A delivery method based on the correlation between delivery strategy and creative matching.

Abstract

This invention discloses a campaign placement method based on the correlation between campaign placement strategy and creative material matching, relating to the field of advertising placement technology. The method includes acquiring strategy time-series data of target creative materials during historical campaign placements, wherein the strategy time-series data includes a sequence of changes in at least one campaign strategy parameter over time; detecting abrupt change points in the strategy time-series data to identify the time-series change patterns of the at least one campaign strategy parameter, wherein the time-series change patterns include the location, magnitude, and direction of the abrupt change points; this invention, through precise detection of abrupt change points in strategy time-series data, combined with the quantitative calculation of matching values ​​using the creative material's emotional tension index, information density index, and inducement intensity index, can accurately determine the cumulative response status of the creative material and delineate risk time windows, predicting the timing of conversion anomalies in advance.

Description

Technical Field

[0001] This invention belongs to the field of advertising delivery technology, specifically a delivery method based on the correlation between delivery strategy and creative material matching. Background Technology

[0002] Existing methods for adjusting digital advertising strategies cannot predict potential conversion anomalies due to varying sensitivities of different types of creatives to strategy changes after sudden shifts in strategy parameters (such as budget adjustments or audience expansion). This results in reactive interventions only being implemented after data declines.

[0003] Specifically, existing technologies lack the ability to dynamically perceive the response status of materials after a sudden change in strategy parameters: on the one hand, they cannot predict whether materials have entered a state of cumulative risk and the lag window between the mutation point and the occurrence of conversion anomalies based on the matching relationship between the intensity of the mutation and the material's own attributes (such as emotional tension); on the other hand, they cannot perform differentiated pre-intervention for different materials within this risk window, resulting in delayed adjustments, excessive or insufficient intervention, and ineffective exposure and wasted costs. Summary of the Invention

[0004] The purpose of this invention is to provide a delivery method based on the relevance of delivery strategy and creative matching, so as to solve the problems mentioned in the background art.

[0005] A delivery method based on the relevance of delivery strategy and creative matching includes: Obtain the strategy time-series data of the target creative during the historical delivery process. The strategy time-series data includes the sequence of changes of at least one delivery strategy parameter over time. The system detects abrupt changes in the strategy time series data and identifies the time series change pattern of at least one deployment strategy parameter. The time series change pattern includes the location, magnitude, and direction of the abrupt change. Based on the mutation point characteristics in the time-series change pattern, combined with the material attributes of the target material, the matching value between the magnitude of the mutation point and the sensitivity of the material attributes is calculated. Based on the matching value, it is determined whether the target material enters the cumulative response state and the time required to reach the critical point. Thus, the risk time window for the conversion effect of the target material to become abnormal after the mutation point is determined. The start time and length of the risk time window are jointly determined by the matching value and the critical point. Within the risk time window, differentiated adjustments are made to the current delivery strategy of the target material. The intensity of the differentiated adjustment is positively correlated with the magnitude of the mutation point, and the starting time of the differentiated adjustment is after the mutation point and before the abnormal conversion effect occurs.

[0006] This invention transforms ad placement adjustments from a passive response to an active prediction by detecting mutation points and defining risk time windows. It enables effective intervention before abnormal conversion results occur, thereby reducing ineffective exposure and cost waste, and improving overall ad placement efficiency.

[0007] In some possible implementations, the matching value is calculated and the cumulative response state is determined based on the mutation point characteristics and material properties, including: Extract the magnitude and direction of mutation points as features of the impact intensity of policy changes; Extract the emotional tension index and information density index from the target material's attributes as features of the material's sensitivity to strategy changes; The influence intensity feature and the response sensitivity feature are weighted and multiplied to obtain the matching value. When the matching value exceeds the preset entry threshold, the material is determined to enter the cumulative response state. Based on the degree to which the matching value exceeds the entry threshold, the lag time required from the mutation point to the critical point is determined, and the risk time window is defined based on the lag time.

[0008] This invention achieves a quantitative match between the intensity of strategy changes and the sensitivity of materials by extracting the magnitude and direction of mutation points and performing a weighted product operation with the emotional tension index and information density index of the material. This improves the accuracy of determining whether the material has entered the cumulative response state and provides a reliable basis for subsequent risk window delineation.

[0009] In some possible implementations, determining the lag time required from the mutation point to reaching the critical point includes: The magnitude of the mutation point is compared with the emotional tension index of the material. The emotional tension index corresponds to the tolerance threshold and the critical threshold obtained based on historical data statistics. The tolerance threshold is lower than the critical threshold. When the amplitude exceeds the tolerance threshold but does not exceed the critical threshold, the material is determined to enter the cumulative response state, and the accumulation rate is determined based on the difference between the amplitude and the tolerance threshold. Calculate the lag time required to reach the critical point based on the cumulative amount required to reach the critical threshold based on the cumulative rate and amplitude; When the amplitude directly exceeds the critical threshold, it is determined that the material has directly reached the critical point, and the lag time is zero or the preset minimum value.

[0010] By adopting the above scheme, by setting tolerance thresholds and critical thresholds, and dynamically calculating the lag time based on the relationship between the magnitude of the mutation and the emotional tension index, the time interval from the mutation point to the occurrence of the anomaly can be accurately predicted, avoiding misjudgment caused by using a uniform standard, and making the delineation of the risk time window more in line with the actual response pattern.

[0011] In some possible implementations, differential adjustments are performed within the risk time window, including: Perform slow-release adjustments at the beginning of the risk time window. Slow-release adjustments include reducing the release intensity in the mutation direction or increasing the release interval. During the middle of the risk time window, monitor the conversion performance metrics of the target creative in real time. If the conversion performance metrics show a downward trend, implement intervention and adjustments in the latter part of the risk time window. Intervention and adjustments include pausing the campaign or switching to buffer creatives. The timing of switching between gradual adjustment and intervention adjustment is dynamically determined based on the rate of decline of conversion effect indicators; the faster the rate of decline, the earlier the switching time.

[0012] By adopting the above approach, phased differentiated adjustments are implemented within the risk time window, including pre-phase slow release, mid-phase monitoring, and post-phase intervention. The adjustment method is dynamically switched according to the rate of decline of conversion effect indicators, thus achieving an organic combination of flexible and rigid intervention and improving the timeliness and accuracy of the adjustment.

[0013] In some possible implementations, the buffer material is a backup material that belongs to the same material group as the target material, but has a lower emotional tension index or information density index than the target material; The duration of switching to buffered material is determined by the magnitude of the mutation point and the duration of the target material's execution before the mutation.

[0014] By adopting the above solution, and by introducing a buffer material mechanism within the same group, and by determining the buffer switching duration in conjunction with the magnitude of the mutation and the duration of the delivery, the delivery content can be smoothly transitioned before an anomaly occurs, effectively mitigating the negative impact of the mutation and ensuring the continuity and stability of the delivery process.

[0015] In some possible implementations, after predicting the risk time window, the following is also included: The actual moment when the conversion effect is abnormal within the risk time window is obtained and compared with the predicted risk time window to obtain the offset. When the offset exceeds the preset tolerance range, extract the external environment data after the mutation point and before the anomaly occurs. The external environment data includes the intensity of competitors' product launches during the same period or holiday factors. External environmental data is used as a correction factor to update the entry threshold and / or critical threshold, which is then used to predict the risk time window for similar materials under similar external environments.

