Advertisement request scheduling method based on real-time resource state constraint

Abstract

The application discloses an advertisement request scheduling method based on real-time resource state constraints, comprising: obtaining a target advertisement exposure request received by an advertisement request server, and obtaining request context information corresponding to the target advertisement exposure request; obtaining real-time state information corresponding to a plurality of advertisement demand platforms; generating a plurality of candidate advertisement traffic request scheduling actions according to the request context information and the real-time state information; according to a preset scheduling rule and a real-time resource constraint condition, constraint evaluation is performed on each candidate advertisement traffic request scheduling action; a target advertisement traffic request scheduling action is determined according to a constraint evaluation result, and the advertisement request server is controlled to send an advertisement request to a corresponding advertisement demand platform according to the target advertisement traffic request scheduling action; an advertisement request processing result is obtained, and advertisement demand platform state information and scheduling constraint parameters are updated according to the advertisement request processing result. The application can realize reasonable allocation of advertisement traffic resources.

Description

Technical Field

[0001] This application relates to the field of internet advertising technology, and in particular to an advertising request scheduling method applied in a programmatic advertising trading system. Specifically, this invention relates to a method for dynamically adjusting the advertising request sending objects and request parameters based on real-time acquisition of advertising demand platform status information by an advertising request server and in conjunction with resource constraints. Background Technology

[0002] With the development of internet applications and mobile terminal technologies, programmatic advertising trading systems have become an important technological foundation for internet advertising. In the process of programmatic advertising trading, when a user access request and an advertising exposure opportunity are generated on the media side, the ad request server usually needs to send ad requests to one or more Demand Side Platforms (DSPs) to obtain the ad bidding results and ultimately determine the display ads.

[0003] In existing advertising transaction systems, advertising request scheduling usually adopts a fixed rule approach, such as: (1) selecting the sending object according to the historical revenue ranking of the advertising demand platform; (2) setting a fixed priority according to the historical response speed of the advertising demand platform; and (3) performing simple filtering according to budget or frequency control rules.

[0004] However, existing technologies have at least the following problems: existing scheduling methods usually make decisions based on a single indicator, failing to comprehensively consider the budget status of advertising demand platforms, request processing capacity, link load, and traffic allocation; existing solutions typically only intercept requests after resource constraints are reached, resulting in a large number of invalid advertising requests being sent, consuming network communication resources and server processing resources; when multiple advertising demand platforms participate in advertising transactions simultaneously, some advertising demand platforms may occupy a high proportion of traffic for extended periods, while other advertising demand platforms cannot obtain reasonable request opportunities; existing technologies lack a mechanism to continuously adjust scheduling parameters based on advertising request processing results, causing the system to be unable to adapt to the real-time changing advertising transaction environment.

[0005] Therefore, a scheduling method is needed that can dynamically adjust the advertising request sending strategy according to the real-time status of the advertising exchange system in order to improve the efficiency of advertising request processing and reduce system resource consumption. Summary of the Invention

[0006] The purpose of this application is to provide an advertising request scheduling method based on real-time resource status constraints. By obtaining advertising request context information and real-time status information of the advertising demand platform, candidate advertising request scheduling actions are generated, and target scheduling actions are dynamically selected based on constraints such as budget, frequency control, link resources, and traffic allocation. This improves the request processing efficiency in the advertising transaction system and achieves reasonable allocation of advertising traffic resources.

[0007] In a first aspect of this application, an advertising request scheduling method based on real-time resource state constraints is provided, applied to an advertising transaction system including an advertising request server and multiple advertising demand platforms, the method comprising: The system obtains the target ad exposure request received by the ad request server and obtains the request context information corresponding to the target ad exposure request. The request context data includes at least media information, ad placement information, device information, geographic information, user characteristic information, and traffic quality information. Obtain real-time status information corresponding to multiple advertising demand platforms. The real-time status information includes at least advertising budget status, request response status, request processing load status, and historical traffic allocation status. Based on the request context information and the real-time status information, multiple candidate advertising traffic request scheduling actions are generated; Based on the preset scheduling rules and real-time resource constraints, the scheduling actions of each candidate advertising traffic request are evaluated and constrained. Based on the constraint evaluation results, the target ad traffic request scheduling action is determined, and the ad request server is controlled to send ad requests to the corresponding ad demand platform according to the target ad traffic request scheduling action. Obtain the advertising request processing result, and update the advertising demand platform status information and scheduling constraint parameters based on the advertising request processing result.

