Intelligent huge advertisement putting method

Abstract

The invention relates to the technical field of Internet advertisement putting, in particular to an intelligent huge advertisement putting method, which comprises the following steps of: acquiring and preprocessing multi-dimensional operation data in real time, constructing a user response prediction model based on a deep interest network, and dynamically weighting a user behavior sequence by utilizing an attention mechanism. A dynamic budget allocation optimization model is established; an objective function is set according to the type of an advertiser; a budget weight is adjusted according to a real-time bidding success rate; a real-time decision engine is deployed; according to the method, each link is monitored through an exposure conversion funnel, an abnormal alarm is set, strategy backtracking analysis is started, and continuous optimization is performed in combination with online learning and model gray release, so that the problem of low efficiency of budget allocation in the traditional technology is solved, intelligent and accurate putting of a huge amount of advertisements is realized, the advertisement conversion rate is improved, and the customer obtaining cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of Internet advertising delivery, and in particular to an intelligent mass advertising delivery method. Background Art

[0002] Internet advertising is an important technology. In the field of digital marketing, intelligent massive advertising technology is a core means to improve advertising conversion rates and reduce customer acquisition costs by processing massive user behavior data in real time and optimizing budget allocation and bidding strategies.

[0003] With the intensification of internet traffic competition and the diversification of advertiser demands, accurately addressing traffic fluctuations, user interest shifts, and budget efficiency in massive advertising campaigns has become a key challenge in achieving marketing goals. However, traditional advertising delivery technologies suffer from the core issues of delayed user interest capture and inefficient dynamic budget allocation. Existing solutions use fixed-weight user response models and fail to incorporate an attention mechanism to dynamically weight user historical behavior. When user interests change rapidly, model prediction bias increases, leading to lower ad click-through rates. Static budget allocation strategies allocate resources according to a fixed ratio. During peak traffic periods, such as e-commerce promotion seasons, the inability to adjust budget weights in real time reduces the acquisition rate of high-value user traffic. Furthermore, traditional methods lack a dynamic bidding pricing mechanism, employing conservative bidding during periods of intense market competition, resulting in low bidding success rates. Furthermore, during low traffic periods, rigid bidding strategies lead to high costs. This crude user interest modeling and lack of dynamic adjustment capabilities ultimately increase the conversion cost of traditional advertising solutions, inefficient budget utilization, and make it difficult to meet advertisers' precise delivery needs in complex traffic environments. To address this technical issue, we propose an intelligent massive advertising delivery method. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent mass advertising delivery method to solve the problems raised in the above background technology.

[0005] 1. Because traditional solutions lag in capturing user interests and have large model prediction biases, this case builds a prediction model based on a deep interest network. Using an attention mechanism to dynamically weight user behavior improves the model's adaptability to changes in user interests and reduces prediction bias.

[0006] 2. Because traditional dynamic budget allocation methods are inefficient and the acquisition rate of high-value traffic is low, this case establishes a dynamic budget allocation optimization model. By adjusting the budget weight according to the real-time bidding success rate, it can improve budget utilization and ensure the acquisition of high-value user traffic.

[0007] To achieve the above objectives, one of the objectives of the present invention is to provide a method for intelligent mass advertising delivery, comprising the following steps:

[0008] S1. Collect multi-dimensional operation data of the advertising delivery system in real time, pre-process it and store it in a distributed time series database;

[0009] S2. Build a user response prediction model based on a deep interest network, using an attention mechanism to dynamically weight the user's historical behavior sequence to generate a user interest vector. An adversarial sample generator is introduced during model training.

[0010] S3. Build a dynamic budget allocation optimization model. This model constructs a constrained linear programming problem based on the advertiser's objective function and allocates budget weights in real time using the Lagrangian dual decomposition algorithm. The objective function integrates real-time bidding costs, expected conversion rates, and user interest matching.

[0011] S4. Deploy a real-time decision-making engine, integrate a streaming computing framework, batch process ad requests according to time windows, dynamically adjust prices based on market competition intensity, and trigger resource expansion when the bidding success rate falls below a preset threshold.

[0012] S5. Implement a multi-dimensional feedback control system, continuously update the parameters of the user response prediction model through the online learning module, and establish a delivery effect tracking matrix. Set abnormal fluctuation alarm thresholds for each link in the exposure conversion funnel, and automatically start the delivery strategy retrospective analysis when the click-through rate of the advertising group is monitored to drop by more than 15%.

[0013] As a further improvement of this technical solution, the step S2 of constructing the user response prediction model includes:

[0014] The user's historical behavior sequence is divided into short-term behavior segments and long-term behavior archives according to the time window. The attention mechanism is used to calculate the correlation weight between the current ad features and each historical behavior segment, and a dynamic weighting coefficient is assigned to the highly correlated segments.

