Real-time bidding method and system based on agent index and double-loop optimization, electronic device and storage medium

Abstract

The application discloses a real-time bidding method and system based on proxy indicators and double-loop optimization, electronic equipment and a storage medium. The method comprises: a preprocessing step of identifying short-term proxy indicators related to long-term value, training a prediction model, and precomputing user value and storing it in a cache; a real-time bidding decision step of obtaining user value from the cache to determine a bidding price when a bidding request is received; and a double-loop optimization feedback step. The double-loop optimization feedback step comprises: a short-period optimization loop that collects real feedback of proxy indicators in the short term to update the prediction model at a high frequency; and a long-period calibration loop that collects real long-term life cycle value data in the long term to calibrate the effectiveness of the proxy indicators at a low frequency. Through the scheme, the adaptability of the system to market changes is improved, and the consistency of the short-term optimization direction and the long-term value target is ensured.

Description

Technical Field

[0001] This application relates to the field of computer data processing technology, and in particular to a real-time bidding method, system, electronic device and storage medium based on proxy indicators and dual-loop optimization. Background Technology

[0002] With the development of the digital marketing industry, Real-Time Bidding (RTB) has become the mainstream method for programmatic advertising buying. To improve long-term ROI, the industry has begun to introduce Lifetime Value (LTV) as an optimization target. Currently, a common improvement approach is to preprocess the computationally intensive user value assessment task offline, classifying users into different value levels. During real-time bidding, only the user level needs to be quickly queried and a bid placed. This alleviates, to some extent, the contradiction between the prediction time of complex models and the millisecond-level response speed of real-time bidding.

[0003] However, the aforementioned existing technologies still have significant drawbacks: their model training and value ranking criteria essentially rely on real long-term user lifetime value data, which can only be obtained after waiting for months or even longer. This results in an extremely long optimization iteration cycle for the entire bidding model, making the system unable to quickly respond to dynamic changes in the market environment, channel quality, and user behavior patterns. When the market changes, the static model trained on old data will continue to make suboptimal decisions, leading to wasted advertising budgets and a decline in return on investment. This constitutes an irreconcilable contradiction between "long-term evaluation" and "timely optimization." Summary of the Invention

[0004] The technical problem this application aims to solve is to provide a real-time bidding solution that can balance decision-making real-time performance, optimization agility, and long-term goal accuracy, addressing the issue that existing technologies rely on long-term user lifetime value data, resulting in excessively long model optimization cycles and an inability to quickly respond to market changes.

[0005] To address the aforementioned technical problems, this application provides a real-time bidding method based on proxy metrics and dual-loop optimization, comprising the following steps: A preprocessing step, including the following three sub-steps: Obtaining proxy metrics: Identifying at least one short-term behavioral metric strongly correlated with the user's long-term lifetime value from historical user behavior data, defining the short-term behavioral metric as a proxy metric for the long-term lifetime value; Model training: Training a prediction model based on historical user data to predict the achievement of the proxy metric; Value assessment and storage: Using the trained prediction model to predict the value of the target user group, obtaining user value assessment results, and storing the mapping relationship between user identifiers and the user value assessment results in a cache; Real-time bidding decision step: When receiving an advertising bidding request containing a user identifier, retrieving the corresponding user value assessment result from the cache based on the user identifier, and determining the bidding price based on the user value assessment result; Dual-loop optimization feedback step, including the following two sub-steps: Short-cycle optimization loop sub-steps: After the advertisement is launched, collect real feedback data on whether users have achieved the agent indicator in a short period of time, and update the prediction model based on the real feedback data; and long-cycle calibration loop sub-steps: Collect real long-term lifetime value data of users in a long period of time, and verify and calibrate the effectiveness of the agent indicator based on the real long-term lifetime value data to ensure that the short-term optimization direction is consistent with the long-term value goal.

[0006] Optionally, the preprocessing step further includes at least one of the following: (1) In the sub-step of obtaining proxy indicators, the step of identifying at least one short-term behavior indicator that is strongly correlated with the long-term lifetime value of the user from historical user behavior data includes: calculating the correlation coefficient between each short-term behavior indicator and the long-term lifetime value of the user from historical user behavior data through regression analysis or correlation analysis, and selecting short-term behavior indicators with correlation coefficients exceeding a preset threshold as the proxy indicators; (2) In the sub-step of model training, the step of training a prediction model aimed at predicting the achievement of the proxy indicators includes: real-time monitoring of key indicators of model training, setting an abnormal threshold corresponding to the key indicators, and executing a multi-level rollback strategy when an abnormality is detected. The multi-level rollback strategy includes: the first level rollback is to enable data cleaning and outlier processing, reload the data and retrain; the second level rollback is to switch to backup computing resources and retrain using simplified model parameters; the third level rollback is to roll back to the model parameters of the previous version, or use the best historical model as the prediction model for the current period. (3) In the value assessment and storage sub-step, obtaining the user value assessment result includes: mapping each user to a pre-set value level according to the prediction result; (4) In the value assessment and storage sub-step, storing the mapping relationship between the user identifier and the user value assessment result in a cache includes: storing the mapping relationship between the user identifier and the user value assessment result in the cache in the form of key-value pairs, wherein the cache is a high-speed memory database. Optionally, in the real-time bidding decision-making step, the cache is a first in-memory database, and the step of obtaining the corresponding user value assessment result from the cache based on the user identifier includes: using the user identifier as a key to query the first in-memory database; if the query fails or the response times out, a multi-level degradation strategy is initiated, the multi-level degradation strategy includes: the first level degradation is to initiate a retry mechanism, setting a maximum preset number of retries and a preset duration for each retry interval; the second level degradation is to enable a local second-level cache and query the user value assessment results accessed within the preset duration; the third level degradation is to call a lightweight backup prediction model and use the basic features in the bidding request to quickly estimate the temporary user value assessment result. Optionally, in the real-time bidding decision-making step, determining the bidding price based on the user value assessment result includes: determining the final bidding price by applying a preset bidding strategy based on the user value assessment result, wherein the bidding strategy is one of the following two: the bidding price equals the base cost multiplied by a grade coefficient, wherein the grade coefficient is determined based on the user value assessment result; or, the bidding price equals the base cost multiplied by a grade coefficient, then multiplied by a time coefficient and a geographic coefficient, wherein the time coefficient is determined based on the time when the current bidding request occurred, and the geographic coefficient is determined based on the geographic region where the current bidding request occurred.Optionally, in the short-cycle optimization loop sub-step, updating the prediction model based on the real feedback data includes: using the real feedback data as new samples to fine-tune or retrain the prediction model, thereby achieving rapid adjustment of the value level classification criteria.