[0008] In a preferred embodiment, the candidate ad request scheduling action includes at least one of the following: Send an advertising request to the designated advertising demand platform; Send an advertising request carrying the first request price parameter to the designated advertising demand platform; Send an advertising request carrying a second request price parameter to the designated advertising demand platform; Send an advertising request carrying a traffic package identifier to the designated advertising demand platform; Execute a request suppression action, wherein the request suppression action is executed when the predicted revenue is lower than a preset threshold or the resource constraint pressure exceeds a preset threshold.

[0009] In a preferred embodiment, the real-time status information of the advertising demand platform is periodically acquired through the status acquisition module, including: remaining budget information of the advertising demand platform, advertising request response time information, advertising request success rate information, current traffic occupancy ratio of the advertising demand platform, and request processing capacity information of the advertising demand platform.

[0010] In a preferred embodiment, the constraint assessment includes budget resource constraint assessment; determining the budget constraint status of the corresponding advertising demand platform based on the deviation between the current budget consumption rate of the advertising demand platform and the preset budget consumption target; and reducing the priority of the corresponding advertising demand platform being selected to send advertising requests when the budget constraint status reaches a preset threshold.

[0011] In a preferred embodiment, the constraint assessment includes a user access frequency constraint assessment; determining the frequency control state based on the number of times the user corresponding to the target ad exposure request displays an ad within a preset time window; and restricting the execution of the corresponding ad request scheduling action when the number of ad displays reaches a preset frequency threshold.

[0012] In a preferred embodiment, the constraint assessment includes request link resource constraint assessment; determining advertising request sending restrictions based on the advertising demand platform response latency, request timeout time, and the number of concurrent requests; and adjusting the advertising request sending ratio when an increase in the request processing pressure of the advertising demand platform is detected.

[0013] In a preferred embodiment, the method further includes: determining a traffic allocation deviation state based on the difference between the actual proportion of advertising traffic obtained by multiple advertising demand platforms and a preset target traffic proportion; adjusting the advertising request scheduling action according to the traffic allocation deviation state; wherein, when an advertising demand platform obtains less traffic than the target traffic proportion, the scheduling priority of sending advertising requests to that advertising demand platform is increased.

[0014] In a preferred embodiment, the advertising request processing result includes at least: whether the advertising demand platform responds, whether it returns the bidding result, whether the advertising request is successful, the advertising display result, the user interaction result, updates based on the advertising request processing result, advertising demand platform status information, resource constraint status information, and advertising request scheduling parameters.

[0015] In a second aspect of this application, a non-transitory computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, which, when executed by a processor, implement the steps of the aforementioned method.

[0016] In a third aspect of this application, a computer program product is provided, wherein computer program product stores computer-executable instructions that, when executed by a processor, implement the steps in the aforementioned method.

[0017] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. It should be understood that the accompanying drawings described below are merely some implementation examples of the present invention, and those skilled in the art can obtain other implementation examples based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating an advertising request scheduling method based on real-time resource state constraints according to one embodiment of this application. Detailed Implementation

[0020] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] One embodiment of this application relates to an advertising request scheduling method based on real-time resource state constraints, applied to an advertising transaction system including an advertising request server and multiple advertising demand platforms, the process of which is as follows: Figure 1 As shown, the method includes steps 1 to 6.

[0023] Step 1: Obtain the target ad exposure request received by the ad request server, and obtain the corresponding request context information. The request context data includes at least media information, ad placement information, device information, geographic information, user characteristic information, and traffic quality information.

[0024] Step 2: Obtain real-time status information for multiple advertising demand platforms. Real-time status information includes at least advertising budget status, request response status, request processing load status, and historical traffic allocation status. The real-time status information of the advertising demand platforms is periodically acquired through the status collection module, including: remaining budget information, advertising request response time information, advertising request success rate information, current traffic usage ratio of the advertising demand platform, and request processing capacity information of the advertising demand platform.

[0025] Step 3: Based on the request context information and real-time status information, generate multiple candidate ad traffic request scheduling actions. The candidate ad request scheduling actions include at least one of the following: sending an ad request to a specified ad demand platform; sending an ad request carrying a first request price parameter to a specified ad demand platform; sending an ad request carrying a second request price parameter to a specified ad demand platform; sending an ad request carrying a traffic package identifier to a specified ad demand platform; or executing a request suppression action. Specifically, the request suppression action is executed when the predicted revenue is lower than a preset threshold or the resource constraint pressure exceeds a preset threshold.