[0015] When the user is a newly registered user, a cross-advertiser transfer learning strategy is adopted to initialize the interest vector based on similar crowd portraits. The adversarial sample generator generates perturbation data by randomly masking the key fields of the user behavior sequence.

[0016] As a further improvement of this technical solution, during the model training:

[0017] Full model training is performed during the low-traffic period in the early morning every day. The training data uses a sliding window to retain behavioral samples from the last 30 days. The incremental learning mode is started during the peak advertising period, and only the last layer of model parameters are updated using real-time streaming data. When the distribution change of advertising categories is monitored to exceed the preset variance threshold, the feature importance re-evaluation module is triggered to automatically adjust the weights of the input feature dimensions.

[0018] As a further improvement of this technical solution, the dynamic budget allocation optimization in step S3 includes:

[0019] Advertisers are divided into two categories: brand exposure type and effect conversion type. The objective functions are set based on the cost per thousand impressions and the cost per conversion respectively.

[0020] 80% of the budget weight is pre-allocated before the start of daily advertising bidding, and the remaining 20% ​​of the budget is dynamically adjusted based on the real-time bidding success rate. When an ad group fails to bid for five consecutive times, its budget weight will be automatically transferred to the substitute ad group of the same advertiser.

[0021] As a further improvement of this technical solution, when the budget weight is adjusted:

[0022] For advertisers who focus on performance conversion, priority will be given to securing the traffic acquisition budget for users with the top 20% historical conversion rates. When the overall traffic utilization rate of the advertising platform is lower than 70%, the budget constraint upper limit for high-value advertisers will be relaxed. During the e-commerce promotion season, a budget overdraft mechanism will be enabled for participating ad groups, allowing single-day consumption to exceed 120% of the daily budget limit.

[0023] As a further improvement of this technical solution, the dynamic price adjustment in step S4 includes:

[0024] The market competition intensity is divided into three levels: stable period, fluctuating period, and fierce battle period:

[0025] During stable periods, a conservative bidding strategy is adopted, bidding at 50% of the estimated conversion value. During volatile periods, a game learning strategy is activated, recording competitor bidding patterns and dynamically matching them. During intense competition, a peak-grabbing strategy is activated, allowing temporary exceeding of the maximum bid limit set by advertisers.

[0026] Resource expansion adopts a hierarchical response mechanism:

[0027] Level 1 capacity expansion: Add 10% computing node replicas to cope with short-term traffic fluctuations;

[0028] Secondary capacity expansion: Enable reserved bidding dedicated server clusters to cope with sustained high loads.

[0029] As a further improvement of this technical solution, when detecting the bidding success rate:

[0030] A separate success rate threshold is set for the first position in the information flow. When the click-through rate of the advertising material continues to decline, the backup creative material is automatically switched to participate in the bidding. For low-value traffic during the night period, the batch bidding package mode is enabled to reduce the cost of a single bid.

[0031] As a further improvement of this technical solution, the delivery effect tracking in step S5 includes:

[0032] The exposure-conversion funnel is divided into four monitoring levels: exposure, visible exposure, click, and conversion. Abnormal fluctuation alarm rules are set for each level:

[0033] Exposure level fluctuations exceeding ±30% of the daily average trigger a yellow alert, while a year-on-year drop in click-through rate exceeding 15% and lasting for two hours triggers a red alert. An attribution analysis report is automatically generated, marking the traffic channels, geographical distribution, and device types associated with the abnormal period.

[0034] As a further improvement of this technical solution, during the strategy backtracking analysis:

[0035] If the CTR drop is limited to a specific region, roll back the recent bid strategy changes for that region. If the CTR drop is occurring across multiple regions simultaneously, perform a creative fatigue check.

[0036] Automatically mark materials with more than one million exposures as "fatigue" and start the A / B testing engine to verify the effectiveness of new material combinations within 30 minutes. When backtracking points to model prediction deviation, freeze the current model and switch to the historical stable version.

[0037] As a further improvement of this technical solution, the online learning module includes:

[0038] The real-time feedback data pipeline is divided into three processing paths. The click log stream is used to update the user's short-term interest vector, the conversion log stream is used to correct the conversion rate estimation deviation, and the negative sample stream is used to optimize the feature cross-weight. The negative sample stream is the exposure that was not clicked.