[0007] Optionally, in the long-cycle calibration loop sub-step, the verification and calibration of the effectiveness of the proxy indicator based on the real long-term lifecycle value data includes: using the real long-term lifecycle value data, calculating the correlation coefficient between each short-term behavior indicator in the candidate indicator library and the long-term value; selecting, based on the correlation evaluation results, one or a group of short-term behavior indicators with the strongest current correlation as the formal proxy indicator for the next optimization cycle; if the selected proxy indicator changes, adjusting the objective function of the prediction model in the offline preprocessing stage and retraining. Optionally, the long-cycle calibration loop sub-step further includes the following steps: (1) Data acquisition monitoring: real-time monitoring of the first key indicator of data acquisition, and setting an abnormal threshold corresponding to the first key indicator; (2) Data acquisition interruption fault tolerance processing: detecting the interruption type and judging the interruption impact range according to the interruption type, and executing a multi-level degradation strategy according to the interruption impact range, the multi-level degradation strategy includes: the first level degradation is to use only the data from the available data source and interpolate to fill the missing data; the second level degradation is to suspend model updates, use historical data to maintain model operation, record user behavior data during the interruption, and collect data again after recovery; the third level degradation is to enable the backup data source or use simulated data to maintain the operation of the optimization loop; (3) Optimization loop suspension and recovery: when the data acquisition interruption exceeds the preset duration, the short-cycle optimization loop is automatically suspended; during the suspension period, the current model continues to be used for value level pre-calculation and storage and real-time bidding decision; when data acquisition is restored and the data collection is completed, the optimization loop operation is automatically restored.

[0008] This application also provides a real-time bidding system based on proxy metrics and dual-loop optimization, for implementing any of the methods described above, comprising: an offline processing module configured to perform the preprocessing steps; a cache module configured to store the mapping relationship between user identifiers and user value assessment results; a real-time decision module configured to perform the real-time bidding decision steps; and a model optimization module configured to perform the dual-loop optimization feedback steps.

[0009] This application also provides an electronic device, including: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform any of the methods described above by executing the executable instructions.

[0010] This application also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program, when executed by a processor, implements any of the methods described above.

[0011] This application has the following beneficial effects:

[0012] 1. Achieved a balance between high precision and high speed: By moving the computationally intensive prediction task of complex models to the offline preprocessing stage, while the real-time bidding decision stage only needs to perform one high-speed cache query, the decision latency is controlled at the millisecond level, thus resolving the contradiction between the high precision of complex models and the high timeliness requirements of real-time bidding.

[0013] 2. Improved system adaptability: By introducing short-term behavioral indicators that are strongly correlated with long-term value as proxy indicators and constructing a short-cycle optimization loop, the model optimization cycle is shortened from several months to several days or weeks, enabling the system to quickly adapt to changes in the market environment and user behavior, and solving the problem of optimization lag caused by reliance on long-term data.

[0014] 3. Ensures the accuracy of long-term goals: Through a long-term calibration loop, the effectiveness of proxy indicators is verified and calibrated regularly using real long-term lifecycle value data. This ensures that the optimization direction of short-term rapid iteration is consistent with the ultimate long-term business goals, avoids deviation from the optimization direction, and guarantees the long-term effectiveness of the optimization strategy. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the structure of a real-time bidding system according to an embodiment of this application.

[0017] Figure 2 This is a flowchart illustrating the real-time bidding method according to an embodiment of this application.

[0018] Figure 3 This is a schematic diagram of the real-time bidding decision interaction according to an embodiment of this application.

[0019] Figure 4 This is a schematic diagram of the dual-loop optimization principle according to an embodiment of this application.

[0020] Key reference numerals:

[0021] 10. Offline processing module; 20. High-speed caching module; 30. Real-time decision-making module; 40. Model optimization module; 50. External data source; 60. Advertising exchange platform; 100. Real-time bidding system. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application. Unless otherwise defined, all technical and scientific terms used in the embodiments of this application have the same meaning as commonly understood by those skilled in the art. The terminology used in the embodiments of this application is for the purpose of describing the embodiments of this application only and is not intended to limit this application. Before further detailed description of the embodiments of this application, the nouns and terms involved in the embodiments of this application are explained as follows.

