Advertisement putting effect real-time optimization method and system based on machine learning

Abstract

The invention discloses an advertisement putting effect real-time optimization method and system based on machine learning, and particularly relates to the technical field of internet advertisements. Comprising a multi-source real-time data acquisition module, a dynamic feature processing module, a multi-dimensional effect evaluation module, an online machine learning model module, a real-time delivery decision and execution module and a visualization and operation management module. According to the invention, through cooperative work of the multi-task online learning model and the reinforcement learning decision model, advertisement putting parameters are optimized in real time; a multi-dimensional evaluation system containing short-term and long-term indexes is adopted, the index association relation is analyzed, and the short-term effect and the long-term value are balanced; the feature calculation frequency is dynamically adjusted by using the self-adaptive time window feature, and the feature accuracy is improved; and continuous optimization of the putting strategy is realized through closed-loop feedback and a strategy iteration mechanism. According to the invention, real-time optimization of the advertisement putting effect is realized, and the rate of return on investment of advertisement putting is obviously improved.

Description

Technical Field

[0001] This invention relates to the field of internet advertising technology, and specifically discloses a method and system for real-time optimization of advertising performance based on machine learning. Background Technology

[0002] In the modern digital marketing field, optimizing ad delivery effectiveness has become a core element for businesses to improve marketing efficiency. With the development of big data and artificial intelligence technologies, machine learning-based ad delivery systems are gradually becoming the industry standard solution.

[0003] Traditional ad optimization methods typically rely on manual experience and offline data analysis. They use pre-set, fixed rules for audience targeting and bid adjustments, or employ single machine learning models to predict click-through rates to guide ad placement decisions. However, traditional ad optimization solutions have significant limitations: on the one hand, rule systems depend on expert experience and struggle to adapt to rapidly changing market environments and user behavior, leading to decision-making delays; on the other hand, single-objective prediction models often prioritize short-term conversion rates, neglecting long-term user value and potentially resulting in inefficient budget allocation.

[0004] Furthermore, traditional feature engineering uses fixed time windows, which cannot adapt to changes in the behavioral patterns of users with varying activity levels, resulting in insufficient feature timeliness. More importantly, existing solutions generally lack real-time performance evaluation and strategy closed-loop mechanisms, making it impossible to quickly verify strategy effectiveness and continuously iterate and optimize. Therefore, developing an advertising delivery system that can optimize campaign performance in real time, balance short-term goals and long-term value, and possess adaptive feature engineering and closed-loop learning capabilities is of great significance for improving the overall effectiveness of digital advertising. Summary of the Invention

[0005] A machine learning-based method for real-time optimization of ad performance includes the following steps:

[0006] S1. Obtain advertising platform data through the real-time data reporting interface of the advertising platform, obtain user behavior data through the front-end monitoring SDK embedded in the advertising media, and obtain business target data through the callback interface that connects with the advertiser's business server; perform normalization processing on the collected advertising platform data, user behavior data and business target data to obtain normalized multi-source real-time data.

[0007] S2. Clean the normalized multi-source real-time data, construct dynamic features, and store the constructed dynamic features in real time to the distributed cache system;

[0008] S3. Based on a multi-dimensional evaluation system that includes short-term and long-term performance indicators, the advertising performance is calculated in real time, and the correlation between short-term and long-term performance indicators is analyzed.

[0009] S4. A multi-task online learning model and a reinforcement learning decision model work together to optimize advertising placement-related parameters in real time; the multi-task online learning model is used to predict the probability of user-related behaviors in parallel, and the reinforcement learning decision model is used to output placement decision parameters.

[0010] S5. Based on the prediction results of the multi-task online learning model and the decision parameters of the reinforcement learning decision model, generate an audience targeting strategy, a dynamic bidding strategy, and a creative selection strategy, and send the generated audience targeting strategy, the dynamic bidding strategy, and the creative selection strategy to the advertising platform for execution in real time.

[0011] S6. The real-time effect data after the execution of the audience targeting strategy, the dynamic bidding strategy and the creative selection strategy are sent back to S1 to update the multi-source real-time data collection content and form a data closed loop; the effects of different delivery strategies are compared through traffic splitting tests, and the strategies are adjusted based on the comparison results to achieve strategy iteration.

