Advertisement production parameter self-learning optimization method and system based on propagation effect feedback

Abstract

The application discloses an advertisement production parameter self-learning optimization method and system based on propagation effect feedback, relates to the technical field of advertisement optimization, and comprises the following steps: S1, collecting feedback production integrated data of an advertisement, and performing pretreatment on the feedback production integrated data; S2, constructing feedback production bucket level samples, performing conversion backfilling to obtain a backfill conversion rate, and evaluating the propagation effect feedback credibility of the feedback production bucket level samples; S3, identifying a material version switching event, forming a credible material version switching response record, and generating a material version updating rule; S4, triggering a material version switching operation, quantitatively evaluating the rule running income after switching, and determining the effectiveness, rollback and running cancellation of the material version updating rule. The application solves the problem that, in the existing advertisement optimization, platform aggregation back transmission and feedback back transmission delay are superimposed, resulting in difficult attribution of material version switching income and unstable updating rule determination.

Description

Technical Field

[0001] This invention relates to the field of advertising placement optimization technology, specifically to a self-learning optimization method and system for advertising production parameters based on feedback on dissemination effects. Background Technology

[0002] In recent years, with the development of mobile internet and programmatic advertising technology, advertising delivery has gradually shifted from manual product selection and fixed rule configuration to a technology route primarily based on data-driven automated delivery and adaptive parameter adjustment. The business rules in the advertising delivery chain have also evolved from static configurations to a set of rules that can be continuously updated based on feedback from dissemination effects. Under a delivery mechanism represented by bill-per-click (CPC), advertising systems typically establish closed-loop control around traffic allocation, recall filtering, bid adjustment, and ranking, and continuously optimize system parameters based on dissemination effect indicators such as impressions, clicks, conversions, and revenue to achieve a dynamic balance between delivery efficiency and revenue goals. Against this backdrop, related technical solutions for advertising system parameter optimization, bidding strategy adjustment, and traffic experimental allocation have emerged, forming various feasible and systematic implementation methods.

[0003] For example, the invention with announcement number CN115730977B relates to a CPC advertising system and a parameter optimization method in the bidding process. The system includes an ad recall module, a filtering module, a bid adjustment module, and a ranking module. The ad recall module is configured to reduce high-priced ads and / or supplement low-priced ads to obtain a first ad recall set when the number / proportion of low-priced ads recalled according to preset conditions is lower than a threshold. The filtering module obtains the estimated CTR of the ads in the first ad recall set and filters all or part of the ads whose estimated CTR is less than the CTR threshold to obtain a second ad recall set. The bid adjustment module is configured to adjust the bids of the ads in the second ad recall set so that the estimated return on investment is not less than the historical statistical return on investment within a preset historical period. The ranking module is configured to rank the ads in the second ad recall set after bid adjustment according to eCPM. This invention can effectively control the average cost per click while improving ad CTR.

[0004] For example, the invention disclosed in CN118333688A provides a method, apparatus, electronic device, and storage medium for optimizing parameters of an advertising system. The method may include: dividing the traffic of the advertising system into a first type of traffic group and a second type of traffic group; determining the transmission parameter values ​​for each first type of traffic group and calculating the revenue parameter values ​​for each first type of traffic group, wherein the number of advertising providers represented by the transmission parameter values ​​is less than the number of advertising providers connected to the advertising system; among the multiple first type of traffic groups, determining the first type of traffic group with the largest revenue parameter value as the optimal experimental group, and determining the transmission parameter values ​​of the optimal experimental group as the optimal transmission parameter values; and determining the optimal transmission parameter values ​​as the transmission parameter values ​​for the second type of traffic group. This is beneficial for improving the parameter optimization efficiency and overall revenue of the advertising system.

[0005] However, existing technologies primarily focus on optimizing recall, filtering, bidding, or traffic delivery parameters. Their dissemination effect feedback is typically relayed back by audience and time buckets based on the aggregation criteria of the advertising platform. Furthermore, the feedback latency of conversion events can cause delays in the attribution of conversion counts and conversion amounts within statistical time buckets. Without delayed conversion backfilling and a unified timeline alignment, the conversion rates within different statistical time buckets are difficult to compare. This makes it difficult to distinguish between apparent fluctuations caused by "real version differences" and "backfilling latency errors" when iterating versions based on bucket-level feedback. Simultaneously, differences in statistical criteria across different advertising platforms and ad placements further amplify the cross-channel drift of backfill conversion rates. When bucket-level samples are directly used for version switch response aggregation under inconsistent statistical backgrounds, differences in criteria and latency noise can easily be mistakenly precipitated as version evolution patterns, leading to unstable version update rule generation, repetitive rule triggering directions, and frequent rollbacks after deployment.

[0006] Therefore, in order to address the above issues, there is an urgent need for a self-learning optimization method and system for advertising production parameters based on feedback on dissemination effects. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a self-learning optimization method and system for advertising production parameters based on feedback on dissemination effects. This solves the problem that the superposition of platform aggregation and feedback transmission delays in existing advertising campaigns makes it difficult to attribute the benefits of material version switching and causes instability in update rule determination.

[0009] Technical solution

[0010] To achieve the above objectives, this invention employs the following technical solution: a self-learning optimization method for advertising production parameters based on dissemination effect feedback, comprising: S1, collecting integrated data on advertising feedback production, and performing time alignment, smoothing and noise reduction, anomaly removal, missing data completion, and standardization on the integrated data to generate preprocessed integrated data on feedback production; S2, constructing feedback production bucket-level samples based on the preprocessed integrated data on feedback production, performing conversion backfilling to obtain the backfilling conversion rate, and evaluating the credibility of the dissemination effect feedback of the feedback production bucket-level samples in conjunction with the feedback backhaul latency, and selecting a feedback production calibration sample set based on the evaluation results; S3, identifying material version switching events based on the feedback production calibration sample set, and forming a credible material version switching response record based on the backfilling conversion rate before and after the switch, and generating material version update rules based on the aggregation statistics of the credible material version switching response records; S4, triggering material version switching operations according to the material version update rules, and quantitatively evaluating the rule operation benefits after the switch to obtain a rule operation benefit evaluation value, and determining the effectiveness, rollback, and cancellation of the material version update rule based on the rule operation benefit evaluation value.

[0011] Furthermore, the specific steps for collecting integrated feedback data from advertisements and performing time alignment, noise reduction, anomaly removal, missing data completion, and standardization on this data to generate pre-processed integrated feedback data are as follows: Real-time collection of integrated feedback data from advertisements, including collection timestamp, platform identifier, ad placement identifier, creative version number, audience bucket identifier, statistical time bucket start timestamp, statistical time bucket duration, impressions, clicks, dwell time, bounce rate, conversion rate, conversion amount, feedback transmission latency, creative file duration, creative frame rate, creative bitrate, screen resolution width, screen resolution height, first frame display duration, shot transition interval, and total number of subtitle characters. The data collection process includes: single-time duration of subtitles, start time of selling points, duration of selling points, and proportion of the main subject to the screen area; sorting of records according to collection timestamps for the collected feedback production data; mapping multi-source asynchronous feedback records to a unified timeline using a linear interpolation time alignment algorithm to construct a unified timeline for feedback production; smoothing the feedback production data using a Kalman filter algorithm; identifying and removing abnormal statistical records using an interquartile range anomaly detection algorithm, and completing the missing segments formed after removal using a forward and backward filling fusion completion algorithm; and standardizing the feedback production data using a Z-score standardization algorithm to eliminate dimensional differences between different physical quantities.

[0012] Furthermore, based on the preprocessed integrated feedback creation data, the specific steps for constructing feedback creation bucket-level samples and performing conversion backfilling to obtain the backfilling conversion rate are as follows: Read the preprocessed integrated feedback creation data, and group various records according to the platform identifier, ad placement identifier, audience bucket identifier, material version number, and statistical time bucket start timestamp. Within each statistical time bucket, summarize the number of impressions, clicks, dwell time, bounces, conversions, and conversion amounts to generate feedback creation bucket-level samples with the statistical time bucket as the smallest granularity, and generate a unique sample number for each feedback creation bucket-level sample. For each feedback creation bucket-level sample, read the statistical time bucket start timestamp and feedback feedback latency to calculate the conversion backfilling timestamp. When the conversion backfilling timestamp falls into a subsequent statistical time bucket, transfer the corresponding number of conversions and conversion amounts from the original statistical time bucket and accumulate them to the target statistical time bucket to obtain the backfilled conversion number and conversion amount. Read the number of impressions from the target statistical time bucket as the backfilled impression number, and perform a ratio calculation between the backfilled conversion number and the backfilled impression number to obtain the backfilling conversion rate.