[0023] (1) Proxy metrics: These are one or more short-term behavioral metrics that are strongly correlated with the long-term lifetime value of users and can be observed within a short period of time (e.g., within a few days or weeks). Proxy metrics serve as leading indicators of long-term value and are used to construct short-cycle optimization loops to achieve rapid model iteration.

[0024] (2) Long-term lifetime value (LTV): refers to the total value that a user contributes to an advertiser from the moment they first interact with the advertiser until a longer period of time (such as a quarter or several months). It is usually measured by quantitative indicators such as cumulative payment amount and number of key actions.

[0025] (3) Dual-loop optimization: refers to a model optimization system that combines short-cycle rapid iteration with long-cycle target calibration. The short-cycle optimization loop updates the model frequently based on feedback data of surrogate indicators; while the long-cycle calibration loop verifies and calibrates the effectiveness of surrogate indicators infrequently based on real long-term lifecycle value data.

[0026] (4) Value level: refers to a pre-defined, discrete value classification to which a user is mapped based on the prediction results output by the prediction model. For example, users can be divided into S-level (highest value), A-level, B-level, C-level (lowest value), etc., for differentiated bidding in real-time bidding.

[0027] (5) Cache: refers to a high-speed memory database used to store the mapping relationship between user identifiers and their corresponding value levels, such as Redis or Memcached. Its purpose is to achieve millisecond-level fast queries in real-time bidding decisions to meet real-time requirements.

[0028] Please see Figure 2 This application provides a real-time bidding method based on proxy metrics and dual-loop optimization, aiming to solve the fundamental problems in existing technologies where the reliance on long-term user lifetime value data for model training leads to excessively long optimization feedback cycles and the system's inability to adapt quickly to market changes. This method achieves a balance between high-precision prediction and millisecond-level response through a three-stage process: a preprocessing step S100, a real-time bidding decision step S200, and a dual-loop optimization feedback step S300.

[0029] The preprocessing step S100 moves the computationally intensive value assessment task forward, preparing for real-time bidding. This preprocessing step comprises three sub-steps: obtaining proxy metrics, model training, and value assessment and storage.

[0030] The sub-step of obtaining proxy metrics involves identifying at least one short-term behavioral metric that is strongly correlated with the user's long-term lifetime value from massive amounts of historical user behavior data, and defining this short-term behavioral metric as a proxy metric for long-term lifetime value. Acquiring long-term lifetime value data (such as total payment amount over 90 days) takes an extremely long time, and directly using it for model optimization can lead to sluggish system response; while short-term behavioral metrics (such as activity level within 7 days) can be obtained quickly. By finding short-term behavioral metrics that are highly correlated with long-term lifetime value as "surrogates," i.e., proxy metrics, a foundation is laid for subsequent rapid optimization, enabling system agility and solving the problem of optimization lag caused by existing technologies relying entirely on long-term lifetime value data.

[0031] In the model training sub-step, the system trains a predictive model based on historical user data, aiming to predict the achievement of the aforementioned proxy metrics. This shifts the optimization objective from the difficult-to-predict long-term lifetime value to specific, quantifiable short-term proxy metrics, enabling the system to indirectly assess the long-term value of users and thus achieve high-precision predictions of user potential. For example, instead of directly predicting a user's spending amount over the next 90 days, the model predicts whether they will "make their first donation within 7 days." This transformation makes the model training objective clearer and feedback data easier to obtain.

[0032] In the value assessment and storage sub-step, the system uses a trained prediction model to perform batch value predictions on the target user group, obtains user value assessment results, and stores the mapping relationship between user identifiers and these results in a high-speed cache. This value assessment and storage sub-step completes the time-consuming model inference process in a non-real-time environment, and stores the final, simplified value assessment results (such as a value level) in a high-speed, low-latency cache, achieving "offline computation, online query." This avoids performing complex model calculations for each bidding request, resolving the contradiction between high-precision, complex models and the millisecond-level response requirements of real-time bidding, while simultaneously achieving high precision and high speed.

[0033] After the preprocessing step, this method proceeds to the real-time bidding decision-making step. When an ad bidding request containing a user identifier is received, the corresponding user value assessment result is retrieved from the cache based on the user identifier, and the bidding price is determined based on the user value assessment result. In this step, the results of the preprocessing step are utilized to simplify the real-time decision-making process into a fast key-value lookup and a simple rule mapping. Since the cache query latency is typically in the millisecond range, the entire real-time bidding decision-making process is extremely short. Through this step, this method ensures that while meeting the high response time requirements of programmatic advertising exchanges (DSPs) (typically 10-100 milliseconds), it can still make accurate bids based on deep insights from complex models.

[0034] To enable the system to continuously learn and evolve, this method also includes a dual-loop optimization feedback step. This dual-loop optimization feedback step comprises a short-cycle optimization loop sub-step and a long-cycle calibration loop sub-step.

[0035] The short-cycle optimization loop sub-step collects real feedback data on whether users have achieved the proxy metrics within a short period (e.g., several days or a week) after the advertisement is launched, and updates the predictive model based on this new feedback data. This sub-step leverages the short-term observability of the proxy metrics to establish a rapid feedback loop. This allows the predictive model to be fine-tuned or retrained frequently, thereby quickly adapting to dynamic changes in the market environment, channel quality, and user preferences. Through this short-cycle loop, the system solves the problem of excessively long optimization cycles in existing technologies, significantly improving adaptability and agility to market changes.