[0012] A machine learning-based real-time advertising performance optimization system includes:

[0013] A multi-source real-time data acquisition module is used to collect data from the advertising platform, user behavior data, and business target data, and to perform the normalization processing on the collected data.

[0014] The dynamic feature processing module is used to receive the regularized data output by the multi-source real-time data acquisition module, clean the regularized data, construct the dynamic features, and store the dynamic features in real time.

[0015] The multi-dimensional performance evaluation module is used to construct the multi-dimensional evaluation system that includes short-term performance indicators and long-term performance indicators, calculate the advertising performance in real time, and analyze the correlation between the indicators.

[0016] The online machine learning model module is used to deploy the multi-task online learning model and the reinforcement learning decision model to realize real-time training, inference and parameter updates of the model;

[0017] The real-time delivery decision and execution module is used to generate the audience targeting strategy, the dynamic bidding strategy and the creative selection strategy based on the output of the online machine learning model module, and to send the delivery strategy to the advertising platform for execution in real time.

[0018] The visualization and operations management module is used to display the campaign performance data, provide a strategy parameter configuration interface, and generate campaign performance review reports.

[0019] The technical effects and advantages of this invention are as follows:

[0020] 1. This invention achieves real-time automated optimization of advertising delivery parameters by constructing a collaborative optimization system that includes a multi-task online learning model and a reinforcement learning decision-making model. This overcomes the lag of traditional manual rule-based decision-making and significantly improves optimization efficiency and response speed. Through a closed-loop feedback mechanism and a strategy iteration mechanism, it ensures that the delivery strategy can continuously evolve and improve itself based on real-time data, significantly enhancing the system's adaptability.

[0021] 2. This invention uses a multi-task online learning model to simultaneously predict click-through rate, conversion rate, and user lifetime value. It also combines reinforcement learning decision-making that integrates immediate and delayed rewards, effectively balancing short-term conversion results with long-term user value. This avoids the problem of unreasonable budget allocation caused by traditional single-objective optimization, thereby significantly improving the advertiser's return on investment in the long run.

[0022] 3. By introducing an adaptive time window feature, this invention can dynamically adjust the feature calculation window according to the user's real-time behavior frequency, which solves the problem that a fixed time window cannot adapt to the differences in activity levels of different users. It significantly improves the accuracy and timeliness of dynamic features in characterizing user intent and provides a higher quality data foundation for model decision-making.

[0023] 4. This invention provides a comprehensive perspective on performance by establishing a multi-dimensional evaluation system that includes short-term performance indicators and long-term value indicators, and by quantitatively analyzing their correlation. When the system detects that short-term indicators are improving but deviating from long-term value, it can trigger an early warning in a timely manner, effectively preventing advertising behavior that unilaterally pursues short-term exposure or clicks and damages the brand's long-term value, thus ensuring the health and sustainability of advertising. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart of the method steps.

[0026] Figure 2 This diagram illustrates how online machine learning models can work together.

[0027] Figure 3 This is a system architecture diagram. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0029] See Figure 1 As shown in the flowchart of the embodiments of the present invention, the specific steps include:

[0030] S1. Obtain advertising platform data through the real-time data reporting interface of the advertising platform, obtain user behavior data through the front-end monitoring SDK embedded in the advertising media, and obtain business target data through the callback interface that connects with the advertiser's business server; perform normalization processing on the collected advertising platform data, user behavior data and business target data to obtain normalized multi-source real-time data.

[0031] The real-time data for core metrics in the advertising platform is collected at a millisecond-level frequency, while the real-time data for non-core metrics in user behavior data is collected at a second-level frequency.

[0032] This embodiment details the steps of multi-source real-time data acquisition and processing. The system first collects data in parallel through three independent channels: First, it acquires core advertising platform data at millisecond-level frequency via the advertising platform's real-time data reporting interface; specific data packets include ad impressions, ad clicks, and spending. Second, it acquires user behavior data at second-level frequency via a front-end monitoring SDK embedded in mobile news applications; specific data packets include user click events on ad placements, dwell time on ad landing pages, and page refresh behavior sequences. Third, it asynchronously acquires business target data via an order callback interface connected to the e-commerce advertiser's business server; specific data packets include user order conversion event identifiers resulting from advertising and corresponding order amounts.

