calculate the sampling priority score can be the amount recovered; in the behavior depth layer, the key field used to calculate the sampling priority score can be the behavior completion depth.

[0117] In other words, the sampling priority score for the core monetization layer is primarily calculated based on the recovered amount field. This field directly reflects the user's economic contribution value. The system will sort users in descending order or perform weighted calculations according to the recovered amount to ensure that high-value users receive priority sampling rights. The sampling priority score for the behavior depth layer is based on the behavior completion depth field. This field quantifies the quantity, frequency, and quality of key behaviors completed by users, comprehensively evaluating user participation and potential value. The system then prioritizes users accordingly.

[0118] S2042, Based on the sampling priority score, sort the users in the core monetization layer, behavior depth layer and user feature layer respectively.

[0119] This step prioritizes users within each value stratum. The system reads the sampling priority scores calculated by S2041 and performs sorting operations within the core monetization layer, behavioral depth layer, and user feature layer. The sorting is in descending order, placing the user with the highest score first. The system employs an efficient sorting algorithm to ensure processing performance under large datasets, establishing an ordered user sequence for subsequent sample extraction.

[0120] S2043, within each value stratum, according to the ranking result, select the number of users whose samples are ranked first in order.

[0121] This step is responsible for completing the final user sampling. Based on the ranking results generated by S2042, the system selects users sequentially from the user with the highest sampling priority score within each value stratum. The selection process continues until the number of selected users reaches the preset sample size for that stratum, ensuring that the quota requirements are accurately met for each stratum.

[0122] Suppose a data feedback task requires extracting two samples from the core monetization layer and one sample from the behavioral depth layer. The system has completed user value segmentation, and user data for each layer is ready.

[0123] In step S2041, the system calculates sampling priority scores according to preset weight configurations. The core monetization layer uses the recovered amount as the calculation basis; in this layer, user C receives 700 points, user A receives 500 points, and user B receives 300 points. The behavior depth layer uses the behavior completion depth as the calculation basis; user E receives 92 points, and user D receives 85 points.

[0124] In step S2042, the system sorts the users within each layer. In the core monetization layer, users are sorted by score from highest to lowest as User C, User A, and User B. In the behavior depth layer, the sorting results are User E and User D.

[0125] In step S2043, the system performs the final selection based on the sample size. The core monetization layer selects the top two users, user C and user A; the behavior depth layer selects user E, who is ranked first. The final user list includes user C, user A, and user E, completing this sampling task.

[0126] In some embodiments, such as Figure 5 As shown, the return quota configuration is dynamically optimized through an automated testing framework, which can be achieved through the following steps:

[0127] S2031 allows for the configuration of multiple test groups, each corresponding to a different backhaul quota configuration.

[0128] The test group can be an experimental unit with a specific backhaul quota configuration, used to compare the actual effects of different configuration schemes. The backhaul quota configuration refers to the sample size ratio scheme allocated to each value stratum.

[0129] This step is responsible for establishing the experimental environment for automated testing. The system first creates multiple independent test groups based on preset configuration parameters. Each test group corresponds to a specific return quota configuration scheme. For example: Test group A uses a quota ratio of 60% for the core monetization layer, 30% for the behavior depth layer, and 10% for the user feature layer; Test group B uses a quota ratio of 50%, 40%, and 10%; Test group C uses a quota ratio of 70%, 20%, and 10%. The system binds these configuration schemes to the corresponding test groups, preparing for subsequent comparative tests.

[0130] S2032, During the testing period, different advertising plans use different test group return quota configurations for data return.

[0131] The testing period refers to a pre-defined timeframe for experimental execution, used to ensure the integrity of data collection. An advertising campaign is an advertising activity unit with its own budget and delivery strategy.

[0132] This step is responsible for conducting the grouped control experiment. Within a preset testing period, the system assigns different advertising campaigns to various test groups created in S2031. Each advertising campaign strictly adheres to the data return quota configuration of its respective test group for data return operations. Through this grouped implementation method, the system can collect actual performance data under multiple quota configurations in parallel, providing a comparative basis for subsequent performance evaluation. After the testing period ends, the system terminates the experiment and enters the data analysis phase.

[0133] After completing the test group configuration, the system enters the test execution phase. Assuming there are three running advertising campaigns (Campaign α, Campaign β, and Campaign γ), the system assigns them to test groups A, B, and C respectively. During the 7-day test period, Campaign α consistently uses the quota configuration of Test Group A (60% core layer, 30% behavior layer, 10% feature layer) for data transmission; Campaign β uses the configuration of Test Group B (50%, 40%, 10%); and Campaign γ uses the configuration of Test Group C (70%, 20%, 10%). Each advertising campaign maintains its normal delivery strategy during the test, only changing its data transmission quota configuration to ensure the uniqueness of the experimental variables. Through this grouped comparison method, the system synchronously collects advertising performance data under different quota configurations.