[0036] The long-term calibration loop sub-step involves collecting real long-term lifetime value data of users over a long period (such as a quarter or several months). Based on this real long-term lifetime value data, the effectiveness of the proxy indicators is verified and calibrated to ensure that short-term optimization directions remain consistent with long-term value goals. The correlation between proxy indicators and long-term value is not static; it may weaken or change with market changes. This long-term calibration loop sub-step avoids the risk of deviation in optimization direction due to over-reliance on proxy indicators, ensuring the long-term robustness and profitability of the entire system.

[0037] Furthermore, in a preferred embodiment, to make the preprocessing step more precise, the preprocessing step further includes at least one of the following four types. Here, "at least one" means that it may include only one type, any two or three types, or even all four types.

[0038] (1) In the sub-step of acquiring proxy indicators, the process of identifying strongly correlated short-term behavioral indicators (identifying at least one short-term behavioral indicator that is strongly correlated with the user's long-term lifetime value from historical user behavior data) can be specified as follows: through regression analysis or correlation analysis (such as calculating the Pearson correlation coefficient), the correlation coefficient between each short-term behavioral indicator and the user's long-term lifetime value is calculated from historical user behavior data, and those short-term behavioral indicators with correlation coefficients exceeding a preset threshold (e.g., 0.7) are selected as proxy indicators. The short-term behavioral indicators may include at least one of the following: the cumulative viewing time within a preset number of days (e.g., 7 days) after user registration, whether the first reward is completed within a preset number of days (e.g., 3 days), whether the first payment is completed within a preset number of days (e.g., 3 days), or the amount paid within a preset number of days (e.g., 7 days).

[0039] (2) In the model training sub-step, the training of a predictive model aimed at predicting the achievement of the proxy metric includes: real-time monitoring of key metrics of model training, wherein the key metrics include at least one of loss function value, gradient norm, validation set accuracy, and training progress. Simultaneously, anomaly thresholds corresponding to the key metrics are set, such as at least one of loss function value exceeding a preset multiple of the historical average, gradient norm exceeding a preset value, validation set accuracy being lower than a preset percentage, or training progress stalling for more than a preset duration. For example, loss function value exceeding 3 times the historical average, gradient norm exceeding 1000, validation set accuracy being lower than 50%, or training progress stalling for more than 10 minutes. When any anomaly is detected, the system automatically executes a multi-level rollback strategy: the first level rollback is for data issues, enabling data cleaning and outlier handling, reloading data, and retraining; the second level rollback is for resource issues, switching to backup computing resources and retraining with simplified model parameters; the third level rollback is for issues with the algorithm itself, directly reverting to the stable model parameters of the previous version, or using the best historical model as the prediction model for the current period. This mechanism greatly enhances the robustness of the offline training process, ensuring that usable predictive models can be generated stably even if problems occur during training.

[0040] (3) In the value assessment and storage sub-step, obtaining the user value assessment result includes: mapping each user to a pre-set value level according to the prediction result. The value level is preferably discrete, such as including S level (highest value), A level (value less than S level), B level (value less than A level), and C level (lowest value).

[0041] (4) In the value assessment and storage sub-step, the mapping relationship between user identifiers and these value levels is stored in the cache in the form of key-value pairs. Specifically, the cache is a high-speed in-memory database, such as Redis or Memcached.

[0042] Furthermore, in the real-time bidding decision-making step, to address potential cache failures, the cache is a first in-memory database, and the step of retrieving the corresponding user value assessment result from the cache based on the user identifier further includes:

[0043] When using the user identifier as the key, a query is performed in the first in-memory database;

[0044] If a query fails or a response times out (e.g., exceeding 5 milliseconds), a multi-level degradation strategy is initiated. This strategy includes: Level 1 degradation involves activating a retry mechanism, setting a maximum preset number of retries and a preset interval between each retry; for example, retrying after 1 millisecond, with a maximum of 2 retries, to handle temporary failures such as network jitter. If the retry fails, Level 2 degradation is initiated; Level 2 degradation involves enabling a local secondary cache and querying user value assessment results accessed within a preset time period. These user value assessment results are stored in the secondary cache in the service's local memory, which may store user value assessment results accessed within the last 24 hours. If the local cache also fails to find the results, Level 3 degradation is initiated, calling a lightweight backup prediction model (such as a logistic regression model) to quickly estimate temporary user value assessment results using basic features (such as channel source and device type) attached to the bidding request. Through this degradation strategy, the system can still complete bidding even in extreme situations where the cache service is unavailable, ensuring business continuity and system robustness.

[0045] In another preferred embodiment, in the real-time bidding decision-making step, the process of determining the bidding price based on the user value assessment result can apply a preset bidding strategy to determine the final bidding price. The bidding strategy can be one of the following two: One bidding strategy is that the bidding price equals the base cost multiplied by a level coefficient determined by the user value assessment result. For example, an S-level user corresponds to a coefficient of 2.0, and an A-level user corresponds to 1.8, thereby achieving focused investment on high-value users. Another more complex bidding strategy is that the bidding price equals the base cost multiplied by the level coefficient, and then multiplied by a time coefficient and a geographic coefficient. The time coefficient is determined based on the time when the current bidding request occurs (e.g., a coefficient of 1.2 during peak hours and 0.8 during off-peak hours), and the geographic coefficient is determined based on the region where the current bidding request occurs (e.g., a coefficient of 1.1 for first-tier cities and 0.9 for third-tier cities). This multi-dimensional bidding strategy enables more refined and contextualized budget allocation, further improving the return on investment for advertising.

