Flow filling rate gap complement prediction and correction method

Abstract

The invention provides a traffic filling rate gap position covering prediction and correction method, which comprises the following steps of: collecting historical and real-time traffic requests, bidding responses and actual filling results, eliminating abnormal data and constructing samples, defining and denoising filling rate gaps, extracting multi-dimensional features such as user groups, media bits, regions and time periods, and performing prediction and correction on the filling rate gaps. Establishing a tendency estimation model to output a complement source selection probability, realizing sample distribution balance through stabilizing inverse probability weight, performing consistency verification on processing identification, a probability range and a sample size, constructing a control group based on a tendency matching and weighting method, and calculating an anti-fact filling rate gap to generate a confidence interval; the method comprises the following steps of: obtaining a mapping function to form a complement source and anti-fact reference mapping table, estimating a causal gain by using a dual-robust learner, obtaining a gap closable degree through the mapping function, and screening out a reliable complement source. According to the method, dynamic prediction and intelligent correction of the gap can be realized in a complex putting environment, and the filling rate stability and the resource utilization efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of advertising placement and traffic control technology, and in particular to a method for predicting and correcting gaps in traffic fill rate. Background Technology

[0002] Existing methods for predicting fill rate gaps and selecting replacement sites typically rely on correlation indicators (such as historical average replacement effectiveness and click-through rate improvement) as the basis for decision-making. However, different replacement sites may cannibalize each other or have negative impacts. For example, a particular replacement site may superficially improve the fill rate, but the traffic it brings may lead to a decrease in overall revenue or conversion rate.

[0003] Existing technologies lack the ability to identify "causal contributions," making it difficult to distinguish between sources that genuinely contribute to filling gaps and those that merely exhibit correlation. While causal inference methods have been applied in medicine and recommender systems in recent years, they have not yet been applied to flow fill rate gap prediction. Therefore, we propose a method for predicting and correcting flow fill rate gaps.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a method for predicting and correcting flow fill rate gaps, thereby solving the technical problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting and correcting flow fill rate gaps includes the following steps: S1. Collect traffic requests, bidding responses and actual fill results, remove abnormal requests, define fill rate gaps and perform noise reduction, extract user segmentation, media position, region and time period features to form training sample set and validation sample set; S2. Statistically analyze the characteristics of the replacement source, establish a tendency estimation model to output the replacement probability, achieve distribution equilibrium based on the stabilized inverse probability weight, and verify the consistency of the processing identifier, probability range and sample size. S3. For the target filler source, select the control sample that did not use the target filler source, calculate the counterfactual fill rate gap by combining propensity matching and weighting methods, and ensure the robustness of the estimation by using confidence intervals and backoff strategies. S4. Estimate the causal gain of the filler source based on the dual robust learner, obtain the gap closure degree using the mapping function, calculate the confidence interval, and screen out reliable filler sources that pass the significance test. S5. Sort the candidate fill sources according to their closure priority, generate correction strategies for fill strength, fill timing and priority configuration, and perform linkage verification in conjunction with frequency control, budget and delay constraints, and gradually go online through the canary release mechanism. S6. Apply a correction strategy to the traffic request, calculate the corrected gap and compare it with the predicted value. If the difference exceeds the threshold, perform small-step feedback update. If a deviation occurs, trigger revaluation and rollback until the convergence condition is met and the closed loop is maintained.

[0007] S1 specifically refers to: Collect historical and real-time traffic requests, bidding responses, and actual fill results, remove HTTP errors and timeout exception requests, and ensure that the sample size is not less than a preset threshold; The fill rate gap is defined as the difference between the target fill rate and the actual fill rate, and noise is removed by filtering. Extract multi-dimensional features such as user segmentation, media location, region, time period, and network status, and encode and standardize them; The samples are divided into training set, validation set and deployment set based on a rolling time window to ensure data distribution consistency. For each request, a set of available replacement sources is registered, and unmatched requests are marked as empty sets to ensure consistency in subsequent calculations.

[0008] S2 specifically refers to: The success rate, backfilling latency, audience segmentation relevance, and revenue indicators of each replacement source are statistically analyzed, and the processing identifier of the request is defined. The model is trained to estimate the probability of selecting the complement source, and the model results are calibrated. The sample distribution is balanced based on the stabilized inverse probability weights, and the overlap constraint is set to prevent the weights from becoming extreme. The system performs consistency checks on the uniqueness of the identifier, the probability range, the average weight, and the total sample size. If the check fails, it performs a rollback or recalculation.

[0009] S3 specifically refers to: Select a set of control samples that do not use the target complement source, and ensure that the sample size meets the preset threshold. A control group was constructed in parallel using propensity matching and weighting methods, and the results of different paths were compared to select the one with smaller variance; the counterfactual fill rate gap was calculated and confidence intervals were generated to measure the performance of the gap under the condition of not using the target fill source; When the confidence interval is too wide or the sample is sparse, a backoff strategy is implemented at the regional, time period, or overall level, and the filler source, its counterfactual gap, and confidence interval are registered in the mapping table for subsequent calculations.