[0033] During the data normalization phase, the system performs anomaly filtering and missing value imputation. Anomaly filtering removes data that does not conform to normal campaign logic. For example, if the system's logic rule is that the cost per click should not exceed 10 yuan, and a record obtained from the advertising platform interface shows a cost of 100 yuan per click, this abnormal data will be filtered out. Missing value imputation supplements missing information in the collected data. For example, if the system detects that a conversion event data obtained from the advertiser's callback interface is missing an order amount field, it automatically uses the average historical order amount of 50 yuan from that advertiser to imput it. After the above normalization process, the raw data from different sources and with varying formats are transformed into normalized multi-source real-time data. This data has a unified time sequence identifier and user session identifier, providing standardized input for subsequent dynamic feature engineering.

[0034] S2. Clean the normalized multi-source real-time data, construct dynamic features, and store the constructed dynamic features in real time to the distributed cache system;

[0035] Constructing dynamic features includes building adaptive time window features, which dynamically adjust the duration of the feature calculation window based on the frequency of user behavior in the user behavior data; the adaptive time window duration... Calculate using the following formula:

[0036] ,

[0037] in, As the baseline window duration, User behavior frequency coefficient , For fixed statistical periods Internally, the system uses a front-end monitoring SDK to track the number of user interactions in advertising scenarios.

[0038] This embodiment specifically illustrates the data cleaning, dynamic feature construction, and storage process. The system first cleans the normalized multi-source real-time data from S1, removing records containing illegal characters or format errors, such as invalid data with empty user device IDs.

[0039] Subsequently, the system enters the dynamic feature construction phase, the core of which is to construct adaptive time window features. The system presets a baseline window duration. The time interval is 60 minutes, with a fixed statistical period. The timeframe is 1 hour. The system uses a front-end monitoring SDK to obtain the number of user interactions with advertisements within the past hour. The number is 90. Based on this data, the system calculates the user behavior frequency coefficient. This equals 90 times divided by 1 hour, meaning the value of k_f is 90. Then, the system uses the formula: Calculate the duration of the adaptive time window That is, 60 minutes divided by 90, the result is approximately 0.67 minutes. Since the calculated k_f value 90 is much greater than 1.5, the system determines that the user is a high-frequency user, and therefore uses this shortened time window of approximately 0.67 minutes to dynamically calculate the user's real-time interest characteristics, such as the frequency of ad clicks within this short window.

[0040] Ultimately, the system will write the dynamic features built for this user, including the calculated adaptive time window value and the click frequency statistics based on this window, into the Redis distributed cache system in real time, using the user as the key, to ensure that the subsequent online machine learning model module can perform millisecond-level feature invocation.

[0041] S3. Based on a multi-dimensional evaluation system that includes short-term and long-term performance indicators, the advertising performance is calculated in real time, and the correlation between short-term and long-term performance indicators is analyzed.

[0042] Short-term performance metrics include real-time click-through rate (CTR), real-time conversion rate (CVR), and cost per conversion (CPA). Long-term performance metrics include real-time estimated lifetime user value. and user repurchase probability Analyzing the relationship between short-term and long-term performance indicators involves calculating the Pearson correlation coefficient r(X, Y) between them.

[0043] This embodiment specifically illustrates the multi-dimensional real-time effect evaluation and correlation analysis process in this application. The system calculates short-term effect indicators in real time based on the normalized multi-source real-time data collected by S1.

[0044] Based on data from the advertising platform, the system detected that an advertising campaign received 10,000 impressions and 200 clicks within the most recent statistical period, calculating a real-time click-through rate (CTR) of 2%. Simultaneously, based on user behavior and business objective data, the system detected 20 conversions during the same period, calculating a real-time conversion rate (CVR) of 10%. Combining this with the total spend of 1000 yuan during that period, the calculated cost per conversion (CPA) was 50 yuan.

[0045] For long-term performance metrics, the system uses the real-time prediction results from the multi-task online learning model in S4. For users who convert during the aforementioned advertising campaign, the model predicts their estimated lifetime value. The price is 300 yuan, and the probability of user repurchase is... It is 15 percent.