[0046] Furthermore, in the short-cycle optimization loop sub-step, updating the prediction model based on real feedback data further includes: using the collected real feedback data (whether the user has achieved the proxy indicator) as new samples to perform online learning or incremental fine-tuning of the prediction model, or for periodic retraining. This high-frequency update (e.g., performed weekly or every few days) enables the model to quickly absorb new knowledge and achieve rapid adjustment of the value level classification criteria, thereby keeping up with the market pace. For example, the short-term period is greater than or equal to 1 day and less than or equal to 7 days, and the high-frequency update includes one execution for each of the short-term periods.

[0047] Similarly, in the long-term calibration loop sub-step, the verification and calibration of the effectiveness of the proxy indicators based on the real long-term lifecycle value data further includes: using the real long-term lifecycle value data collected over a long period, calculating the correlation coefficient between each short-term behavioral indicator in the candidate indicator library and the long-term value; and selecting, based on the correlation evaluation results, the short-term behavioral indicator or group with the strongest current correlation as the official proxy indicator for the next optimization cycle (e.g., the next quarter). If the selected proxy indicator changes, the objective function of the prediction model in the offline preprocessing stage is adjusted and retrained. For example, the long term includes greater than or equal to 1 month and less than or equal to 12 months, and the low-frequency verification and calibration is performed once for each long term. This low-frequency (e.g., once per quarter or every few months) strategic calibration ensures that the foundation of the entire optimization system is solid and accurate.

[0048] In a particularly preferred embodiment, in order to address the potential interruption of data acquisition during the dual-loop optimization feedback process, the long-cycle calibration loop sub-step further includes the following steps.

[0049] (1) Data acquisition monitoring: Real-time monitoring of the first key indicator of data acquisition; the first key indicator may be, for example, data acquisition success rate, data acquisition delay, data acquisition volume and data quality indicator, and set an abnormal threshold corresponding to the first key indicator. The abnormal threshold includes acquisition success rate lower than a preset percentage, acquisition delay exceeding the expected time by a preset multiple, acquisition volume lower than the expected amount by a preset percentage, and data quality indicator lower than a preset percentage.

[0050] Once an interruption is detected, proceed to (2) Data Acquisition Interruption Fault Tolerance Processing. First, detect the interruption type (e.g., data source failure, network failure, system failure) and determine the scope of the interruption's impact based on the interruption type (e.g., partial or complete data source interruption). Execute a multi-level degradation strategy based on the scope of the interruption's impact: Level 1 degradation involves using only available data sources and interpolating to fill in missing data; Level 2 degradation involves pausing model updates, using historical data to maintain model operation, recording user behavior data during the interruption period, and re-acquiring data after recovery; Level 3 degradation involves enabling a backup data source or using simulated data to maintain the operation of the optimization loop. In a specific implementation, for partial interruptions, missing data can be interpolated to fill in; for short-term complete interruptions, model updates can be paused and data recorded for subsequent re-acquisition; for long-term complete interruptions, a backup data source can be enabled or simulated data can be used to maintain the operation of the optimization loop.

[0051] (3) Optimization Loop Pause and Resumption: When data acquisition is interrupted for more than a preset duration (e.g., 24 hours), the short-cycle optimization loop is automatically paused to avoid making erroneous updates based on incomplete data. During the pause, the system continues to use the current model for value level pre-calculation and storage and real-time bidding decisions. When data acquisition resumes and supplementary acquisition is completed, the optimization loop operation is automatically resumed.

[0052] The embodiments of this application will be described in more detail below with reference to the accompanying drawings and specific scenarios. It should be understood that these embodiments are only used to more clearly illustrate the technical solutions of this application, and do not constitute any limitation thereon.

[0053] In a specific application scenario, the technical solution provided in this application can be implemented as a complete real-time bidding system. The technical principle of this system lies in achieving efficient bidding by pre-calculating user value levels offline and storing them in a high-speed cache, combined with real-time fast querying and mapping.

[0054] Please see Figure 1 The system 100 structurally includes an offline processing module 10, a high-speed caching module 20, a real-time decision-making module 30, and a model optimization module 40. These modules work together to interact with external data sources 50 and an advertising exchange platform 60.

[0055] Offline processing module 10 is responsible for performing preprocessing steps. In a specific execution cycle, the data analysis team first performs regression analysis on historical data, identifying two metrics from external data source 50: "cumulative viewing time within 7 days of user registration" and "whether the first donation was completed within 3 days." These metrics are highly positively correlated with the "90-day user lifetime value," and are therefore defined as proxy metrics. Subsequently, offline processing module 10 starts a model training task at 2 AM daily, using a deep neural network model with 5 hidden layers. This model takes multi-dimensional user features as input and outputs predicted values ​​for the two proxy metrics. This process corresponds to... Figure 2 Steps S110 and S120 in the process.

[0056] After successful model training, the offline processing module 10 uses the model to perform batch predictions on the target user group and calculates a comprehensive value score based on the prediction results. This score then categorizes users into four discrete value levels: S-level (top 5%), A-level (5%-20%), B-level (20%-60%), and C-level (bottom 40%). Finally, the offline processing module 10 writes the user's device identifier and its corresponding value level (e.g., 'user_id_123' -> 'A') as key-value pairs to the Redis cluster managed by the caching module 20. This process corresponds to... Figure 2 Step S130 in the process.