[0026] In this embodiment, the request suppression action does not signify abandonment of the advertising business strategy, but rather a control operation on the resource status of the advertising request chain. When the advertising request server detects that there are currently no advertising demand platforms that meet the resource constraints, sending advertising requests to multiple advertising demand platforms will increase the request processing overhead of the advertising request server, increase the number of network communications, and increase the pressure on the advertising demand platforms to handle invalid requests. Therefore, this invention, by executing the request suppression action, causes the advertising request server to terminate the invalid request chain during the request sending phase. This reduces the number of invalid communications in the advertising transaction system and improves server resource utilization efficiency.

[0027] In this embodiment, the traffic allocation deviation state is determined based on the difference between the actual proportion of advertising traffic obtained by multiple advertising demand platforms and the preset target proportion of traffic. The advertising request scheduling action is then adjusted according to the traffic allocation deviation state. Specifically, when the traffic obtained by an advertising demand platform is lower than the target proportion, the scheduling priority of sending advertising requests to that platform is increased.

[0028] In this implementation, traffic fairness constraints are used to avoid imbalances in the distribution of request resources among multiple advertising demand platforms due to long-term fixed scheduling strategies. Specifically, the advertising request server periodically calculates the actual proportion of advertising requests received by each advertising demand platform and compares it with a preset target traffic proportion. When an advertising demand platform is detected to have a consistently lower traffic proportion than the target, the execution opportunity of its corresponding candidate scheduling action is increased. When an advertising demand platform is detected to have a consistently higher traffic proportion than the target, the execution opportunity of its corresponding candidate scheduling action is decreased. Through this method, advertising request resources can be dynamically adjusted according to the status of multiple advertising demand platforms, improving the overall resource utilization balance of the advertising transaction system.

[0029] In this embodiment, the candidate ad traffic request scheduling action is not only used to select an ad demand platform, but also to describe the request processing path that the ad request server can execute for the current ad exposure request.

[0030] Specifically, each candidate ad traffic request scheduling action includes at least: the request sending target, i.e., the target ad demand platform; the request configuration parameters, i.e., the parameter information carried by the ad request; and the request execution method, i.e., sending the request or suppressing the request. The request sending target is used to determine the ad demand platform to which the ad request server will send the current ad impression request; the request configuration parameters are used to adjust the processing method of the ad request in the ad transaction chain; and the request execution method is used to determine whether the current ad impression request enters the bidding process of the ad demand platform.

[0031] By generating multiple candidate ad traffic request scheduling actions, the ad request server can select different request processing paths based on the current ad exchange system status, rather than using a fixed ad demand platform selection method.

[0032] Step 4: Based on the preset scheduling rules and real-time resource constraints, perform constraint evaluation on the scheduling actions of each candidate ad traffic request. The constraint evaluation includes budget resource constraint evaluation, user access frequency constraint evaluation, and request link resource constraint evaluation.

[0033] In the budget resource constraint assessment, the budget constraint status of the corresponding advertising demand platform is determined based on the deviation between the current budget consumption rate and the preset budget consumption target. When the budget constraint status reaches a preset threshold, the priority of the corresponding advertising demand platform in sending ad requests is reduced. In the user access frequency constraint assessment, the frequency control status is determined based on the number of ad impressions for the user corresponding to the target ad exposure request within a preset time window. When the number of ad impressions reaches a preset frequency threshold, the execution of the corresponding ad request scheduling action is restricted. In the request link resource constraint assessment, ad request sending restrictions are determined based on the ad demand platform's response latency, request timeout time, and number of concurrent requests. When an increase in the request processing pressure of the ad demand platform is detected, the ad request sending ratio is adjusted.

[0034] In this embodiment, the constraint evaluation is not used to rank advertising performance metrics, but rather to determine whether candidate advertising traffic request scheduling actions meet the current resource status requirements of the advertising exchange system. Specifically, the advertising request server checks the resource status of the corresponding advertising demand platform for each candidate advertising traffic request scheduling action. When the advertising demand platform corresponding to a candidate scheduling action has resource limitations, the execution conditions of that candidate scheduling action are adjusted. Specifically: when the budget consumption rate of the advertising demand platform exceeds a preset range, the proportion of continuing to send advertising requests to that advertising demand platform is reduced; when the request response time of the advertising demand platform increases or the request processing load increases, the frequency of sending requests to that advertising demand platform is reduced; when the number of times a user's advertisement is displayed reaches a frequency control threshold, the execution of the corresponding advertising request scheduling action for that user is restricted; when the historical traffic allocation ratio of the advertising demand platform deviates from the target ratio, traffic balancing adjustments are made to the corresponding advertising request scheduling action.

[0035] By employing the above methods, the ad request server can consider the system resource status in advance during the ad request sending stage, thus avoiding passive restrictions after the request enters the subsequent processing link.