[0046] The system then analyzes the correlation between short-term and long-term performance metrics, specifically calculating the Pearson correlation coefficient r between short-term performance metric X and long-term performance metric Y. The system selects data from the five most recent statistical periods for calculation, where real-time click-through rate (CTR) is used for short-term performance metric X, and real-time user lifetime value (PWV) is used for long-term performance metric Y. The CTR values ​​for the five time periods were 1.8%, 2.1%, 1.9%, 2.3%, and 2.0%, respectively; corresponding to... The values ​​were 280 yuan, 320 yuan, 290 yuan, 350 yuan, and 310 yuan, respectively. The system calculated the Pearson correlation coefficient (r) for these five sets of data, and the result was approximately 0.98, indicating that the CTR and... There is a strong positive correlation between them. According to the preset rules, an alert is triggered when the correlation coefficient r is less than 0.3 and the short-term indicator increases by more than 10%. In this example, the correlation coefficient is much greater than 0.3, so no alert is triggered, indicating that the current campaign strategy has not harmed long-term user value while increasing short-term attention.

[0047] S4. A multi-task online learning model and a reinforcement learning decision model work together to optimize advertising-related parameters in real time; the multi-task online learning model is used to predict the probability of user-related behaviors in parallel, and the reinforcement learning decision model is used to output the advertising decision parameters.

[0048] The multi-task online learning model simultaneously undertakes three tasks: click-through rate (CTR) prediction, conversion rate (CVR) prediction, and lifetime value (LTV) prediction. A weighted loss function is used to calculate the model's loss value, as shown in the following formula:

[0049] ,

[0050] in, This represents the total loss value of the multi-task model. These are the loss weights for the CTR prediction task, CVR prediction task, and LTV prediction task, respectively. The loss value for the CTR prediction task. The loss value for the CVR prediction task. The loss value for the LTV prediction task.

[0051] Reinforcement learning decision-making models model the advertising delivery process as a Markov decision process. The reward function of a Markov decision process includes immediate reward and delayed reward, as shown in the following formula:

[0052] ,

[0053] in, Let be the total reward value at time t. The instantaneous reward at time t. Let γ be the delayed reward at time t, and γ be the reward discount factor.

[0054] See Figure 2 The diagram illustrates the collaborative operation of the online machine learning model provided in this embodiment of the invention. As shown, the multi-task online learning model and the reinforcement learning decision-making model do not work in isolation but rather collaborate closely. Dynamic features serve as shared input, simultaneously supplied to both models. The multi-task online learning model executes click-through rate prediction, conversion rate prediction, and user lifetime value prediction tasks in parallel. Its output, such as the probability of repeat purchases, representing long-term value metrics, is fed into the reward function calculation unit of the reinforcement learning decision-making model to calculate the decision reward that integrates long-term value. This collaborative mechanism ensures that the reinforcement learning decision-making not only focuses on immediate conversion revenue but also considers long-term user value, thereby achieving better ad placement decisions and outputting the optimal bidding coefficient. This design is key to achieving the core objective of balancing short-term effects and long-term value in this solution.

[0055] This embodiment details the online machine learning model optimization process. The system employs a multi-task online learning model and a reinforcement learning decision model working collaboratively. First, the multi-task online learning model receives dynamic features output from the dynamic feature processing module of S2 and executes three prediction tasks in parallel. This model uses a weighted loss function. Calculate the total loss value. Specifically, the model calculates the loss value for the CTR prediction task based on dynamic features and the exposure and click data collected by S1. The loss value is 0.25; based on dynamic features and click and conversion data collected by S1, the loss value for the CVR prediction task is calculated. The loss value is 0.18; based on dynamic features and historical user consumption data collected by S1, the loss value for the LTV prediction task is calculated. It is 1200.

[0056] Based on the advertiser's current business objective of prioritizing long-term return on investment, the system dynamically sets loss weights. It is 0.3. It is 0.3. It is 0.4. Then, according to the formula:

[0057] ,

[0058] The total loss value was calculated. The total loss is 480.129. The model updates its parameters in real time based on this total loss value.

[0059] Meanwhile, the reinforcement learning decision model models the advertising delivery process as a Markov decision process. At time t, the model state includes current user characteristics, competitor advertising environment, and remaining advertising budget. The decision-maker outputs the optimal bid coefficient based on this state. The model's reward function... From instant rewards Delayed rewards Composition. Instant rewards are calculated based on business objective data collected by S1, representing the revenue per user conversion at time t. The cost per user conversion is 200 yuan. If it is 50 yuan, then It equals 200 minus 50, which is 150 yuan.