[0039] The model version grayscale release mechanism means that the new model is first run in a 1% traffic environment for comparative testing, and then fully deployed when the key indicators are improved beyond the confidence interval.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] In an intelligent mass advertising delivery method, through multi-dimensional data preprocessing and deep interest network modeling, it can dynamically capture changes in user interests, improve the accuracy of model predictions, and lay the foundation for precise delivery. The dynamic budget allocation optimization model combines advertiser type and real-time bidding data to adjust budget weights, improve budget utilization efficiency, and ensure the acquisition of high-value traffic. The real-time decision engine dynamically adjusts prices according to the market competition intensity, and cooperates with the resource expansion mechanism to improve the bidding success rate and the system's ability to cope with traffic fluctuations. The multi-dimensional feedback control system timely detects delivery anomalies and optimizes strategies through exposure conversion funnel monitoring and strategy backtracking analysis. The online learning module and model grayscale release mechanism continuously optimize the model to ensure stable improvement in delivery effect. This method realizes the precise capture of user interests, efficient budget allocation, intelligent bidding regulation and continuous optimization of delivery effect, improves advertising conversion rate, reduces customer acquisition cost, and meets the precise delivery needs of advertisers in complex traffic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall workflow diagram of the present invention. DETAILED DESCRIPTION

[0043] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0044] See also Figure 1 As shown, this embodiment provides a method for intelligent mass advertising delivery, including the following steps:

[0045] S1. Collect multi-dimensional operation data of the advertising delivery system in real time, pre-process it and store it in a distributed time series database;

[0046] S2. Build a user response prediction model based on a deep interest network, using an attention mechanism to dynamically weight the user's historical behavior sequence to generate a user interest vector. An adversarial sample generator is introduced during model training.

[0047] S3. Build a dynamic budget allocation optimization model. This model constructs a constrained linear programming problem based on the advertiser's objective function and allocates budget weights in real time using the Lagrangian dual decomposition algorithm. The objective function integrates real-time bidding costs, expected conversion rates, and user interest matching.

[0048] S4. Deploy a real-time decision-making engine, integrate a streaming computing framework, batch process ad requests according to time windows, dynamically adjust prices based on market competition intensity, and trigger resource expansion when the bidding success rate falls below a preset threshold.

[0049] S5. Implement a multi-dimensional feedback control system, continuously update the parameters of the user response prediction model through the online learning module, and establish a delivery effect tracking matrix. Set abnormal fluctuation alarm thresholds for each link in the exposure conversion funnel, and automatically start the delivery strategy retrospective analysis when the click-through rate of the advertising group is monitored to drop by more than 15%.

[0050] In order to solve the problems of delayed user interest capture and large model prediction deviation in traditional solutions, it is necessary to build a user response prediction model based on a deep interest network. The construction of the user response prediction model in step S2 includes:

[0051] The user's historical behavior sequence is divided into short-term behavior segments and long-term behavior archives according to the time window, where the short-term is 7 days and the long-term is 7 to 30 days. The segmentation is achieved by parsing the timestamp field in the user behavior log. The attention mechanism is used to calculate the association weight between the current advertising features and each historical behavior segment. The advertising features include five parameters: advertiser ID, product category ID, price range code, brand ID, and promotion label. This group of parameters is automatically assigned by the advertising delivery system when the advertisement is created and stored in the distributed time series database. The association weight between the advertising feature vector and the behavior segment feature vector is calculated using the cosine similarity algorithm. High-correlation segments with a weight value greater than 0.7 are assigned a 1.5-fold dynamic weighting coefficient, and low-correlation segments with a weight value less than 0.3 are assigned a 0.5-fold coefficient. Finally, the user interest vector is weighted to generate the user interest vector. For newly registered users, the interest vector is initialized through the following technical process:

[0052] When the user is a newly registered user, a cross-advertiser transfer learning strategy is adopted. First, the age, gender, and region fields in the user registration information and the operating system and device model parameters in the device fingerprint are extracted to form a 10-dimensional feature vector. Then, based on the cosine similarity algorithm, a set of similar people with a similarity greater than 0.8 to the feature vector is retrieved from the advertising platform's population portrait library. Finally, a weighted average operation is performed on the interest vectors of similar people, with the weight being the similarity value between the individual people and the new user, to generate the initial interest vector of the new user. The adversarial sample generator processes the user behavior sequence according to the following rules:

[0053] The product ID field in the sequence is randomly perturbed with a mask probability of 0.3. Specifically, a random number in the interval [0,1) is generated. When the value is less than 0.3, the corresponding product ID is replaced with a special placeholder "[MASK]" to generate perturbed training data. Through the above optimization, the system achieves accurate modeling of user interests and provides a solid model foundation for precise advertising delivery.

[0054] During model training, to balance model training accuracy and online service performance, a time-segment differentiation strategy is adopted for model training. The specific implementation is as follows:

[0055] Taking into account the low number of advertising requests during the early morning hours, such as 0:00-6:00, this period is selected to perform full model training to reduce the resource usage of online services. The training data retains the user behavior samples of the last 30 days through a sliding window mechanism. Specifically, by parsing the timestamp field in the behavior log, the behavioral data of clicks, exposures, and conversions within 30 days are filtered out, and repeated clicks and missing values ​​of the same user at the minute level are eliminated. The distributed training framework is used to update the weights of the entire network to ensure that the model captures long-term user interest trends. During peak advertising hours, such as 9:00-22:00, to ensure that the model responds quickly to real-time changes in user interests, the incremental learning mode is started. The latest behavioral data is obtained through the real-time stream data processing engine, such as the log message queue. Only the parameters of the last fully connected layer of the model are updated. Specifically, the real-time data is aggregated into a 5-minute time window, and the gradient is calculated using the stochastic gradient descent algorithm. The weights of the last layer are incrementally updated, while the parameters of the previous feature extraction layer remain unchanged. When a sudden increase in real-time clicks on a certain product category is detected, incremental learning can increase the weight of the corresponding feature within 10 minutes, enabling model predictions to adapt more quickly to traffic changes. Compared to full retraining, parameter update latency is reduced from hours to minutes. To address dynamic changes in the advertising category structure, the system monitors fluctuations in the advertising category distribution in real time. By calculating the hourly impression share of each advertising category, the variance of the category distribution between the current hour and the past 24 hours is calculated. When the variance exceeds a preset threshold, the feature importance reassessment module is triggered. This module uses a tree-based feature attribution algorithm to reorder the importance of input features and adjust the input weights of each feature in descending order. For example, if the original weight of the "mobile phone category" feature is 0.2, after detecting a 20% increase in the overall traffic share of the 3C category, its weight is reassessed and increased to 0.25. At the same time, the weights of non-relevant categories are reduced. This improves the match between the model input features and the current traffic structure, forming a model training system that balances stability and flexibility.

[0056] In order to achieve efficient dynamic budget allocation and solve the resource waste problem of traditional static allocation strategy when traffic fluctuates, the specific implementation method of dynamic budget allocation optimization in step S3 is as follows:

[0057] Based on the core delivery goals set by advertisers in the system background, advertisers are divided into two categories: brand exposure and effect conversion. Objective functions constrained by cost per thousand impressions and cost per conversion are set respectively. Advertisers who choose "increase brand awareness" are automatically classified as brand exposure. The system sets an objective function constrained by cost per thousand impressions. Cost per thousand impressions is calculated based on historical advertising delivery data. Advertisers who choose "promote product conversion" are classified as effect conversion. An objective function constrained by cost per conversion is set. Cost per conversion is calculated based on historical conversion data and expected ROI. The objective function also integrates real-time bidding cost, expected conversion rate and user interest matching. The constrained linear programming problem is transformed into an unconstrained optimization problem through the Lagrangian dual decomposition algorithm to achieve real-time allocation of budget weights. The real-time bidding cost is taken from the average bid in the current bidding environment, the expected conversion rate is predicted by the model in step S2, and the user interest matching is based on the cosine similarity between the advertising features and the user interest vector.

[0058] Before the daily ad auction begins, the system pre-allocates 80% of the budget weight to each advertiser's primary ad group based on historical delivery performance and objective function priority. This pre-allocation process determines the initial budget ratio of each ad group by solving a linear programming problem. The remaining 20% ​​of the budget enters the dynamic adjustment pool and is dynamically adjusted based on the real-time bidding success rate. The real-time bidding success rate is the number of successful bids for a certain ad group divided by the total number of participations, with a time window of 15 minutes:

[0059] When an ad group fails to bid for five consecutive times, the system automatically triggers the budget transfer mechanism, extracting 10% from the current budget weight of the ad group and transferring it to the substitute ad group preset by the same advertiser. The substitute ad group is specified by the advertiser in the background and is usually a backup material combination ranked in the top three in historical conversion rate, ensuring that the budget is tilted towards ad groups with high bidding success rates, thereby improving budget utilization. Through the above-mentioned classification objective function and dynamic budget adjustment mechanism, the system realizes precise and real-time budget allocation, reducing the budget waste rate.

[0060] To further improve budget efficiency, a differentiated strategy will be implemented for performance-driven advertisers during the budget weight adjustment phase. The specific implementation methods are as follows:

[0061] Considering that 20% of high-value users usually contribute 80% of conversions, the system calculates the historical conversion rate of each user by advertiser dimension based on the user conversion data of the past 7 days, and screens out the top 20% user groups. The historical conversion rate is the number of conversions / number of exposures. For effect-conversion advertisers, priority is given to securing the budget for the traffic of users with the top 20% historical conversion rate. Specifically, the conversion records corresponding to the user ID are queried through the distributed database, the conversion rate is calculated and sorted by advertiser group, and a whitelist of high-value users is generated. When allocating budgets, a budget weight bonus is set for the traffic requests of this type of users. When a high When a valuable user requests an ad, the budget allocation ratio of the corresponding ad group will be automatically increased by 15%, and the real-time bidding system will give priority to meeting its traffic acquisition needs. When the overall traffic utilization rate of the advertising platform is lower than 70%, it indicates that the current traffic resources are not fully utilized. At this time, the system will automatically relax the budget constraint upper limit for high-value advertisers, that is, the effect-conversion advertisers with the top 30% historical conversion rate. The overall traffic utilization rate is the actual used traffic / total available traffic. Specifically, the traffic fill rate of each ad position is monitored in real time, and the overall traffic utilization rate of the platform is counted hourly. When it is lower than 70% for 2 consecutive hours, it will trigger The constraint relaxation process increases the single conversion cost constraint threshold of effect conversion advertisers by 20% to avoid traffic waste due to rigid bidding strategies. During the e-commerce promotion season, by connecting to the e-commerce platform calendar or monitoring promotion keyword identification, such as "Double 11" and "618", the budget overdraft mechanism is enabled for the participating ad groups. The system first marks the corresponding ad group in the background according to the promotion time range submitted by the advertiser. When the single-day consumption reaches 100% of the daily budget, the overdraft approval process is automatically triggered, allowing continued consumption to 120% of the daily budget, but the hourly overdraft limit does not exceed 10% of the daily budget. For example, the daily budget of an advertising group is 100,000 yuan. When the consumption reaches 100,000 yuan, the system allows it to continue to consume to 120,000 yuan, but at most 10,000 yuan per hour. The conversion effect is monitored in real time during the overdraft period. If the conversion rate is lower than 80% of the historical average for 2 consecutive hours, the overdraft will be automatically stopped and an early warning will be sent, achieving a balance between sales and costs. Through the above budget weight adjustment strategy, the system has achieved refined operation of budget resources: high-value users are given priority to improve conversion efficiency, constraints are relaxed during low-traffic periods to improve resource utilization, and the overdraft mechanism during the promotion season seizes traffic peak opportunities to improve budget utilization efficiency.

[0062] To cope with the dynamic changes in market competition intensity and traffic fluctuations in advertising delivery, the specific implementation methods of dynamic price adjustment and resource expansion in step S4 are as follows:

[0063] The system divides the market competition intensity into three levels: stable period, fluctuating period and fierce competition period by monitoring the number of bidding participants and the average bid fluctuation range of advertising space in real time.

[0064] When the number of bidders for the same ad space is less than 5 and the average bid fluctuation is less than or equal to 10%, it is determined to be a stable period. At this time, a conservative bidding strategy is adopted, and a bid is made at 50% of the conversion value predicted by the model in step S2. For example, if the estimated conversion value of a certain click is 20 yuan, the bid is 10 yuan to control the cost. When the number of bidders is between 5 and 15 and the average bid fluctuation is between 10% and 30%, it is determined to be a fluctuation period. The game learning strategy is started, and the bidding time interval of competitors in the past 30 minutes is analyzed. The system generates a dynamically matching bidding strategy based on the price change pattern. For example, if it detects that a competitor is raising their bid by 5% every 10 minutes, the system will adjust the bidding rhythm to improve the bidding success rate. When the number of bidders exceeds 15 and the average bid fluctuation is greater than 30%, it is determined to be a fierce competition period and the peak preemption strategy is activated, allowing temporary exceeding of the maximum bid limit set by the advertiser, but not exceeding 120% of the original limit. For example, if an advertiser sets a maximum bid of 30 yuan, it can be temporarily increased to 36 yuan during the fierce competition period to ensure the acquisition of high-value traffic.

[0065] To ensure the stability of the bidding system during traffic fluctuations, a hierarchical response resource expansion strategy is adopted. The resource expansion adopts a hierarchical response mechanism:

[0066] When the number of bid requests is monitored to increase by 10% to 20% year-on-year within 10 minutes, a first-level capacity expansion is triggered: the container orchestration system automatically adds 10% of the compute node replicas. For example, if the original deployment was 100 nodes, 10 nodes would be added to cope with short-term traffic fluctuations. This process is completed within 5 minutes, ensuring that the system response latency remains within 50 milliseconds.

[0067] When the request volume increases by more than 20% year-on-year and lasts for more than 30 minutes, initiate secondary capacity expansion: enable the reserved bidding dedicated server cluster, a physical server cluster with a capacity of 30% of the daily deployment, and direct traffic to the newly added cluster through the load balancer. For example, if 1,000 container nodes are used daily, 300 physical servers will be activated during secondary capacity expansion to avoid bidding failures due to insufficient resources.

[0068] To further improve bidding efficiency and reduce costs, a differentiated processing strategy is implemented during the bidding success rate detection phase. The specific implementation methods are as follows:

[0069] Taking into account the first position in the information flow, that is, the first ad position that users see after opening the app, the traffic value is significantly higher than that of other positions, the system sets a success rate threshold for it separately. By analyzing historical data, the bidding success rate threshold for the first position is set to 70%, and that for ordinary positions is 50%. This threshold is determined by the high exposure and conversion potential of the first position, and the bidding success of the first position is monitored in real time. When the success rate of an ad group in the first position is lower than 70% for three consecutive times, the system automatically increases the bid coefficient of the ad group in the first position to 1.3 times, that is, it participates in the bidding at 130% of the original bid. If it is lower than the threshold for five consecutive times, it is determined that the current material is not attractive enough in the first position, and the backup creative process is triggered. The system automatically selects the creative with the highest historical click-through rate from the advertiser's preset backup creative library to replace the current creative in the bidding. The click-through rate is calculated as clicks / impressions. The advertiser uploads 3-5 sets of backup creatives to the backup creative library in the background to avoid traffic waste caused by material fatigue, so that the traffic acquisition ability of the ad group in the first position remains stable. For low-value traffic during the night time period (i.e., 23:00-6:00) with low user activity, the batch bidding package mode is enabled to reduce the cost of a single bid. The system executes according to the following process:

[0070] First, through user behavior data analysis, the traffic value coefficient of each ad space during the night time period is determined. When the traffic value coefficient is lower than 0.5, the packaging mode is triggered. Then, 10 to 20 similar ad requests from the same advertiser are packaged into a bidding unit, and a unified bid is used to participate in the bidding. The unified bid takes the average bid of the packaged requests. After the bidding is successful, it is allocated to each ad request according to the preset rules. For example, 15 ad requests with bids of 5-8 yuan are packaged and a unified bid of 6.5 yuan is used to participate in the bidding. The cost of a single bidding is reduced from 0.1 yuan / time to 0.05 yuan / time, achieving low-cost traffic coverage. Through the above-mentioned bidding success rate detection and optimization strategy, the system realizes the precise maintenance of high-value traffic and cost control of low-value traffic.

[0071] In order to promptly detect anomalies in the advertising delivery process and accurately locate problems, the delivery effect tracking in step S5 is implemented through a four-level monitoring hierarchy and anomaly alarm mechanism. The specific implementation method is as follows:

[0072] Based on the user behavior path of ad delivery, the exposure-conversion funnel is divided into four monitoring levels: exposure, viewable exposure, click, and conversion. Impressions are collected through ad request log statistics. Viewable exposures are monitored by front-end tracking of the duration of an ad on the user's screen. Durations greater than or equal to one second are considered viewable. Clicks are collected from interaction logs of users clicking on ads. Conversions are transmitted back through the back-end data interface based on the advertiser's conversion goals, forming a complete data monitoring chain for conversion. Alarm rules for abnormal fluctuations are set for each level of the four-level funnel. Exposure levels are compared with the average of the previous three days. Fluctuations exceeding ±30% trigger a yellow alert. This threshold is determined by analyzing the standard deviation of historical data. For example, if an ad group averages 100,000 daily impressions, an alert is triggered if the real-time exposure suddenly increases to 130,000 or drops to 70,000. A red alert is triggered if the click-through rate drops by more than 15% compared to the same period of the previous day and persists for two hours. The click-through rate is calculated as clicks divided by viewable exposures. When an alert is triggered, the system automatically generates an attribution analysis report and identifies the cause of the anomaly through the following steps:

[0073] First, extract the traffic channel data for the abnormal period, that is, the 2-hour period that triggered the red alert, such as information flow A and search B, and count the exposure share and click-through conversion rate of each channel. Then analyze the regional distribution data to compare the changes in regional traffic during the abnormal period and the normal period. Finally, parse the device type data to see if there are any performance anomalies on specific devices. For example, a drop in the click-through rate of an ad group triggers a red alert. The report shows that the traffic share from the "information flow C" channel increased from 20% to 40% during the abnormal period, but the click-through rate of this channel was only 60% of that of the normal channel. Therefore, the decline in channel quality is located as the cause of the anomaly, and the accuracy of anomaly positioning is improved. Through the above-mentioned four-level funnel monitoring and attribution analysis mechanism, the system realizes full-link visualization of advertising delivery effects and rapid positioning of anomalies, providing data support for the optimization of delivery strategies.

[0074] To accurately identify the cause of the decline in click-through rate and quickly implement optimization measures, a scenario-based processing mechanism is adopted during strategy retrospective analysis. The specific implementation method is as follows:

[0075] If a drop in click-through rate occurs only in a specific region, the system first retrieves the bid strategy change records for that region from the past seven days, stored in the strategy version management database. This includes operation logs for bid modifier adjustments and changes to target demographic screening criteria. If a bid strategy change is detected within the 24 hours prior to the drop in click-through rate in that region, a strategy rollback process is automatically triggered, restoring the bid parameters for that region to the pre-change version and simultaneously recording the rollback time and impact range. If a drop in click-through rate occurs simultaneously in more than three provincial-level administrative regions, the system determines that creative fatigue may be the problem and automatically initiates creative fatigue detection. The system first calculates the cumulative impressions of each creative and aggregates them by creative ID from the ad delivery log. Creatives with more than one million impressions are automatically marked as "fatigue" and a status tag is added to the creative management system. The A / B testing engine then launches, selecting two or three similar creatives from the backup creative library, pre-uploaded by the advertiser and marked as "backup," and constructing a test plan within 30 minutes.