[0010] S4 specifically refers to: A dual robust learner is constructed to estimate the causal gain of the complement source, and a meta-learner is trained based on pseudo-labels to obtain stable estimation results; Based on hierarchical statistics, variance estimation of causal gain is performed, and the causal gain is converted into gap closure degree using a mapping function; confidence intervals are calculated using the variance propagation method, and a stability criterion is set to identify unreliable estimates. The closure property is tested for significance, and the complement sources that pass the significance test and meet the confidence interval requirements are selected. The causal gain, closure property and confidence interval are then bound together to generate a list of complement sources.

[0011] S5 specifically refers to: The selected gap filling sources are sorted according to the gap closureability to form a candidate set, and the gap filling strength, gap filling timing and priority configuration are set for the candidate gap filling sources. The correction strategy is linked with frequency control, budget, and delay constraints for verification. When the constraints are not met, the padding strength or candidate set is adjusted. A version signature is generated for the correction strategy that passes the verification and it is released in a gray-scale manner, gradually expanding the application scope according to a preset ratio. If the fill rate gap does not improve as expected during the grayscale process, roll back to the previous stable version and record the relevant parameter information.

[0012] S6 specifically refers to: The correction strategy is applied in real time during traffic requests, and the request processing latency is monitored. When the latency exceeds the preset threshold, the padding strength is automatically reduced and the anomaly is recorded. Calculate the corrected fill rate gap and compare it with the original gap. If the gap does not shrink or an anomaly occurs, trigger a warning. If the deviation between the corrected result and the predicted value exceeds a threshold, perform a small-step feedback update and adjust only the model calibration layer to ensure stability. A reassessment process is triggered when there are consecutive deviations or a decline in service level achievement rate. If the reassessment does not converge, it will revert to a stable version. When the preset convergence conditions are met, the current strategy is maintained, and the counterfactual benchmark and closure list are updated periodically to achieve closed-loop stability.

[0013] The beneficial effects of this invention are as follows: This invention establishes a propensity estimation model and combines it with a stable inverse probability weight to make the sample distribution of different feature layers tend to be balanced, effectively reducing estimation bias. It can accurately predict the gap level under large-scale heterogeneous flow and provide a reliable basis for subsequent strategy generation.

[0014] This invention introduces counterfactual benchmark construction and a dual robust causal learner, which can accurately measure causal gain by comparing "used" and "unused" complement sources. Through significance testing and confidence interval evaluation, the selected complement sources are guaranteed to be statistically reliable, thereby enhancing the interpretability and scientific rigor of strategy decisions.

[0015] This invention transforms causal gain into an intuitive indicator of fill-in effect through gap closure mapping, and generates fill-in strength, triggering timing, and priority configuration in real time based on this indicator. Under the constraint verification mechanism, the system can automatically adapt to fluctuations in different time periods, regions, or network environments, achieving dynamic optimization and precise deployment of strategies.

[0016] This invention designs a tiered anomaly handling process, including degradation triggered by latency monitoring, small-step feedback triggered by causal prediction deviation, and reassessment rollback triggered by SLA decline. When the model or strategy deviates from the normal trajectory, the system can quickly roll back to a stable version, ensuring the continuity and robustness of the ad delivery process.

[0017] This invention introduces frequency control limits, budget deviation thresholds, and latency restrictions during the correction strategy generation stage, and achieves unified adjustment through a shadow pricing mechanism. This not only ensures improved fill rate but also achieves a balance between revenue and cost, improving overall resource utilization efficiency and platform profitability.

[0018] This invention establishes a convergence criterion in the closed-loop feedback phase. Convergence is determined when the SLA achievement rate, over / underfill ratio, and gap indicators all continuously meet threshold conditions, and the strategy execution is maintained. Simultaneously, the counterfactual benchmark and closure list are automatically updated daily; if the indicators deviate again, a reassessment and rollback are triggered. This mechanism ensures that the system not only converges quickly in the short term but also maintains dynamic stability during long-term operation. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a flow fill rate gap prediction and correction method according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: As Figure 1 As shown, this embodiment provides a method for predicting and correcting flow fill rate gaps, including the following steps: S1. Collect traffic requests, bidding responses and actual fill results, remove abnormal requests, define fill rate gaps and perform noise reduction, extract user segmentation, media position, region and time period features to form training sample set and validation sample set; S2. Statistically analyze the characteristics of the replacement source, establish a tendency estimation model to output the replacement probability, achieve distribution equilibrium based on the stabilized inverse probability weight, and verify the consistency of the processing identifier, probability range and sample size. S3. For the target filler source, select the control sample that did not use the target filler source, calculate the counterfactual fill rate gap by combining propensity matching and weighting methods, and ensure the robustness of the estimation by using confidence intervals and backoff strategies. S4. Estimate the causal gain of the filler source based on the dual robust learner, obtain the gap closure degree using the mapping function, calculate the confidence interval, and screen out reliable filler sources that pass the significance test. S5. Sort the candidate fill sources according to their closure priority, generate correction strategies for fill strength, fill timing and priority configuration, and perform linkage verification in conjunction with frequency control, budget and delay constraints, and gradually go online through the canary release mechanism. S6. Apply a correction strategy to the traffic request, calculate the corrected gap and compare it with the predicted value. If the difference exceeds the threshold, perform small-step feedback update. If a deviation occurs, trigger revaluation and rollback until the convergence condition is met and the closed loop is maintained.