[0060] The delayed reward is calculated based on the user repurchase probability predicted by a multi-task online learning model and the user's historical repurchase data collected by S1. The preset statistical period T for the delayed reward is 7 days, predicting the user repurchase probability within the next 7 days. Sequence and corresponding repurchase income The weighted sum of the sequences is 80 yuan.

[0061] The reward discount factor γ is preset to 0.9. Therefore, the total reward value at time t is... Yuan.

[0062] The reinforcement learning decision-making model uses this reward value to update its decision network, thereby optimizing the bid coefficient, a key delivery decision parameter. The two models work together in this way to achieve real-time optimization of ad delivery parameters.

[0063] S5. Based on the prediction results of the multi-task online learning model and the decision parameters of the reinforcement learning decision model, generate audience targeting strategy, dynamic bidding strategy and creative selection strategy, and send the generated audience targeting strategy, dynamic bidding strategy and creative selection strategy to the advertising platform for execution in real time.

[0064] Audience targeting strategies are based on user conversion probabilities output by multi-task online learning models. Filter user conversion probability Higher than the preset threshold Users are considered a high-value user group;

[0065] Dynamic bidding strategy combines the optimal bidding coefficient output by a reinforcement learning decision model. With remaining budget Adjust ad bids in real time using the following formula:

[0066] ,

[0067] in Based on the offer, This is the total budget for the campaign;

[0068] Creative optimization strategy is based on the real-time click-through rate of different ad creatives. The exposure percentage for each ad creative is allocated using the following formula:

[0069] ,

[0070] in This represents the percentage of exposure for the c-th creative. This represents the real-time click-through rate of the c-th creative.

[0071] This embodiment details the generation and execution process of the real-time delivery strategy. The system first uses the user conversion probability output by the multi-task online learning model... Generate audience targeting strategies. Building upon the S4 example, the model predicts the conversion probability of users. The value is 0.72. This is the system's preset threshold. It is 0.6.

[0072] Since 0.72 is higher than 0.6, the system marks this user as a high-value user and categorizes them into the high-value user group. Simultaneously, based on the user's demographic attributes and interest tags, the system expands to include potential user groups with similar characteristics, forming the final target audience.

[0073] Next, the system generates a dynamic bidding strategy. The reinforcement learning decision model outputs the current optimal bidding coefficient. The value is 1.2. The system obtains the current remaining budget from the business target data collected by S1. The total budget for the campaign is 8,000 yuan. The base bid is 10,000 yuan. The price is 1.5 yuan. According to the bid adjustment formula:

[0074] ,

[0075] Calculate the real-time bid .

[0076] Simultaneously, the system generates a creative optimization strategy. The system uses the real-time click-through rate of different ad creatives calculated in real-time by S3. To allocate exposure. Currently, there are three creatives participating in the campaign, and their real-time click-through rates are... The percentages for Creative A, Creative B, and Creative C are 1.8%, 2.2%, and 1.5%, respectively. The sum of the real-time click-through rates for all Creatives is 5.5%. Based on the exposure share distribution formula, Creative A's exposure share... Similarly, Creative B accounted for 40% of the exposure, while Creative C accounted for 27.27%.

[0077] Ultimately, the system will execute the generated audience targeting strategy, dynamic bidding strategy, and creative optimization strategy in real time through the advertising platform's delivery management API, achieving precise control and optimization of the advertising delivery process.

[0078] S6. The real-time performance data of the audience targeting strategy, dynamic bidding strategy and creative selection strategy after execution are sent back to S1 to update the multi-source real-time data collection content and form a data closed loop; the effects of different delivery strategies are compared through traffic splitting tests, and the strategies are adjusted based on the comparison results to achieve strategy iteration.

[0079] Traffic splitting testing involves dividing the campaign traffic into a test group and a control group, and comparing the campaign revenue of the two groups in real time; the difference in campaign revenue... Calculate using the following formula:

[0080] ,

[0081] in, For the test group's return on investment, The return on investment is the control group.