[0016] By adopting the above approach, the threshold is dynamically adjusted by introducing external environmental data (such as the intensity of competitor advertising and holiday factors), making the prediction results of the risk time window more in line with the actual market environment, improving the adaptability and accuracy of the prediction model, and providing a more reliable reference for the adjustment of similar materials in the future.

[0017] In some possible implementations, abnormal conversion results include various types of anomalies, with different types of anomalies corresponding to different critical points and risk time windows: The critical point for a surge in user negative feedback anomalies is located within the first conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle. The critical point for a conversion rate decline anomaly is located within the second conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle; The critical point for cost surge anomalies is located within the third conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle; Differentiated adjustments are made based on the conversion cycle within which the risk time window falls, selecting the corresponding adjustment strategy.

[0018] In this invention, different critical points and risk window positions are defined according to different anomaly types (surge in negative user feedback, decline in conversion rate, sudden increase in cost), which realizes refined identification of anomaly types and targeted adjustment strategy selection, thereby improving the efficiency and effectiveness of dealing with different anomaly scenarios.

[0019] In some possible implementations, mutation point detection is performed on the policy time series data, including: Set up a sliding window and calculate the mean and fluctuation range of the delivery strategy parameters within each window; When the mean of the current window deviates continuously from the mean of the previous window, and the deviation exceeds a multiple of the fluctuation range of the previous window, mark the starting point of the deviation as a mutation point, and record the direction of the deviation as a positive mutation or a negative mutation.

[0020] This invention provides a mutation point detection method based on a combination of sliding window and fluctuation amplitude comparison, which can accurately identify the mutation location, amplitude and direction of strategy parameters, providing a reliable data foundation for subsequent risk assessment and ensuring the stability and accuracy of the detection results.

[0021] In some possible implementations, material attributes include quantitative metrics obtained through material content analysis: The emotional tension index is determined based on the density and intensity of emotional expression elements in the source text or image; The information density index is determined based on the number of core information points appearing within a unit of time in the material; The induction intensity index is determined based on the frequency and prominence of elements in the material that guide user action; Material attributes are extracted and stored in advance when the target material is added to the database, and are used for subsequent matching with mutation point features.

[0022] By adopting the above scheme, the sensitivity of materials to policy changes is digitally represented by quantifying material attributes such as emotional tension index, information density index, and induction intensity index. This provides a scientific and reusable input dimension for matching value calculation and improves the reliability of the correlation analysis between material attributes and policy changes.

[0023] In some possible implementations, after the differential adjustment is performed, the following is also included: Record the strategy combination, timing of execution, and results of this adjustment; The records will serve as a reference for selecting adjustment strategies when similar mutation points occur in the same group of materials in the future; "Materials in the same group" refers to a collection of materials with similar emotional tension or information density indices.

[0024] This invention records the combination of adjustment strategies, the timing of execution, and the effects, and establishes a reference mechanism for materials in the same group. This enables the accumulation and reuse of adjustment experience, reduces the cost of repeated trial and error, and improves the intelligence level of strategy recommendation and the efficiency of deployment adjustment.

[0025] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: This invention accurately detects abrupt changes in strategy time-series data and combines the emotional tension index, information density index, and induction intensity index of the material to quantify and calculate the matching value. This allows for accurate determination of the cumulative response status of the material and delineation of risk time windows, enabling early prediction of when conversion anomalies will occur.

[0026] By implementing segmented, differentiated adjustments that are positively correlated with the magnitude of mutations within the risk time window, the impact of strategy fluctuations can be mitigated before anomalies occur, effectively avoiding problems such as a surge in negative user feedback, a decline in conversion rates, and a sudden increase in campaign costs.

[0027] Simultaneously, by leveraging a peer-to-peer material reference mechanism and dynamically adjusting thresholds based on the external environment, the accuracy of risk prediction and the applicability of adjustment strategies are continuously improved. The overall solution reduces the risk of ad placement anomalies while enhancing the stability and conversion efficiency of material placement, achieving refined ad delivery and maximized returns. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the method structure of the present invention. Detailed Implementation

[0029] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0030] Please see Figure 1 This application provides a delivery method based on the relevance of delivery strategy and creative matching, including: Step 1: Obtain policy time series data First, it is necessary to collect the strategy time-series data of the target creative during the historical campaign process. This data needs to fully cover the changes in key strategy parameters within at least one campaign period. Specifically, the target materials should cover various advertising formats such as images, videos, and text. It is recommended to set the time range of historical ad placement records to the past 30 days, and the data collection frequency should be in hours. This will ensure the continuity of time-series data and accurately capture short-term fluctuations in parameters.

[0031] It's important to note that the selection of campaign strategy parameters must revolve around the core factors influencing conversion rates. In this example, these specifically include budget parameters based on daily spending, audience targeting parameters encompassing age ranges, gender, and geographic location, bid parameters calculated per click, and campaign time slot parameters divided into 24-hour periods. These parameters directly relate to the exposure opportunities and reach quality of the creative materials and serve as the core basis for subsequent anomaly detection and strategy adjustment.

[0032] The strategy time-series data is obtained through the backend data interface of the campaign platform, with a sampling frequency of once per hour. Each time node records one data record, including timestamp, budget parameters, bid parameters, target audience parameters, and campaign time parameters. Outliers (such as those deviating from three standard deviations of the historical mean) are removed, and missing values ​​are supplemented using conventional interpolation methods.

[0033] It should be understood that the uniformity of data time granularity is crucial. In this embodiment, the recording nodes of all parameters are kept completely synchronized to avoid deviations in subsequent mutation detection and trend analysis due to time misalignment.

[0034] Step 2: Mutation point detection and temporal pattern recognition The core objective of this step is to accurately identify mutation points from strategy time-series data, clarifying the location, magnitude, and direction of the mutations, thus providing a foundation for subsequent risk assessment. Mutation point detection of strategy time-series data includes the following specific procedures.

[0035] First, a sliding window is set up, and the mean and fluctuation range of the delivery strategy parameters within each window are calculated. The length of the sliding window is not a fixed value, but is dynamically adjusted according to the characteristics of the parameter type. Budget parameters and bid parameters have a more direct impact on conversion results, and parameter changes are highly time-sensitive. Therefore, the window length is set to six time nodes, corresponding to six hours, to improve the sensitivity of capturing sudden change signals.

[0036] The impact of changes in the targeted audience parameters and the time period parameters is relatively mild. The window length is set to twelve time nodes, corresponding to twelve hours, which can effectively filter out short-term noise interference. The window uses a sliding method with a step size of one time node to traverse the entire data, meaning that there are most overlapping nodes between adjacent windows, ensuring data continuity and consistency in mutation detection.

[0037] The mean within each window is calculated by the arithmetic mean of all parameter values ​​within the window, while the fluctuation range is quantified by the sample standard deviation. This indicator reflects the dispersion of parameter values ​​within the window and is an important benchmark for judging whether subsequent deviations are abrupt changes.

[0038] In this embodiment, assuming a budget parameter window contains six nodes with values ​​of 5000 yuan, 5100 yuan, 4950 yuan, 5050 yuan, 4980 yuan, and 5020 yuan respectively, the calculated average is approximately 5016.67 yuan, with a fluctuation range of approximately 57.45 yuan. This result clearly shows the average level and fluctuation range of the budget parameter during this period.

[0039] Furthermore, the judgment rule for continuous deviation is set as follows: for the time series data of parameter P, all windows are traversed with a sliding window length L (budget / bid parameter L=6, orientation / time parameter L=12) and a step size of 1.