[0022] S1 specifically includes the following sub-steps: S110. Data Acquisition and Preprocessing: Acquire historical and real-time traffic request data, bidding response data, and actual fill result data, and label them with unique request identifiers (UUID) and timestamps to form the raw dataset; the target fill rate is set according to the dimension of "media position × country / region × time period (5-minute sliding window)," and the actual fill rate is statistically analyzed at the same granularity. When an abnormal response occurs (HTTP5xx or response latency greater than 1500ms), the request is discarded. If the percentage of abnormal data exceeds 10%, the sampling window needs to be reset to ensure that the total sample size is not less than 105 records, thus ensuring the effectiveness of the modeling statistics.

[0023] S120. Gap Definition and Denoising: The non-negative portion of the difference between the target fill rate and the actual fill rate is defined as the fill rate gap Δ, and the unit is specified as percentage points (pp). When the target fill rate is less than or equal to the actual fill rate, Δ is 0. The value of Δ is retained to two decimal places, and Hampel filtering (window size 7) is used to remove extreme values ​​from the time series of Δ. The reason for choosing a window size of 7 is that cross-validation showed that it is robust within the range of 5-9, and 7 provides the best performance.

[0024] S130. Feature Extraction: For each traffic request, extract request features and context features, including user segmentation, media location, region, time period, and network status; when necessary, it can also be extended to device type and operating system features; all categorical features are uniformly represented by one-hot encoding, and continuous features are represented by standardization or embedding to ensure the comparability of different features in the model.

[0025] S140. Sample Construction and Segmentation: The feature vector of each traffic request is associated with its corresponding fill rate gap Δ to form a gap sample set. The gap sample set is segmented using a rolling window method: T0-28d to T0-7d is the training set, T0-7d to T0-1d is the validation set, and T0 is the live set. Each rolling window lasts for one day to cover dynamic changes. The ratio of the training set to the validation set is 4:1, and stratified sampling is used to ensure a balanced distribution of samples across different regions, media segments, and user groups.

[0026] S150, Registration of Supplementary Source Set: For each traffic request in the gap sample set, register its corresponding supplementary source set S; the sources of the supplementary source set include demand-side platforms, direct-purchase advertisers, and proprietary backfill pools. When a traffic request fails to match any supplementary source in the corresponding time period, its supplementary source set is registered as an empty set, and a default identifier is marked in the sample to ensure that data inconsistency will not occur due to missing data during subsequent modeling.

[0027] Sample size (≥10) 5 (strip); feature expansion and encoding method; segmentation method (rolling window, stratified sampling); boundary conditions (default identifier for missing filler sources).

[0028] To avoid ambiguity in formula interpretation, this invention provides a unified definition for all involved variables as follows (units or dimensions are shown in the table): The sampling strategy details are as follows: the sliding window length is 28 days, the step size is 1 day; the sample segmentation ratio is 4:1:1 (training set / validation set / launch set); the stratified sampling dimension is "region × media position × user group"; the lower limit of a single stratum is 500 samples to ensure modeling stability.

[0029] S2 specifically includes the following sub-steps: S210, Filling Source Feature Statistics: For each filling source in the gap sample set, its filling performance within the historical 7-day rolling window is statistically analyzed to form a filling source feature vector; The characteristics of the replacement source should include at least the following: Replacement success rate: the proportion of the number of times a valid ad is successfully returned out of the total number of requests, calculated as a weighted average after daily statistics; Average replacement latency: in milliseconds (ms), calculated according to the 95th quantile; Audience segment association strength: calculated using the Pearson correlation coefficient, ranging from -1 to 1; Revenue-related indicators: such as eCPM and conversion rate, calculated as a weighted average based on the number of impressions; If the 7-day data is insufficient, the data will be rolled back to the 30-day window to ensure statistical robustness.

[0030] S220. Processing Identifier Definition: Each traffic request is labeled with a processing identifier T, and its value is determined according to the following rules: When the request selects only one supplementary source, T = the identifier of that supplementary source; when the request selects multiple supplementary sources in parallel, T = the identifier of the first supplementary source that returns a valid advertisement; if the first response source returns an invalid advertisement, it will be deferred to the next supplementary source that successfully returns an advertisement; when the request does not select any supplementary source, T = "not selected".

[0031] The above rules ensure that all samples are clearly defined in causal inference and cover the boundary cases of invalid responses.

[0032] S230. Propensity Model Training and Calibration: Under the combined influence of request features and filler source features, train a propensity estimation model to output an estimated probability e(X) of each filler source being selected. Model type: Logistic regression or Gradient Boosting Tree (XGBoost); Objective function: Cross-entropy loss; Feature selection: All features are screened using chi-square test or information gain to ensure significance p < 0.05; Training sample size: No less than 10 5 strip; Calibration method: Plattscaling, adjusting the output probability to the [0,1] interval; Performance lower limit: AUC on the validation set ≥ 0.70. If it is lower than this threshold, parameter tuning or feature expansion is required until the requirements are met.

[0033] S240, Stabilization Weights and Overlap Constraints: After obtaining the propensity score, the stabilization inverse probability weight (SIPW) is calculated for each sample: in This represents the weight of the i-th sample. It is used to readjust the importance of samples in weighted regression or causal estimation; T represents the treatment variable, such as whether a complement source is selected; T=1: the sample received a complement source; T=0: the sample did not receive the complement source; t is a specific treatment state (e.g., t=1 indicates use, t=0 indicates no use); P(T=t) represents the marginal probability of receiving treatment t at the population level. This is a constant reflecting the population proportion; Representing the features of a given sample When processing t, the propensity score is typically estimated by a logistic regression or machine learning model.

