Flow filling rate gap filling prediction and correction method

Abstract

The application provides a flow filling rate gap filling prediction and correction method, comprising the following steps: collecting historical and real-time flow 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 clustering, media bit, region and time period, establishing a tendency estimation model to output a filling source selection probability, realizing sample distribution balance through a stable inverse probability weight, performing consistency verification on processing identification, probability range and sample size, constructing a control group based on tendency matching and weighting method, calculating counterfactual filling rate gaps to generate a confidence interval, forming a filling source and counterfactual benchmark mapping table, estimating causal gain by using a double robust learner, obtaining gap closability through a mapping function, and screening out reliable filling sources. The method can realize dynamic prediction and intelligent correction of gaps in a complex delivery environment, and improve the stability of filling rate and resource utilization efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of advertisement delivery and traffic regulation, and particularly relates to a traffic filling rate gap filling prediction and correction method. BACKGROUND

[0002] In the existing filling rate gap prediction and filling selection method, the correlation index (such as historical average filling effect, click rate improvement) is usually taken as the basis for decision-making. However, there may be mutual cannibalization or negative influence between different filling sources. For example, a certain filling source may have improved the filling rate on the surface, but the traffic introduced by it may have caused the overall revenue or conversion rate to decline.

[0003] The prior art lacks identification of "causal contribution", and it is difficult to distinguish between filling sources that truly help fill the gap and sources that only show correlation. Causal inference methods have been applied in medicine and recommendation systems in recent years, but have not been applied in the field of traffic filling gap prediction. Therefore, we propose a traffic filling rate gap filling prediction and correction method.

[0004] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to overcome the shortcomings of the prior art and provide a traffic filling rate gap filling prediction and correction method to solve the technical problems mentioned in the background.

[0006] To achieve the above purpose, the present application provides the following technical solutions:

[0007] A traffic filling rate gap filling prediction and correction method, comprising the following steps:

[0008] S1, collect traffic requests, bidding responses and actual filling results, eliminate abnormal requests, define filling rate gaps and perform denoising processing, extract user clustering, media bit, region and time period features, form a training sample set and a verification sample set;

[0009] S2, statistics of filling source features are performed, a propensity estimation model is established to output filling probability, distribution balance is realized based on stable inverse probability weight, and consistency verification is performed on processing identifier, probability range and sample size;

[0010] S3, for a target filling source, select a control sample that does not use the target filling source, calculate the counterfactual filling rate gap by combining propensity matching and weighting method, and ensure the estimation robustness through confidence interval and rollback strategy;

[0011] S4, estimating the fill-in source causal gain based on the dual-robust learner, obtaining the gap closable degree by using a mapping function, and calculating a confidence interval to screen out reliable fill-in sources with significant significance;

[0012] S5, ranking the candidate fill-in sources according to the closable degree priority, generating a correction strategy for fill-in strength, fill-in timing and priority configuration, and combining frequency control, budget and delay constraints for linkage verification, and gradually putting into operation through the gray release mechanism;

[0013] S6, applying the correction strategy in the traffic request, calculating the corrected gap and comparing it with the predicted value, if the difference exceeds the threshold, performing small-step feedback update, if deviation occurs, triggering re-estimation and rollback, until the convergence condition is met and the closed loop is stable.

[0014] S1 is specifically:

[0015] Collect historical and real-time traffic requests, bidding responses and actual filling results, eliminate HTTP errors, timeout exception requests and ensure that the sample size is not less than the preset threshold;

[0016] Define the filling rate gap as the difference between the target filling rate and the actual filling rate, and perform denoising processing through filtering;

[0017] Extract multi-dimensional features such as user clustering, media position, region, time period and network status, and perform encoding and standardization;

[0018] Based on the rolling time window, the samples are divided into training set, validation set and online set to ensure data distribution consistency;

[0019] And register the available fill-in source set for each request, and mark the empty set for unmatched requests to ensure consistency in subsequent calculations.

[0020] S2 is specifically:

[0021] Statistical features of success rate, backfill delay, audience cluster correlation and revenue indicators for each fill-in source, define the processing identifier of the request;

[0022] Train the propensity estimation model to output the selection probability of the fill-in source, and calibrate the model results;

[0023] Based on the stable inverse probability weight, the sample distribution is balanced, and the overlap constraint is set to prevent weight extreme;

[0024] Conduct consistency check on the uniqueness of the processing identifier, the probability range, the weight mean and the total sample size, and perform rollback or recalculation if the check fails.

[0025] S3 is specifically:

[0026] The control sample set not using the target source is selected, and the sample size is ensured to meet a preset threshold;

[0027] The control group is constructed by parallelizing the propensity matching and the weighting method, and the variance of different path result selection is compared to obtain a smaller one; the counterfactual filling rate gap is calculated and a confidence interval is generated to measure the performance of the gap under the condition of not using the target source;

[0028] When the confidence interval is too wide or the sample is sparse, a fallback strategy is performed at the regional, time period or overall level, and the target source and its counterfactual gap and confidence interval are registered in a mapping table for subsequent calculation.