[0040] Calculate the difference between the current window mean μcurr and the previous window mean μprev, Δμ = |μcurr - μprev|, and the fluctuation range σprev (standard deviation) of the previous window.

[0041] When Δμ>k·σprev (k is the threshold coefficient, budget / bid parameter k=3, orientation / time parameter k=2.5), the window is recorded as a suspected deviation window.

[0042] If two consecutive windows (i.e., window i and window i+1) are both identified as suspected deviations, the starting time of window i is marked as the abrupt change point, and the abrupt change magnitude A = (μcurr - μprev) / μprev (relative rate of change), with the direction determined by the relationship between μcurr and μprev. If only a single window conforms, it is considered noise and no abrupt change point is marked. Finally, the starting time node of the window that first meets the deviation condition is marked as the mutation point. If the mean of the current window is greater than the mean of the previous window, it is determined to be a positive mutation and the parameter value is increased. If the mean of the current window is less than the mean of the previous window, it is determined to be a negative mutation and the parameter value is decreased.

[0043] Let's assume another scenario: the mean of the preceding window budget parameter is 5000 yuan, with a fluctuation range of 50 yuan, corresponding to a deviation threshold of 150 yuan. The mean of the current first window is 5200 yuan, with a deviation range of 200 yuan, exceeding the threshold. The mean of the current second window is 5250 yuan, with a deviation range of 250 yuan, also exceeding the threshold. This satisfies the requirement of two consecutive deviations. Therefore, the starting time point of the current first window is marked as the mutation point and determined to be a positive mutation.

[0044] The magnitude of the abrupt change is calculated using the relative rate of change index, which is the percentage change in the parameter value at the abrupt change point compared to the parameter value at the previous point. For example, if the budget parameter at the abrupt change point is 5150 yuan and the parameter at the previous point is 5000 yuan, then the magnitude is 3%.

[0045] Understandably, mutations in different strategy parameters need to be detected independently. For example, the mutation point of the budget parameter may occur at one time point, while the mutation point of the bid parameter may occur at another time point. Subsequent analysis needs to handle the mutation characteristics of each parameter separately to avoid cross-interference.

[0046] As a complete example of this step, in the time series data of the target material's budget parameters, the average value of the preceding window is 5000 yuan, the fluctuation range is 40 yuan, and the threshold is 120 yuan. The average value of the current first window is 6200 yuan, and the deviation range is 1200 yuan, exceeding the threshold. The average value of the current second window is 6300 yuan, and the deviation range is 1300 yuan, still exceeding the threshold. Therefore, the starting time node of the current first window is marked as the mutation point, with a mutation point parameter value of 6100 yuan. The parameter value of the previous node is 5000 yuan, and the calculated amplitude is 22%, with the direction being a positive mutation.

[0047] Step 3: Matching value calculation, cumulative response determination, and risk time window delineation. This step requires combining the characteristics of mutation points with the attributes of the materials, and using quantitative matching values ​​to determine whether the materials have entered the cumulative response state, thereby determining the lag time and risk time window required to reach the critical point, providing a time basis for the implementation of adjustment strategies.

[0048] Specifically, the magnitude and direction of the mutation point features are the core inputs, while the material attributes focus on the key dimensions that affect the sensitivity of the parameter mutation response, including material type, material lifecycle, and core promotional selling points. These attributes directly determine the material's adaptability to strategy changes and are an important foundation for matching value calculation.

[0049] The matching value is calculated and the cumulative response status is determined based on the characteristics of the mutation point and the attributes of the material, including the following specific contents: First, extract the magnitude and direction of the mutation point as the influence intensity features of the strategy change. To achieve comparability and fusion of data from different dimensions, the magnitude needs to be standardized, mapping the original magnitude value to the range of 0 to 1. For example, when the magnitude is 50%, the standardized value is 0.25, and when the magnitude is -30%, the standardized value is 0.15.

[0050] Meanwhile, considering that negative mutations typically have a faster impact on transformation effects, a directional coefficient is introduced. The coefficient is 1.0 for positive mutations and 1.3 for negative mutations (e.g., 0.15 after standardization when the amplitude is -30%, multiplied by 1.3 yields an influence strength feature of 0.195). This coefficient strengthens the weight of the negative mutation's influence. The comprehensive quantitative value of the influence strength feature is the product of the standardized amplitude and the directional coefficient, ranging from 0 to 1.3. A larger value indicates a greater impact from the strategy change.

[0051] Furthermore, the emotional tension index and information density index are extracted from the target material's attributes as features of the material's sensitivity to strategy changes. The emotional tension index measures the intensity of the emotional resonance the material evokes in users, calculated using a public sentiment dictionary combined with natural language processing technology. First, the material text is segmented, and then matched with positive, negative, and neutral words from the sentiment dictionary. The weight of each sentiment word is calculated as: sentiment polarity (+1 / -1) × sentiment intensity (1-5 points). Finally, the weighted sum of sentiment words is multiplied by 10 to obtain the final index.

[0052] For image or video materials, a pre-trained ResNet-50 model is used for emotion classification. The classification confidence score is used as the emotion intensity, which is then weighted and fused with the text emotion index (if text is present, the weight is 0.6 for text and 0.4 for video; if it is purely video, the video result is used directly). The value ranges from 0 to 10, with higher values ​​indicating stronger emotional intensity. For example, promotional materials containing strong emotional phrases like "limited-time discount" and "don't miss out," have an index of 8.2. Product display materials, which mainly describe features, have an index of 3.5.

[0053] The information density index uses different calculation methods for different material types, and all calculations are based on rules that can be automated: Image material: The proportion of effective information area is calculated using a foreground segmentation algorithm (such as U²-Net) to determine the proportion of main pixels (R); the number of text information items is detected and counted using OCR (such as PaddleOCR) (N). The index = R × 5 + N × 1.5.

[0054] Video footage: The number of camera cuts is automatically counted every 10 seconds (S) using a camera boundary detection algorithm (such as TransNetV2). The audio text information entropy is calculated by transcribing the audio into text using ASR (such as Whisper) and then calculating the character-level information entropy H of the text (H = -∑pilogpi, where pi is the probability of each character's occurrence). If the audio text is empty, H takes a default value of 0.5 (based on the average of historical silent videos). The index is calculated as S × 2 + H × 3.

[0055] Text Material: Keyword density is calculated by extracting core words using TF-IDF and then determining the word frequency percentage K; the number of information levels is automatically calculated based on HTML tags (h1 / h2 / p, etc.) or Markdown levels, resulting in an index of K × 6 + L × 2.

[0056] All indices are normalized to the range of 0 to 10. For example, if a video clip switches camera angles 3 times every 10 seconds and the audio-visual information entropy is 1.2, the calculated index is 9.6.

[0057] The material attributes also include a persuasive strength index, which is determined based on the frequency of action-oriented elements (such as "Buy Now" or "Click to Claim") and their visual / auditory prominence (such as button size, color contrast, and vocal tone intensity). This index is then weighted, summed, and normalized to a range of 0-10. Specifically, this is achieved by quantitatively analyzing the text, visuals, and audio of the material, statistically analyzing the frequency of action-oriented elements, assigning weights based on their visual and auditory prominence, and then weighting and normalizing the sum to a range of 0-10. Higher values ​​indicate stronger persuasiveness in guiding user action. The implementation can utilize conventional image processing and speech analysis methods in this field, which will not be elaborated upon here.