[0013] Furthermore, the specific steps for evaluating the credibility of feedback in creating bucket-level samples based on feedback return latency are as follows: Using the platform identifier, ad placement identifier, and audience bucket identifier as joint grouping keys, and under the condition of the same statistical time bucket duration, a return conversion rate sequence is constructed based on the return conversion rate in the order of the statistical time bucket start timestamps. The sliding window statistical method is then used to calculate the sliding window mean and standard deviation of the return conversion rate, yielding the return conversion rate mean and standard deviation. The absolute value of the difference between the return conversion rate and the return conversion rate mean is calculated to obtain the return conversion rate offset. The return conversion rate offset is divided by the sum of the return conversion rate standard deviation and the minima to obtain the standardized offset value. The feedback return latency is divided by the duration of the statistical time bucket, and the negative of the resulting ratio is used as the exponent to perform an exponential operation on the natural constant e to obtain the return latency decay term. The standardized offset value is multiplied by the return latency decay term to obtain the sample credibility evaluation value.

[0014] Furthermore, the specific steps for selecting the feedback to create a calibration sample set based on the evaluation results are as follows: Using the material version number as the grouping key, read the sample credibility evaluation value in the order of the start timestamp of the statistical time bucket. Mark the feedback creation bucket-level samples that are below the credibility threshold within N consecutive sliding windows as unstable feedback and remove them; otherwise, mark them as stable feedback and retain them. Extract the backfill conversion rate, sample credibility evaluation value and feedback creation integrated data corresponding to the retained feedback creation bucket-level samples and associate them to form the feedback creation calibration sample set.

[0015] Furthermore, the specific steps for identifying material version switching events based on feedback-based calibration sample sets and forming reliable material version switching response records based on the backfill conversion rates before and after the switch are as follows: Read the feedback-based calibration sample set, sort the feedback-based bucket-level samples according to the material version number and the start timestamp of the statistical time bucket. When the material version number differs between adjacent statistical time buckets, mark the start timestamp of the corresponding statistical time bucket as a material version switching event and record the material version number before and after the switch. For each material version switching event, extract the backfill conversion rate and sample reliability evaluation value within the M consecutive statistical time buckets before and after the switch, and the M statistical time buckets after the switch. When the reliability evaluation values ​​of all samples corresponding to the statistical time buckets before and after the switch are higher than the reliability threshold, calculate the interval mean of the backfill conversion rate before and after the switch, calculate the difference between the interval mean of the backfill conversion rate after the switch and the interval mean of the backfill conversion rate before the switch, and obtain the change in backfill conversion rate. Associate the change in backfill conversion rate with the material version number before and after the switch to form a reliable material version switching response record.

[0016] Furthermore, the specific steps for generating material version update rules based on the aggregated statistics of trusted material version switching response records are as follows: Using the combination of version numbers before and after the material version switch as the grouping key, aggregate statistics are performed on the trusted material version switching response records; when the same material version switching relationship occurs K times, and the signs of the changes in backfill conversion rate corresponding to the K occurrences are consistent, the current material version switching relationship is marked as a stable version evolution relationship; the arithmetic mean of all backfill conversion rate changes corresponding to the stable version evolution relationship is taken to obtain the evolution gain value, and the material version number before the switch, the material version number after the switch, and the evolution gain value are associated to generate material version update rules.

[0017] Furthermore, the specific steps for triggering a creative version switching operation based on the creative version update rules and quantifying the operational benefits of the switched rules to obtain the operational benefit evaluation value are as follows: Read the creative version update rules, match them according to the platform identifier, ad placement identifier, audience bucket identifier, and creative version switching relationship, and associate the matched creative version update rules with the currently running creative version number to generate a creative version update task to be verified; for each creative version update task to be verified, without changing the platform identifier, ad placement identifier, and audience bucket identifier, trigger the corresponding creative version switching operation, and calculate the corresponding backfill within W consecutive statistical time buckets after triggering. Conversion rate; simultaneously extract the backfill conversion rate of W consecutive statistical time buckets before the material version switch and calculate the mean to obtain the baseline mean backfill conversion rate; for the backfill conversion rate in the W consecutive statistical time buckets after the update, calculate the difference between it and the baseline mean backfill conversion rate to obtain the backfill conversion rate gain term for each statistical time bucket; calculate the ratio of the feedback return latency corresponding to each statistical time bucket to the duration of the statistical time bucket, square the obtained ratio and add it to a constant, then take the reciprocal to obtain the latency risk attenuation term; multiply the backfill conversion rate gain term by the corresponding latency risk attenuation term to obtain the weighted benefit term for each statistical time bucket; sum the weighted benefit terms in the W consecutive statistical time buckets and take the arithmetic mean to obtain the rule operation benefit evaluation value.

[0018] Furthermore, the specific steps for determining the effectiveness, rollback, and cancellation of material version update rules based on the rule operation benefit assessment value are as follows: The rule operation benefit assessment value is compared with the operation benefit threshold. When the rule operation benefit assessment value is not less than the operation benefit threshold, the current material version update rule is deemed to have passed verification, marked as an effective version evolution rule, and its effectiveness timestamp is recorded. Otherwise, the current material version update rule is deemed to have failed verification, rolled back to the material version number before the switch, and marked as an invalid rule. Operation monitoring is continuously performed on the effective version evolution rule. When a reversal in the direction of the backfill conversion rate is detected within N consecutive statistical time buckets, or when the sample credibility assessment value is lower than the credibility threshold within N consecutive statistical time buckets, the corresponding effective version evolution rule is automatically cancelled, and the material version is restored to its state before cancellation.

[0019] The second aspect of this invention provides a self-learning optimization system for advertising production parameters based on dissemination effect feedback, comprising: a feedback data acquisition and preprocessing module, a cross-channel calibration and traceability module, a material version switching response learning module, and an update rule verification and controlled effectiveness module. The feedback data acquisition and preprocessing module collects integrated feedback production data for the advertisement and performs time alignment, smoothing and noise reduction, anomaly removal, missing data completion, and standardization on the integrated feedback production data to generate preprocessed integrated feedback production data. The cross-channel calibration and traceability module constructs bucket-level samples for feedback production based on the preprocessed integrated feedback production data, performs conversion backfilling to obtain the backfill conversion rate, and considers the feedback transmission latency. The system evaluates the credibility of feedback in creating bucket-level samples and selects a calibration sample set based on the evaluation results. A media version switching response learning module identifies media version switching events based on the calibration sample set and generates credible media version switching response records based on the backfill conversion rate before and after the switch. It then generates media version update rules based on the aggregated statistics of these credible media version switching response records. Finally, an update rule verification and controlled activation module triggers media version switching operations based on the media version update rules and quantifies the benefit of the rule operation after the switch, obtaining a rule operation benefit evaluation value. Based on this evaluation value, it determines whether the media version update rule is effective, rolled back, or canceled.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) A self-learning optimization method and system for advertising production parameters based on feedback of dissemination effect. By introducing a conversion backfilling mechanism, the conversion error caused by feedback backfilling delay is backfilled and corrected in the statistical time bucket dimension, so that the backfilling conversion rate can stably reflect the real conversion contribution within the same time bucket, and improve the correspondence between bucket-level feedback and material version exposure window.

[0023] (2) A self-learning optimization method and system for advertising production parameters based on feedback of dissemination effect. By constructing a credibility assessment of dissemination effect feedback based on the statistical background of conversion rate and feedback transmission delay, and selecting feedback to create a calibration sample set, the subsequent learning process is carried out only on the credible bucket-level samples, reducing the risk of mislearning platform caliber differences and noise fluctuations as version differences.

[0024] (3) A self-learning optimization method and system for advertising production parameters based on feedback of dissemination effect. By identifying material version switching events and comparing the backfill conversion rate before and after the switching, a reliable material version switching response record is formed. Then, the response record is aggregated to generate material version update rules, so that the version evolution law is transformed from "single point fluctuation" to "reproducible switching response relationship", thereby improving the stability and transferability of rule generation.

[0025] (4) A self-learning optimization method and system for advertising production parameters based on feedback of dissemination effect. The method verifies and judges the material version update rules by using the rule-based operation benefit evaluation value, and provides a mechanism for activation, rollback and operation cancellation. This transforms the version update from a one-time launch into a quantifiable, verifiable and revocable controlled evolution process, reducing the continuous revenue loss caused by the long-term solidification of erroneous rules. Attached Figure Description

[0026] Figure 1 A flowchart for a self-learning optimization method for advertising production parameters based on feedback on dissemination effects;

[0027] Figure 2 A structural diagram of a self-learning optimization system for advertising production parameters based on feedback on dissemination effects;

[0028] Figure 3 A bar chart showing the evolution of rules for existing versions based on the evaluation of benefits from rule-based operations;

[0029] Figure 4 Flowchart for verifying and controlling the implementation of material version update rules. Detailed Implementation

[0030] 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.