[0029] S4 specifically is:

[0030] A double-robust learner is constructed to estimate the causal gain of the source, and a meta-learner is trained based on pseudo-labels to obtain stable estimation results;

[0031] On the basis of hierarchical statistics, the variance of the causal gain is estimated, and the mapping function is used to convert the causal gain into the closable degree of the gap; the confidence interval is calculated by the variance propagation method, and the stability criterion is set to identify unreliable estimates;

[0032] The closable degree is subjected to a significance test, and the target sources that pass the significance test and meet the requirements of the confidence interval are selected, and the causal gain, closable degree and confidence interval are bound to generate a list of target sources.

[0033] S5 specifically is:

[0034] The target sources selected are sorted according to the closable degree of the gap to form a candidate set, and the target sources are configured with a source strength, a source timing and a priority;

[0035] The correction strategy is linked and verified with frequency control, budget and delay constraints, and when the constraints are not met, the source strength or the candidate set is adjusted; the version signature of the correction strategy that passes the verification is generated and is released in a gray scale, and the application range is gradually expanded according to a preset proportion;

[0036] If the filling rate gap does not reach the expected improvement during the gray scale process, the system is rolled back to the last stable version, and the relevant parameter information is recorded.

[0037] S6 specifically is:

[0038] The correction strategy is applied in real time in the traffic request, and the request processing delay is monitored, and when the delay exceeds a preset threshold, the source strength is automatically reduced and the exception is recorded;

[0039] The corrected filling rate gap is calculated and compared with the one before correction, and a warning is triggered when the gap is not reduced or an anomaly occurs; small step feedback update is performed when the deviation between the correction result and the predicted value exceeds the threshold value, and only the model calibration layer is adjusted to ensure stability;

[0040] When there is a continuous deviation or the service level achievement rate decreases, a re-estimation process is triggered, and if the re-estimation does not converge, it is rolled back to the stable version;

[0041] When the preset convergence condition is met, the current strategy is maintained for execution, and the counterfactual benchmark and closable degree list are periodically updated to achieve closed-loop stability.

[0042] The beneficial effects of the present application are:

[0043] The present application establishes a tendency estimation model and combines a stabilized inverse probability weight to make the sample distribution of different characteristics stratified and balanced, effectively reducing the estimation bias, and accurately predicting the gap level under large-scale heterogeneous traffic, providing a reliable basis for subsequent strategy generation.

[0044] The present application introduces a counterfactual benchmark construction and a double robust causal learner, which can accurately measure causal gains under the control of "used" and "unused" fill-in sources. Through significance test and confidence interval evaluation, it ensures that the selected fill-in sources are statistically reliable, thereby enhancing the interpretability and scientificity of strategy decision-making.

[0045] The present application maps the closable degree of the gap to convert the causal gain into an intuitive fill-in effect indicator, and generates fill-in strength, trigger timing and priority configuration in real time according to the indicator. Under the constraint checking mechanism, the system can automatically adapt to fluctuations in different time periods, regions or network environments, and realize dynamic optimization and precise delivery of strategies.

[0046] The present application designs a hierarchical abnormality processing procedure, including degradation triggered by latency monitoring, small step feedback triggered by causal prediction deviation, and re-estimation rollback triggered by SLA decrease. When the model or strategy deviates from the normal trajectory, the system can quickly roll back to the stable version, ensuring the continuity and robustness of the advertisement delivery process.

[0047] The present application introduces a frequency control upper limit, a budget deviation threshold and a latency limit in the strategy correction generation stage, and realizes unified adjustment through a shadow price mechanism. Therefore, not only the filling rate is improved, but also a balance between revenue and cost is achieved, improving the overall resource utilization efficiency and platform revenue capacity.

[0048] The application sets a convergence criterion in the closed-loop feedback stage. When the SLA achievement rate, over / under filling ratio and gap index all continuously meet the threshold condition, it is determined that the convergence is achieved, and the strategy execution is maintained. At the same time, the counterfactual benchmark and the closable degree list are automatically updated daily. If the index deviates again, the re-estimation rollback is triggered. This mechanism ensures that the system not only converges quickly in the short term, but also maintains dynamic stability in the long-term operation. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 A flow filling rate gap filling prediction and correction method is shown in the figure. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0051] Embodiment one: as shown, the embodiment provides a flow filling rate gap filling prediction and correction method, including the following steps: Figure 1

[0052] S1, collect flow requests, bidding responses and actual filling results, eliminate abnormal requests, define filling rate gaps and perform denoising processing, extract user clustering, media positions, regions and time period features, form training sample sets and verification sample sets;