[0058] It should be further explained that the induction intensity index = 0.6 × frequency of text guiding words + 0.3 × area ratio of image guiding elements + 0.1 × frequency of speech guiding words, and the result is linearly normalized to 0-10. Text guiding words are matched and counted using a preset keyword library (such as "Buy Now"); image guiding elements are identified using an object detection model (such as YOLO) to recognize buttons, arrows, etc.; speech is transcribed using ASR and counted in the same way as text.

[0059] The comprehensive quantitative value of response sensitivity characteristics is a weighted sum of the emotional tension index, information density index, and inducement strength index, normalized to the range of 0 to 1. The weight of the emotional tension index is 0.4, the weight of the information density index is 0.3, and the weight of the inducement strength index is 0.3. This weighting is based on historical campaign data statistics to ensure a consistent correlation with conversion effectiveness. For example, when the emotional tension index is 8.2, the information density index is 7.0, and the inducement strength index is 6.5, the comprehensive quantitative value is 8.2 multiplied by 0.4 plus 7.0 multiplied by 0.3 plus 6.5 multiplied by 0.3, then divided by 10 equals 0.743, which is 0.74 after normalization.

[0060] Next, a weighted product of the influence strength feature and the response sensitivity feature is performed to obtain the matching value. The matching value is the product of the influence strength feature, the response sensitivity feature, and the strategy parameter weights, where the strategy parameter weights are set according to the degree of influence of different parameters on the conversion effect. It is clear that the formula for calculating the matching value is: M = I × S × W, where I is the influence intensity feature (standardized value 0~1.3), S is the response sensitivity feature (normalized value 0~1), and W is the strategy parameter weight (budget 1.0, bid 0.9, targeting 0.8, time period 0.7). This product result M reflects the degree of matching between the intensity of strategy change and the sensitivity of the material.

[0061] In this embodiment, a weighted product operation is used to obtain the matching value.

[0062] It is understandable that other weighted calculations can be used in other implementations, as long as the degree of matching between the intensity of strategy changes and the sensitivity of creative materials can be quantified. The budget parameter directly determines the exposure, with a weight of 1.0. The bidding parameter affects competitiveness, with a weight of 0.9. The audience targeting parameter relates to reach accuracy, with a weight of 0.8. The campaign time parameter affects the matching degree of user activity, with a weight of 0.7.

[0063] When the matching value exceeds the preset entry threshold (in this embodiment, the preset entry threshold is 0.2, which is determined based on historical data sensitivity analysis), the material is determined to enter the cumulative response state. The cumulative response state refers to a quantitative decay process in which the material's conversion potential enters due to a sudden change in strategy: Let the initial conversion potential at the mutation point t~0~ be P~0~ (taking the average conversion rate of the 7 days before the mutation as the benchmark), and the matching value M between the mutation amplitude and the emotional tension index of the material determines the decay rate k (for example, k=0.5×M, and the specific coefficient can be optimized through historical data fitting or A / B testing). Then, the remaining conversion potential at any time t is P(t)=P~0~−k·(t−t~0~); when P(t) decays to 0, the critical point is reached.

[0064] Based on the degree to which the matching value exceeds the entry threshold, the lag time required from the mutation point to reaching the critical point is determined, and a risk time window is defined based on the lag time. Specifically, determining the lag time required from the mutation point to reaching the critical point includes: The magnitude of the mutation point is compared with the emotional tension index of the material. The emotional tension index corresponds to the tolerance threshold and the critical threshold obtained based on historical data statistics. The tolerance threshold is lower than the critical threshold. It should be understood that the tolerance threshold is the maximum range of strategic mutations that the material can adapt to, while the critical threshold is the threshold at which the conversion effect is about to deteriorate abnormally. Both are based on statistical data of strategic mutation response of 12,000 sets of materials with different emotional tension indices in the past 180 days. The statistical method is as follows: the materials in each emotional tension index range are arranged in ascending order of mutation range, and the 25th percentile is taken as the tolerance threshold (i.e., 75% of the materials have not experienced abnormalities within this range), and the 75th percentile is taken as the critical threshold (i.e., 25% of the materials have experienced abnormalities within this range). A segmented mapping method is used to bind it to the emotional tension index.

[0065] When the emotional tension index is between 0 and 3, the tolerance threshold is 30%, and the critical threshold is 60%. When the emotional tension index is between 3 and 6, the tolerance threshold is 45%, and the critical threshold is 80%.

[0066] When the emotional tension index is between 6 and 10, the tolerance threshold is 60%, and the critical threshold is 100%. This mapping logic conforms to the actual rule that the higher the emotional tension index, the stronger the material's ability to withstand strategic changes. For example, an emotional tension index of 8.2 corresponds to a tolerance threshold of 60% and a critical threshold of 100%. An emotional tension index of 3.5 corresponds to a tolerance threshold of 45% and a critical threshold of 80%.

[0067] Abnormal conversion results include various types of anomalies, and different types of anomalies correspond to different critical points and risk time windows. Based on statistical analysis of historical anomaly data (500 sets of materials that experienced conversion anomalies in the past 180 days), 83% of anomalies caused by a surge in negative user feedback occurred in the first conversion cycle, 76% of anomalies caused by a decline in conversion rate occurred in the second conversion cycle, and 71% of anomalies caused by a sudden increase in cost occurred in the third conversion cycle. Therefore, the above correspondence was established.

[0068] It should be understood that the user negative feedback surge type anomaly refers to that the occurrence frequency of negative feedback behaviors such as user complaints, bad reviews, unfollowing, etc. within a unit time increases by more than 50% compared with the average value before the mutation. Its critical point is within the first conversion cycle after the mutation point. In this embodiment, the conversion cycle is set to 2 hours, which is statistically obtained based on the industry average user decision-making cycle and is derived from the statistical analysis of historical delivery data: Select 10,000 advertisement click logs within the past 180 days, and statistically analyze the time interval from the first exposure to the conversion (click / purchase) of users. Calculate that the median is about 1.8 hours and the average value is about 2.1 hours. After comprehensive consideration, it is rounded up to 2 hours as the standard conversion cycle. In practical applications, this cycle value can be adjusted according to different industries (such as fast-moving consumer goods, finance), but it is necessary to ensure that the corresponding relationship between the anomaly type and the cycle ordinal number remains unchanged. That is, from 0 to 2 hours after the mutation point, the corresponding risk time window covers this conversion cycle, the starting moment is the mutation point, and the length is 2 hours.

[0069] The conversion rate decline type anomaly refers to that the core conversion indicators such as click conversion rate and order placement conversion rate decrease by more than 15% compared with the average value before the mutation. Its critical point is within the second conversion cycle after the mutation point. From 2 to 4 hours after the mutation point, the corresponding risk time window covers this conversion cycle, the starting moment is 2 hours after the mutation point, and the length is 2 hours.

[0070] The cost sudden increase type anomaly refers to that the cost per click and the cost per conversion increase by more than 20% compared with the average value before the mutation. Its critical point is within the third conversion cycle after the mutation point. From 4 to 6 hours after the mutation point, the corresponding risk time window covers this conversion cycle, the starting moment is 4 hours after the mutation point, and the length is 2 hours.