[0031] Please see Figures 1-4This invention provides a technical solution: a self-learning optimization method for advertising production parameters based on dissemination effect feedback, comprising: S1, collecting integrated data of advertising feedback production, and performing time alignment, smoothing and noise reduction, anomaly removal, missing data completion and standardization on the integrated data of feedback production to generate preprocessed integrated data of feedback production; S2, constructing feedback production bucket-level samples based on the preprocessed integrated data of feedback production, performing conversion backfilling to obtain the backfilling conversion rate, and evaluating the credibility of the dissemination effect feedback of the feedback production bucket-level samples in combination with the feedback backhaul latency, and selecting a feedback production calibration sample set based on the evaluation results; S3, identifying material version switching events based on the feedback production calibration sample set, and forming a credible material version switching response record based on the backfilling conversion rate before and after the switch, and generating material version update rules based on the aggregation statistics of the credible material version switching response records; S4, triggering material version switching operations according to the material version update rules, and quantitatively evaluating the rule operation benefits after the switch to obtain a rule operation benefit evaluation value, and determining the effectiveness, rollback and cancellation of the material version update rule based on the rule operation benefit evaluation value.

[0032] Specifically, the steps for collecting integrated feedback data from advertisements and performing time alignment, noise reduction, anomaly removal, missing data completion, and standardization on this data to generate pre-processed integrated feedback data are as follows: Real-time collection of integrated feedback data from advertisements, including collection timestamp, platform identifier, ad placement identifier, creative version number, audience bucket identifier, statistical time bucket start timestamp, statistical time bucket duration, impressions, clicks, dwell time, bounce rate, conversion rate, conversion amount, feedback transmission latency, creative file duration, creative frame rate, creative bitrate, screen resolution width, screen resolution height, first frame rendering duration, camera transition interval, and text... The data includes the total number of subtitle characters, the duration of each subtitle's single dwell time, the start time of the selling point's appearance, the duration of the selling point's appearance, and the proportion of the main subject's area on the screen. The data collection timestamp is obtained by the access server writing events to the event log under network time protocol synchronization conditions. The platform identifier is output by the ad delivery interface in the returned response field and written by the interface collection program. The ad slot identifier is output by the ad request in the ad slot configuration field and written by the request log collection program. The material version number is output by the material distribution list in the material version field and written by the delivery distribution log collection program. The audience bucket identifier is generated by the audience bucket mapping table after the user identifier is matched and written by the delivery request tracking collection program. The statistical time bucket start timestamp is obtained by the statistical time... The bucket generation program aligns the duration of the statistical time bucket with the integer boundary and writes the data; the duration of the statistical time bucket is obtained by writing a fixed bucket width parameter into the statistical time bucket generation program; the number of exposures is obtained by the exposure event tracking program counting exposure event records within the statistical time bucket; the number of clicks is obtained by the click event tracking program counting click event records within the statistical time bucket; the dwell time is calculated by the page dwell time tracking program using the difference between the entry and exit timestamps; the number of bounces is obtained by the session bounce judgment tracking program counting session records when the dwell time is below the bounce judgment threshold; the number of conversions is obtained by the conversion feedback interface tracking program counting conversion feedback records within the statistical time bucket; the conversion amount is obtained from the transaction... The order feedback interface collection program summarizes the order amount field within the statistical time bucket; the feedback feedback latency is calculated by subtracting the event occurrence timestamp from the reception timestamp in the feedback access log; the media file duration is obtained by parsing the media file header information using the media information probe tool; the media frame rate is obtained by parsing the video stream parameters using the media information probe tool; the media bitrate is obtained by parsing the bitstream statistics parameters using the media information probe tool; the screen resolution width is obtained by parsing the video stream width parameters using the media information probe tool; the screen resolution height is obtained by parsing the video stream height parameters using the media information probe tool; the first frame rendering duration is calculated by the player's first frame rendering timing and data collection program using the difference between the playback start timestamp and the first frame rendering completion timestamp;The shot switching interval is calculated by the shot segmentation detection program based on the time difference between adjacent switching points after triggering the switching point according to the inter-frame histogram difference; the total number of subtitle characters is obtained by the subtitle file parser performing character counting on the subtitle text; the single subtitle dwell time is obtained by the subtitle file parser reading the subtitle in-point timecode and out-point timecode and calculating the difference; the starting time of the selling point is obtained by the selling point timecode annotation acquisition tool reading the mark injection point timecode; the duration of the selling point is obtained by the selling point timecode annotation acquisition tool reading the mark exit point timecode and calculating the difference with the mark injection point timecode; the proportion of the subject area in the frame is obtained by the subject detection program outputting the subject bounding box area on the keyframe and dividing it by the total number of pixels in the frame. The area ratio is used to obtain the data. During data collection and writing, each piece of feedback data is bound to the collection timestamp while retaining the physical unit representation of the original numerical fields. The start timestamp and duration of the statistical time bucket are written into the record header field to define the bucket-level statistical scope boundary. The feedback transmission latency is written into the record header field to define the transmission link latency boundary. The platform identifier, ad placement identifier, audience bucket identifier, and material version number are written into the index field to define the sample affiliation boundary. For the collected feedback data, records are sorted according to the collection timestamp. During the sorting process, the collection timestamp is used as the primary key to perform ascending order while maintaining the original write data of records within the same collection timestamp. The input order remains unchanged to avoid secondary aggregation of bucket-level summary fields during the rearrangement process. A linear interpolation time alignment algorithm is used to map multi-source asynchronous feedback records to a unified time axis. A unified time axis is constructed by creating feedback data. During the mapping process, the time step boundary of the unified time axis is determined by the start timestamp and duration of the statistical time bucket. Each time step on the unified time axis is used as the alignment target time. The collection timestamps of the two adjacent feedback data records before and after the target time are located, and linear interpolation coefficients are calculated. Linear interpolation is then performed on the corresponding field values ​​using the interpolation coefficients to obtain the alignment value. During the alignment process, the platform identifier, ad placement identifier, audience bucket identifier, and element identifier are adjusted. The material version number is checked for consistency and cross-identity splicing records are removed to avoid data with different affiliations being incorrectly inserted into the same time step. The feedback production integrated data is smoothed using the Kalman filter algorithm. In the smoothing process, the unified time axis of feedback production is used as the input sequence index to construct a state vector. The state transition matrix describes the smooth evolution relationship between adjacent time steps. The observation vector carries the alignment value of the feedback production integrated data on the unified time axis. The observation noise covariance describes the bucket-level statistical backhaul jitter amplitude. The process noise covariance describes the slow change amplitude of the actual propagation effect. At each time step, prediction update and correction update are performed sequentially and the smoothed feedback production integrated data is output.For the integrated feedback production data, an interquartile range (IQR) anomaly detection algorithm is used to identify and remove abnormal statistical records. During the identification process, using the unified timeline of feedback production as a benchmark, the record sequence within a fixed-length sliding window is extracted, and the first and third quartiles are calculated. The IQR is calculated, and upper and lower anomaly boundaries are constructed. Records falling outside the anomaly boundaries are marked as abnormal statistical records, and their corresponding header and value fields are removed. During the removal process, unmarked records under the same collection timestamp are retained to avoid overall data loss. The missing segments formed after removal are then filled using a forward-filling and backward-filling fusion algorithm. During the completion process, the preceding valid record of the missing segment is used as the source of the forward-filling value, and the following valid record is used as the source of the backward-filling value. For each time step within the missing segment, forward-filling and backward-filling values ​​are generated separately. The imputed value is obtained by weighting and fusing the imputed value with the missing value based on the positional distance within the missing segment, ensuring continuity between the imputed value and adjacent valid records at the beginning and end of the missing segment. The Z-score standardization algorithm is used to numerically standardize the feedback production integrated data, eliminating dimensional differences between different physical quantities. In the standardization process, the platform identifier, ad placement identifier, audience bucket identifier, and material version number are used as grouping keys. Within each group, a standardized input sequence is extracted according to the unified feedback production timeline. The mean and standard deviation of the input sequence are calculated, and a standardization mapping function is constructed. For each time step, the input sequence value is subtracted from the mean and divided by the standard deviation to obtain the standardized value. If the standard deviation is zero, the minimum term is used to replace the standard deviation to avoid division by zero. The standardized feedback production integrated data is output while retaining the original field names to ensure data name consistency in subsequent steps.

[0033] This implementation plan unifies and solidifies the data collection chain of integrated feedback production by imposing constraints on field attribution, time boundaries, and physical units. It also establishes a verifiable record header and index field structure on a unified timeline for feedback production. This ensures that the integrated feedback production data has a stable and consistent expression foundation across different platforms, ad slots, audience groups, and creative version numbers. Consequently, the entire process—including feedback production bucket-level sample construction, conversion backfilling, credibility assessment of dissemination effect feedback, identification of creative version switching events, and evaluation of rule operation benefits—can remain traceable, comparable, and recalculated under the same data structure. This reduces the risk of judgment bias and rule misapplication caused by mixed data sources, unclear time boundaries, and inconsistent field scales.