[0053] S2, statistics of the gap filling source features are performed, a propensity estimation model is established to output the gap filling probability, distribution balance is realized based on the stabilized inverse probability weight, and consistency verification is performed on the processing identifier, probability range and sample size;

[0054] S3, for the target gap filling source, select the control sample without using the target gap filling source, calculate the counterfactual filling rate gap by combining the propensity matching and weighting method, and ensure the estimation robustness through the confidence interval and the rollback strategy;

[0055] S4, estimate the gap filling source causal gain based on the double robustness learner, obtain the gap closable degree by using the mapping function, calculate the confidence interval, and screen out the reliable gap filling source with significant passing;

[0056] S5, sort the candidate gap filling sources according to the closable degree priority, generate the correction strategy of the gap filling strength, gap filling opportunity and priority configuration, and perform linkage verification combined with the frequency control, budget and delay constraints, and gradually put into operation through the gray release mechanism;

[0057] ​S6, apply the correction strategy in 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 deviation occurs, trigger re-estimation and fallback, until the convergence condition is met and the closed loop is stable.

[0058] S1 specifically includes the following sub-steps:

[0059] S110, data acquisition and preprocessing: acquire historical and real-time traffic request data, bidding response data and actual filling result data, respectively mark unique request identification (UUID) and timestamp, form the original data set; the target filling rate is set according to the dimension of "media bit × country / region × time period (5 minute sliding window)", and the actual filling rate is counted according to the same granularity;

[0060] When an abnormal response occurs (HTTP5xx or response delay greater than 1500ms), the request is rejected. If the proportion of abnormal data exceeds 10%, the sampling window needs to be reset to ensure that the total sample size is not less than 105, and the modeling statistics are effective.

[0061] S120, gap definition and denoising: define the non-negative part of the difference between the target filling rate and the actual filling rate as the filling rate gap Δ, and specify the unit as percentage point (pp); when the target filling rate is less than or equal to the actual filling rate, Δ takes the value of 0. The value of Δ is retained to two decimal places, and Hampel filtering (window size 7) is used to remove outliers in the time series of Δ. The reason for choosing a window size of 7 is that through cross-validation, it is found that within the range of 5-9, it can remain stable, and 7 has the best effect.

[0062] S130, feature extraction: for each traffic request, extract request features and context features, including user grouping, media bit, region, time period and network status; if necessary, it can also be extended to device type and operating system features; all category type features are uniformly represented by one-hot encoding, and continuous type features are represented by standardization or embedding to ensure the comparability of different features in the model.

[0063] S140, sample construction and division: associate the feature vector of each traffic request with its corresponding filling rate gap Δ to form a gap sample set. The gap sample set is divided in a rolling window manner: T0-28d to T0-7d as the training set, T0-7d to T0-1d as the validation set, and T0 as the online set. Each time the sliding window is moved by 1 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 that the sample distribution of different regions, media bits and user groups is balanced.

[0064] S150, register the corresponding complement source set S for each traffic request in the gap sample set; the source of the complement source set includes demand side platform, direct sampling advertiser and self-backfill pool. When a certain traffic request fails to match any complement source in the corresponding period, its complement source set is registered as an empty set, and a default identifier is marked in the sample to ensure that the subsequent modeling process will not be inconsistent due to missing data.

[0065] Sample size (≥10 5 Bar); feature expansion and encoding method; splitting method (rolling sliding window, stratified sampling); boundary condition (default identifier for missing complement source).

[0066] To avoid formula interpretation ambiguity, the application uniformly defines all involved variables as follows (unit or dimension as shown in the table):

[0067]

[0068] The sampling strategy is as follows: the sliding window length is 28 days, and the step length is 1 day; the sample splitting ratio is 4:1:1 (training set / validation set / online set); the stratified sampling dimension is "region x media bit x user group"; the lower limit of the single stratified sample is 500, to ensure the stability of modeling.

[0069] S2 specifically includes the following sub-steps:

[0070] S210, complement source feature statistics: for each complement source in the gap sample set, the complement performance in the historical 7-day rolling window is counted to form a complement source feature vector;

[0071] The complement source feature at least includes: complement success rate: the number of times of successfully returning effective advertisements accounts for the proportion of request times, and the weighted average is taken after daily statistics; average backfill delay: in milliseconds (ms), calculated by the 95th percentile; audience group association strength: calculated by Pearson correlation coefficient, ranging from-1 to 1; revenue-related indicators: such as eCPM, conversion rate, weighted by exposure; if the 7-day data is insufficient, fall back to 30-day window to ensure statistical robustness.

[0072] S220, processing identifier definition: label each traffic request with a processing identifier T, whose value rule is as follows: when the request selects only one complement source, T = the identifier of the complement source; when the request selects multiple complement sources in parallel, T = the identifier of the first complement source that returns effective advertisements; if the first response source returns invalid advertisements, then extend to the next successful complement source; when the request does not select any complement source, T = "not selected".