[0057] To ensure the stability of the offline processing workflow, the offline processing module 10 integrates a model training failure fault tolerance mechanism. This module maintains a model version repository, for example, storing the 10 most recent versions of deep neural network models. Suppose that during a training task on a certain day, training fails due to an anomaly in data collection from external data source 50. The system will automatically detect an anomaly where the loss function value exceeds five times the historical average and classify it as data anomaly. It will first attempt to initiate a data cleaning process to remove outliers, then reload the data and retrain. If retraining still fails, the system will automatically revert to the previous successful model version (i.e., the model from the previous day) from the version repository and use this reverted model to complete the value level pre-calculation task for the day, thus ensuring that the data in the cache module 20 is updated on time. Simultaneously, the system will trigger an alarm to notify the technical team and provide a detailed failure analysis report.

[0058] Once offline preprocessing is complete, the system enters real-time operation mode. Please also refer to... Figure 1 and Figure 3 When the advertising exchange platform 60 sends a bidding request to the system, the request is first received by the real-time decision module 30. The real-time decision module 30 parses the request and extracts the user's device identifier. Then, as... Figure 3 As shown, the real-time decision module 30 uses the device identifier to initiate a query request to the cache module 20 to obtain the user's value level. Since the cache module 20 uses a Redis cluster, this query typically takes about 1 millisecond and successfully returns the pre-calculated value level, such as 'A'. This process corresponds to... Figure 2 Step S210 in the process.

[0059] After obtaining the value level, the real-time decision-making module 30 applies a preset bidding strategy to determine the bidding price. For example, a simple strategy could be: Bid = Average Cost Per Thousand Impressions (CPM) * 1.8. Here, "bid" is the final determined bidding price, and "Average Cost Per Thousand Impressions" is the average cost (i.e., base cost) for a specific channel based on historical data. Assuming the average CPM is 10 yuan and the coefficient for level 'A' is 1.8, the calculated final bidding price is 18 yuan. The real-time decision-making module 30 then constructs a bidding response and returns it to the advertising exchange platform 60. The entire decision-making process, from receiving the request to sending the response, is strictly controlled within 10 milliseconds, fully meeting the requirements of real-time bidding. This process corresponds to... Figure 2 Step S230 in the process.

[0060] Similarly, to cope with unexpected situations in the real-time environment, the real-time decision module 30 also integrates a cache fault tolerance mechanism. If a query to the cache module 20 times out due to network jitter, the system will automatically retry within 1 millisecond. If the retry still fails (e.g., a Redis cluster crash), the system will immediately query a secondary cache maintained in the local memory of the real-time decision module 30 (e.g., an LRU cache storing the most recently accessed user level data). If the local cache also fails, the system will initiate the final degradation scheme: calling a lightweight logistic regression model preloaded in memory. This model quickly estimates that the user is temporarily at level A based only on basic characteristics such as the 'premium channel' source attached to the bidding request, and completes the bidding accordingly. The additional time for the entire degradation process does not exceed 3 milliseconds, ensuring that even in the extreme case of a main cache failure, the total response time can still be controlled within 15 milliseconds, greatly improving the robustness of the system.

[0061] Completing ad campaigns is not the end, but rather the beginning of an optimization loop. Please refer to [link / reference]. Figure 1 and Figure 4 The model optimization module 40 is responsible for executing the dual-loop optimization feedback step. In the short-cycle optimization loop (such as...) Figure 4 In the inner ring (as shown), the model optimization module 40 collects real data on the "7-day viewing time" and "3-day tipping" of new users from external data source 50 weekly. This data is used as new training samples to fine-tune the prediction model in the offline processing module 10. This high-frequency update enables the system to quickly correct its value assessment of users from different channels and adapt rapidly to market changes. This process corresponds to... Figure 2 Step S310 in the process.

[0062] In long-cycle calibration loops (such as...) Figure 4 In the outer ring (as shown), the model optimization module 40 obtains the user's actual "90-day user lifetime value" data from external data source 50 every quarter. This data is used to reassess and adjust the weights and effectiveness of the proxy metrics themselves. For example, it calculates whether the correlation between "7-day viewing time" and long-term value remains stable. This ensures that short-term optimization goals do not deviate from long-term business value goals. This process corresponds to... Figure 2 Step S320 in the process.

[0063] During the dual-loop optimization process, the model optimization module 40 also implemented a data acquisition interruption fault tolerance mechanism. For example, during data acquisition in a certain week, a third-party API failure caused a 12-hour interruption in the acquisition of "rewards within 3 days" data. The system triggered an alarm after detecting that the data acquisition success rate had dropped to 60%, and determined that part of the data source was interrupted. At this time, the system activated the first-level degradation strategy. For the missing "rewards within 3 days" data, it interpolated and filled the gaps using the average of the user's historical reward behavior, and continued to perform short-cycle optimization using the filled data. After 12 hours, the interface recovered, and the system automatically triggered a data replenishment task to fill in the missing data and re-evaluate the effectiveness of the model update, ensuring that model performance was not significantly affected. This ensured the continuity of the optimization process.