[0034] Specifically, the steps for constructing feedback production bucket-level samples based on preprocessed feedback production integrated data and performing conversion backfilling to obtain the backfilling conversion rate are as follows: Read the preprocessed feedback production integrated data, and jointly group various records according to the platform identifier, ad placement identifier, audience bucket identifier, creative version number, and statistical time bucket start timestamp. During the joint grouping process, the platform identifier is used as the platform affiliation boundary, the ad placement identifier as the resource placement affiliation boundary, the audience bucket identifier as the audience statistical caliber boundary, the creative version number as the creative affiliation boundary, and the statistical time bucket start timestamp as the bucket-level time boundary. Furthermore, a non-empty validation is performed on the grouping key field to remove missing affiliation fields. Record; summarize impressions, clicks, dwell time, bounces, conversions, and conversion amounts within each statistical time bucket. During the summarization process, verify whether records fall within the current statistical time bucket boundary using the duration of the statistical time bucket. Records that do not fall within the current statistical time bucket boundary are removed to prevent cross-bucket mixing. Maintain consistency of field names before and after summarization to ensure recalculation for subsequent comparisons. Generate bucket-level samples of feedback with the statistical time bucket as the smallest granularity, and generate a unique sample number for each bucket-level sample. During the generation process, perform concatenation encoding and additional verification using the platform identifier, ad placement identifier, audience bucket identifier, creative version number, and statistical time bucket start timestamp. A unique sample number is generated, enabling the unique sample number to be used to locate the corresponding bucket-level belonging boundary and time boundary. A bucket-level sample is created for each feedback. The start timestamp of the statistical time bucket and the feedback return delay are read, and the conversion backfill timestamp is calculated. During the calculation, the start and end intervals of the current statistical time bucket are determined by the start timestamp and duration of the statistical time bucket. The end timestamp of the current statistical time bucket is obtained by adding the start timestamp to the duration. The conversion backfill timestamp is obtained by adding the start timestamp to the feedback return delay. Records with a feedback return delay less than zero are removed to prevent the backfill direction from reversing. When the conversion backfill timestamp falls into the subsequent statistical... When creating a time bucket, the corresponding conversion count and conversion amount are transferred from the original statistical time bucket and accumulated to the target statistical time bucket. During the transfer process, the starting timestamp of the target statistical time bucket is recalculated based on the conversion backfill timestamp, and the target statistical time bucket is located. The conversion count and conversion amount corresponding to the feedback in the bucket-level sample in the original statistical time bucket are deducted and accumulated in the corresponding feedback in the bucket-level sample in the target statistical time bucket. If a negative value occurs after deduction, the negative value is corrected to zero to avoid damage to the statistical caliber. The conversion count and conversion amount after backfilling are obtained. After obtaining the conversion count after backfilling, the original field names are retained to ensure that the data names used in subsequent credibility evaluation of dissemination effect feedback are consistent.The number of exposures in the target statistical time bucket is read as the backfill exposure count. During the reading process, the unique sample number used to create the bucket-level samples in the feedback is used to locate the exposure count corresponding to the current statistical time bucket. If the exposure count is zero, it is replaced with a minor term to avoid division by zero in the ratio calculation. The ratio of the backfilled conversion count to the backfilled exposure count is calculated to obtain the backfilled conversion rate. This backfilled conversion rate is written to a derived field of the feedback bucket-level sample creation so that a backfilled conversion rate sequence can be constructed subsequently based on the start timestamp of the statistical time bucket.

[0035] In this implementation plan, feedback production data is solidified into uniquely identifiable feedback production bucket-level samples at the statistical time bucket granularity. A conversion backfilling mechanism is used to introduce the latency of the backhaul link into the correction process of the bucket-level statistical caliber. This ensures that the backfilling conversion rate maintains a stable correspondence with the number of exposures within the same statistical time bucket. Consequently, subsequent credibility assessments of dissemination effect feedback and material version switching response learning based on the backfilling conversion rate have higher comparability and consistency. This reduces the risk of misjudging version evolution relationships caused by cross-bucket drift of conversion events and improves the reusability of the judgment input for material version update rules under different advertising platform identifiers, ad placement identifiers, and audience bucket identifiers.

[0036] Specifically, the steps for evaluating the credibility of feedback in creating bucket-level samples based on feedback return latency are as follows: Using the platform identifier, ad placement identifier, and audience bucket identifier as joint grouping keys, and under the condition of the same statistical time bucket duration, a return conversion rate sequence is constructed based on the return conversion rate and in the order of the statistical time bucket start timestamps. During the construction process, only records with continuously increasing start timestamps of the statistical time buckets are retained, and records with reverse timestamps are removed to ensure that the return conversion rate sequence meets the monotonic time constraint. Furthermore, a sliding window statistical method is used to calculate the sliding window mean and standard deviation of the return conversion rate on the return conversion rate sequence. During the calculation, the sliding window width is limited to between three and fifteen consecutive statistical time buckets, and the sliding window step size is limited to one statistical time bucket. The window width is kept fixed to form a consistent local statistical caliber, and the sliding window mean is used to represent... The statistical background of the current joint group's backfill conversion rate within a local time range is analyzed. The sliding window standard deviation characterizes the natural fluctuation intensity within this statistical background, yielding the mean and standard deviation of the backfill conversion rate. The absolute value of the difference between the backfill conversion rate and the mean is calculated to obtain the backfill conversion rate offset. This offset quantifies the deviation of the current feedback bucket-level sample from the local statistical background of the joint group and eliminates the influence of the deviation direction on the deviation intensity measurement. The backfill conversion rate offset is divided by the sum of the standard deviation and the minterm to obtain the standardized offset value. This standardized offset value maps the offset to a dimensionless comparable scale under different platform identifiers, ad placement identifiers, and audience bucket identifiers. The minterm is a small but non-zero positive real number used to avoid numerical instability caused by division by zero during calculation, and its value range is [insert range here]. arrive The feedback latency is divided by the duration of the statistical time bucket. The negative of this ratio is used as the exponent, and an exponential operation is performed on the natural constant e to obtain the feedback latency attenuation term. This term utilizes the monotonically decreasing property of the exponential function to map the feedback latency to a bounded attenuation coefficient. This ensures that the feedback latency attenuation term continuously decreases as the feedback latency increases and maintains a smooth penalty strength in the high latency range, preventing the feedback latency from causing a jump in credibility near the threshold. The standardized offset value is multiplied by the feedback latency attenuation term to obtain the sample credibility assessment value. This value is used to simultaneously characterize the deviation of the backfill conversion rate from the statistical background and the degree of credibility attenuation caused by the feedback latency within the same numerical carrier, and serves as the input for subsequent selection of feedback to create a calibration sample set. The standardized offset value primarily reflects the degree of anomaly in the backfill conversion rate under the joint grouping local statistical background. When differences in material versions lead to real changes in the dissemination effect, the deviation of the backfill conversion rate from the mean backfill conversion rate is considered. The deviations typically remain in the same direction and are reproducible within consecutive statistical time buckets, thus ensuring a stable deviation structure for standardized offset values ​​within adjacent sliding windows. Feedback latency primarily reflects the risk intensity of conversion feedback migration across statistical time buckets. When the deviation is mainly caused by mis-bucket migration due to feedback latency, the deviation is more likely to be concentrated in statistical time buckets with larger feedback latency and exhibit drift as the feedback arrival position changes, making the same offset less statistically significant under high latency conditions. By continuously suppressing the offset intensity of high-latency buckets through a backhaul latency attenuation term, the backfill conversion rate offset formed under high latency conditions is given a lower confidence weight, while the offset formed under low latency conditions and consistent across windows receives a higher confidence weight. Thus, without changing the backfill conversion rate sequence construction rules, a separable numerical interval is created at the sample confidence assessment level between stable offsets caused by differences in real material versions and mis-bucket noise caused by feedback latency.

[0037] The specific formula for calculating the sample credibility assessment value is as follows:

[0038] ;

[0039] In the formula, This represents the sample reliability assessment value. Indicates the backfill conversion rate. This represents the average backfill conversion rate. This represents the standard deviation of the backfill conversion rate. Indicates minterms, This indicates the feedback transmission delay. This indicates the duration of the statistical time bucket.

[0040] In this implementation plan, the conversion rate is placed within a local statistical reference system with a fixed window caliber under the constraints of the platform identifier, ad placement identifier, and audience bucket identifier. The conversion rate offset and feedback transmission delay are jointly mapped to the same quantification scale for the credibility of dissemination effect feedback. This ensures that the sample credibility assessment value can maintain stable discriminative consistency under different conversion rate fluctuation intensities and different feedback transmission delay levels. This improves the ability to distinguish the usability of feedback bucket-level samples, reduces the probability of misjudgment caused by short-term abnormal spikes, and ensures that the subsequent identification of material version switching events and generation of material version update rules based on the feedback production calibration sample set have a more stable input foundation.