[0034] This weighting formula is used for sample reweighting to ensure that the selected and unselected samples are more balanced in feature distribution during subsequent counterfactual estimation and causal gain calculation. Its role is to eliminate the bias caused by the uneven distribution of features such as region, time period, and media position in the traffic samples; to ensure that the estimation of the effect of the supplementary source is close to the unbiased result of "random experiment"; and to improve the robustness of counterfactual gap calculation and causal gain estimation, avoiding overfitting or outliers dominating the results.

[0035] Suppose that the selection probability of a certain replacement source in the overall sample is P(T=1)=0.3, and for a certain sample feature... The model predicts its propensity score as =0.1; Ordinary IPW weight: 1 / 0.1=10, which is too large and can easily cause variance explosion. Stabilized SIPW weight: 0.3 / 0.1=3, which has a significantly convergent value and is more stable.

[0036] P(T=t) is estimated using empirical frequency and smoothed using Laplace (with an increment of 1) to avoid the denominator being zero; to prevent extreme weighting, an overlap constraint is set: when If the sample is not within the range of [0.05, 0.95], the sample should be truncated or removed. If the removal rate exceeds 5%, the sample window should be expanded (from 7 days to 30 days) or the feature dimension should be increased to restore the overlap.

[0037] S250. Consistency Verification: After writing the processing identifier and preference score, perform a consistency verification on the sample set with preference information, including: checking whether the processing identifier of each request is unique and matches the set of replacement sources; checking whether the preference score is within the range [0,1] and the sum of the probabilities of each replacement source is 1; checking whether the mean of the SIPW weights is close to 1 (allowing fluctuations not exceeding ±0.05); and checking whether the total number of samples is still greater than 10 after removal. 5 strip.

[0038] If the validation fails, the following rollback mechanism will be executed: Mild bias (e.g., weight mean bias ≤ 0.1): recalibrate the propensity score model; Moderate bias (e.g., exclusion ratio 5-10%): expand the statistical window or add feature dimensions and recalculate; Severe bias (e.g., total sample size < 10...). 5 ): Revert to the previous stable version of the dataset and log any anomalies. Only after verification can the output sample set with biased information proceed to the next step, S310–S350, for counterfactual benchmark construction.

[0039] S3 specifically includes the following sub-steps: S310. Control Sample Selection: For each target filler source s in the gap sample set, select traffic requests that did not use the filler source from the sample set with bias information to construct a control sample set.

[0040] The size of the control sample set must be ≥5000; If there are fewer than 5,000 records, the statistical window will be expanded sequentially (from 7 days to 30 days), or adjacent regions / time periods will be merged until the threshold is met; if there are still fewer, the data will be marked as sparse stratification and a fallback strategy will be implemented (see S340).

[0041] S320. Parallel Construction of Matching and Weighting: Constructing counterfactuals using a dual-path parallel method on the control sample set: Path 1: Nearest neighbor tendency matching; matching method: k=5 nearest neighbors, Mahalanobis distance metric; ensure that each treatment sample is paired with at least 5 control samples.

[0042] Path 2: SIPW weighting: using stabilized inverse probability weights calculated using S240 Weight the control group samples to ensure that the overall distribution is comparable to that of the treatment group.

[0043] If the difference in counterfactual gap estimates between the two paths is greater than 1 percentage point (pp), the path with smaller variance is selected as the preferred result, and the other result is retained for cross-validation.

[0044] S330, Counterfactual Benchmark Estimation: On the preferred set of controls, calculate the counterfactual fill rate gap under the condition of not using the target filler source s. : in This indicates a counterfactual gap, which is "the difference between the system fill rate and the target fill rate when a certain fill source is not used"; This represents the actual fill rate. This represents the expected filling rate when the treatment variable T=0 (i.e., no filler source is selected) given sample features X; it is a counterfactual estimate, usually obtained through propensity modeling, matching, or weighting methods; it expresses "what the system would behave if a certain filler source were not used"; This indicates the system's preset target fill rate (e.g., 95%). This is the performance metric the platform aims to achieve.

[0045] The fill rate is estimated based on the propensity score model when no fill source is used. Then, compare it with the target fill rate. Compare the two; the difference between them is the counterfactual gap. That is, if This indicates that the gap still exists without a replacement source and requires further correction; if This indicates that when no filler source is used, the fill rate already meets or exceeds the target, and the gap problem is not significant.

[0046] The above formula forms the core of the counterfactual baseline construction, used to determine the potential severity of the gap. Combined with the effect estimation of the filler source, it can form a "filler source-counterfactual gap mapping table." Through this baseline, subsequent calculations of causal gain and closure are more valuable, avoiding distorted decisions based solely on observations. Bootstrap resampling (B=1000, B=500 when sample size > 106) is used for calculation. The 95% confidence interval [L,U] is calculated and rounded to two decimal places.