[0076] Allocate traffic between the original and new creatives in a 4:6 ratio, while keeping other delivery parameters consistent. Compare the click-through rate differences between each creative group in real time. For example, if a creative has accumulated 1.2 million impressions and its click-through rate has dropped in multiple regions, and A / B testing shows that the new creative has a 20% higher click-through rate than the original, the new creative will be automatically switched to. If attribution analysis indicates model prediction deviation, such as if the model's estimated click-through rate deviates from the actual click-through rate by more than 20%, the system will retrieve historical stable versions from the model version management library, select versions with key indicator fluctuations of less than 5% over the past 7 days, and perform the following operations:

[0077] First, the parameter update of the current model is frozen and the reception of real-time training data is stopped. Then, the advertising request traffic is gradually switched to the historical stable version through the load balancer. The switching process adopts a four-level traffic increase strategy of 10%, 30%, 70%, and 100%. Each stage lasts for 5 minutes to ensure a smooth switching process, guarantee the stability of the delivery effect, and ensure the continuous optimization of the advertising delivery effect.

[0078] To achieve continuous optimization and stable iteration of the model, the online learning module uses a three-way feedback data pipeline and a phased release mechanism to ensure model performance. The specific implementation is as follows:

[0079] Considering the differentiated value of different types of feedback data for model optimization, the real-time feedback data pipeline is divided into click log stream, conversion log stream, and negative sample stream for processing. The click log stream collects real-time behavioral data of users clicking on ads, including ad ID, user ID, and click time. By analyzing the product category and ad features corresponding to the click behavior, a sliding window algorithm is used to update the user's short-term interest vector. The conversion log stream obtains event data of users completing conversions, compares the actual conversion results with the model's estimated conversion rate, and corrects the conversion rate estimation deviation using the gradient descent algorithm. The negative sample stream collects data on ads that were exposed but not clicked. The feature crossover algorithm analyzes the reasons for non-clicks, optimizes the feature crossover weights, and improves the model's predictive ability for non-click scenarios. To ensure the stability of the new model after launch, a grayscale release mechanism with a 1% traffic environment is adopted:

[0080] After the new model training is completed, a 1% traffic ratio is first configured in the advertising delivery system, and traffic diversion is achieved through user ID hash modulo. For example, users with an ID ending in 0 enter the new model test group, while keeping other delivery parameters consistent. During the grayscale test, the click-through rate is monitored in real time and used as a key indicator. When the key indicator of the new model improves by more than 95% confidence interval, the full deployment process is started. If the indicator does not meet expectations or an abnormality occurs, the grayscale test is terminated and rolled back to the original model, avoiding the effect fluctuations that may be caused by the launch of the new model. Through the above-mentioned online learning and grayscale release mechanism, the system realizes dynamic optimization and stable iteration of the model, providing continuous model capability support for the precise delivery of massive advertisements.

[0081] The above-mentioned intelligent massive advertising delivery method, the present invention collects multi-dimensional operation data in real time and pre-processes it, constructs a user response prediction model based on a deep interest network, uses the attention mechanism to dynamically weight user behavior sequences, combines the adversarial sample generator to improve the robustness of the model, establishes a dynamic budget allocation optimization model, sets the objective function according to the advertiser type, adjusts the budget weight according to the real-time bidding success rate, deploys a real-time decision engine, dynamically adjusts prices according to the market competition intensity and triggers resource expansion, implements multi-dimensional feedback control, monitors each link through the exposure conversion funnel, sets abnormal alarms and starts strategy backtracking analysis, combines online learning with model grayscale release for continuous optimization, and solves the problem of inefficient budget allocation in traditional technologies, realizes intelligent and precise delivery of massive advertisements, improves advertising conversion rates and reduces customer acquisition costs.

[0082] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent mass advertising delivery, characterized in that: The following steps are involved: S1. Collect multi-dimensional operation data of the advertising delivery system in real time, pre-process it and store it in a distributed time series database; S2. Build a user response prediction model based on a deep interest network, using an attention mechanism to dynamically weight the user's historical behavior sequence to generate a user interest vector. An adversarial sample generator is introduced during model training. S3. Build a dynamic budget allocation optimization model. This model constructs a constrained linear programming problem based on the advertiser's objective function and allocates budget weights in real time using the Lagrangian dual decomposition algorithm. The objective function integrates real-time bidding costs, expected conversion rates, and user interest matching. S4. Deploy a real-time decision-making engine, integrate a streaming computing framework, batch process ad requests according to time windows, dynamically adjust prices based on market competition intensity, and trigger resource expansion when the bidding success rate falls below a preset threshold. S5. Implement a multi-dimensional feedback control system, continuously update the parameters of the user response prediction model through the online learning module, and establish a delivery effect tracking matrix. Set abnormal fluctuation alarm thresholds for each link in the exposure conversion funnel, and automatically start the delivery strategy retrospective analysis when the click-through rate of the advertising group is monitored to drop by more than 15%.