[0064] In another preferred embodiment, to further optimize computing resources, a scenario-adaptive model selection mechanism can be introduced in the offline processing stage. In this case, the offline processing module 10 maintains a "model pool," rather than just a single deep neural network model. This model pool may include: a high-precision deep neural network model for processing existing users with rich historical behavior and sufficient information; a medium-precision gradient boosting decision tree (GBDT) model for processing new users with less information but clear source channels; and a lightweight rule engine for processing users from low-value channels or inactive periods. When performing batch predictions, the offline processing module 10 first performs user segmentation based on user attributes (such as registration duration and source channel tags), routing different user groups to the most suitable model in the model pool for prediction. Then, all prediction results are uniformly mapped to value levels and stored in the cache module 20. This approach optimizes the allocation of computing resources, using higher-precision models for the most valuable prediction tasks, ensuring the prediction accuracy for high-value users while reducing the overall pre-computation cost and time.

[0065] Furthermore, during the real-time bidding decision-making stage, the bidding strategy can be further refined. The real-time decision-making module 30 can use a multi-dimensional set of rules to determine the bid. For example, the bidding formula can be designed as: Bid = Base Cost * Grade Coefficient * Time Coefficient * Regional Coefficient. Here, "bid" is the final determined bidding price, "base cost" is the preset benchmark bid, the grade coefficient is determined by the queried value grade (2.0 for S-level, 1.5 for A-level, etc.), and the time and regional coefficients are dynamically adjusted according to the time (e.g., 1.2 during peak hours) and region (e.g., 1.1 for first-tier cities). For example, if an A-level user triggers a bidding request in a first-tier city during prime time in the evening, their final bid will be "base cost * 1.5 * 1.2 * 1.1", achieving more refined budget allocation. Meanwhile, when a brand new user (whose query in the cache module 20 fails) triggers a request, the backup mechanism is activated. The real-time decision module 30 calls the lightweight logistic regression model, estimates the user's "high-quality channel" source as a temporary A-level, and completes the bidding, ensuring coverage of new users and the capture rate of bidding opportunities.

[0066] Furthermore, the dual-loop optimization system can be further refined, particularly in the long-term calibration loop, where a dynamic evaluation and selection mechanism for the proxy metrics themselves can be introduced. The model optimization module 40 can maintain a "candidate proxy metric library," containing multiple short-term behavioral metrics potentially related to long-term value, such as "next-day retention rate," "viewing time within 7 days," "first payment completed within 3 days," and "payment amount within 7 days." During each quarterly long-term calibration, the model optimization module 40 uses the latest collected real long-term lifecycle value data to recalculate the correlation coefficient between each metric in the candidate library and long-term value. If the analysis reveals that, with increased market competition, the weight of "first payment completed within 3 days" decreases, while the correlation between "payment amount within 7 days" and long-term value significantly increases, the system will automatically select "payment amount within 7 days" as the new core proxy metric and notify the offline processing module 10 to adjust its prediction model's objective function and retrain it. This ensures the effectiveness of the proxy metrics themselves, enabling the entire prediction and optimization system to adapt to deep-seated, long-term market and user behavior changes.

[0067] This application also provides a real-time bidding system 100 based on proxy metrics and dual-loop optimization, the structure of which is as follows: Figure 1As shown. The system 100 includes: an offline processing module 10, configured to perform the aforementioned preprocessing steps, including acquiring agent metrics, training a prediction model, and performing value assessment and storage; a cache module 20, configured to store the mapping relationship between user identifiers generated by the offline processing module 10 and user value assessment results; a real-time decision module 30, configured to perform the aforementioned real-time bidding decision steps, including acquiring value assessment results from the cache module 20 and determining the bidding price; and a model optimization module 40, configured to perform the aforementioned dual-loop optimization feedback steps, including a short-cycle optimization loop and a long-cycle calibration loop. The specific functions and interaction relationships of each module have been described in detail in the aforementioned method embodiments and will not be repeated here.

[0068] This application also provides an electronic device, which can be a server, a personal computer, a mobile terminal, etc. The electronic device includes a processor and a memory, the memory storing executable instructions of the processor. The processor is configured to implement the method as described in any one of claims 1 to 7 by executing the executable instructions stored in the memory.

[0069] This application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program can implement the method as described in any one of claims 1 to 7. The computer-readable storage medium can be volatile or non-volatile, such as read-only memory (ROM), random access memory (RAM), solid-state drive (SSD), optical disk, or magnetic disk.

[0070] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A real-time bidding method based on proxy metrics and dual-loop optimization, characterized in that, Includes the following steps: The preprocessing step includes the following three sub-steps: The sub-step for obtaining proxy metrics is as follows: Identify at least one short-term behavior metric that is strongly correlated with the user's long-term lifetime value from historical user behavior data, and define the short-term behavior metric as a proxy metric for the long-term lifetime value. Model training sub-step: Based on historical user data, train a prediction model aimed at predicting the achievement of the agent metric; Value assessment and storage sub-steps: Use the trained prediction model to predict the value of the target user group, obtain the user value assessment results, and store the mapping relationship between the user identifier and the user value assessment results in the cache; Real-time bidding decision steps: When an ad bidding request containing a user identifier is received, the corresponding user value assessment result is obtained from the cache according to the user identifier, and the bidding price is determined according to the user value assessment result; The dual-loop optimization feedback process includes the following two sub-steps: Short-cycle optimization loop sub-steps: After ad placement, collect real feedback data on whether users have achieved the aforementioned agency metrics within a short period; update the prediction model based on the real feedback data; and... Long-term calibration loop sub-steps: Collect real long-term lifetime value data of users over a long period of time, and verify and calibrate the effectiveness of the proxy indicators based on the real long-term lifetime value data to ensure that the short-term optimization direction is consistent with the long-term value goal.