[0082] This embodiment specifically illustrates the effect closed-loop feedback and strategy iteration process. The system first transmits the real-time effect data generated by the audience targeting strategy, dynamic bidding strategy, and creative optimization strategy generated and executed by S5, including newly generated advertising platform data, user behavior data, and business goal data, back to the multi-source real-time data acquisition module of S1 through the data pipeline to update the collected content and form a continuously optimized data closed loop.

[0083] Simultaneously, the system performed a traffic split test to evaluate the effectiveness of the new strategy. The system divided the total delivery traffic into two groups, with 30% of the traffic serving as the test group. This test group used the new delivery strategy generated by S5, including strategies based on user conversion probability. The test group used an audience targeting strategy of 0.72, a dynamic bidding strategy with a bid of 1.44 yuan, and a creative selection strategy where Creative B accounted for 40% of the impressions. The remaining 70% of traffic served as a control group, using historical delivery strategies. After a complete test period, the system calculated the ROI for both groups using business objective data collected by S1. The test group received the total ROI. The total cost of the campaign is 7,500 yuan. The return on investment is 3000 yuan. The control group achieved a total benefit. The total cost of the campaign is 14,000 yuan. The return on investment is 7,000 yuan. .

[0084] The system calculates the difference in revenue from deployment based on the formula. This refers to the increase in ROI of the test group relative to the control group. The calculation process is as follows: Since the calculated ΔROI is 25%, which is greater than the preset 10% promotion threshold, the system determines that the new strategy is effective and automatically promotes the newly generated audience targeting strategy, dynamic bidding strategy, and creative optimization strategy to all traffic campaigns, completing the strategy iteration. If the system finds that the fluctuation range of key indicators such as click-through rate exceeds the preset range of 15% in subsequent monitoring, it will automatically roll back the strategy to the previous stable version and send an anomaly alarm message to the operations personnel to ensure the stable operation of the system.

[0085] See Figure 3 The diagram shows the architecture of the real-time advertising performance optimization system based on machine learning provided in this embodiment of the invention. The system comprises six core modules: a multi-source real-time data acquisition module, a dynamic feature processing module, an online machine learning model module, a multi-dimensional performance evaluation module, a real-time advertising decision-making and execution module, and a visualization and operations management module. It constructs a complete and collaborative technical system from multi-source data perception and intelligent model processing to strategy execution and management. The data flow originates from the multi-source real-time data acquisition module, is processed by the dynamic feature processing module, and then supplied to the online machine learning model module for model optimization and the multi-dimensional performance evaluation module for performance analysis. The output of the online machine learning model module drives the real-time advertising decision-making and execution module to generate and execute strategies. Strategy performance data is fed back to the data acquisition end, forming a closed loop. The visualization and operations management module interacts with each functional module to achieve full-process monitoring and management.

[0086] A machine learning-based real-time advertising performance optimization system includes:

[0087] The multi-source real-time data acquisition module is used to collect data from advertising platforms, user behavior data, and business objective data, and to process and organize the collected data.

[0088] The dynamic feature processing module is used to receive regularized data output from the multi-source real-time data acquisition module, clean the regularized data, construct dynamic features, and store the dynamic features in real time.

[0089] The multi-dimensional performance evaluation module is used to build a multi-dimensional evaluation system that includes short-term and long-term performance indicators, calculates the advertising performance in real time and analyzes the correlation between indicators.

[0090] The online machine learning model module is used to deploy multi-task online learning models and reinforcement learning decision models, enabling real-time training, inference, and parameter updates of the models;

[0091] The real-time delivery decision and execution module is used to generate audience targeting strategies, dynamic bidding strategies, and creative optimization strategies based on the output of the online machine learning model module, and to send the delivery strategies to the advertising platform for execution in real time.

[0092] The visualization and operations management module is used to display campaign performance data, provide strategy parameter configuration interfaces, and generate campaign performance review reports.

[0093] The dynamic feature processing module includes a stream processing submodule, which is used to execute adaptive time window duration. Calculation;

[0094] The online machine learning model module includes an online training submodule, which is used to call the weighted loss function. With reward function The computational interface;

[0095] The real-time delivery decision and execution module includes a strategy calculation submodule, which is used to calculate the bid adjustment formula, the exposure ratio formula, and the ROI improvement formula ΔROI.