[0134] S2033, collect advertising performance metrics for each test group, and select the optimal backhaul quota configuration based on the advertising performance metrics to update the global backhaul quota configuration. The advertising performance metrics include long-term return on investment and / or the diversity of new user characteristic distribution.

[0135] Among them, advertising performance metrics refer to data indicators that measure the effectiveness of advertising campaigns, including long-term return on investment and the diversity of new user characteristics. The global backhaul quota configuration is an optimized configuration scheme that will be rolled out across the entire platform.

[0136] After the testing period is completed, the system collects advertising performance data for each test group. Using the example from S2032, the system compares and analyzes the performance metrics of test groups A, B, and C. If test group B maintains a good long-term return on investment while exhibiting the best diversity in new user characteristics, the system selects the corresponding quota configuration (50% for the core layer, 40% for the behavioral layer, and 10% for the characteristic layer) as the optimal solution and updates it to the global quota configuration. Subsequently, all advertising campaigns will use this validated optimal configuration for data feedback.

[0137] Based on the same inventive concept, this application also provides a data device based on user value stratification. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more user value-stratified data backhaul device embodiments provided below can be found in the limitations of the user value-stratified data backhaul method above, and will not be repeated here.

[0138] In some embodiments, such as Figure 6 As shown, a data backhaul device based on user value stratification is provided, comprising:

[0139] Module 21 is used to acquire the set of users who completed the target conversion event within a preset time window and their corresponding user attribute data.

[0140] The segmentation module 22 is used to sequentially apply preset hierarchical judgment rules to segment each user into a unique corresponding value layer based on the user attribute data; the value layer includes at least a core monetization layer, a behavior depth layer, and a user feature layer.

[0141] Configuration module 23 is used to determine the sample extraction quantity for each of the core monetization layer, behavior depth layer, and user feature layer according to the preset return quota configuration.

[0142] Extraction module 24 is used to extract a corresponding number of users from the core monetization layer, behavior depth layer, and user feature layer respectively, based on the sample extraction quantity of each value layer, and merge them to generate a user list to be returned.

[0143] The return module 25 is used to return the user list to the media platform.

[0144] In some embodiments, the partitioning module 22 may include:

[0145] The core monetization layer unit 221 is used to apply the first judgment rule and, based on the monetization data in the user attribute data, classify users whose recovered amount is higher than the first threshold or whose recovered amount ranks in the top N percentiles into the core monetization layer, where N is a real number greater than 0.

[0146] The behavior depth layer unit 222 is used to apply a second judgment rule to users who have completed at least one preset key deep behavior into the behavior depth layer based on the behavior sequence data in the user attribute data.

[0147] User feature layer unit 223 is used to classify users who are not classified into the core monetization layer and the behavior depth layer into the user feature layer.

[0148] In some embodiments, when classifying users into the user feature layer, the user feature layer unit 223 performs sampling based on the static attribute data in the user attribute data, so that the sampled users cover different regions and / or different device models.

[0149] In some embodiments, the partitioning module 22 may further include:

[0150] The new user layer unit 224 is used to apply the third judgment rule after applying the second judgment rule, and according to the user attribute data, to classify users who are visiting for the first time and complete the target conversion or key deep behavior in the first session into the new user layer; and to process the new user layer as an independent value layer.

[0151] In some embodiments, the extraction module 24 may include:

[0152] The calculation unit 241 is used to calculate the sampling priority score of each user in the core monetization layer, behavior depth layer and user feature layer respectively, based on the number of samples extracted from each value layer and the weight configuration set for the key fields in the user attribute data.

[0153] The sorting unit 243 is used to sort users in the core monetization layer, behavior depth layer and user feature layer according to the sampling priority score.

[0154] Selection unit 243 is used to select, in order of sorting results, the number of users whose samples are at the top of the sorting hierarchy within each value stratum.

[0155] In some embodiments, in the core monetization layer, the key field used by the calculation unit 241 to calculate the sampling priority score is the amount recovered; in the behavior depth layer, the key field used to calculate the sampling priority score is the behavior completion depth.

[0156] In some embodiments, the configuration module 23 may include:

[0157] Configuration unit 231 is used to configure multiple test groups, each test group corresponding to a return quota configuration;

[0158] The backhaul unit 232 is used to backhaul data for different advertising plans using different test group backhaul quota configurations during the test period.