[0047] S340, Robustness and Sparsity Backoff: If If the confidence interval width (UL) is greater than 2pp, it is marked as an unstable estimate. When an unstable estimate occurs or S310 determines that the stratification is sparse, the following rollback is performed in sequence: remove the regional dimension and retain only the media bit × time period; then remove the time period dimension and retain only the media bit; if it is still unstable, rollback to the overall control baseline; during rollback, the complement strength parameter u is reduced by 50% of the original value to avoid overfitting. To ensure strategy stability, this invention sets an upper limit on the daily rollback trigger frequency. Once this limit is exceeded, subsequent triggers will be automatically postponed to the next day and recorded in the exception log, effectively avoiding strategy oscillations caused by frequent rollbacks and improving the engineering stability of the correction process.

[0048] S350, Mapping Registration: Register each filler source s with its... The confidence interval [L, U] and hierarchical information are bound together to generate a "complementary source-counterfactual benchmark mapping table". The mapping table is stored in the policy database and supports daily updates; if... If the deviation from the historical average by ±2σ is reached for 3 consecutive days, counterfactual baseline reconstruction is triggered; the mapping table supports backtracking queries for the most recent 30 days for policy drift monitoring.

[0049] S4 specifically includes the following sub-steps: S410. Causal learner construction: Based on the "filler source-counterfact benchmark" mapping table, a dual robust learner (DR-Learner) is used to estimate the causal gain value g of each filler source on the fill rate gap Δ.

[0050] Phase 1: Fitting the conditional expectation functions for the treatment group and the control group respectively. , and tendency score .

[0051] Phase Two: Constructing Pseudo-Labels in This represents the pseudo-label value of the i-th sample. As a processing identifier variable for the sample, This is the observed fill rate result for this sample. In sample features The tendency score for choosing a replacement source under certain conditions. and These represent the expected values ​​of the filling rate prediction under the conditions of using and not using the filler source, respectively. The above formula integrates the weighted residual terms of the treatment group and the control group, as well as the prediction difference term. Through dual correction by the propensity score model and the outcome regression model, it achieves a dual robust estimate of the causal gain at the sample level. When either the propensity model or the outcome model can be accurately fitted, the overall estimate remains consistent, thereby effectively reducing the estimation error caused by sample distribution bias.

[0052] Based on the pseudo-label calculation results, the system can further utilize a meta-learner to model the causal effect of the gap filling source, obtaining the gap improvement contribution value under different flow characteristics. This method can achieve stable causal gain estimation in heterogeneous flow environments, ensuring the scientific validity and reliability of the gap filling strategy generation.

[0053] Meta-learner training: with To train LightGBM with supervised signals, the parameters are set as follows: maximum depth ≤ 8, learning rate = 0.05, minimum number of samples per leaf node ≥ 50; convergence and repeatability: random seed is set to 2025, mean squared error on the validation set is ≤ 0.05; if not satisfied, retrain until convergence.

[0054] S420, Stratified Statistics and Variance Estimation: Perform stratified statistics on the causal gain value g according to the requested feature and context feature, and output the mean, variance and sample size of each stratum; Sample lower limit: The sample size of a single stratum must be ≥1000, otherwise merge the samples; Merging rule order: merge adjacent regions first → merge time periods → merge media positions last to ensure distribution balance; Variance estimation: Bayesian robust estimation is adopted, and the prior distribution is set to inverse gamma (α=3, β=2); Result precision: All statistical results are retained to two decimal places.

[0055] S430. Closure Mapping and Confidence Interval: The causal gain value g is normalized to the gap closure degree C, calculated using the following formula: Where C represents the gap closure degree, which is used to characterize the ability of a certain filling source to compensate for the current filling rate gap, and the value ranges from 0 to 1; The Sigmoid mapping function is defined as follows: , used to normalize the input value to the interval (0,1); g is the causal gain calculated by the causal learning model, representing the actual improvement of the fill rate gap by the fill source; This is the smoothing coefficient, used to control the steepness of the Sigmoid function's mapping. Its value ranges from [1.5, 2.5] percentage points (pp), with a default value of 2pp.

[0056] This function mapping allows causal gains of different magnitudes to be uniformly converted into comparable standardized closure indices. When the causal gain g is large, the Sigmoid function output approaches 1, indicating that the filler source has high closure potential; when the causal gain is small or negative, the output value approaches 0, indicating that the filler source has limited effect on gap correction. This can be achieved by adjusting the smoothing parameters. It can adaptively adjust the mapping sensitivity according to traffic fluctuations, thereby achieving dynamic balance in the strategy generation process.

[0057] This calculation method can effectively solve the problem of inconsistent dimensions of causal effects among different complement sources, enabling the evaluation of complement effects to be carried out on a unified scale, thereby improving the stability and interpretability of complement source selection and strategy generation.

[0058] Variance propagation: Variance is propagated using the Delta method to obtain confidence intervals. Stability criterion: If the interval width If the result is not found, it is determined to be an unstable estimate and the process enters the S340 sparse rollback procedure.

[0059] S440. Significance and Reliability Screening: The complement source is screened based on the closure confidence interval and statistical test. Judgment Criterion: Only when... The replacement source is included in the candidate set only when the two-tailed t-test result of the causal gain g is greater than 0.1 and the result satisfies p < 0.05; Replacement sources that fail to meet the conditions are marked as "unreliable" and excluded in subsequent policy generation; if the candidate set is empty, it degenerates into greedy sorting based on eCPM as a fallback solution.