[0071] When the amplitude exceeds the tolerance threshold but does not exceed the critical threshold, it is determined that the material enters the cumulative response state, and according to the difference ΔA between the amplitude and the tolerance threshold, the cumulative rate R is determined by the formula R = base(E) × ΔA, where base(E) is the rate coefficient related to the emotional tension index E, and is calculated by piecewise linear interpolation: When 0 ≤ E ≤ 3, base(E) = 2% / h / %; When 3 < E ≤ 6, base(E) = 2% + (E - 3) × ((3% - 2%) / (6 - 3)) = 2% + (E - 3) × 0.333% / h / %; When 6 < E ≤ 10, base(E) = 3% + (E - 6) × ((4% - 3%) / (10 - 6)) = 3% + (E - 6) × 0.25% / h / %; If E is exactly equal to the interval boundary (e.g., E=3.0), then the right-hand interval coefficient (3%) is taken. ΔA is the difference between the mutation amplitude A and the tolerance threshold Atol (if A≤Atol, then R=0, and it does not enter the accumulation state). This formula ensures that R changes continuously within the interval. The accumulation rate quantifies the speed at which the material accumulates towards the critical point, in percentages per hour.

[0072] Based on the accumulation rate R and the accumulation amount Qleft required for the amplitude to reach the critical threshold (i.e., the difference between the critical threshold and the current amplitude), calculate the lag time Tlag required to reach the critical point, which is Tlag = Qleft / R, in hours, and keep one decimal place.

[0073] It should be noted that the cumulative amount is the remaining distance between the amplitude of the mutation point and the critical threshold.

[0074] If the lag time is less than 1 hour, it will be taken as 1 hour to ensure the holding adjustment time. For example, if the emotional tension index is 8.2, the amplitude is 80%, the critical threshold is 100%, the cumulative amount is 20%, and the cumulative rate is 80% per hour, then the lag time is 0.25 hours, which will be corrected to 1 hour.

[0075] With an emotional tension index of 3.5, an amplitude of 50%, a critical threshold of 80%, a cumulative amount of 30%, and a cumulative rate of 15% per hour, the lag time is 2.0 hours.

[0076] When the amplitude directly exceeds the critical threshold, the material is determined to have reached the critical point, and the lag time is zero or a preset minimum value. In this embodiment, the preset minimum value is 0.5 hours. This duration is the minimum execution cycle of the strategy adjustment command, ensuring that the system has sufficient time to complete the adjustment operation. For example, if the emotional tension index is 8.2 and the amplitude is 110%, exceeding the critical threshold of 100%, then the lag time is 0.5 hours. If the emotional tension index is 3.5 and the amplitude is 85%, exceeding the critical threshold of 80%, the lag time is also 0.5 hours.

[0077] The start time of the risk time window is the larger of 0.5 hours and the result of multiplying the lag duration by 0.2 and rounding down. The rounding function ensures that the start time is an integer hour segment. This setting guarantees that the start time is after the abrupt change point while allowing for basic adjustment preparation time. For example, when the lag duration is 2.0 hours, multiplying the lag duration by 0.2 equals 0.4 hours, and the start time is 0.5 hours. When the lag duration is 1.0 hour, multiplying the lag duration by 0.2 equals 0.2 hours, and the start time is 0.5 hours. When the lag duration is 0.5 hours, the start time is simply 0.5 hours.

[0078] The window length is the difference between the lag time and the start time. If the length is less than 1 hour, it is forcibly set to 1 hour to avoid insufficient adjustment time. For example, if the lag time is 2.0 hours, the start time is 0.5 hours, and the window length is 1.5 hours, then the lag time of 1.0 hour, the start time of 0.5 hours, and the window length of 0.5 hours are corrected to 1 hour. Similarly, if the lag time of 0.5 hours, the start time of 0.5 hours, and the window length of 0 hours are corrected to 1 hour.

[0079] In this embodiment, we assume that the mutation point is a positive mutation of the budget parameter with an amplitude of 80%, the target material is video material with emotional resonance, emotional tension index of 8.2, information density index of 8.3, and induction intensity index of 7.1.

[0080] First, the threshold is matched. The emotional tension index of 8.2 falls within the range of 6 to 10, with a tolerance threshold of 60% and a critical threshold of 100%. Since the amplitude exceeds the tolerance threshold by 80% but does not reach the critical threshold, it is determined to enter the cumulative response state. The difference is calculated at 20%, the rate coefficient is 4% per percent of the difference per hour, and the cumulative rate is 80% per hour. The cumulative amount is 20%, with a lag time of 0.25 hours, adjusted to 1 hour. The starting time is 0.5 hours, the window length is 0.5 hours, adjusted to 1 hour, and the final risk time window is 0.5 to 1.5 hours after the mutation point. If the mutation triggers a surge in negative user feedback, the risk time window is adjusted to 0 to 2 hours after the mutation point, matching the conversion cycle of the anomaly type.

[0081] Step 4: Implement differentiated adjustments within the risk time window This step requires differentiated adjustments within the defined risk time window, based on the characteristics of the mutation and the response of the creative materials. It is crucial to ensure that the adjustment intensity matches the magnitude of the mutation, and that the adjustment is made before any abnormal conversion results occur. Differential adjustments are made according to the conversion cycle in which the risk time window falls. If the risk time window is in the first conversion cycle and there is a surge in user negative feedback, a strong, gradual adjustment strategy is adopted. The campaign intensity is reduced by 1.2 times the adjustment coefficient, and the campaign interval is increased to 1.5 times the adjustment coefficient. Simultaneously, negative feedback indicators are monitored in real time, and intervention is triggered immediately upon the appearance of a surge trend.

[0082] If the risk window falls within the second conversion cycle and the conversion rate declines abnormally, a standard adjustment strategy should be adopted. Slow-release adjustments should be performed using the standard adjustment coefficient, with priority given to switching buffer creatives to avoid traffic loss due to paused campaigns.

[0083] If the risk window falls within the third conversion cycle and there is a sudden increase in costs, a light adjustment strategy will be adopted. The campaign intensity will be reduced by 0.8 times the adjustment coefficient, with a focus on monitoring cost indicators. Campaigning will only be paused if the cost exceeds the preset limit.

[0084] Specifically, the focus of differentiated adjustments should be on parameters that have the greatest impact on conversion rates. If multiple strategy parameters experience sudden changes simultaneously, they should be sorted from highest to lowest matching value, with priority given to parameters with higher matching values, as changes in these parameters have a more significant impact on creative conversion rates.

[0085] The adjustment intensity is quantified by an adjustment coefficient, which is positively correlated with the absolute value of the abrupt change amplitude. The calculation logic is 0.1 plus the absolute value of the amplitude divided by 200 and then multiplied by 0.9, while limiting the maximum value of the coefficient to 0.8. This setting ensures that the adjustment intensity increases with the amplitude of the abrupt change, while avoiding excessive adjustment that could impact normal deployment. For example, when the amplitude is 50%, the adjustment coefficient is 0.55. When the amplitude is 80%, the adjustment coefficient is 0.82, and we take 0.8. When the amplitude is -30%, the adjustment coefficient is 0.37.

[0086] Implement differentiated adjustments within the risk time window, including the following specific procedures: This is achieved by implementing mitigation adjustments at the beginning of the risk time window. These adjustments include reducing the intensity of delivery in the direction of mutation or increasing the delivery interval. The initial period of the risk time window accounts for 40% of the total length; for example, if the window length is 1 hour, the initial period is 0 to 0.4 hours. If the window length is 1.5 hours, the initial period is 0 to 0.6 hours. The core idea of ​​mitigation adjustments is to gently reduce the impact of mutations rather than drastic interventions, and the specific methods are designed differently based on the parameter type.