[0041] Specifically, the steps for selecting feedback to create a calibration sample set based on the evaluation results are as follows: Using the material version number as the grouping key, read the sample reliability evaluation values ​​in order of the statistical time bucket start timestamp. During the reading process, perform time continuity verification on feedback bucket-level samples under the same material version number and remove duplicate records of the statistical time bucket start timestamp to avoid stability judgment bias caused by duplicate statistical records. Mark feedback bucket-level samples below the reliability threshold within N consecutive sliding windows as unstable feedback and remove them, where N is limited to two to eight consecutive sliding windows. During the marking process, use the statistical time bucket start timestamp as the sliding window traversal index, perform threshold comparison on each sample reliability evaluation value within each sliding window, and count the number of occurrences below the reliability threshold. When all sample reliability evaluation values ​​within a sliding window are below the reliability threshold, trigger the unstable feedback marking and write the feedback bucket-level samples covered by that sliding window into the removal set. Otherwise, it is marked as stable feedback and retained. During the retention process, bucket-level samples of feedback marked as stable feedback are written into the retention set and the corresponding unique sample number is recorded for subsequent backtracking. The backfill conversion rate, sample credibility assessment value and feedback production integrated data corresponding to the retained feedback production bucket-level samples are extracted and correlated. During the correlation process, the unique sample number is used as the correlation key to perform one-to-one matching verification between the backfill conversion rate and the sample credibility assessment value, and the verification failure record is removed to ensure that the backfill conversion rate, sample credibility assessment value and feedback production integrated data have consistent belonging boundaries under the same statistical time bucket start timestamp. A feedback production calibration sample set is formed. During the formation process, the fields of backfill conversion rate, sample credibility assessment value, statistical time bucket start timestamp, statistical time bucket duration, delivery platform identifier, ad position identifier, audience bucket identifier and material version number are completely retained to support the subsequent identification of material version switching events and quantitative evaluation of rule operation benefits.

[0042] In this implementation scheme, a continuous sliding window stability judgment is performed on the sample credibility assessment value along the dimension of the material version number. A strong consistency association is established between the backfill conversion rate, the sample credibility assessment value, and the feedback production integrated data using a unique sample number. This enables the feedback production calibration sample set to maintain traceable sample convergence characteristics along the time evolution link of the material version number, thereby reducing sample set fluctuations caused by short-term abnormal backloads, improving the sensitivity of subsequent material version switching event identification to real version change signals, and ensuring that the material version update rule generation process has a more stable evidentiary basis.

[0043] Specifically, the steps for identifying material version switching events based on feedback-based calibration sample sets and forming reliable material version switching response records based on the backfill conversion rates before and after the switch are as follows: Read the feedback-based calibration sample set, sort the feedback bucket-level samples according to the material version number and the start timestamp of the statistical time bucket. During the sorting process, use the start timestamp of the statistical time bucket as the primary key to perform ascending order and use the material version number as the secondary key to maintain consistency in version attribution under the same statistical time bucket start timestamp, thus avoiding version switching boundary drift caused by time reordering; when the material version number differs between adjacent statistical time buckets, mark the corresponding statistical time bucket start timestamp as a material version switching event. During the tagging process, the start timestamps of two adjacent statistical time buckets are recorded simultaneously, and the time interval between them is verified to be consistent with the duration of the statistical time buckets to ensure that the material version switch event occurs at the boundary of consecutive buckets. The material version numbers before and after the switch are also recorded. During the recording process, the material version number before the switch is written to the preceding version field of the switch event, and the material version number after the switch is written to the following version field of the switch event. The start timestamp of the corresponding statistical time bucket is written to the switch event time field to form a traceable event index. For each material version switch event, the backfill conversion rate and sample availability within the M consecutive statistical time buckets before and after the switch are extracted. The reliability assessment value, where M is limited to a range of three to twelve consecutive statistical time buckets, is calculated by backtracking M statistical time buckets backward and forward M statistical time buckets from the switching event time field as the center, respectively, to create bucket-level samples and construct pre-switching and post-switching interval sample sequences. Simultaneously, the continuity of the starting timestamps of the statistical time buckets within the interval is verified, and switching events with missing buckets are removed to avoid interference with the interval mean calculation. When the reliability assessment values ​​of all samples corresponding to the statistical time buckets before and after the switch are higher than the reliability threshold, the interval mean of the backfill conversion rate before and after the switch is calculated. The mean value of the backfill conversion rate before the switch is obtained by summing the backfill conversion rates within the sample sequence of the interval before the switch and dividing by the number of samples in the corresponding interval. The mean value of the backfill conversion rate after the switch is obtained by summing the backfill conversion rates within the sample sequence of the interval after the switch and dividing by the number of samples in the corresponding interval. This means that the interval mean can represent the stable level of the propagation effect before and after the switch event. The difference between the mean value of the backfill conversion rate after the switch and the mean value of the backfill conversion rate before the switch is calculated to obtain the change in backfill conversion rate. The change in backfill conversion rate is used to quantify the intensity of the change in propagation effect caused by the switch of material version number under the same platform identifier, the same ad position identifier, and the same audience bucket identifier, and to reduce the impact of single bucket fluctuations on the judgment.The change in conversion rate is correlated with the version numbers of the materials before and after the switch. During this correlation, the switch event time field is used as the correlation time index, and the material version numbers before and after the switch are used as the correlation version index. The change in conversion rate is written to the change field of the correlation record. This forms a reliable material version switch response record. During the record's formation, the switch event time field, the material version number before the switch, the material version number after the switch, the average conversion rate interval before the switch, the average conversion rate interval after the switch, and the change in conversion rate are retained to support subsequent aggregation statistics and evolutionary gain quantification of material version update rules.

[0044] In this implementation scheme, by limiting the change of material version number to the continuous boundary of adjacent statistical time buckets, and using the average of the backfill conversion rate intervals of a fixed length before and after the switch as the quantification carrier of the switch response, the reliable material version switch response record can suppress the impact of occasional fluctuations in a single statistical time bucket on a time scale, and strengthen the correspondence stability between the change of material version number and the change of backfill conversion rate on an evidentiary scale. This improves the identification accuracy of real material version evolution signals, reduces the risk of misjudging short-term traffic disturbances as material version switch benefits, and ensures that the generation of subsequent material version update rules has a more repeatable statistical basis.

[0045] Specifically, the steps for generating material version update rules based on the aggregated statistics of trusted material version switching response records are as follows: Using the combination of version numbers before and after the material version switch as the grouping key, aggregate statistics are performed on the trusted material version switching response records. During the aggregation statistics process, the records are merged using the material version number before and after the switch as a composite index, and the switching event time field within the same group is sorted in ascending order to maintain the time traceability of the evolution chain. When the same material version switching relationship occurs K times, and the signs of the changes in backfill conversion rate corresponding to the K occurrences are consistent, the current material version switching relationship is marked as a stable version evolution relationship. The value of K is limited to between three and twelve occurrences. During the process, the sign of the change in backfill conversion rate for each trusted material version switching response record within the group is determined, and a change direction marker is generated. The number of dominant direction markers is counted and compared with the number of trusted material version switching response records within the group to obtain the proportion of consistency in the same direction. When the proportion of consistency in the same direction is equal to 1 and the number of trusted material version switching response records within the group is not less than K, a stable version evolution relationship marker is triggered. The arithmetic mean of all backfill conversion rate changes corresponding to the stable version evolution relationship is taken to obtain the evolution gain value. The material version number before the switch, the material version number after the switch, and the evolution gain value are associated to generate a material version update rule. The material version update rule includes a rule number and the material version number before the switch. The following parameters are included: Post-switch material version number, Evolutionary Gain value, Rule Evidence Count, Consistency Ratio, Switching Response Time Span, Rule Generation Time Field, Applicable Platform Identifier Set, Applicable Ad Placement Identifier Set, Applicable Audience Bucket Identifier Set, and Statistical Time Bucket Duration. The rule number is generated by concatenating and encoding the pre-switch material version number, post-switch material version number, and rule generation time field, with an added checksum. The Rule Evidence Count is the number of credible material version switching response records within the group. The Consistency Ratio is the ratio of the number of change direction markers corresponding to the dominant direction to the number of credible material version switching response records within the group. The Switching Response Time Span is the time span between the earliest and latest switching event times within the group. The difference is defined as follows: the applicable platform identifier set is the identifier sequence obtained by deduplicating the platform identifiers corresponding to the trusted material version switch response records within the group and arranging them in lexicographical order; the applicable ad slot identifier set is the identifier sequence obtained by deduplicating the ad slot identifiers corresponding to the trusted material version switch response records within the group and arranging them in lexicographical order; the applicable audience bucket identifier set is the identifier sequence obtained by deduplicating the audience bucket identifiers corresponding to the trusted material version switch response records within the group and arranging them in lexicographical order; and the duration of the statistical time bucket is the duration corresponding to the maximum occurrence of the statistical time bucket duration corresponding to the trusted material version switch response records within the group. This is used to constrain the material version update rules to maintain consistent statistical standards during the matching phase of the material version update task to be verified.

[0046] In this implementation plan, trusted material version switching response records are precipitated into reusable business rules. Within these business rules, the material version number before switching, the material version number after switching, the evolution gain value, the number of recurrences, and the duration of the statistical time bucket are fixed. This allows subsequent material version updates to be reused by manually comparing switching records one by one. At the same time, it enables version management and audit tracking of rule activation and deactivation based on the same business rule object, thereby improving the maintainability and manageability of the material version evolution process and reducing the risk of state distortion caused by rule dispersion during rule iteration.