[0036] Step 5: Determine the target ad traffic request scheduling action based on the constraint evaluation results, and control the ad request server to send ad requests to the corresponding ad demand platform according to the target ad traffic request scheduling action.

[0037] Step 6: Obtain the ad request processing result and update the ad demand platform status information and scheduling constraint parameters based on the ad request processing result. The ad request processing result includes at least: whether the ad demand platform responded, whether it returned the bidding result, whether the ad request was successful, the ad display result, the user interaction result, the status information of the ad demand platform updated based on the ad request processing result, the resource constraint status information, and the ad request scheduling parameters.

[0038] In this implementation, the request response status is updated based on the advertising demand platform's response result; the request link status is updated based on the advertising request processing time; the traffic allocation status is updated based on the advertising request sending ratio; and the resource constraint status is updated based on the advertising request execution result. The updated status information is used for generating candidate scheduling actions and evaluating constraints for the next advertising exposure request. Through this feedback and update process, the advertising request scheduling method forms a real-time status feedback closed loop, improving the system's adaptability to changes in the advertising transaction environment.

[0039] Compared with existing fixed-rule scheduling methods, this invention: comprehensively considers the real-time status of advertising demand platforms before sending ad requests, reducing the sending of invalid requests to platforms with low processing capabilities; incorporates budget resources, user frequency control resources, request link resources, and traffic allocation status into a unified scheduling process, improving ad request processing efficiency; reduces invalid network communication and server resource consumption through a request suppression mechanism; enables the ad request scheduling process to adapt to the real-time changing ad trading environment through a feedback update mechanism; and improves the overall resource utilization efficiency of the ad trading system through dynamic traffic adjustment among multiple ad demand platforms.

[0040] To better understand the technical solution of this application, a specific example is provided below. The details listed in this example are mainly for ease of understanding and are not intended to limit the scope of protection of this application.

[0041] This invention aims to model DSP selection, base price / guide price, and traffic package selection as a unified online decision-making problem. Under constraints such as budget, frequency control, and geographic / target audience, it achieves a balance between exploration and utilization through contextual multi-armed slot machine to maximize long-term transaction revenue and effective conversion. At the system level, it ensures stable budgeting, compliant frequency control, fair ecosystem coverage, and controllable launch risks. Implementation embodiments of this invention can employ the following engineering constraints and implementation details, specifically: Candidate set truncation: During the candidate generation stage, candidates are first filtered according to compliance and rules, and then truncated according to historical availability / response rate / budget status, so that each score is calculated only on a limited set of DSPs, price tiers and traffic packages.

[0042] Score calculation can be linearized: prioritize the use of fast-computable linear / generalized linear predictors or lightweight models; complex models can be distilled offline into lightweight scorers or called in a hierarchical / segmented manner to avoid high-overhead inference per request.

[0043] Matrix operations can be incrementally scaled: Matrix inversion operations involved in LinUCB (Linear Context Bandit) can be incrementally updated or periodically batch recalculated using the Sherman-Morrison formula, Cholesky decomposition, or by request-by-request inversion, avoiding the unacceptable overhead caused by inversion calculations.

[0044] Online updates can be batched: Under high QPS, mini-batch updates or asynchronous updates can be used to decouple model parameter updates from request paths to ensure latency and stability.

[0045] Logs and Privacy: The identifiers required for frequency control and attribution can be hashed / anonymized; log fields are collected minimally with access control and auditing to reduce compliance risks.

[0046] Execution Example (Use Case) The following example, using a single ad impression request, illustrates how this invention forms a complete closed loop from input data, candidate generation, model prediction, constraint scoring, action selection, request dispatch to feedback updates. The following values ​​are for illustrative purposes only and do not represent actual business performance data.

[0047] Scene Input Suppose that the media receives an ad impression request at a certain time, and this request corresponds to a mobile app in-feed ad slot. The system receives the following input information.

[0048] 1. Exposure request context feature x_t, including: Media type: Mobile app; Ad placement type: In-feed ad placement; Device type: Android; Region: The area where a specific target is deployed; Time period: Evening rush hour; User identifier: Anonymized or hashed user identifier; User segmentation: Internal system aggregation and segmentation, such as "highly active users" and "users with recent click behavior"; Traffic quality score: It is obtained by combining historical clicks, dwell time, conversions, and invalid traffic risk.

[0049] 2. DSP Status Characteristics: Assume the currently available DSPs include DSP_A, DSP_B, and DSP_C. The system reads the budget, consumption rate, response status, and historical performance of each DSP in real time. DSP_A: Remaining budget is sufficient, recent win-rate is moderate, and conversion quality is high; DSP_B: Remaining budget is tight, recent win-rate is high, but frequency control is nearing its limit; DSP_C: The remaining budget is sufficient. The recent win-rate is low, but the historical conversion rate of this type of traffic package is good.