[0060] S450, Result Binding and Inventory Generation: This involves binding the filtered filler sources with their causal gain value g, gap closure degree C, and interval. Binding is performed to generate a "Complementary Source - Closure List". Storage mechanism: The list is stored in the policy database, and the update frequency is consistent with the counterfactual baseline (once a day); Retrospective capability: Supports querying the list changes of the most recent 30 days for monitoring drift; Drift Trigger: If the C of a certain filler source decreases by more than 20% for three consecutive days, the re-evaluation mechanism is triggered, and S410–S430 is re-executed; if the re-evaluation still fails to converge, it will immediately revert to the list of the previous stable version and issue an alarm.

[0061] S5 specifically includes the following sub-steps: S510, Selection of Replacement Sources and Formation of Candidate Strategies: Based on the "Completion Source-Closeability List", select the replacement source with the highest closureability C value and that passes the significance test as the primary selection target, and sort the second-best replacement sources according to their C values ​​to form a candidate set.

[0062] Number limit: The candidate set shall retain a maximum of Top-3 to avoid overly complex strategies; A fallback mechanism: If the candidate set is empty, it degenerates into a backfilling strategy based on eCPM greedy sorting; Additional information: Each complement source in the candidate set must carry a causal gain value g and a closability interval. And hierarchical identification to ensure subsequent traceability.

[0063] S520, Gap Correction Strategy Parameterization: For each gap filler source in the candidate set, generate a corresponding gap correction strategy, including: Filling strength in The complement strength parameter is used to control the proportional weight allocated to the complement source during the strategy execution phase; C is the gap closure index calculated based on causal gain, and its value ranges from 0 to 1. This is the amplification factor, used to adjust the response sensitivity of the filler intensity; This is the upper limit of the fill strength, used to limit the magnitude of strategy adjustments and prevent excessive corrections from causing system oscillations.

[0064] When the closure degree C is low, the system proportionally reduces the padding strength to avoid overreacting to noisy samples; when the closure degree C is high, the system proportionally increases the padding strength, but not exceeding the maximum value. This ensures the controllability and stability of the replacement and adjustment process.

[0065] In this embodiment, A value of 0.6 indicates that a single padding source can account for a maximum of 60% of the total requests. A value of 1.2 is used to slightly amplify the response amplitude under high closability conditions, thereby improving the overall fill rate correction effect. Through this function constraint mechanism, the system can adaptively adjust the filling strength under different flow scenarios, achieving refined and robust gap correction.

[0066] If C ≥ 0.8, then u = 0.6 is directly taken; if C ≤ 0.2, then u = 0.2 is adjusted down. Filling timing control: Trigger threshold: Triggered when Δ ≥ 2pp; Cooldown time: The interval between two triggers of the same filling source is ≥ 3 minutes; Peak suppression: When Δ ≥ 5pp, only the filling source with the highest C value is enabled, and other candidate sources are frozen for 10 minutes.

[0067] Filling priority configuration: sort by C value as priority; when the difference between the C values ​​of two filling sources is <0.05, the filling source with lower average delay is selected first.

[0068] S530, Constraint Linkage and Feasibility Domain Check: The generated gap correction strategy needs to be linked and verified with three types of constraints: frequency control, budget, and delay. Frequency control constraint: The number of exposures for the same user in 24 hours ≤ 10; if the prediction exceeds the limit, the fill strength u will be reduced to 50% of the original value. Budget constraint: The daily budget deviation must not exceed ±3%; if it exceeds the limit, the overspending fill source will be frozen to the next cycle. Delay constraint: Request-response latency ≤ 800ms; if it exceeds the limit, the priority will be reduced by 1 level, and it will be marked as "high latency".

[0069] Unified shadow price Write the budget and delay constraints into the Lagrange dual function, if If the value is greater than θ (θ=100), the overall strategy is downgraded, and only the Top-1 replacement source is retained.

[0070] S540, Strategy Binding and Gray-Scale Preparation: The corrected strategy, verified by constraints, needs to be aligned with the target fill rate r, the actual fill rate r, and the predicted closure degree. The three elements are bound together to generate the corresponding policy signature hash. The canary release process uses four phases: 10% → 30% → 50% → 100%, each phase lasting 30 minutes. Rollback condition: If Δ decreases by less than 0.5pp within a phase, the process rolls back to the previous phase. Consistency guarantee: The next phase can only proceed when the hash comparison passes and the version numbers are consecutive.

[0071] S550, Distribution and Execution Ready: The bound gap correction strategy is written into the strategy distribution module and pushed to the real-time engine.

[0072] Metadata records include: strategy version number, generation time, candidate set, fill strength parameter, priority sequence, and constraint verification results; Consistency check: If metadata is missing or hash does not match, the distribution will be blocked and the process will be returned to S510 for recalculation; Anomaly Handling: If the number of distribution failures is ≥3 times / hour, an alarm will be triggered and the system will automatically revert to the most recent stable strategy.

[0073] S6 specifically includes the following sub-steps: S610 Real-time Correction Execution: Apply the gap correction strategy that has passed grayscale verification to real-time traffic requests, and execute the gap filling operation one by one according to the gap filling strength u, gap filling timing control and gap filling priority configuration.

[0074] Latency monitoring: Request processing latency is monitored in real time in milliseconds (ms); Degradation conditions: If a single delay is greater than 1000ms, or three consecutive delays are greater than 800ms, the degradation mechanism will be triggered: the padding strength u will be reduced by 30%, and the abnormal event will be recorded. Exception logging: All degradation trigger events are written to the log as input for subsequent feedback updates.