[0087] The budget parameter's ad intensity is reduced by 30% of the adjustment factor. For example, with an adjustment factor of 0.8, the reduction is 24%. If the current budget of 10,000 yuan corresponds to an ad intensity of 1,000 yuan per hour, the adjusted ad intensity will be reduced to 760 yuan per hour. The bid parameter's ad intensity is reduced by 25% of the adjustment factor. With an adjustment factor of 0.8, the reduction is 20%, from 3.0 yuan per click to 2.4 yuan per click. The targeted audience parameter reduces ad intensity by narrowing the coverage area, with a reduction of 20% of the adjustment factor. With an adjustment factor of 0.8, the reduction is 16%, from 10 million reachable individuals to 8.4 million. The ad time parameter is adjusted by reducing exposure during non-core time periods and increasing the ad interval. The interval increase is 10 minutes multiplied by the adjustment factor. With an adjustment factor of 0.8, the interval increases by 8 minutes, changing from once every 15 minutes to once every 23 minutes.

[0088] During the middle of the risk window, monitor the conversion performance metrics of the target creatives in real time. If the conversion performance metrics show a downward trend, implement intervention adjustments in the latter part of the risk window. These intervention adjustments include pausing the campaign or switching to buffered creatives. The middle of the window accounts for 30%, and the latter part accounts for 30%. For example, for a 1-hour campaign, the middle segment is 0.4 to 0.7 hours, and the latter segment is 0.7 to 1.0 hours. For a 1.5-hour campaign, the middle segment is 0.6 to 1.05 hours, and the latter segment is 1.05 to 1.5 hours.

[0089] The conversion performance metrics selected are click-through rate (CTR), conversion rate, and return on investment (ROI). These three metrics comprehensively reflect the performance of the creative materials from three dimensions: exposure quality, conversion efficiency, and campaign revenue. The monitoring frequency is set to once every 5 minutes to ensure timely capture of trend changes. Each time monitoring is performed, the ratio of the current metric to the mean before the change is calculated. When this ratio is not higher than -5%, a downward trend is identified. When this ratio is not higher than -10%, a significant downward trend is identified, requiring stronger intervention and adjustments.

[0090] Buffer materials are backup materials belonging to the same material group as the target material, but with a lower emotional tension index or information density index. Specifically, a material group is defined as a collection of related materials built around the same promotional theme, product, or service. For example, a promotional material group for an electronic product might include product feature videos, promotional posters, and user testimonials, ensuring consistency in the promotional theme after a switch and avoiding user confusion. Buffer materials must meet two core conditions: first, they must belong to the same material group as the target material; second, their emotional tension index or information density index must be lower than the target material, with a difference of at least 1.0. This ensures that buffer materials are less sensitive to policy changes than the target material, effectively mitigating fluctuations in conversion rates. For example, if the target material has an emotional tension index of 8.2 and an information density index of 8.3, a backup material with an emotional tension index of 6.8 and an information density index of 7.2 within the same group could be selected as a buffer material, satisfying both difference requirements. If no material within the group meets both conditions, the material with the lower emotional tension index should be prioritized; if none is found, the material with the lower information density index should be selected.

[0091] The intervention and adjustment method is selected based on the degree of the downward trend. When the difference ratio is between -10% and -5%, a switch to buffer creatives is executed. The switch adopts a smooth transition strategy, gradually changing the exposure ratio of the target creative and buffer creatives from 100%:0% to 0%:100% within 30 minutes to avoid traffic fluctuations caused by sudden switching. When the difference ratio is below -10%, a pause operation is executed. The pause duration is the remaining time window of risk. If the remaining time is less than 30 minutes, it is extended to 30 minutes. During the pause, no exposure resources are allocated to this creative to minimize losses.

[0092] The duration of switching to buffered content is determined by the magnitude of the abrupt change and the duration of the target content's runtime before the abrupt change. Let the runtime before the abrupt change be in days, and the magnitude of the abrupt change be its absolute value. The calculation logic for the switching duration is as follows: multiply the absolute value of the magnitude by 0.05% per day, add the natural logarithm of the runtime plus 1 multiplied by 1.2 days, and take the larger of this result and 2 days to ensure the duration is no less than 2 days. This design logic dictates that the larger the abrupt change and the shorter the runtime of the content, the weaker the stability, and the longer the buffering duration.

[0093] Furthermore, the specific segmented adjustment rules are as follows: When the campaign duration is no more than 3 days, the coefficient of the natural logarithm is adjusted to 1.5 days during the initial campaign period, as the new materials are less stable and require a longer buffer. When the absolute value of the magnitude is not less than 100%, the coefficient of the magnitude is adjusted to 0.08 days per percentage point, as the impact of a large change is more lasting. For example, if the absolute value of the magnitude is 80%, the campaign duration is 5 days, and the stable campaign period is in progress, the switching duration will be approximately 6.15 days, rounded down to 6 days. If the campaign duration is 2 days and the absolute value of the magnitude is 120%, the switching duration will be approximately 11.25 days, rounded down to 11 days. If the absolute value of the magnitude is 30% and the campaign duration is 10 days, the switching duration will be approximately 4.38 days, rounded down to 4 days.

[0094] Finally, the switching point between gradual adjustment and intervention adjustment is dynamically determined based on the rate of decline of the conversion effect indicator; the faster the decline rate, the earlier the switching point. The decline rate is the difference between the current monitored difference ratio and the previous monitored difference ratio, divided by the 5-minute interval between the two monitoring periods.

[0095] It should be understood that the risk time window comprises 40% in the early stage, 30% in the middle stage, and 30% in the later stage. Therefore, the start time of the middle stage = the start time of the window + the window length × 40%. The reference position for the switching time is the start time of this middle stage. The baseline position for the switching time is the start of the middle of the risk time window. The switching time offset is determined through a preset mapping relationship: Offset = Decrease rate × 10 minutes, where the decrease rate is the difference between the current monitored difference ratio and the previous monitored difference ratio divided by the monitoring interval (5 minutes), expressed as a percentage per minute. The offset is limited to the range of [-15 minutes, 5 minutes]. If the calculated value exceeds this range, the boundary value of the range is used. The final switching time = baseline position time + offset. If no downward trend is observed (difference ratio higher than -5%), the gradual adjustment is maintained until the end of the window without switching, to avoid excessively early or late adjustments that could lead to failure.

[0096] The final switching time is the baseline time plus the offset. For example, if the window length is 1 hour, the baseline time is 24 minutes. If the current difference ratio is -8% in a certain monitoring, the previous one was -3%, the rate of decrease is -1% per minute, and the offset is -10 minutes, then the switching time is 14 minutes, which is still in the early stage, and intervention and adjustment are initiated in advance.

[0097] If the descent rate is -0.3% per minute, the offset is -3 minutes, and the switching time is 21 minutes.

[0098] If the descent rate is -0.1% per minute, the offset is -1 minute, and the switching time is 23 minutes.

[0099] If no downward trend emerges and the difference ratio is higher than -5%, the gradual adjustment will continue until the window ends.

[0100] Step 5: Adjust the recording and reference mechanism for materials in the same group. After the differentiated adjustments are implemented, the following also applies: Record the strategy combination, timing of execution, and effects of this adjustment. The strategy combination should be recorded in detail, including the specific measures adopted in this adjustment, such as the type of adjustment parameters, the intensity ratio of the mitigation adjustment, the method of intervention, the key attributes of the buffer materials, the anomaly type, and the corresponding adjustment strategy, to ensure that the adjustment logic can be accurately reproduced in the future.