[0073] The above rules ensure that all samples have a clear definition in causal inference, and cover the boundary conditions of invalid responses.

[0074] S230, propensity model training and calibration: train propensity estimation model under the joint action of request features and supply source features, output the probability estimation value e(X) of each supply source selection. Model type: logistic regression or gradient boosting tree (XGBoost); objective function: cross-entropy loss; feature selection: all features are screened by chi-square test or information gain, ensuring significance p < 0.05; training sample size: not less than 10 5 bars;

[0075] Calibration method: Platt scaling, adjust the output probability to the interval [0, 1]; performance lower limit: AUC ≥ 0.70 on the validation set. If lower than the threshold, re-tune the parameters or expand the features until the requirements are met.

[0076] S240, stabilization weight and overlap constraint: after obtaining the propensity score, calculate the stabilized inverse probability weight (SIPW) for each sample:

[0077]

[0078] where represents the weight of the i-th sample. It is used to adjust the importance of the sample in weighted regression or causal estimation; T represents the treatment variable, such as whether to select a certain supply source; T = 1: the sample has accepted a certain supply source; T = 0: the sample has not accepted the supply source; t is a specific treatment state (such as t = 1 for use, t = 0 for non-use); P(T = t) represents the marginal probability of accepting treatment t on the overall level. It is a constant, reflecting the overall proportion; represents the propensity score of accepting treatment t given the sample features , which is usually estimated by a logistic regression or machine learning model.

[0079] The weight formula is used for sample re-weighting to ensure that the selected and non-selected samples of the supply source are more balanced in feature distribution in subsequent counterfactual estimation and causal gain calculation. Its role is to eliminate the bias caused by uneven distribution of regional, time period, media bit, etc. in the traffic sample; ensure that the estimation of supply source effect is close to the unbiased result of "random experiment"; improve the robustness of counterfactual gap calculation and causal gain estimation, and avoid overfitting or abnormal value dominating the result.

[0080] Suppose the selection probability of a certain supply 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 easy to cause variance explosion. Stabilized SIPW weight: 0.3 / 0.1 = 3, which converges significantly and is more stable.

[0081] P(T=t) is estimated by empirical frequency and smoothed by Laplace (adding 1) to avoid zero denominator; to prevent extreme weight, overlap constraint is set: when If not in the interval [0.05, 0.95], the sample is truncated or removed; if the removal ratio exceeds 5%, the sample window needs to be expanded (from 7 days to 30 days) or the feature dimension needs to be increased to restore the overlap.

[0082] S250, consistency check: after completing the processing identifier and propensity score writing, the sample set with propensity information is checked for consistency, including: checking whether each request processing identifier is unique and matches the complement source set; checking whether the propensity score is in the interval [0, 1] and the sum of the probabilities of each complement source is 1; checking whether the mean of the SIPW weight is close to 1 (allowing a fluctuation of no more than ±0.05); checking whether the total sample quantity after removal is still greater than 10 5 .

[0083] If the check fails, the following rollback mechanism is executed: mild deviation (such as weight mean deviation ≤0.1): recalibrate the propensity model; moderate deviation (such as removal ratio 5-10%): recalculate after expanding the statistical window or increasing the feature dimension; severe deviation (such as total sample quantity <10 5 ): rollback to the last stable version of the data set, and record the abnormal log. After passing the check, the sample set with propensity information output can enter the next step S310-S350 of counterfactual benchmark construction.

[0084] S3 specifically includes the following sub-steps:

[0085] S310, control sample selection: for each target complement source s in the gap sample set, select a traffic request from the sample set with propensity information that does not use the complement source to construct a control sample set.

[0086] The size of the control sample set must be ≥5000;

[0087] If it is less than 5000, the statistical window is expanded in turn (from 7 days to 30 days), or adjacent regions / time periods are merged until the threshold is met; if it is still insufficient, it is marked as sparse stratification and enters the rollback strategy (see S340).

[0088] S320, matching and weighting parallel construction: on the control sample set, a double-path parallel method is used to construct counterfactuals:

[0089] Path 1: nearest neighbor propensity matching; matching method: k=5 nearest neighbors, Mahalanobis distance metric; ensure that each processing sample is paired with at least 5 control samples.

[0090] Path 2: SIPW weighting: using the stabilized inverse probability weights calculated in S240 Control samples are weighted; ensure the overall distribution is comparable to the treatment group.

[0091] If the counterfactual gap estimate difference between the two paths is > 1 percentage point (pp), the path with smaller variance is chosen as the preferred result, and the other result is kept for cross-validation.