2. The method as described in claim 1, characterized in that, The preprocessing step further includes at least one of the following: (1) In the sub-step of obtaining proxy indicators, the sub-step of identifying at least one short-term behavior indicator that is strongly correlated with the long-term lifecycle value of the user from historical user behavior data includes: calculating the correlation coefficient between each short-term behavior indicator and the long-term lifecycle value of the user from historical user behavior data through regression analysis or correlation analysis, and selecting the short-term behavior indicator whose correlation coefficient exceeds a preset threshold as the proxy indicator. (2) In the model training sub-step, the training of a predictive model aimed at predicting the achievement of the proxy indicator includes: real-time monitoring of key indicators of model training, setting an anomaly threshold corresponding to the key indicator, and executing a multi-level rollback strategy when an anomaly is detected. The multi-level rollback strategy includes: the first level rollback is to enable data cleaning and outlier handling, reload the data and retrain; the second level rollback is to switch to standby computing resources and retrain using simplified model parameters; the third level rollback is to roll back to the model parameters of the previous version, or use the best historical model as the predictive model for the current period. (3) In the value assessment and storage sub-step, obtaining the user value assessment result includes: mapping each user to a pre-set value level based on the prediction result; (4) In the value assessment and storage sub-step, storing the mapping relationship between the user identifier and the user value assessment result in the cache includes: storing the mapping relationship between the user identifier and the user value assessment result in the cache in the form of key-value pairs, wherein the cache is a high-speed memory database.

3. The method as described in claim 1, characterized in that, In the real-time bidding decision-making step, the cache is a first in-memory database, and the step of retrieving the corresponding user value assessment result from the cache based on the user identifier includes: Use the user identifier as the key to perform a query in the first memory database; If the query fails or the response times out, a multi-level degradation strategy is initiated. The multi-level degradation strategy includes: the first level degradation is to initiate a retry mechanism, setting a maximum preset number of retries and a preset duration for each retry interval; the second level degradation is to enable a local second-level cache and query the user value assessment results accessed within the preset duration; the third level degradation is to call a lightweight backup prediction model and use the basic features in the bidding request to quickly estimate the temporary user value assessment results.

4. The method as described in claim 1, characterized in that, In the real-time bidding decision-making step, determining the bidding price based on the user value assessment result includes: determining the final bidding price by applying a preset bidding strategy based on the user value assessment result, wherein the bidding strategy is one of the following two: The bidding price equals the base cost multiplied by a tier coefficient, which is determined based on the user value assessment result; or... The bidding price equals the base cost multiplied by the grade coefficient, then multiplied by the time coefficient and the region coefficient. The time coefficient is determined based on the time when the current bidding request occurs, and the region coefficient is determined based on the region where the current bidding request occurs.

5. The method as described in claim 1, characterized in that, In the short-cycle optimization loop sub-step, updating the prediction model based on the real feedback data includes: The real feedback data is used as new samples to fine-tune or retrain the prediction model, thereby enabling rapid adjustment of the value level classification criteria.

6. The method as described in claim 1, characterized in that, In the long-cycle calibration loop sub-step, the verification and calibration of the effectiveness of the proxy indicator based on the real long-term lifecycle value data includes: Using the real long-term lifecycle value data, calculate the correlation coefficient between each short-term behavior indicator and long-term value in the candidate indicator library. Based on the correlation evaluation results, select the short-term behavior indicator or group with the strongest current correlation as the formal proxy indicator for the next optimization cycle. If the selected proxy metric changes, adjust the objective function of the prediction model in the offline preprocessing stage and retrain it.

7. The method according to any one of claims 1 to 6, characterized in that, The long-cycle calibration loop sub-step also includes the following steps: (1) Data acquisition monitoring: Real-time monitoring of the first key indicator of data acquisition, and setting the abnormal threshold corresponding to the first key indicator; (2) Data acquisition interruption fault tolerance handling: detect the interruption type and determine the interruption impact range according to the interruption type, and execute a multi-level degradation strategy according to the interruption impact range. The multi-level degradation strategy includes: the first level degradation is to use only the data from the available data source and interpolate to fill the missing data; the second level degradation is to suspend model updates, use historical data to maintain model operation, record user behavior data during the interruption, and re-collect data after recovery; the third level degradation is to enable the backup data source or use simulated data to maintain the operation of the optimization loop. (3) Optimize loop pause and resume: When data acquisition is interrupted for more than the preset duration, the short-cycle optimization loop is automatically paused. During the suspension, the current model will continue to be used for value level pre-calculation and storage and real-time bidding decisions; Once data acquisition is restored and supplementary acquisition is completed, the optimized loop operation will automatically resume.

8. A real-time bidding system based on proxy metrics and dual-loop optimization, characterized in that, For implementing the method according to claims 1 to 7, comprising: The offline processing module is configured to perform the preprocessing steps; The cache module is configured to store the mapping relationship between user identifiers and user value assessment results; The real-time decision-making module is configured to execute the real-time bidding decision-making steps. The model optimization module is configured to execute the dual-loop optimization feedback step.

9. An electronic device, characterized in that, include: processor; Memory for storing the executable instructions of the processor; The processor is configured to perform the method of any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor as described in claim 9, it implements the method of any one of claims 1 to 7.

Images