2. The intelligent mass advertising delivery method according to claim 1, characterized in that: The step S2 of constructing the user response prediction model includes: The user's historical behavior sequence is divided into short-term behavior segments and long-term behavior archives according to the time window. The attention mechanism is used to calculate the correlation weight between the current ad features and each historical behavior segment, and a dynamic weighting coefficient is assigned to the highly correlated segments. When the user is a newly registered user, a cross-advertiser transfer learning strategy is adopted to initialize the interest vector based on similar crowd portraits. The adversarial sample generator generates perturbation data by randomly masking the key fields of the user behavior sequence.

3. The intelligent mass advertising delivery method according to claim 2, characterized in that: When the model is trained: Full model training is performed during the low-traffic period in the early morning every day. The training data uses a sliding window to retain behavioral samples from the last 30 days. The incremental learning mode is started during the peak advertising period, and only the last layer of model parameters are updated using real-time streaming data. When the distribution change of advertising categories is monitored to exceed the preset variance threshold, the feature importance re-evaluation module is triggered to automatically adjust the weights of the input feature dimensions.

4. The intelligent mass advertising delivery method according to claim 1, characterized in that: The dynamic budget allocation optimization in step S3 includes: Advertisers are divided into two categories: brand exposure type and effect conversion type. The objective functions are set based on the cost per thousand impressions and the cost per conversion respectively. 80% of the budget weight is pre-allocated before the start of daily advertising bidding, and the remaining 20% ​​of the budget is dynamically adjusted based on the real-time bidding success rate. When an ad group fails to bid for five consecutive times, its budget weight will be automatically transferred to the substitute ad group of the same advertiser.

5. The intelligent mass advertising delivery method according to claim 4, characterized in that: When the budget weights are adjusted: For advertisers who focus on performance conversion, priority will be given to securing the traffic acquisition budget for users with the top 20% historical conversion rates. When the overall traffic utilization rate of the advertising platform is lower than 70%, the budget constraint upper limit for high-value advertisers will be relaxed. During the e-commerce promotion season, a budget overdraft mechanism will be enabled for participating ad groups, allowing single-day consumption to exceed 120% of the daily budget limit.

6. The intelligent mass advertising delivery method according to claim 1, characterized in that: The dynamic price adjustment in step S4 includes: The market competition intensity is divided into three levels: stable period, fluctuating period, and fierce battle period: During stable periods, a conservative bidding strategy is adopted, bidding at 50% of the estimated conversion value. During volatile periods, a game learning strategy is activated, recording competitor bidding patterns and dynamically matching them. During intense competition, a peak-grabbing strategy is activated, allowing temporary exceeding of the maximum bid limit set by advertisers. Resource expansion adopts a hierarchical response mechanism: Level 1 capacity expansion: Add 10% computing node replicas to cope with short-term traffic fluctuations; Secondary capacity expansion: Enable reserved bidding dedicated server clusters to cope with sustained high loads.

7. The intelligent mass advertising delivery method according to claim 6, characterized in that: When testing the bidding success rate: A separate success rate threshold is set for the first position in the information flow. When the click-through rate of the advertising material continues to decline, the backup creative material is automatically switched to participate in the bidding. For low-value traffic during the night period, the batch bidding package mode is enabled to reduce the cost of a single bid.

8. The intelligent mass advertising delivery method according to claim 1, characterized in that: The delivery effect tracking in step S5 includes: The exposure-conversion funnel is divided into four monitoring levels: exposure, visible exposure, click, and conversion. Abnormal fluctuation alarm rules are set for each level: Exposure level fluctuations exceeding ±30% of the daily average trigger a yellow alert, while a year-on-year drop in click-through rate exceeding 15% and lasting for two hours triggers a red alert. An attribution analysis report is automatically generated, marking the traffic channels, geographical distribution, and device types associated with the abnormal period.

9. The intelligent mass advertising delivery method according to claim 8, characterized in that: When back-analyzing the strategy: If the click-through rate drop is limited to a specific region, roll back the recent bid strategy changes for that region. If the click-through rate drops across multiple regions simultaneously, perform a creative fatigue check. Automatically mark materials that have been exposed more than one million times as "fatigue-marked" and start the A / B testing engine to verify the effect of new material combinations within 30 minutes. When backtracking points to model prediction deviation, freeze the current model and switch to the historical stable version.

10. The intelligent mass advertising delivery method according to claim 1, characterized in that: The online learning module includes: The real-time feedback data pipeline is divided into three processing paths. The click log stream is used to update the user's short-term interest vector, the conversion log stream is used to correct the conversion rate estimation deviation, and the negative sample stream is used to optimize the feature cross-weight. The negative sample stream is the exposure that was not clicked. The model version grayscale release mechanism means that the new model is first run in a 1% traffic environment for comparative testing, and then fully deployed when the key indicators are improved beyond the confidence interval.

Images