[0075] S620. Calculation and Comparison of Execution Results: Collect the fill results after the correction execution and calculate the actual fill rate gap. and the gap before correction In comparison, the reduction in output gap .

[0076] Update frequency: Results are updated every 5 minutes. Numerical precision: Round to two decimal places; Judgment criteria: If If the gap is ≤0.5pp, it is considered to be effectively closed; if If this occurs, an abnormal warning will be triggered immediately, and the data will be pushed to the monitoring system.

[0077] S630, Small Step Feedback and Calibration Update: [This will...] Compare with the predicted closure C, if the difference is... This will trigger a small-step feedback update: Triggering condition: Triggered when the difference between three consecutive 5-minute windows is greater than the threshold; Update method: Only adjust the calibration layer parameters of the bias model and the causal learner, with a learning rate η=0.05; Scope limitation: The update only applies to the last two layers to avoid catastrophic drift; Warm start strategy: If the difference is only a one-time noise, no update will be triggered.

[0078] S640. Reassessment Trigger and Rollback: The policy reassessment process is triggered if any of the following conditions are met: The average difference between the measured C and the predicted C was >1.5pp for three consecutive windows; 95th percentile gap Two consecutive windows exceeded the historical average +2σ; SLA achievement rate has been below 90% for three consecutive windows.

[0079] Post-trigger processing: Re-enter S410–S450 to calculate causal gain and closure degree, and rebuild and correct the policy in S510–S550; if convergence is still not achieved after two re-evaluations, revert to the most recent stable version of the policy and issue an alarm.

[0080] S650. Convergence Criterion and Loop Maintenance: The system is considered convergent and stable when the following conditions are met: SLA achievement rate ≥ 95% within six consecutive windows (30 minutes); overfill / underfill ratio < 0.1 within six consecutive windows; actual... The 95th percentile value is ≤2pp.

[0081] After convergence: maintain the current strategy; update the "supplementary source-counterfactual benchmark" and "supplementary source-closure list" once a day; if the subsequent monitoring indicators deviate from the convergence conditions again, automatically return to S640 to trigger reassessment and rollback.

[0082] To avoid frequent system revaluations due to short-term fluctuations, a daily limit of 3 revaluation triggers is set, with any exceeding this limit being processed later. All trigger events will be logged for subsequent strategy stability analysis and model adaptive adjustment; ensuring the system is dynamic but does not fall into an oscillating state, thus maintaining controllability and stability in high-frequency bidding scenarios.

[0083] Example 2: This example provides a flow fill rate gap prediction and correction system, including: The data acquisition module is used to collect historical and real-time traffic request data, bidding response data, and actual fill result data. Abnormal requests are removed, and the difference between the target fill rate and the actual fill rate is defined as the fill rate gap. The gap is then denoised using filtering. At the same time, multi-dimensional features such as user segmentation, media position, region, time period, and network status are extracted. The samples are divided into training set, validation set, and online set according to the rolling time window, and the corresponding fill source set is registered. The tendency estimation and sample balancing module is used to statistically analyze the success rate, backfilling delay, revenue indicators, and audience segmentation association characteristics of each replacement source. Based on the request characteristics and replacement source characteristics, it trains a tendency estimation model to output the replacement source selection probability and achieves sample distribution balancing based on stable inverse probability weights. This module is also used to perform consistency checks and rollback processing on the uniqueness of the processing identifier, probability range, weight mean, and sample size. The counterfactual gap estimation module is used to select a set of control samples that do not use the target filler source, construct a control group in parallel using propensity matching and weighting methods, calculate the counterfactual filling rate gap and confidence interval under the condition of not using the filler source, and execute a regional or time-based backoff strategy when the sample is insufficient or the estimation is unstable. The causal gain calculation module is used to construct a pseudo-label training meta-learner based on a dual robust learner on the basis of the counterfactual gap, estimate the causal gain of the filler source in the gap dimension, and use the Sigmoid mapping function to convert the causal gain into gap closure degree, calculate the confidence interval by combining the variance propagation method, screen reliable filler sources according to the significance test and stability criterion, and generate a "filler source-closure degree list". The strategy generation module is used to sort candidate filler sources by their closure degree, calculate filler strength, trigger timing and priority configuration, and perform linkage verification with constraints such as frequency control, budget, and latency; when the shadow price item exceeds the set threshold, the strategy is downgraded to retain only the top-1 filler sources; a gray release plan is generated for the verified strategies, and a strategy signature is formed for version management. The closed-loop execution and feedback module is used to apply correction strategies in real time during traffic requests, monitor request processing latency and trigger automatic degradation, collect the actual fill rate gap after correction execution and compare it with the predicted results, execute small-step feedback updates when the deviation exceeds a preset threshold, trigger reassessment and rollback when there is continuous deviation, maintain the execution of the current strategy when the convergence criterion is met, and periodically update the counterfactual benchmark and the list of fill sources to achieve closed-loop stability.