[0101] The timing of implementation needs to record the specific range of the risk window, the corresponding conversion cycle, the start and end times of mitigation adjustments, and the trigger point for intervention adjustments, clearly defining the matching relationship between the adjustments and the risk window. The effectiveness of implementation is quantified through multi-dimensional indicators, including changes in click-through rate, conversion rate, and return on investment before and after the adjustment, the duration of the anomaly, and the final change in campaign revenue, comprehensively reflecting the actual value of the adjustment.

[0102] The records will serve as a reference for selecting adjustment strategies when similar mutation points occur in the same group of materials later. The core of the reference mechanism is to establish a correspondence between mutation characteristics, anomaly types, adjustment strategies, and execution effects. Subsequently, when materials in the same group encounter mutations, the system will automatically retrieve historical records.

[0103] If the parameter type, amplitude range, material sensitivity attributes, and anomaly type of the mutation are consistent with or similar to the historical records, and there are no significant differences in the external environment, then the combination of adjustment strategies and execution timing with the best performance in the historical records should be recommended first to reduce the cost of repeated trial and error.

[0104] It should be understood that "same group of materials" refers to a set of materials whose emotional tension index and information density index are similar to the target material. The specific criteria are: the difference in emotional tension index ≤ 1.0 and the difference in information density index ≤ 1.0. If a material lacks one index, the other index is used as the criterion, and this is noted in the record. This similarity threshold is determined based on historical data clustering results: K-means clustering is performed on the indices of 10,000 groups of materials in the material library, and the average of the maximum distances within each cluster is used as the threshold benchmark. For example, if a material has an emotional tension index of 7.8, and the difference between it and the target material with an emotional tension index of 8.2 is 0.4, it belongs to the same group. Similarly, if a material has an information density index of 7.5, and the difference between it and the target material with an information density index of 8.3 is 0.8, it is also classified into the same group. This setting ensures that the response characteristics of materials within the same group are consistent to policy changes, while avoiding the problem of low historical data reuse caused by overly strict grouping.

[0105] Furthermore, to enhance the practicality of the reference mechanism, historical records will be categorized and stored according to material groups, mutation types, anomaly types, and external environments, and a multi-level retrieval index will be established.

[0106] For example, in the emotional resonance-based content group, records of adjustments are specifically stored for instances of positive budget parameter shifts, fluctuations of 70% to 90%, abnormal conversion rate declines, and adjustments made during non-holiday periods. This allows for quick matching of the optimal reference solution when subsequent content in the same group encounters similar scenarios. If multiple adjustment cases with similar scenarios exist in the history, they will be sorted by their execution results, with the most effective solution being recommended first.

[0107] If there is no perfectly matching case, the adjustment logic based on similar cases will be used for adaptation and optimization. For example, if the budget was reduced by 24% when the magnitude was 80% in a historical case, it can be reduced by 22.5% when the magnitude is 75% in the current case.

[0108] In this embodiment, the adjusted records will be categorized into the material group with an emotional tension index of 6 to 8 and an information density index of 7 to 9. These will be tagged with scenario labels such as positive budget parameter shift, 70% to 90%, conversion rate decline anomaly, non-holiday period, and moderate competitor ad intensity. If subsequently, another emotionally resonant material in the same group, with an emotional tension index of 7.3, an information density index of 7.8, and an inducement intensity index of 6.8, experiences a positive budget parameter shift of 78% and triggers a conversion rate decline anomaly, and the external environment is similar, the system will automatically recommend a standard adjustment strategy, a gradual adjustment reducing the budget by 24%, a switch to buffer materials approximately 0.7 hours after the shift point, an emotional tension index in the 6 to 7 range, and a continuous buffer for 6 days. The system will also refer to the results of this implementation to set expected goals, improving adjustment efficiency and accuracy.

[0109] Step 6: Dynamic Correction of Risk Time Window Prediction After predicting the risk time window, the process also includes: obtaining the actual time when the conversion effect is abnormal within the risk time window, comparing it with the predicted risk time window, and obtaining the offset. It should be understood that the actual time of an anomaly occurrence is determined through real-time monitoring data, specifically the time when the conversion performance indicator first meets the criteria for the corresponding anomaly type. The offset is the time difference between this actual time and the start time of the predicted risk time window. If the actual time is earlier than the start time, the offset is negative; if it is later than the end time of the window, the offset is positive; and if it is within the window range, the offset is zero. For example, if the predicted risk time window is 2 to 4 hours after the mutation point, and the anomaly is a conversion rate decline, and the actual anomaly occurs 1.8 hours after the mutation point, the offset is -0.2 hours. If the actual anomaly occurs 4.2 hours after the mutation point, the offset is 0.2 hours.

[0110] When the offset exceeds the preset tolerance range, extract the external environment data after the mutation point and before the anomaly occurs. The external environment data includes the intensity of competitors' product launches during the same period or holiday factors. In this embodiment, the tolerance range is set to ±0.3 hours. This range is based on historical deployment data and covers the normal fluctuation range. If the deviation exceeds this range, it indicates that external environmental factors have affected the timing of abnormal conversion results and need to be incorporated into the correction system.

[0111] The intensity of competitor advertising was obtained through publicly available third-party data interfaces (such as AppGrowing and ADX), quantified as the percentage change in competitor ad impressions relative to the average of the same period 7 days prior to the mutation point. Holiday factors were represented using binary labels (statutory holidays and the day before and after are 1, otherwise 0), with an added historical effect coefficient (e.g., 1.5 for Spring Festival, 1.3 for National Day) for subsequent correction. For example, if competitor advertising intensity increased by 40% from the mutation point to the anomaly, and this occurred during a holiday period, it would be marked as 1; these two external environmental data points would then be extracted as the basis for correction.

[0112] The entry threshold or critical threshold is multiplicatively corrected based on the external environment data, using the following formula: New threshold = Original threshold × (1 + Competitor placement intensity change ratio × α) × (1 + Holiday effect coefficient × β), where α and β are preset correction weights (α = 0.2, β = 0.3), and the holiday effect coefficient is 1 for holidays and 0 for non-holidays. The corrected threshold is stored in a parameter database for predicting the risk time window of similar materials under similar external environments. The correction logic is designed differently based on the type of external environment data.

[0113] Regarding competitor ad intensity, if it increases compared to the average, it indicates intensified market competition, and abnormal conversion rates may occur earlier. In this case, the entry threshold should be appropriately lowered and the critical threshold raised. For example, if the original entry threshold was 0.2, and competitor ad intensity increases by 40%, it should be revised to 0.18. The original critical threshold of 100%, corresponding to an emotional tension index range of 6 to 10, should be revised to 105%, advancing the risk window and initiating adjustments earlier. If competitor ad intensity decreases, the entry threshold should be appropriately raised and the critical threshold lowered to avoid excessive intervention.

[0114] Regarding holiday factors, if marked as 1, during holiday periods, user behavior fluctuates significantly, increasing the uncertainty of abnormal conversion results. Therefore, the entry threshold needs to be lowered and the critical threshold raised. For example, the entry threshold could be revised from 0.2 to 0.17, and the critical threshold from 100% to 108%. If marked as 0, during non-holiday periods, the original threshold should be maintained or slightly adjusted.

[0115] The revised threshold will be stored in the parameter database. When encountering similar materials, materials of the same type, core promotional selling points, the same anomaly type, and similar external environments, the updated threshold will be automatically called to predict the risk time window, continuously improving the accuracy of the prediction.