[0047] Specifically, the steps for triggering a material version switch based on the material version update rules and quantitatively evaluating the benefits of the switched rules to obtain the rule benefit evaluation value are as follows: Figure 4As shown, the process reads the creative version update rules and matches them according to the platform identifier, ad placement identifier, audience bucket identifier, and creative version switching relationship. During the matching process, the platform identifier, ad placement identifier, audience bucket identifier, and the creative version switching number before and after the switch are used as the joint matching key. The process also verifies that the duration of the statistical time bucket recorded in the creative version update rules matches the duration of the statistical time bucket in the current delivery scenario to ensure consistency in the calculation of the rule's revenue assessment value. Finally, the matched creative version update rules are associated with the currently running creative version number. During this association process, the currently running creative version number is compared with the creative version number before the switch in the creative version update rules. Consistency verification is performed. When the consistency verification passes, an associated flag is written and an associated timestamp is recorded to avoid repeated switching triggers under conditions where the version has changed. A task for updating the creative version to be verified is generated. During the generation process, the following fields are written: platform identifier, ad placement identifier, audience bucket identifier, creative version number before switching, creative version number after switching, duration of the statistical time bucket, and rule number, for unique positioning within the subsequent verification interval. For each task for updating the creative version to be verified, without changing the platform identifier, ad placement identifier, and audience bucket identifier, the corresponding creative version switching operation is triggered. During the triggering process, the trigger timestamp of the switching operation and the creative version number after switching are written to the operation record, and the operation record is linked to the task for verification. The rule number of the verification material version update task is associated to form a traceable verification link; and the corresponding backfill conversion rate is calculated for each of the W consecutive statistical time buckets after the trigger, where the value of W is limited to three to twelve consecutive statistical time buckets. During the calculation, the number of conversions and the number of exposures after backfilling in each statistical time bucket after the trigger are read in the order of the start timestamp of the statistical time bucket, and the ratio is calculated to obtain the backfill conversion rate. The backfill conversion rate is used to characterize the dissemination effect level of the material version after the switch under the conditions of fixed delivery platform identifier, fixed ad position identifier, and fixed audience bucket identifier; at the same time, the backfill conversion rate of the W consecutive statistical time buckets before the material version switch is extracted and the average is calculated to obtain the backfill conversion rate benchmark. The mean is calculated by backtracking the start timestamps of W statistical time buckets based on the switching operation trigger timestamp and reading the corresponding conversion rate. The baseline mean of the conversion rate is used to characterize the stable dissemination effect of the material version before the switch under the same platform identifier, the same ad slot identifier, and the same audience bucket identifier, so that subsequent revenue evaluation can be compared with the version switch as the only variable. For the conversion rate of the conversion rate in the W consecutive statistical time buckets after the update, the difference between it and the baseline mean of the conversion rate is calculated to obtain the conversion rate gain term for each statistical time bucket. The conversion rate gain term is used to quantify the increase in dissemination effect after the switch relative to before the switch and maintains a positive or negative direction to characterize the trend of revenue improvement and revenue decline.The feedback latency corresponding to each statistical time bucket is compared to the duration of the statistical time bucket. The square of the ratio is then added to a constant, and the reciprocal is taken to obtain the latency risk attenuation term. The ratio operation maps the feedback latency to a dimensionless ratio consistent with the duration of the statistical time bucket for comparison across different durations. The squaring operation applies a non-linear penalty to larger feedback latency ratios to enhance the ability to distinguish high-latency conditions. The constant provides a lower bound for the denominator to prevent distortion of the weighted benefit term when the ratio approaches zero. The reciprocal is used to map high latency to low weight and low latency to high weight, thus improving the rule-based benefit assessment. The valuation is more sensitive to backhaul stability during calculation; the backfill conversion rate gain term is multiplied by the corresponding latency risk attenuation term to obtain the weighted revenue term for each statistical time bucket. The weighted revenue term is used to suppress the revenue contribution of statistical time buckets with high feedback backhaul latency and avoid false gains caused by high latency backhaul, while maintaining the original scale of revenue increment. The weighted revenue terms within W consecutive statistical time buckets are summed and the arithmetic mean is taken to obtain the rule operation revenue evaluation value. The rule operation revenue evaluation value is used to characterize the net revenue level brought by material version switching within the observation window of W consecutive statistical time buckets and serves as a quantitative input for subsequent material version update rule effectiveness and rollback determination.

[0048] The specific formula for calculating the benefit assessment value of rule operation is as follows:

[0049] ;

[0050] In the formula, This represents the assessed value of the benefits of running the rule. This represents the backfill conversion rate within the nth statistical time bucket after the trigger. This represents the baseline average of the backfill conversion rate. This represents the feedback latency within the nth statistical time bucket. This indicates the duration of the statistical time bucket. This indicates the number of time buckets used for statistics.

[0051] In this embodiment, Table 1 shows the statistical results of the rule operation benefit evaluation values ​​and related input data for the five material version update rules in the verification window. Specifically: Material version update rule R1: The number of statistical time buckets is 3, the duration of the statistical time bucket is 300, the baseline average conversion rate is 0.020, the conversion rates of the first to third statistical time buckets after triggering are 0.035, 0.034 and 0.036 respectively, the feedback transmission latency of the first to third statistical time buckets is 15, 20 and 10 respectively, and the corresponding rule operation benefit evaluation value is 0.0150. Material version update rule R2: The number of statistical time buckets is 3, the duration of the statistical time bucket is 300, the baseline average conversion rate is 0.030, the conversion rates in the first to third statistical time buckets after triggering are 0.043, 0.046, and 0.044 respectively, the feedback latency in the first to third statistical time buckets is 25, 30, and 20 respectively, and the corresponding rule operation benefit evaluation value is 0.0142. Material version update rule R3: The number of statistical time buckets is 3, the duration of the statistical time bucket is 300, the baseline average conversion rate is 0.018, the conversion rates in the first to third statistical time buckets after triggering are 0.031, 0.033, and 0.030 respectively, the feedback latency in the first to third statistical time buckets is 12, 18, and 15 respectively, and the corresponding rule operation benefit evaluation value is 0.0133. Material version update rule R4: The number of statistical time buckets is 3, the duration of each statistical time bucket is 300, the baseline average conversion rate is 0.025, and the conversion rates within the first to third statistical time buckets after triggering are 0.039, 0.041, and 0.040, respectively. The feedback latency within the first to third statistical time buckets is 35, 40, and 30, respectively, resulting in a rule operation benefit assessment value of 0.0148. Material version update rule R5: The number of statistical time buckets is 3, the duration of each statistical time bucket is 300, the baseline average conversion rate is 0.040, and the conversion rates within the first to third statistical time buckets after triggering are 0.058, 0.060, and 0.057, respectively. The feedback latency within the first to third statistical time buckets is 20, 25, and 22, respectively, resulting in a rule operation benefit assessment value of 0.0182. The data in Table 1 is used to quantify and compare the coupled impact of the conversion rate gain and feedback transmission delay risk of different material version update rules within the same verification window. It also provides a direct input basis for comparing the rule operation benefit evaluation value with the operation benefit threshold to determine the effectiveness, rollback, and operation cancellation of the material version update rule.

[0052] Table 1. Data Table of Rule Operation Benefits Assessment

[0053] like Figure 3The chart shows a comparison of the rule performance evaluation values ​​for five material version update rules. The bar chart uses the material version update rule number as the x-axis and the rule performance evaluation value as the y-axis. The bar height represents the rule performance evaluation value obtained by comprehensively quantifying the conversion rate gain and latency risk attenuation term after triggering the material version switch, based on continuous statistical time bucket backfilling. Blue bars represent material version update rules with performance evaluation values ​​not less than the performance benefit threshold, while orange bars represent those with performance evaluation values ​​less than the performance benefit threshold. The black dashed line marks the performance benefit threshold position for intuitive comparison. Specifically, the performance evaluation values ​​of material version update rules R1, R4, and R5 are all above the performance benefit threshold, corresponding to the blue bars, and can be used as input for determining effective version evolution rules; the performance evaluation values ​​of material version update rules R2 and R3 are below the performance benefit threshold, corresponding to the orange bars, and can be used as input for determining rollback and failure handling. Figure 3 By using a visualization method that includes threshold dashed line comparison, two types of bar color differentiation, and bar top numerical annotation, the revenue quantification results of different material version update rules within the same verification window are presented in a structured manner. This provides a directly referable basis for determining the effectiveness, rollback, and cancellation of material version update rules based on the revenue evaluation value in subsequent operations.

[0054] In this implementation plan, by keeping the platform identifier, ad placement identifier, and audience bucket identifier unchanged, a peer-to-peer comparison is conducted before and after the creative version switch. A recalculated revenue baseline is constructed using the average conversion rate benchmark. At the same time, the feedback transmission delay is embedded into the quantification process of the rule operation revenue evaluation value in a non-linear decay manner. This allows the rule operation revenue evaluation value to stably represent the net revenue contribution brought by the creative version switch under a unified statistical time bucket duration caliber, reducing the judgment of inflated revenue caused by feedback lag and improving the reliability of the effective determination and the consistency of rollback determination of the creative version update rule in the actual delivery process.