[0096] This embodiment specifically illustrates a machine learning-based real-time optimization system for ad performance. A multi-source real-time data acquisition module continuously collects ad platform data, user behavior data, and business objective data through ad platform interfaces, front-end monitoring SDKs, and advertiser callback interfaces, and performs data normalization to provide the system with real-time data sources. The dynamic feature processing module receives this normalized data, and its internal stream processing submodule executes an adaptive time window duration. Calculations, such as using a baseline window duration For 60 minutes and user behavior frequency coefficient The value is 90, and the calculation is as follows: It takes approximately 0.67 minutes to process and build dynamic features based on this, which are then stored in the Redis distributed cache.

[0097] The multi-dimensional performance evaluation module constructs an evaluation system, calculating short-term performance metrics such as click-through rate and conversion rate, and long-term performance metrics such as user lifetime value prediction in real time, and analyzing their correlations. The online machine learning model module deploys and runs multi-task online learning models and reinforcement learning decision models. Its internal online training submodule calls a weighted loss function during model updates. The calculation interface uses preset loss weights and real-time calculated loss values ​​for each task to calculate the total loss value. The value is 480.129; simultaneously, during the training of the reinforcement learning decision model, the reward function is invoked. The calculation interface combines the immediate reward of 150 yuan and the delayed reward of 80 yuan to calculate the total reward value. It costs 222 yuan.

[0098] The real-time delivery decision and execution module generates a delivery strategy based on the model output. Its internal strategy calculation submodule executes key formula calculations: using the bid adjustment formula, combined with a base bid of 1.5 yuan, an optimal bid coefficient of 1.2, and a budget ratio of 0.8, it calculates the real-time bid. The cost was 1.44 yuan; using the exposure percentage formula, based on the click-through rates of creatives A, B, and C, their respective exposure percentages were calculated to be 32.73%, 40%, and 27.27%; during the strategy iteration phase, the ROI improvement formula was used, based on the test group... 2.5 and the control group The value is 2.0, and the increase in revenue is calculated. The figure is 25%. Finally, the visualization and operations management module displays all campaign performance data and strategy review reports to operations personnel, completing the entire optimization process.

[0099] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0100] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A machine learning-based method for real-time optimization of advertising performance, characterized in that, Includes the following steps: S1. Obtain advertising platform data through the real-time data reporting interface of the advertising platform, obtain user behavior data through the front-end monitoring SDK embedded in the advertising media, and obtain business target data through the callback interface that connects with the advertiser's business server. The collected advertising platform data, user behavior data, and business objective data are processed and standardized to obtain standardized multi-source real-time data. S2. Clean the normalized multi-source real-time data, construct dynamic features, and store the constructed dynamic features in real time to the distributed cache system. S3. Based on a multi-dimensional evaluation system that includes short-term and long-term performance indicators, the advertising performance is calculated in real time, and the correlation between short-term and long-term performance indicators is analyzed. S4. A multi-task online learning model and a reinforcement learning decision model work together to optimize advertising placement-related parameters in real time; the multi-task online learning model is used to predict the probability of user-related behaviors in parallel, and the reinforcement learning decision model is used to output placement decision parameters. S5. Based on the prediction results of the multi-task online learning model and the decision parameters of the reinforcement learning decision model, generate an audience targeting strategy, a dynamic bidding strategy, and a creative selection strategy, and send the generated audience targeting strategy, the dynamic bidding strategy, and the creative selection strategy to the advertising platform for execution in real time. S6. The real-time effect data after the execution of the audience targeting strategy, the dynamic bidding strategy and the creative selection strategy are sent back to S1 to update the multi-source real-time data collection content and form a data closed loop; the effects of different delivery strategies are compared through traffic splitting tests, and the strategies are adjusted based on the comparison results to achieve strategy iteration.

2. The method for real-time optimization of advertising performance based on machine learning according to claim 1, characterized in that, The real-time data corresponding to the core indicators in the advertising platform data is collected at a millisecond-level frequency, while the real-time data corresponding to the non-core indicators in the user behavior data is collected at a second-level frequency.

3. The method for real-time optimization of advertising performance based on machine learning according to claim 1, characterized in that, The construction of dynamic features includes constructing adaptive time window features, wherein the adaptive time window features dynamically adjust the duration of the feature calculation window based on the frequency of user behavior in the user behavior data; the duration of the adaptive time window... Calculate using the following formula: ， in, As the baseline window duration, User behavior frequency coefficient , For fixed statistical periods The system uses the front-end monitoring SDK to obtain the number of times users interact with advertisements.