[0116] In this embodiment, assuming the predicted risk time window is 0 to 2 hours after the mutation point, and a surge in negative user feedback is observed, the actual anomaly occurs 0.1 hours after the mutation point, with an offset of -0.4 hours, exceeding the tolerance range of ±0.3 hours. External environmental data reveals that competitor advertising intensity increased by 50% compared to the average during the same period, and this occurred during a holiday, marked as 1. Therefore, the entry threshold is revised from 0.2 to 0.16, and the critical threshold for the emotional tension index range of 6 to 10 is revised from 100% to 110%.

[0117] If subsequent emotional resonance-based, similar, or core promotional selling point-price advantage materials are deployed, and if an anomaly such as a surge in negative user feedback, a 40% to 60% increase in competitor ad intensity, holiday periods, or similar external environments occurs, the system will automatically use a modified threshold to calculate the matching value and define a risk time window, making the prediction results more consistent with the actual scenario.

[0118] The innovation of this application lies in its precise detection of abrupt changes in strategy time-series data. By combining this with quantitative calculations of matching values ​​based on the emotional tension index, information density index, and induction intensity index of the creative materials, the cumulative response status of the creative materials can be accurately determined, and risk time windows can be defined, allowing for early prediction of when conversion anomalies will occur. By performing segmented, differentiated adjustments positively correlated with the magnitude of the abrupt change within the risk time window, the impact of strategy fluctuations can be mitigated before anomalies occur, effectively avoiding problems such as a surge in negative user feedback, a decline in conversion rates, and a sudden increase in campaign costs. Simultaneously, relying on a reference mechanism with the same group of creative materials and dynamically adjusting thresholds based on the external environment, the accuracy of risk prediction and the applicability of the adjustment strategy are continuously improved.

[0119] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A delivery method based on the correlation between delivery strategy and creative matching, characterized in that, include: Obtain the strategy time-series data of the target creative during the historical delivery process. The strategy time-series data includes the sequence of changes of at least one delivery strategy parameter over time. The system detects abrupt changes in the strategy time series data and identifies the time series change pattern of at least one deployment strategy parameter. The time series change pattern includes the location, magnitude, and direction of the abrupt change. Based on the mutation point characteristics in the time-series change pattern, combined with the material attributes of the target material, the matching value between the magnitude of the mutation point and the sensitivity of the material attributes is calculated. Based on the matching value, it is determined whether the target material enters the cumulative response state and the time required to reach the critical point. Thus, the risk time window for the conversion effect of the target material to become abnormal after the mutation point is determined. The start time and length of the risk time window are jointly determined by the matching value and the critical point. Within the risk time window, differentiated adjustments are made to the current delivery strategy of the target material. The intensity of the differentiated adjustment is positively correlated with the magnitude of the mutation point, and the starting time of the differentiated adjustment is after the mutation point and before the abnormal conversion effect occurs.

2. The delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, The matching value is calculated and the cumulative response status is determined based on the characteristics of the mutation point and the material attributes, including: Extract the magnitude and direction of mutation points as features of the impact intensity of policy changes; Extract the emotional tension index and information density index from the target material's attributes as features of the material's sensitivity to strategy changes; The influence intensity feature and the response sensitivity feature are weighted and multiplied to obtain the matching value. When the matching value exceeds the preset entry threshold, the material is determined to enter the cumulative response state. Based on the degree to which the matching value exceeds the entry threshold, the lag time required from the mutation point to the critical point is determined, and the risk time window is defined based on the lag time.

3. A delivery method based on the correlation between delivery strategy and creative matching according to claim 2, characterized in that, Determine the lag time required from the mutation point to reach the critical point, including: The magnitude of the mutation point is compared with the emotional tension index of the material. The emotional tension index corresponds to the tolerance threshold and the critical threshold obtained based on historical data statistics. The tolerance threshold is lower than the critical threshold. When the amplitude exceeds the tolerance threshold but does not exceed the critical threshold, the material is determined to enter the cumulative response state, and the accumulation rate is determined based on the difference between the amplitude and the tolerance threshold. Calculate the lag time required to reach the critical point based on the cumulative amount required to reach the critical threshold based on the cumulative rate and amplitude; When the amplitude directly exceeds the critical threshold, it is determined that the material has directly reached the critical point, and the lag time is zero or the preset minimum value.

4. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, Implement differentiated adjustments within the risk time window, including: Perform slow-release adjustments at the beginning of the risk time window. Slow-release adjustments include reducing the release intensity in the mutation direction or increasing the release interval. During the middle of the risk time window, monitor the conversion performance metrics of the target creative in real time. If the conversion performance metrics show a downward trend, implement intervention and adjustments in the latter part of the risk time window. Intervention and adjustments include pausing the campaign or switching to buffer creatives. The timing of switching between gradual adjustment and intervention adjustment is dynamically determined based on the rate of decline of conversion effect indicators; the faster the rate of decline, the earlier the switching time.

5. A delivery method based on the correlation between delivery strategy and creative matching according to claim 4, characterized in that, Buffer material is a backup material that belongs to the same material group as the target material, but has a lower emotional tension index or information density index than the target material; The duration of switching to buffered material is determined by the magnitude of the mutation point and the duration of the target material's execution before the mutation.

6. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, Following the risk prediction time window, it also includes: The actual moment when the conversion effect is abnormal within the risk time window is obtained and compared with the predicted risk time window to obtain the offset. When the offset exceeds the preset tolerance range, extract the external environment data after the mutation point and before the anomaly occurs. The external environment data includes the intensity of competitors' product launches during the same period or holiday factors. External environmental data is used as a correction factor to update the entry threshold and / or critical threshold, which is then used to predict the risk time window for similar materials under similar external environments.

7. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, Anomalies in conversion results can be categorized into several types, each corresponding to different critical points and risk time windows: The critical point for a surge in user negative feedback anomalies is located within the first conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle. The critical point for a conversion rate decline anomaly is located within the second conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle; The critical point for cost surge anomalies is located within the third conversion cycle after the mutation point, and the corresponding risk time window covers this conversion cycle; Differentiated adjustments are made based on the conversion cycle within which the risk time window falls, selecting the corresponding adjustment strategy.

8. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, The strategy time series data is subjected to mutation point detection, including: Set up a sliding window and calculate the mean and fluctuation range of the delivery strategy parameters within each window; When the mean of the current window deviates continuously from the mean of the previous window, and the deviation exceeds a multiple of the fluctuation range of the previous window, mark the starting point of the deviation as a mutation point, and record the direction of the deviation as a positive mutation or a negative mutation.

9. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, Material attributes include quantitative indicators obtained through content analysis: The emotional tension index is determined based on the density and intensity of emotional expression elements in the source text or image; The information density index is determined based on the number of core information points appearing within a unit of time in the material; The induction intensity index is determined based on the frequency and prominence of elements in the material that guide user action; Material attributes are extracted and stored in advance when the target material is added to the database, and are used for subsequent matching with mutation point features.

10. A delivery method based on the correlation between delivery strategy and creative matching according to claim 1, characterized in that, After the differentiated adjustments are implemented, the following also applies: Record the strategy combination, timing of execution, and results of this adjustment; The records will serve as a reference for selecting adjustment strategies when similar mutation points occur in the same group of materials in the future; "Materials in the same group" refers to a collection of materials with similar emotional tension or information density indices.

Images