[0055] Specifically, the steps for determining the effectiveness, rollback, and cancellation of a creative version update rule based on its operational revenue assessment value are as follows: The operational revenue assessment value is compared with the operational revenue threshold. During the comparison, the rule number corresponding to the creative version update task to be verified is read and bound to the comparison timestamp to distinguish the judgment records of the same creative version update rule in different campaign periods. When the operational revenue assessment value is not less than the operational revenue threshold, the current creative version update rule is deemed to have passed verification. Upon successful verification, the operational revenue assessment value is written into the rule judgment record, and the rule judgment record is associated with the campaign platform identifier, ad placement identifier, and audience bucket identifier, enabling subsequent operational monitoring to be conducted within the same campaign platform. The criteria for reusing the rules are determined under the constraints of platform identifiers, identical ad placement identifiers, and identical audience bucket identifiers. The current material version update rule is marked as an effective version evolution rule, and its effective timestamp is recorded. During the recording process, the effective timestamp is bound to the material version number after the switch and written into the effective status record to solidify the baseline version after effectiveness. Otherwise, the current material version update rule is determined to have failed verification. When the determination fails, the rule's operational benefit assessment value is written into the failure determination record, and the failure timestamp is recorded. The process reverts to the material version number before the switch. During the revert, the material version number before the switch in the material version update task to be verified is read, and the material version switch operation is executed. The revert operation trigger timestamp is written into the revert record, and the revert record is associated with the rule number. To facilitate tracing of unverified trigger links; and to mark the current material version update rule as an invalid rule, during the marking process, the invalid status is bound to the platform identifier, ad slot identifier, and audience bucket identifier and written into the rule status table to avoid repeated triggering of the same invalid rule under the same platform identifier, ad slot identifier, and audience bucket identifier; continuous operation monitoring is performed on the effective version evolution rule, and during the operation monitoring process, the return conversion rate and sample credibility assessment value corresponding to the effective version evolution rule are continuously read according to the start timestamp of the statistical time bucket, while keeping the duration of the statistical time bucket unchanged. When the direction of change of the return conversion rate is reversed in the subsequent N consecutive statistical time buckets, or the sample credibility assessment value is reversed in the subsequent N consecutive statistical time buckets, the system will take action. When all values ​​within a statistical time bucket are below the confidence threshold, the corresponding effective version evolution rule is automatically revoked. The value of N is limited to three to twelve consecutive statistical time buckets. When the direction of the backfill conversion rate change is reversed, the backfill conversion rate of the previous statistical time bucket is used as a reference to calculate the difference between the backfill conversion rate of the current statistical time bucket and determine the sign of the difference. When the sign of the difference in N consecutive statistical time buckets is inconsistent with the sign of the change in backfill conversion rate recorded during the generation stage of the effective version evolution rule, a reversal judgment is triggered. When the confidence assessment value of a detected sample is continuously below the confidence threshold, the confidence assessment value of the sample is compared with the confidence threshold one by one according to the starting timestamp of the statistical time bucket and the number of times it is below the threshold is accumulated. When the number of times it is below the threshold reaches N, a revocation judgment is triggered.After the revocation decision is triggered, a revocation timestamp is recorded, and a revocation reason marker is written into the revocation record to distinguish between reverse revocation and low-confidence revocation. The media version is then restored to its pre-revocation state. During the restoration process, the pre-effective media version number is read from the effective version evolution rule's effective status record, and a media version switching operation is performed. The restoration operation trigger timestamp is written into the restoration record, and the restoration record is associated with the rule number.

[0056] In this implementation plan, an integrated traceability link is established by combining the judgment result of the rule operation revenue assessment value with the rule number, the platform identifier, the ad slot identifier, the audience bucket identifier, and the effective timestamp. After the rule takes effect, the consistency of the direction of change of the backfill conversion rate and the stability of the sample credibility assessment value are used as the revocation criteria under the constraint of continuous statistical time bucket. This ensures that the material version update rule has a verifiable state loop and recalculated revocation trigger conditions within the operation cycle, thereby improving the operational security of the material version update rule under dynamic traffic fluctuations, reducing the risk of false continuation caused by short-term rebound in revenue, and ensuring that the evolution process of the material version number can be controlled and stable under the constraint of traceable records.

[0057] like Figure 2 As shown, the second aspect of this invention provides a self-learning optimization system for advertising production parameters based on dissemination effect feedback, including: a feedback data collection and preprocessing module, a cross-channel calibration and traceability module, a material version switching response learning module, and an update rule verification and controlled effectiveness module. The feedback data collection and preprocessing module collects integrated feedback production data for advertising and performs time alignment, smoothing and noise reduction, anomaly removal, missing data completion, and standardization on the integrated feedback production data to generate preprocessed integrated feedback production data. The cross-channel calibration and traceability module constructs bucket-level samples for feedback production based on the preprocessed integrated feedback production data, performs conversion backfilling to obtain the backfill conversion rate, and combines feedback feedback. The latency evaluation assesses the propagation effect and credibility of feedback in creating bucket-level samples, and selects a calibration sample set based on the evaluation results. The media version switching response learning module identifies media version switching events based on the calibration sample set and generates credible media version switching response records based on the backfill conversion rate before and after the switch. It then generates media version update rules based on the aggregation and statistics of these credible media version switching response records. The controlled effectiveness verification module for update rules triggers media version switching operations based on the media version update rules, quantifies the benefit of the rule operation after the switch, obtains a rule operation benefit evaluation value, and determines the effectiveness, rollback, or cancellation of the media version update rule based on the rule operation benefit evaluation value.

[0058] This implementation plan uses integrated feedback data as a unified data carrier to connect the entire chain of preprocessing, calibration, reliable response learning, and controlled effectiveness determination. It forms progressive quantitative constraints with backfill conversion rate, sample credibility assessment value, and rule operation benefit assessment value. At the same time, the material version update rules are solidified in the form of business rules with constraints on applicable platform identifiers, applicable ad slot identifiers, and audience bucket identifiers. This makes the evolution process of material version numbers have a traceable evidence loop and a recalculated judgment basis, thereby improving the consistency of update decisions under the condition of different identifiers across different platform, reducing the risk of mislearning caused by aggregated backhaul, and ensuring that the self-learning optimization of ad production parameters can achieve controllable iteration under stable feedback constraints.

[0059] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0060] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A self-learning optimization method for advertising production parameters based on feedback on dissemination effects, characterized in that, Includes the following steps: S1 collects integrated data on advertising feedback and performs time alignment, noise reduction, anomaly removal, missing data completion and standardization on the integrated data to generate pre-processed integrated data on advertising feedback. S2, construct feedback production bucket-level samples based on the preprocessed feedback production integrated data, perform conversion backfilling to obtain the backfilling conversion rate, and evaluate the propagation effect and feedback credibility of the feedback production bucket-level samples in combination with the feedback back transmission latency, and select the feedback production calibration sample set based on the evaluation results. S3, based on feedback, create a calibration sample set to identify material version switching events, and based on the backfill conversion rate before and after the switch, form a reliable material version switching response record, and based on the aggregation statistics of the reliable material version switching response record, generate material version update rules; S4 triggers a material version switching operation based on the material version update rules, and quantitatively evaluates the benefits of the rule operation after the switch to obtain the rule operation benefit evaluation value. Based on the rule operation benefit evaluation value, it determines whether the material version update rule is effective, rolled back, or canceled.

2. The self-learning optimization method for advertising production parameters based on communication effect feedback as described in claim 1, characterized in that: The specific steps for collecting integrated feedback data from advertisements and performing time alignment, noise reduction, anomaly removal, missing data completion, and standardization on the integrated feedback data to generate preprocessed integrated feedback data are as follows: Real-time collection of integrated data for ad feedback production. This integrated data includes collection timestamp, platform identifier, ad placement identifier, material version number, audience bucket identifier, statistical time bucket start timestamp, statistical time bucket duration, number of impressions, number of clicks, dwell time, number of bounces, number of conversions, conversion amount, feedback transmission latency, material file length, material frame rate, material bitrate, screen resolution width, screen resolution height, first frame presentation duration, camera transition interval, total number of subtitle characters, single subtitle dwell time, starting time of selling point appearance, duration of selling point appearance, and the proportion of the main subject to the screen area. For the collected feedback production integrated data, the records are sorted according to the collection timestamp; a linear interpolation time alignment algorithm is used to map the multi-source asynchronous feedback records to a unified time axis to construct a unified time axis for feedback production; and a Kalman filter algorithm is used to smooth the feedback production integrated data. The integrated feedback production data is analyzed using an interquartile range anomaly detection algorithm to identify and remove abnormal statistical records. The missing segments formed after removal are then filled using a forward-filling and backward-filling fusion completion algorithm. The Z-score standardization algorithm is used to perform numerical standardization on the integrated feedback production data to eliminate dimensional differences between different physical quantities.