[0084] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0085] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0086] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0087] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0088] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0089] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0090] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0091] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0092] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0093] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting and correcting gaps in flow rate filling rate, characterized in that, Includes the following steps: S1. Collect traffic requests, bidding responses and actual fill results, remove abnormal requests, define fill rate gaps and perform noise reduction, extract user segmentation, media position, region and time period features to form training sample set and validation sample set; S2. Statistically analyze the characteristics of the replacement source, establish a tendency estimation model to output the replacement probability, achieve distribution equilibrium based on the stabilized inverse probability weight, and verify the consistency of the processing identifier, probability range and sample size. S3. For the target filler source, select the control sample that did not use the target filler source, calculate the counterfactual fill rate gap by combining propensity matching and weighting methods, and ensure the robustness of the estimation by using confidence intervals and backoff strategies. S4. Estimate the causal gain of the filler source based on the dual robust learner, obtain the gap closure degree using the mapping function, calculate the confidence interval, and screen out reliable filler sources that pass the significance test. S5. Sort the candidate fill sources according to their closure priority, generate correction strategies for fill strength, fill timing and priority configuration, and perform linkage verification in conjunction with frequency control, budget and delay constraints, and gradually go online through a canary release mechanism.

2. The method for predicting and correcting flow fill rate gaps according to claim 1, characterized in that, It also includes S6, applying a correction strategy in traffic requests, calculating the corrected gap and comparing it with the predicted value. If the difference exceeds the threshold, a small-step feedback update is performed. If a deviation occurs, a revaluation and rollback are triggered until the convergence condition is met and the closed-loop stability is maintained.

3. The method for predicting and correcting flow fill rate gaps according to claim 1, characterized in that, S1 specifically refers to: Collect historical and real-time traffic requests, bidding responses, and actual fill results, remove HTTP errors and timeout exception requests, and ensure that the sample size is not less than a preset threshold; The fill rate gap is defined as the difference between the target fill rate and the actual fill rate, and noise is removed by filtering. Extract multi-dimensional features such as user segmentation, media location, region, time period, and network status, and encode and standardize them; The samples are divided into training set, validation set and deployment set based on a rolling time window to ensure data distribution consistency. For each request, a set of available replacement sources is registered, and unmatched requests are marked as empty sets to ensure consistency in subsequent calculations.

4. The method for predicting and correcting flow fill rate gaps according to claim 1, characterized in that, S2 specifically refers to: The success rate, backfilling latency, audience segmentation relevance, and revenue indicators of each replacement source are statistically analyzed, and the processing identifier of the request is defined. The model is trained to estimate the probability of selecting the complement source, and the model results are calibrated. The sample distribution is balanced based on the stabilized inverse probability weights, and the overlap constraint is set to prevent the weights from becoming extreme. The system performs consistency checks on the uniqueness of the identifier, the probability range, the average weight, and the total sample size. If the check fails, it performs a rollback or recalculation.

5. The method for predicting and correcting flow fill rate gaps according to claim 1, characterized in that, S3 specifically refers to: Select a set of control samples that do not use the target complement source, and ensure that the sample size meets the preset threshold. A control group was constructed in parallel using propensity score matching and weighting methods, and the results of different paths were compared to select the one with smaller variance. Calculate the counterfactual fill rate gap and generate confidence intervals to measure the gap performance under conditions where no target fill source is used; When the confidence interval is too wide or the sample is sparse, a backoff strategy is implemented at the regional, time period, or overall level, and the filler source, its counterfactual gap, and confidence interval are registered in the mapping table for subsequent calculations.

6. The method for predicting and correcting flow fill rate gaps according to claim 5, characterized in that, S4 specifically refers to: A dual robust learner is constructed to estimate the causal gain of the complement source, and a meta-learner is trained based on pseudo-labels to obtain stable estimation results; Based on hierarchical statistics, variance estimation of causal gain is performed, and the causal gain is converted into gap closure degree using a mapping function; Confidence intervals are calculated using the variance propagation method, and a stability criterion is set to identify unreliable estimates; The closure property is tested for significance, and the complement sources that pass the significance test and meet the confidence interval requirements are selected. The causal gain, closure property and confidence interval are then bound together to generate a list of complement sources.

7. The method for predicting and correcting flow fill rate gaps according to claim 6, characterized in that, S5 specifically refers to: The selected gap filling sources are sorted according to the gap closureability to form a candidate set, and the gap filling strength, gap filling timing and priority configuration are set for the candidate gap filling sources. The correction strategy is linked with frequency control, budget and delay constraints for verification. When the constraints are not met, the fill strength or candidate set is adjusted. Generate version signatures for the verified correction strategies and conduct canary releases, gradually expanding the application scope according to a preset ratio; If the fill rate gap does not improve as expected during the grayscale process, roll back to the previous stable version and record the relevant parameter information.

8. The method for predicting and correcting flow fill rate gaps according to claim 2, characterized in that, S6 specifically refers to: The correction strategy is applied in real time during traffic requests, and the request processing latency is monitored. When the latency exceeds the preset threshold, the padding strength is automatically reduced and the anomaly is recorded. Calculate the corrected fill rate gap and compare it with the original gap. Trigger a warning if the gap does not shrink or an anomaly occurs. If the deviation between the corrected result and the predicted value exceeds a threshold, a small-step feedback update is performed, adjusting only the model calibration layer to ensure stability. A reassessment process is triggered when there are consecutive deviations or a decline in service level achievement rate. If the reassessment does not converge, it will revert to a stable version. When the preset convergence conditions are met, the current strategy is maintained, and the counterfactual benchmark and closure list are updated periodically to achieve closed-loop stability.

Images