[0050] 3. Constraint State Budget constraints: Each DSP must not exceed its budget consumption target in the current hour, current day, or current campaign period; Frequency capping constraint: The number of times the same anonymous user displays the same advertiser or campaign within a time window W shall not exceed F_max; QPS / Concurrency Constraint: Each DSP has a request rate limit within the current time window; Fair coverage constraint: The system aims to prevent certain DSPs from being unable to obtain their target traffic packets for an extended period, while also preventing a single DSP from consuming an excessively high proportion of premium traffic over a long period.

[0051] Budget, frequency control, QPS / quota, and DSP coverage fairness are all abstracted into a unified multi-dimensional resource consumption model. This model incorporates constraints endogenously into the action scoring function through shadow pricing, rather than handling them with hard truncation at the end. Shadow pricing can be interpreted as the "price of resource scarcity." The more scarce the resource, the higher its shadow price, and the more the system tends to choose actions that conserve that resource; when resources are plentiful, the system is more proactive in exploring and raising prices. Shadow pricing transforms budget constraints into opportunity costs, reducing cost fluctuations from "premature exhaustion / near-deadline rushes," and improving budget achievement rate and cost smoothness.

[0052] Candidate action generation The system first generates a set of candidate actions based on compliance rules, geographic targeting, audience targeting, blacklists and whitelists, budget availability, and whether frequency control has reached its limit.

[0053] Suppose that the following actions occur after candidate generation: Action a1=(DSP_A,price_level_1,package_news_high); Action a2 = (DSP_A, price_level_2, package_news_high); Action a3 = (DSP_B, price_level_1, package_news_high); Action a4 = (DSP_C, price_level_1, package_news_high); The action a5=NULL indicates that no request is sent, i.e., a request suppression action.

[0054] Where: price_level_1 represents a lower base price / guide price; price_level_2 represents a higher base price / guide price; package_news_high represents the high-quality traffic package of the information flow to which the current exposure belongs; the NULL action is used to suppress invalid sending when the expected revenue is low, frequency control is tight, budget is tight, or QPS is tight.

[0055] Model prediction For each candidate action, the system calls the prediction module and outputs the following predicted value: win_prob_hat(a,x): The probability that the DSP will bid and win the bid under this action; price_hat(a,x): The expected transaction price under the winning bid condition; conv_prob_hat(a,x): The probability that this action will result in a valid conversion; c_hat_budget(a,x): Budget spending forecast; c_hat_freq(a,x): Frequency control consumption prediction; c_hat_qps(a,x): QPS or link cost prediction; c_hat_fair(a,x): Fair coverage deviates from prediction.

[0056] Example prediction results are shown in Table 1 below. Table 1 The values ​​in the table are examples of the process. Budget consumption can be approximated as "estimated probability of winning the bid multiplied by estimated transaction price"; frequency control consumption can be approximated as "estimated probability of winning the bid multiplied by display count increment"; a positive fair deviation indicates that the DSP is currently receiving relatively excessive traffic, and continuing to select it will increase the deviation; a negative fair deviation indicates that the DSP is currently below the target coverage, and selecting this DSP can improve coverage balance.

[0057] Reward Construction The system constructs the expected return of candidate actions based on business objectives. If the objective is "revenue + conversion value", then the following can be defined: revenue_hat(a,x) = win_prob_hat(a,x) multiplied by price_hat(a,x); conv_value_hat(a,x) = win_prob_hat(a,x) multiplied by conv_prob_hat(a,x) and then multiplied by value_per_conv; r_hat(a,x) = revenue_hat(a,x) + beta multiplied by conv_value_hat(a,x).

[0058] If the goal is Return On Ad Spend (ROAS), it can be expressed as a difference: roas_reward_hat(a,x) = Estimated Conversion Value - rho multiplied by Estimated Ad Spend. Here, rho represents the spend penalty coefficient. The larger the rho, the more the system favors actions with low cost and high conversion value.

[0059] If the target is cost per conversion (CPA), then the formula can be: cpa_reward_hat(a,x) = estimated number of conversions - gamma multiplied by estimated ad spend. Here, gamma represents the penalty intensity on spending. In this example, the system uses a combined return of "revenue + conversion value" and treats budget, frequency control, QPS, and fair coverage as constraints.

[0060] Constraint opportunity cost calculation The system maintains a dual variable lambda for each resource dimension to represent the current stress level of that resource. Example: lambda_budget: budget stress level; lambda_freq: frequency control stress level; lambda_qps: QPS / link resource stress level; lambda_fair: fairness coverage deviation penalty intensity.