[0092] S330, Counterfactual benchmark estimation: on the preferred control set, calculate the counterfactual fill rate gap without using the target source s condition :

[0093]

[0094] Where represents the counterfactual gap, i.e., the difference between the system fill rate and the target fill rate without using a certain source; is the actual fill rate; represents the expected fill rate when the treatment variable T = 0 (i.e., no source is selected) given the sample characteristics X; this is a counterfactual estimate, usually obtained through propensity score models, matching, or weighting methods; it expresses "what the system would look like if it did not use a certain source"; represents the target fill rate set by the system (e.g., 95%). This is the performance indicator that the platform hopes to achieve.

[0095] Estimate the fill rate without using the source based on the propensity score model ; then, compare it with the target fill rate ; the difference between the two is the counterfactual gap ; that is, if , it means that without the source, the gap still exists and needs to be further corrected; if , it means that without using the source, the fill rate has met or exceeded the target, and the gap problem is not significant.

[0096] The above formula is the core of the counterfactual benchmark construction, used to judge the potential severity of the gap; combined with the effect estimation of the source, a "source-counterfactual gap mapping table" can be formed; through this benchmark, the subsequent causal gain and closable degree calculation have more reference value, avoiding making distorted decisions based only on observations; Bootstrap resampling (B = 1000, sample size > 106, take B = 500) is used to calculate the 95% confidence interval [L, U] of , and keep two decimal places.

[0097] S340, Robustness and sparse rollback: if When the confidence interval width (U-L) > 2pp, it is marked as unstable estimation; when unstable estimation occurs or S310 determines that the stratification is sparse, the following fallbacks are sequentially performed: removing the regional dimension and only keeping the media bit x period; removing the period dimension and only keeping the media bit; if it is still unstable, fallback to the overall control benchmark; when fallback, the site strength parameter u is adjusted by 50% of the original value to avoid overfitting;

[0098] To ensure the stability of the strategy, the application sets an upper limit for the daily fallback trigger frequency When the upper limit is exceeded, the subsequent trigger will be automatically delayed to the next day for execution and will be recorded in association with the abnormal log, effectively avoiding the policy shock caused by frequent fallbacks and improving the engineering stability of the correction process.

[0099] S350, mapping registration: binding each site source s with its , confidence interval [L, U] and stratification information to generate a "site source- counterfactual benchmark mapping table". The mapping table is stored in the strategy database and supports daily updates; if Deviation from the historical mean ± 2σ for 3 consecutive days, trigger counterfactual benchmark reconstruction; the mapping table supports backtracking query of the last 30 days for policy drift monitoring.

[0100] S4 specifically includes the following sub-steps:

[0101] S410, causal learner construction: based on the "site source-counterfactual benchmark" mapping table, a double robust learner (DR-Learner) is used to estimate the causal gain value g of each site source in the filling rate gap Δ.

[0102] First stage: respectively fitting the conditional expectation functions 、 and propensity score of the treatment group and the control group.

[0103] Second stage: constructing pseudo labels:

[0104]

[0105] Wherein represents the pseudo label value of the i-th sample, is the treatment identification variable of the sample, is the observed filling rate result of the sample, is the propensity score of selecting the site source under the sample characteristics , and respectively represent the filling rate prediction expected value under the condition of using and not using the supply source; the above formula integrates the weighted residual term of the treatment group and the control group and the prediction difference term, and through the double correction of the propensity score model and the result regression model, the double robust estimation of the sample level causal gain is realized. When any one of the propensity model or the result model can be accurately fitted, the overall estimation still maintains consistency, thereby effectively reducing the estimation error caused by sample distribution bias.

[0106] Based on the pseudo-label calculation result, the system can further utilize the meta-learner to model the causal effect of the supply source, and obtain the gap improvement contribution value under different traffic feature conditions. This way can realize stable causal gain estimation in a heterogeneous traffic environment, and ensure the scientificity and reliability of the generated supply strategy.

[0107] Meta-learner training: train LightGBM with as the supervision signal, with the parameter settings being: maximum depth ≤ 8, learning rate = 0.05, and minimum sample number of leaf node ≥ 50; convergence and repeatability: the random seed is set to 2025, and the mean square error (MSE) on the validation set is ≤ 0.05; if not, retrain until convergence.

[0108] S420, hierarchical statistics and variance estimation: the causal gain value g is statistically stratified according to the request feature and the context feature, and the mean, variance and sample size of each stratification are output; the lower limit of the sample: the sample size of a single stratification must be ≥ 1000, otherwise merging is performed; the merging rule sequence: preferentially merging adjacent regions → then merging time periods → finally merging media positions, to ensure balanced distribution; variance estimation: Bayesian robust estimation is used, with the prior distribution being set as inverse gamma (α = 3, β = 2); result accuracy: all statistical results are retained to two decimal places.