4. The method for real-time optimization of advertising performance based on machine learning according to claim 1, characterized in that, The short-term performance metrics include real-time click-through rate (CTR), real-time conversion rate (CVR), and cost per acquisition (CPA). The long-term performance metrics include real-time estimated user lifetime value. and user repurchase probability The analysis of the correlation between the short-term effect index and the long-term effect index involves calculating the Pearson correlation coefficient r(X, Y) between the short-term effect index and the long-term effect index.

5. The method for real-time optimization of advertising performance based on machine learning according to claim 4, characterized in that, The multi-task online learning model simultaneously performs three tasks: click-through rate (CTR) prediction, conversion rate (CVR) prediction, and lifetime value (LTV) prediction. A weighted loss function is used to calculate the model's loss value, as shown in the following formula: ， in, This represents the total loss value of the multi-task model. These are the loss weights for the CTR prediction task, CVR prediction task, and LTV prediction task, respectively. The loss value for the CTR prediction task. The loss value for the CVR prediction task. The loss value for the LTV prediction task.

6. The method for real-time optimization of advertising performance based on machine learning according to claim 5, characterized in that, The reinforcement learning decision model models the advertising delivery process as a Markov decision process. The reward function of the Markov decision process includes immediate reward and delayed reward, as shown in the following formula: ， in, Let be the total reward value at time t. The instantaneous reward at time t. Let γ be the delayed reward at time t, and γ be the reward discount factor.

7. The method for real-time optimization of advertising performance based on machine learning according to claim 1, characterized in that, The audience targeting strategy is based on the user conversion probability output by the multi-task online learning model. Filter the user conversion probability Higher than the preset threshold Users are considered a high-value user group; The dynamic bidding strategy combines the optimal bidding coefficient output by the reinforcement learning decision model. With remaining budget Adjust ad bids in real time using the following formula: ， in Based on the offer, This is the total budget for the campaign; The creative selection strategy is based on the real-time click-through rate of different ad creatives. The exposure percentage for each ad creative is allocated using the following formula: ， in This represents the percentage of exposure for the c-th creative. This represents the real-time click-through rate of the c-th creative.

8. The method for real-time optimization of advertising performance based on machine learning according to claim 1, characterized in that, The traffic splitting test involves dividing the deployed traffic into a test group and a control group, and comparing the deployment revenue of the test group and the control group in real time; the difference in deployment revenue... Calculate using the following formula: ， in, For the test group's return on investment, The return on investment is the control group.

9. A machine learning-based real-time optimization system for advertising performance, characterized in that, The system is used to implement the real-time optimization method for advertising performance based on machine learning as described in any one of claims 1 to 8, the system comprising: A multi-source real-time data acquisition module is used to collect data from the advertising platform, user behavior data, and business target data, and to perform the normalization processing on the collected data. The dynamic feature processing module is used to receive the regularized data output by the multi-source real-time data acquisition module, clean the regularized data, construct the dynamic features, and store the dynamic features in real time. The multi-dimensional performance evaluation module is used to construct the multi-dimensional evaluation system that includes short-term performance indicators and long-term performance indicators, calculate the advertising performance in real time, and analyze the correlation between the indicators. The online machine learning model module is used to deploy the multi-task online learning model and the reinforcement learning decision model to realize real-time training, inference and parameter updates of the model; The real-time delivery decision and execution module is used to generate the audience targeting strategy, the dynamic bidding strategy and the creative selection strategy based on the output of the online machine learning model module, and to send the delivery strategy to the advertising platform for execution in real time. The visualization and operations management module is used to display the campaign performance data, provide a strategy parameter configuration interface, and generate campaign performance review reports.

10. The real-time optimization system for advertising performance based on machine learning according to claim 9, characterized in that, The dynamic feature processing module includes a stream processing submodule, used to execute the adaptive time window duration. The calculation; the online machine learning model module includes an online training submodule, used to call the weighted loss function. With the reward function The calculation interface; the real-time delivery decision and execution module includes a strategy calculation submodule, which is used to calculate the bid adjustment formula, the exposure ratio formula and the ROI increase formula ΔROI.

Images