[0061] Assuming that within the current time window: DSP_B's budget is being consumed rapidly, therefore lambda_budget is higher for DSP_B; the current user frequency capping for the activity corresponding to DSP_B is close to the upper limit, therefore lambda_freq is higher for DSP_B; DSP_C is currently below the target traffic coverage, therefore c_hat_fair is negative, and choosing DSP_C can reduce the fairness deviation; the overall QPS is acceptable, therefore lambda_qps is low.

[0062] The system calculates a comprehensive score for each action: score(a,x) = r_hat(a,x) - lambda_budget multiplied by c_hat_budget(a,x) - lambda_freq multiplied by c_hat_freq(a,x) - lambda_qps multiplied by c_hat_qps(a,x) - lambda_fair multiplied by c_hat_fair(a,x). This score reflects that: the higher the expected benefit of the action, the higher the score; the greater the resource consumption such as budget, frequency control, and QPS, the lower the score; if a DSP is currently below the fair coverage target, the fairness item can improve its score.

[0063] Action selection Assuming the calculation yields the comprehensive score shown in Table 2 below: Table 2 Although DSP_B had a higher initial projected benefit, its overall score was significantly lowered due to its rapid budget depletion, frequency capping nearing its limit, and high fairness deviation. DSP_C did not have the highest initial benefit, but it received the highest overall score due to its ample budget, lower frequency control pressure, and current coverage share below the target.

[0064] Therefore, the system selects: a = (DSP_C, price_level_1, package_news_high), which means sending a request to DSP_C, carrying the corresponding base price / guide price and traffic package identifier.

[0065] Request dispatch and logging The system sends an ad request to DSP_C, carrying the following information: exposure request context; package_news_high traffic package identifier; base price or guide price corresponding to price_level_1; necessary compliance, region, audience, and media placement fields. Simultaneously, the system logs: context feature x_t; selected action a_t; candidate action set A(x_t); the propensity of the selected action, i.e., the probability of historical strategy selection of this action, used for importance weighting and counterfactual estimation; constraint status such as budget, frequency control, QPS, and fair coverage; whether a DSP response has been received; whether a bid has been placed; whether the bid has been won; the transaction price; whether a timeout has occurred; and subsequent conversion feedback or quality feedback.

[0066] Propensity is crucial for offline counterfactual evaluation. If IPS / DR is needed to evaluate a new strategy, the probability of the old strategy choosing that action must be known. IPS (Inverse Propensity Scoring): An importance sampling value estimation method. It weights historical samples according to the ratio of "probability of choosing the new strategy / probability of choosing the old strategy." When the probability of the old strategy is very small, it can lead to high variance; in engineering, it is often combined with weight trimming or SNIPS stabilization. SNIPS (Self-Normalized IPS): A stabilization method that normalizes IPS weights, typically using the formula "sum of weighted rewards / sum of weights." DR (Doubly Robust): A doubly robust estimation method that integrates the reward prediction model and the IPS residual correction term; when either the propensity or the reward model is correct, the estimation remains consistent, usually more stable than pure IPS.

[0067] Feedback Update If DSP_C successfully returns a bid and ultimately wins the bid, the system receives feedback on the transaction price; if subsequent clicks, conversions, or value feedback from the advertiser occur, the system receives feedback on delayed conversions.

[0068] The system performs the following updates based on feedback: 1. Update the reward model: update win_prob_hat; update price_hat; update conv_prob_hat; apply attribution window or time decay processing to delayed conversions; 2. Update Bandit parameters: update the parameters corresponding to the selected action online or in batches; if tiered Bandit is used, update the DSP layer parameters and price / traffic package layer parameters simultaneously; 3. Update dual variables: if a DSP's current budget consumption exceeds the target pace, increase lambda_budget; if a user or user group is close to the frequency capping limit, increase lambda_freq; if a DSP consistently falls below the target coverage share, reduce its fair penalty or provide fair compensation; if QPS or timeout rate increases, increase lambda_qps; 4. Update audit reports: record why DSP_C was selected in this decision; display the revenue forecast, constraint cost, and comprehensive score of each candidate action; support review by operations, business, or technical personnel.

[0069] Offline assessment and online access control When the system is ready to launch a new policy, instead of directly replacing the old policy entirely, it uses historical logs to perform an offline counter-policy evaluation (OPE).

[0070] 1. IPS evaluation calculates the following for each historical sample: Weight = Probability of the new strategy choosing this action / Probability of the old strategy choosing this action; The estimated return of the new strategy = weights multiplied by the actual observed return; If some weights are too large, then weight pruning or SNIPS can be used to reduce variance.