[0109] S430, closable degree mapping and confidence interval: the causal gain value g is normalized to the gap closable degree C, and the calculation formula is:

[0110]

[0111] Where C represents the gap closable degree, used to represent the compensation ability of a supply source to the current filling rate gap, with the value range being 0-1; is a Sigmoid mapping function, defined as , 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 amplitude of the supply source to the filling rate gap; is a smoothing coefficient, used to control the mapping steepness of the Sigmoid function, with the value range being [1.5, 2.5] percentage points (pp), and the default value being 2 pp.

[0112] Through the function mapping, the causal gain with different magnitudes can be uniformly converted into a comparable standardized closability index. When the causal gain g is large, the Sigmoid function output tends to 1, indicating that the filling source has high closing potential; when the causal gain is small or negative, the output value is close to 0, indicating that the filling source has limited effect on the gap correction. By adjusting the smoothing parameter , the mapping sensitivity can be adaptively adjusted according to the flow fluctuation, realizing the dynamic balance of the strategy generation process.

[0113] The calculation method can effectively solve the problem of inconsistent dimensions of causal effects between different filling sources, so that the filling effect evaluation is carried out on a unified scale, thereby improving the stability and interpretability of the filling source screening and strategy generation.

[0114] Variance propagation: propagate the variance through the Delta method to get the confidence interval Stability criterion: if the interval width , it is determined as unstable estimation, and enters the S340 sparse rollback process.

[0115] S440, significance and reliability screening: screening the filling source according to the closability confidence interval and statistical test: determination condition: only when >0.1 and the two-tailed t-test result of causal gain g satisfies p<0.05, the filling source enters the candidate set;

[0116] The filling source that does not pass the condition is marked as "unreliable" and is excluded in the subsequent strategy generation; if the candidate set is empty, it degenerates into a greedy sorting based on eCPM as a fallback solution.

[0117] S450, result binding and list generation: binding the filling source that passes the screening with its causal gain value g, gap closability C and interval , and generating a "filling source-closability list". Storage mechanism: the list is stored in the strategy database, and the update frequency is consistent with the counterfactual benchmark (once a day); backtracking ability: supporting list change query for the last 30 days, used for monitoring drift;

[0118] Drift trigger: if the C of a certain filling source decreases by more than 20% for three consecutive days, the re-estimation mechanism is triggered, and S410-S430 is re-executed; if the re-estimation still does not converge, immediately rollback to the list of the last stable version and issue an alarm.

[0119] S5 specifically includes the following sub-steps:

[0120] S510, the preferred position source is preferably formed with the candidate strategy: on the basis of "the list of position source-closability", the position source with the highest closability C value and passing the significance test is selected as the main preferred object, and the secondary preferred position source is sorted according to the C value to form a candidate set.

[0121] Quantity limit: the candidate set is limited to top-3 at most to avoid overcomplication of the strategy;

[0122] Bottom mechanism: if the candidate set is empty, it is degraded to a backfill strategy based on eCPM greedy sorting;

[0123] Additional information: each position source in the candidate set needs to carry the causal gain value g, the closability interval and the hierarchical identifier to ensure subsequent traceability.

[0124] S520, gap correction strategy parameterization: for each position source in the candidate set, generate the corresponding gap correction strategy, including:

[0125] Position intensity

[0126]

[0127] Wherein represents the position intensity parameter, which is used to control the proportional weight allocated to the position source in the strategy execution phase; C is the gap closability index calculated according to the causal gain, whose value range is 0-1; is the amplification coefficient, which is used to adjust the response sensitivity of the position intensity; is the upper limit value of the position intensity, which is used to limit the adjustment amplitude of the strategy to prevent system shock caused by excessive correction.

[0128] When the closability C is low, the system reduces the position intensity in proportion to avoid overreaction to noise samples; when the closability C is high, the system increases the position intensity in proportion, but does not exceed the maximum value , so as to ensure the controllability and stability of the position adjustment process.

[0129] In this embodiment, 0.6 is taken, which means that a single position source can occupy 60% of the total request amount at most; 1.2 is taken to slightly amplify the response amplitude in the case of high closability to improve the overall filling rate correction effect. Through the function constraint mechanism, the system can adaptively adjust the position intensity in different traffic scenarios to realize the fine and robust gap correction.

[0130] If C ≥ 0.8, directly take u = 0.6; if C ≤ 0.2, down-regulate to u = 0.2. Filling opportunity control: trigger threshold: trigger when Δ ≥ 2pp; cooling time: interval between two triggers of the same filling source ≥ 3 minutes; peak suppression: when Δ ≥ 5pp, only enable the filling source with the highest C value, and freeze other candidate sources for 10 minutes.

[0131] Filling priority configuration: sort by C value as priority; when the C value difference between two filling sources is < 0.05, prefer the filling source with lower average delay.