[0159] The data collection unit 233 is used to collect advertising performance indicators for each test group and select the optimal backhaul quota configuration based on the advertising performance indicators to update the global backhaul quota configuration. The advertising performance indicators include long-term return on investment and / or the diversity of new user characteristic distribution.

[0160] Each module in the aforementioned data backhaul device based on user value stratification can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0161] In some embodiments, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the data backhaul method based on user value stratification as described in any of the above embodiments.

[0162] In some embodiments, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the data backhaul method based on user value stratification as described in any of the above embodiments.

[0163] In some embodiments, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the data backhaul method based on user value stratification as described in any of the above embodiments.

[0164] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0165] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0166] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A data feedback method based on user value stratification, characterized in that, The method includes: Get the set of users who complete the target conversion event within a preset time window and the corresponding user attribute data; The preset stratification rules are applied sequentially to classify each user into a unique corresponding value stratification based on the user attribute data; the value stratification includes at least a core monetization stratification, a behavior depth stratification, and a user feature stratification. Based on the preset feedback quota configuration, determine the number of samples to be extracted for each value layer of the core monetization layer, behavior depth layer, and user feature layer; Based on the number of samples extracted from each value layer, a corresponding number of users are extracted from the core monetization layer, the behavior depth layer, and the user feature layer, and then merged to generate a user list to be returned. The returned user list will be sent back to the media platform.

2. The method according to claim 1, characterized in that, The preset stratification rules are applied sequentially to classify each user into a unique corresponding value stratum based on the user attribute data, including: Applying the first judgment rule, based on the monetization data in the user attribute data, users whose recovered amount is higher than the first threshold or whose recovered amount ranks in the top N percentiles are classified into the core monetization layer, where N is a real number greater than 0; For users not included in the core monetization layer, the second judgment rule is applied, and users who have completed at least one preset key deep behavior are included in the behavior depth layer based on the behavior sequence data in the user attribute data. Users who are not classified into the core monetization layer and the behavior depth layer are classified into the user feature layer.

3. The method according to claim 2, characterized in that, When classifying users into the user feature layer, sampling is performed based on the static attribute data in the user attribute data so that the sampled users cover different regions and / or different device models.

4. The method according to claim 2, characterized in that, The method further includes: After applying the second judgment rule, the third judgment rule is applied, and based on the user attribute data, users who are visiting for the first time and complete the target conversion or key deep behavior in the first session are classified into a new user layer. The new user layer will be treated as a separate value layer.

5. The method according to claim 1, characterized in that, Based on the sample extraction quantity for each value layer, the corresponding number of users extracted from the core monetization layer, behavior depth layer, and user feature layer include: Based on the number of samples extracted from each value layer and the weight configuration set for the key fields in the user attribute data, the sampling priority score of each user in their respective value layer is calculated from the core monetization layer, behavior depth layer and user feature layer. Based on the sampling priority scores, users in the core monetization layer, behavior depth layer, and user feature layer are sorted respectively. Within each value stratum, based on the ranking results, the number of users whose samples are ranked first are selected sequentially.

6. The method according to claim 5, characterized in that, In the core monetization layer, the key field used to calculate the sampling priority score is the amount recovered; In the behavior depth layer, the key field used to calculate the sampling priority score is the behavior completion depth.

7. The method according to claim 1, characterized in that, The backhaul quota configuration is dynamically optimized through an automated testing framework, and the specific steps include: Configure multiple test groups, each test group corresponding to a different return quota configuration; During the testing period, different advertising campaigns used different test group return quota configurations for data return; The advertising performance metrics of each test group are collected, and the optimal backhaul quota configuration is selected based on the advertising performance metrics to update the global backhaul quota configuration. The advertising performance metrics include long-term return on investment and / or the diversity of new user characteristic distribution.

8. A data backhaul device based on user value stratification, characterized in that, The device includes: The acquisition module is used to acquire the set of users who have completed the target conversion event within a preset time window and the corresponding user attribute data; The segmentation module is used to sequentially apply preset hierarchical judgment rules to segment each user into a unique corresponding value layer based on the user attribute data; the value layer includes at least a core monetization layer, a behavior depth layer, and a user feature layer; The configuration module is used to determine the number of samples extracted for each of the core monetization layer, the behavior depth layer, and the user feature layer according to the preset return quota configuration. The extraction module is used to extract a corresponding number of users from the core monetization layer, the behavior depth layer, and the user feature layer respectively, based on the number of samples extracted from each value layer, and merge them to generate a user list to be returned. The return module is used to return the user list to the media platform.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images