[0071] 2. The DR evaluation first uses a reward model to predict the returns under the new strategy, and then uses IPS to correct the model error: DR = Model Predicted Return + Weights Multiplied by (Actual Return - Model Predicted Return of Old Actions); This method is generally more stable than pure IPS when at least one of the reward model or propensity is more accurate.

[0072] 3. The rollout of new access control strategies will only be permitted gradually if the following conditions are met: The projected returns will be no less than those of the old strategy. Frequency capping does not increase the risk of violations; Budget pacing bias does not worsen; QPS / timeout rate does not worsen; DSP coverage fairness metrics did not deteriorate.

[0073] This execution example demonstrates the differences between the present invention and traditional rule-based hierarchical schemes: 1. Traditional rules may only select DSP_B based on short-term benefits or win-rate; this invention will take into account budget, frequency capping, fair coverage and QPS cost, so it will not mechanically select the DSP with the highest short-term benefits.

[0074] 2. Traditional rules typically intercept requests at the end of the process after budget or frequency control has been triggered; this invention incorporates the opportunity cost of constraints in advance during the action scoring stage, which can reduce invalid requests and post-event interception.

[0075] 3. Traditional rules are difficult to explain "why a certain DSP is selected or not selected"; this invention can output the estimated benefits, resource consumption, dual variables and comprehensive scores of candidate actions, forming an auditable explanation chain.

[0076] 4. Traditional deployment strategies mainly rely on small-scale trial and error; this invention uses propensity logs and IPS / DR for offline counterfactual evaluation, which can reduce deployment risks.

[0077] 5. For advertiser objectives such as ROAS / CPA, this invention can be uniformly incorporated into the contextual bandit decision framework through reward function reconstruction, without the need to design a completely independent rule system for each objective.

[0078] This invention offers the following technical effects and advantages: By exploring / utilizing balance to continuously discover high-value combinations (DSP-traffic packages-price tiers) ignored by rules, it improves overall eCPM, revenue, and effective conversions; frequency control resource consumption is explicitly included in costs during the decision-making stage, reducing invalid requests and invalid bids caused by end-point hard truncation, and increasing the effective revenue per request; it outputs the target share, actual share, deviation, and corresponding shadow price for each DSP, forming an auditable and interpretable chain, avoiding long-term squeezing of tail DSPs and improving ecosystem stability; layered bandit and constraint-based intrinsic scoring can reduce invalid sending and high timeout requests, lower QPS pressure and link costs, and improve overall system stability and availability; this invention can achieve learning and optimization without relying on plaintext personal identifiers: the log layer only stores aggregated features required by the business, anonymized or hashed identifiers, and decision-related statistics (such as propensity, transaction / timeout / conversion feedback, etc.). It can also implement minimal collection, access control, and audit records for sensitive fields, thereby reducing data exposure and improving compliance traceability.

[0079] In summary, this invention targets programmatic advertising scenarios, unifying the decisions regarding "which DSP to send a request to, what floor price / guide price to set, and which traffic package to select" into an online decision-making problem. It achieves an adaptive balance between exploration and exploitation through contextual multi-armed slot machines, and intrinsically integrates multi-dimensional constraints such as budget, frequency control, QPS / quota, and ecosystem fairness coverage into action scoring to maximize long-term returns and effective conversions, while improving budget rhythm stability, frequency control compliance, and interpretability and auditability. This method supports layered bandits to reduce action space sparsity and cold start risks; in engineering, it can be extended to various dispatch forms such as "selecting one DSP at a time," "selecting multiple DSPs for concurrent sending," or "selecting empty actions to suppress sending," and establishes an online gate through OPE to reduce strategy iteration fluctuations.

[0080] Accordingly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the various method embodiments of this application. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media do not include transient computer-readable media, such as modulated data signals and carrier waves.

[0081] Furthermore, embodiments of this application also provide an advertising request scheduling system based on real-time resource state constraints, including a memory for storing computer-executable instructions and a processor; the processor is used to implement the steps in the above-described method embodiments when executing the computer-executable instructions in the memory. The processor may be a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Digital Signal Processor (DSP), Microcontroller Unit (MCU), Neural Processing Unit (NPU), Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), or other programmable logic devices. The aforementioned memory may be read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or solid-state drive, etc. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0082] Furthermore, embodiments of this application also provide a computer program product, including computer-executable instructions that, when executed by a processor, implement the steps in the above-described method embodiments.