[0132] S530, constraint linkage and feasible region check: the generated gap correction strategy needs to be linked and checked with three types of constraints: frequency control constraint: the 24h exposure times of the same user ≤ 10; if the prediction exceeds the limit, the filling intensity u is down-regulated to 50% of the original value. Budget constraint: the single-day budget deviation cannot exceed ±3%; if it exceeds, freeze the overspending filling source until the next period. Delay constraint: request-response delay ≤ 800ms; if it exceeds the limit, reduce the priority by 1 level and mark "high delay".

[0133] Unified shadow price : write the budget and delay constraints into the Lagrange dual function, if > θ (θ = 100), the overall strategy is downgraded, and only the Top-1 filling source is retained.

[0134] S540, strategy binding and grayscale preparation: the correction strategy that passes the constraint check needs to be bound with the target fill rate r, the actual fill rate r, and the predicted closable degree , generate the corresponding strategy signature hash. Grayscale release process: adopt four stages of 10%→30%→50%→100%, each stage for 30 minutes; rollback condition: if Δ decreases by < 0.5pp within the stage, roll back to the previous stage; consistency guarantee: when the hash comparison passes and the version number is continuous, it can enter the next stage.

[0135] S550, distribution and execution readiness: write the bound gap correction strategy into the strategy distribution module and push it to the real-time engine.

[0136] Metadata record: including strategy version number, generation time, candidate set, filling intensity parameter, priority sequence, constraint check result;

[0137] Consistency check: if the metadata is missing or the hash does not match, prevent issuance and return to S510 for recalculation;

[0138] Exception handling: if the distribution fails ≥ 3 times / hour, trigger an alarm and automatically roll back to the latest stable strategy.

[0139] S6 specifically includes the following sub-steps:

[0140] S610, Real-time correction execution: Apply the gap correction strategy that has passed the grayscale verification to the real-time traffic request, and perform the gap filling operation according to the gap filling intensity u, gap filling timing control, and gap filling priority configuration.

[0141] Latency monitoring: Real-time monitoring of request processing latency in milliseconds (ms);

[0142] Degradation condition: If the single-time latency is > 1000 ms, or the continuous three-time latency is > 800 ms, trigger the degradation mechanism: reduce the gap filling intensity u by 30%, and record the abnormal event;

[0143] Abnormal record: All degradation triggering events are written into the log as the input for subsequent feedback update.

[0144] S620, Execution result calculation and comparison: Collect the filling results after correction execution, calculate the actual filling rate gap after correction , and compare it with the gap before correction , and output the gap reduction amplitude .

[0145] Update frequency: The result is updated every 5 minutes;

[0146] Numerical accuracy: to two decimal places;

[0147] Decision criteria: If ≤ 0.5pp, it is considered that the gap is effectively closed; if , immediately trigger an abnormal warning and push the data to the monitoring system.

[0148] S630, Small-step feedback and calibration update: Compare with the predicted closure C, if the difference , trigger the small-step feedback update:

[0149] Trigger condition: Trigger when the difference in the last three 5-minute windows is greater than the threshold;

[0150] Update method: Only adjust the calibration layer parameters of the tendency model and the causal learner, learning rate η = 0.05;

[0151] Range limit: The update only acts on the last two layers to avoid catastrophic drift;

[0152] Warm start strategy: If the difference deviation is only one-time noise, do not trigger the update.

[0153] S640, Re-estimation trigger and rollback: If any of the following conditions is met, trigger the strategy re-estimation process:

[0154] The average difference between the measured C and the predicted C is greater than 1.5pp for three consecutive windows;

[0155] 95th percentile gap The actual fill rate is greater than the historical average + 2σ for two consecutive windows;

[0156] The SLA achievement rate is less than 90% for three consecutive windows.

[0157] Post-trigger processing: re-enter S410-S450 to calculate the causal gain and closability, and rebuild the revised strategy in S510-S550; if the re-estimation does not converge after two times, fall back to the latest stable version of the strategy and issue an alarm.

[0158] S650, convergence criterion and closed loop maintenance: when the following conditions are met, the system is determined to be convergent and stable: the SLA achievement rate is ≥95% for six consecutive windows (30 minutes); the over / underfill ratio is <0.1 for six consecutive windows; the actual 95th percentile value is ≤2pp.

[0159] After convergence: maintain the current strategy execution; update the "repositioning source- counterfactual benchmark" and "repositioning source-closability list" once a day; if the subsequent monitoring indicators deviate from the convergence conditions again, automatically return to S640 to trigger re-estimation and rollback.

[0160] To avoid frequent re-estimation of the system due to short-term fluctuations, the number of daily re-estimation triggers is also set to an upper limit of 3, and the excess part is delayed for processing. All trigger events are recorded in the log for subsequent policy stability analysis and model adaptive adjustment; ensure that the system has dynamic but will not fall into a state of shock, so as to maintain controllability and steady state in high-frequency bidding scenarios.