3. The self-learning optimization method for advertising production parameters based on communication effect feedback as described in claim 1, characterized in that: The specific steps for constructing feedback bucket-level samples based on preprocessed feedback data and performing conversion backfilling to obtain the backfilling conversion rate are as follows: Read the pre-processed feedback production integrated data, and jointly group various records according to the platform identifier, ad placement identifier, audience bucket identifier, material version number and statistical time bucket start timestamp. In each statistical time bucket, summarize the number of exposures, clicks, dwell time, bounces, conversions and conversion amount, and generate feedback production bucket-level samples with the statistical time bucket as the smallest granularity, and generate a unique sample number for each feedback production bucket-level sample. For each feedback, create a bucket-level sample, read the start timestamp of the statistical time bucket and the feedback feedback delay, and calculate the conversion backfill timestamp. When the conversion backfill timestamp falls into the subsequent statistical time bucket, the corresponding number of conversions and conversion amount are transferred from the original statistical time bucket and accumulated to the target statistical time bucket to obtain the number of conversions and conversion amount after backfilling; the number of exposures in the target statistical time bucket is read as the number of exposures after backfilling, and the ratio of the number of conversions after backfilling to the number of exposures after backfilling is calculated to obtain the backfilling conversion rate.

4. The self-learning optimization method for advertising production parameters based on communication effect feedback as described in claim 1, characterized in that: The specific steps for evaluating the propagation effect and reliability of feedback in creating bucket-level samples by combining feedback return latency are as follows: Using the platform identifier, ad slot identifier, and audience bucket identifier as the joint grouping key, and under the condition of the same statistical time bucket duration, a backfill conversion rate sequence is constructed based on the backfill conversion rate in the order of the start timestamp of the statistical time bucket. The sliding window statistical method is then used on the sequence to calculate the sliding window mean and sliding window standard deviation of the backfill conversion rate, thus obtaining the backfill conversion rate mean and backfill conversion rate standard deviation. Calculate the absolute value of the difference between the backfill conversion rate and the mean backfill conversion rate to obtain the backfill conversion rate offset; divide the backfill conversion rate offset by the sum of the standard deviation of the backfill conversion rate and the minima to obtain the standardized offset value. Divide the feedback transmission delay by the duration of the statistical time bucket, use the negative of the resulting ratio as the exponent, and perform an exponential operation on the natural constant e to obtain the transmission delay attenuation term; multiply the standardized offset value by the transmission delay attenuation term to obtain the sample reliability assessment value.

5. The self-learning optimization method for advertising production parameters based on communication effect feedback as described in claim 1, characterized in that: The specific steps for creating a calibration sample set based on the evaluation results are as follows: Using the material version number as the grouping key, the sample credibility assessment value is read in the order of the start timestamp of the statistical time bucket. Feedback production bucket-level samples that are below the credibility threshold within N consecutive sliding windows are marked as unstable feedback and removed; otherwise, they are marked as stable feedback and retained. The backfill conversion rate, sample credibility assessment value and feedback production integrated data corresponding to the retained feedback production bucket-level samples are extracted and correlated to form a feedback production calibration sample set.

6. The self-learning optimization method for advertising production parameters based on communication effect feedback according to claim 1, characterized in that: The specific steps for creating a calibration sample set based on feedback to identify material version switching events, and forming a reliable material version switching response record based on the backfill conversion rate before and after the switch are as follows: Read feedback to create a calibration sample set, sort the feedback to create bucket-level samples according to the material version number and the start timestamp of the statistical time bucket, and when the material version number is different between adjacent statistical time buckets, mark the start timestamp of the corresponding statistical time bucket as a material version switch event, and record the material version number before and after the switch. For each media version switching event, the backfill conversion rate and sample credibility assessment value are extracted from the M consecutive statistical time buckets before the switch and the M statistical time buckets after the switch. When the credibility assessment values ​​of all samples corresponding to the statistical time buckets before and after the switch are higher than the credibility threshold, the interval mean of the backfill conversion rate before and after the switch is calculated respectively. The difference between the interval mean of the backfill conversion rate after the switch and the interval mean of the backfill conversion rate before the switch is calculated to obtain the change in backfill conversion rate. The change in backfill conversion rate is then associated with the media version number before and after the switch to form a credible media version switching response record.

7. The self-learning optimization method for advertising production parameters based on communication effect feedback according to claim 1, characterized in that: The specific steps for generating material version update rules based on aggregated statistics of trusted material version switching response records are as follows: Using the combination of version numbers before and after the material version switch as the grouping key, aggregate and statistically analyze the reliable material version switch response records; when the same material version switch relationship occurs K times, and the signs of the changes in the backfill conversion rate corresponding to the K times are consistent, mark the current material version switch relationship as a stable version evolution relationship; The evolution gain value is obtained by taking the arithmetic mean of all backfill conversion rate changes corresponding to the stable version evolution relationship. The material version number before the switch and the material version number after the switch are associated with the evolution gain value to generate material version update rules.

8. The self-learning optimization method for advertising production parameters based on communication effect feedback according to claim 1, characterized in that: The specific steps for triggering a material version switching operation based on the material version update rule, and for quantitatively evaluating the benefit of the rule operation after the switch to obtain the rule operation benefit evaluation value are as follows: Read the material version update rules, match the material version update rules according to the platform identifier, ad slot identifier, audience bucket identifier and material version switching relationship, and associate the matched material version update rules with the currently running material version number to generate a material version update task to be verified; For each material version update task to be verified, without changing the platform identifier, ad slot identifier, and audience bucket identifier, the corresponding material version switch operation is triggered, and the corresponding return conversion rate is calculated for each of the W consecutive statistical time buckets after the trigger. At the same time, the return conversion rate of the W consecutive statistical time buckets before the material version switch is extracted and the average is calculated to obtain the baseline average return conversion rate. For the backfill conversion rate within W consecutive statistical time buckets after the update, calculate the difference between it and the baseline mean of the backfill conversion rate to obtain the backfill conversion rate gain term for each statistical time bucket; calculate the ratio between the feedback backhaul latency corresponding to each statistical time bucket and the duration of the statistical time bucket, square the obtained ratio, add it to a constant, and take the reciprocal to obtain the latency risk attenuation term; multiply the backfill conversion rate gain term by the corresponding latency risk attenuation term to obtain the weighted benefit term for each statistical time bucket; The weighted revenue items within W consecutive statistical time buckets are summed and their arithmetic average is taken to obtain the rule operation revenue evaluation value.

9. The self-learning optimization method for advertising production parameters based on communication effect feedback according to claim 1, characterized in that: The specific steps for determining the effectiveness, rollback, and cancellation of material version update rules based on the rule-based revenue assessment value are as follows: The rule's operational benefit assessment value is compared with the operational benefit threshold. If the rule's operational benefit assessment value is not less than the operational benefit threshold, the current material version update rule is determined to have passed verification. The current material version update rule is marked as an effective version evolution rule, and the effective timestamp is recorded. Otherwise, the current material version update rule is determined to have failed verification. The rule is rolled back to the material version number before the switch, and the current material version update rule is marked as an invalid rule. The evolution rules of the effective version are continuously monitored. When the direction of the change in the backfill conversion rate is reversed in the subsequent N consecutive statistical time buckets, or when the sample confidence assessment value is lower than the confidence threshold in the N consecutive statistical time buckets, the corresponding effective version evolution rule is automatically revoked and the material version is restored to the state before revocation.

10. A self-learning optimization system for advertising production parameters based on feedback on dissemination effects, characterized in that: include: The module includes a feedback data acquisition and preprocessing module, a cross-channel caliber calibration and traceability module, a material version switching response learning module, and an update rule verification and controlled effectiveness module. The feedback data acquisition and preprocessing module is used to collect integrated feedback production data of advertisements, and to perform time alignment, smoothing and noise reduction, anomaly removal, missing data completion and standardization on the integrated feedback production data to generate preprocessed integrated feedback production data. The cross-channel calibration and traceability module is used to construct feedback production bucket-level samples based on preprocessed feedback production integrated data, perform conversion backfilling to obtain backfilling conversion rate, and evaluate the propagation effect and feedback credibility of feedback production bucket-level samples in combination with feedback back transmission delay, and select feedback production calibration sample set based on evaluation results. The material version switching response learning module is used to create a calibration sample set based on feedback to identify material version switching events, and to form a reliable material version switching response record based on the backfill conversion rate before and after the switching. Based on the aggregation statistics of the reliable material version switching response record, material version update rules are generated. The update rule verification controlled effectiveness module is used to trigger a material version switching operation based on the material version update rule, and to quantitatively evaluate the rule operation benefits after the switch to obtain a rule operation benefit evaluation value. Based on the rule operation benefit evaluation value, the effectiveness, rollback, and operation cancellation of the material version update rule are determined.

Images