[0083] The various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which should be considered as having been recorded in this specification), unless such a combination of technical features is technically infeasible. For example, in one example, feature A+B+C is disclosed, and in another example, feature A+B+D+E is disclosed. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; it is impossible to use both simultaneously. Feature E can be technically combined with feature C. Therefore, the solution A+B+C+D should not be considered as having been recorded because it is technically infeasible, while the solution A+B+C+E should be considered as having been recorded.

[0084] All references to this specification are considered to be incorporated integrally into the disclosure of this application so that they can serve as the basis for modifications if necessary. Furthermore, it should be understood that the above descriptions are merely preferred embodiments of this specification and are not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of one or more embodiments of this specification.

[0085] In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An advertising request scheduling method based on real-time resource state constraints, characterized in that, The method, applied to an advertising transaction system that includes an ad request server and multiple ad demand platforms, comprises: The system obtains the target ad exposure request received by the ad request server and obtains the request context information corresponding to the target ad exposure request. The request context data includes at least media information, ad placement information, device information, geographic information, user characteristic information, and traffic quality information. Obtain real-time status information corresponding to multiple advertising demand platforms. The real-time status information includes at least advertising budget status, request response status, request processing load status, and historical traffic allocation status. Based on the request context information and the real-time status information, multiple candidate advertising traffic request scheduling actions are generated; Based on the preset scheduling rules and real-time resource constraints, the scheduling actions of each candidate advertising traffic request are evaluated and constrained. Based on the constraint evaluation results, the target advertising traffic request scheduling action is determined, and the advertising request server is controlled to send advertising requests to the corresponding advertising demand platform according to the target advertising traffic request scheduling action. Obtain the advertising request processing result, and update the advertising demand platform status information and scheduling constraint parameters based on the advertising request processing result.

2. The advertising request scheduling method according to claim 1, characterized in that, The candidate ad request scheduling action includes at least one of the following: Send an advertising request to the designated advertising demand platform; Send an advertising request carrying the first request price parameter to the designated advertising demand platform; Send an advertising request carrying a second request price parameter to the designated advertising demand platform; Send an advertising request carrying a traffic package identifier to the designated advertising demand platform; Execute a request suppression action, wherein the request suppression action is executed when the predicted revenue is lower than a preset threshold or the resource constraint pressure exceeds a preset threshold.

3. The advertising request scheduling method according to claim 1, characterized in that, The real-time status information of the advertising demand platform is acquired periodically through the status acquisition module, including: remaining budget information of the advertising demand platform, advertising request response time information, advertising request success rate information, current traffic usage ratio of the advertising demand platform, and request processing capacity information of the advertising demand platform.

4. The advertising request scheduling method according to claim 1, characterized in that, The constraint assessment includes budgetary resource constraint assessment; Based on the deviation between the current budget consumption rate and the preset budget consumption target of the advertising demand platform, determine the budget constraint status of the corresponding advertising demand platform; When the budget constraint reaches a preset threshold, the priority of the corresponding advertising demand platform for sending advertising requests is reduced.

5. The advertising request scheduling method according to claim 1, characterized in that, The constraint assessment includes user access frequency constraint assessment; determining the frequency control status based on the number of times the user corresponding to the target ad exposure request displays ads within a preset time window; and restricting the execution of the corresponding ad request scheduling action when the number of ad displays reaches a preset frequency threshold.

6. The advertising request scheduling method according to claim 1, characterized in that, The constraint assessment includes request link resource constraint assessment; determining advertising request sending restrictions based on advertising demand platform response latency, request timeout time, and the number of concurrent requests; and adjusting the advertising request sending ratio when increased request processing pressure is detected on the advertising demand platform.

7. The advertising request scheduling method according to claim 1, characterized in that, Also includes: Based on the difference between the actual proportion of advertising traffic obtained by multiple advertising demand platforms and the preset target proportion of traffic, the traffic allocation deviation status is determined; The advertising request scheduling action is adjusted according to the traffic allocation deviation status; wherein, when the traffic obtained by an advertising demand platform is lower than the target traffic ratio, the scheduling priority of sending advertising requests to that advertising demand platform is increased.

8. The advertising request scheduling method according to claim 1, characterized in that, The advertising request processing result includes at least: whether the advertising demand platform responded, whether the bidding result was returned, whether the advertising request was successful, the advertising display result, the user interaction result, updates based on the advertising request processing result, the advertising demand platform status information, the resource constraint status information, and the advertising request scheduling parameters.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer-executable instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 8.

10. A computer program product comprising computer-executable instructions, characterized in that, When executed by a processor, the computer-executable instructions implement the steps of the method according to any one of claims 1 to 8.

Images