[0161] Embodiment two: the embodiment provides a traffic filling rate gap repositioning prediction and revision system, comprising:

[0162] A data acquisition module is used to collect historical and real-time traffic request data, bidding response data and actual filling result data, to exclude abnormal requests, define the difference between the target filling rate and the actual filling rate to form a filling rate gap, and perform denoising processing on the gap through filtering; at the same time, multi-dimensional features such as user grouping, media position, region, time period, network state are extracted, the samples are divided into training set, validation set and online set according to the rolling time window, and the corresponding repositioning source set is registered;

[0163] The propensity estimation and sample balancing module is configured to statistically analyze the success rate, backfilling time delay, revenue index, and audience group correlation characteristics of each supply source, train a propensity estimation model based on the request characteristics and supply source characteristics to output a supply source selection probability, and achieve sample distribution balancing based on a stabilized inverse probability weight; the module is also configured to perform consistency checking and rollback processing of processing identification uniqueness, probability range, weight mean, and sample size;

[0164] The counterfactual gap estimation module is configured to select a control sample set that does not use the target supply 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 supply source, and perform a rollback strategy at the regional or time period level when the sample is insufficient or the estimation is unstable.

[0165] The causal gain calculation module is configured to construct a pseudo-label meta-learner based on a double-robust learner based on the counterfactual gap, estimate the causal gain of the supply source in the gap dimension, convert the causal gain into a gap closable degree using a Sigmoid mapping function, calculate the confidence interval using a variance propagation method, filter reliable supply sources according to a significance test and stability criterion, and generate a "supply source-closable degree list".

[0166] The correction strategy generation module is configured to sort candidate supply sources according to closable degree, calculate supply strength, trigger timing, and priority configuration, and perform linkage checking with constraints such as frequency control, budget, and time delay; perform strategy degradation when the shadow price item exceeds a set threshold, and only retain the Top-1 supply source; generate a gray release plan for the verified strategy, and form a strategy signature for version management.

[0167] The closed-loop execution and feedback module is configured to apply the correction strategy in real time in the traffic request, monitor the request processing time delay and trigger automatic degradation, collect the actual filling rate gap after correction execution and compare it with the prediction result, perform small-step feedback update when the deviation exceeds a preset threshold, trigger re-estimation and rollback when there is continuous deviation, maintain the current strategy execution when the convergence criterion is met, and periodically update the counterfactual benchmark and supply source list to achieve closed-loop stability.

[0168] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula of the most recent real situation, and the preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0169] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of 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, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center through a wired (for example, infrared, wireless, microwave, etc.) manner. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0170] Those of ordinary skill in the art can realize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether the 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 the present application.

[0171] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device, and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0172] In several embodiments provided in the present application, it should be understood that the disclosed system, device, and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the shown or discussed ones can be indirect coupling or communication connection through some interfaces, devices, or modules, which can be electrical, mechanical, or other forms.

[0173] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment.

[0174] In addition, the functional modules in each embodiment of the present application can be integrated into one processing module, or each module can exist physically alone, or two or more modules can be integrated into one module.

[0175] The functions, if realized in the form of software function modules and sold or used as independent products, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various program code storage media.

[0176] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0177] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application, any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application, should be included in the protection scope of the present application.

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 control samples that did not use the target filler source, calculate the counterfactual fill rate gap using propensity matching and weighting methods, and ensure the robustness of the estimation through confidence intervals and backoff strategies; S3 specifically includes: 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. The results of different paths were compared and the path with smaller variance was selected. Path 1 was nearest neighbor propensity matching, and Path 2 was weighted using the stabilized inverse probability weights in S2. If the difference between the counterfactual gaps of the two paths was less than 1 percentage point, the path with smaller variance was selected. Calculate the counterfactual fill rate gap and generate confidence intervals to measure the gap performance under conditions where no target fill source is used; The counterfactual fill rate gap represents the difference between the system fill rate and the target fill rate when a certain fill source is not used. The calculation formula is: This indicates a counterfactual fill rate gap. This represents the actual fill rate. This represents the expected fill rate when the processing variable T=0, i.e., no fill source is selected, given sample features X. This indicates the system's preset target fill rate; When the confidence interval is too wide or the sample is sparse, a backoff strategy at the regional, time period or overall level is implemented, and the filler source, its counterfactual gap and confidence interval are registered in the mapping table to generate a filler source and counterfactual benchmark mapping table for subsequent calculations. S4. Based on the mapping table of the complement source and the counterfactual benchmark, the causal gain of the complement source is estimated based on the dual robust learner, the gap closure degree is obtained by using the mapping function, and the confidence interval is calculated to screen out reliable complement 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 4, 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.

6. The method for predicting and correcting flow fill rate gaps according to claim 5, 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.

7. 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