Precision marketing-oriented intelligent delivery targeted penetration decision-making method

Abstract

The invention discloses an intelligent delivery targeted penetration decision-making method for precision marketing, particularly relates to the technical field of intelligent delivery targeted penetration decision-making, and comprehensively identifies potential anomalies through unified processing of a standardized index set and dynamic association construction of multi-level thresholds; by means of the constraint of a delivery strategy oscillation index on a delivery behavior, dynamic bidding, material optimization and flow distribution can be smoothly adjusted in a fluctuating environment, the controllability of budget consumption and the continuity of a delivery effect are ensured, and a multi-dimensional penetration strategy of competition, scenes, crowds and materials is combined, so that the delivery efficiency is improved. According to the method, high-value user groups are more accurately reached, the material matching degree is improved, the return on investment and the conversion rate are stably kept increased, oscillation and resource waste in the putting process are reduced, and the cross-platform collaborative optimization capability and the risk prevention and control level are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent delivery target penetration decision-making, and more particularly to an intelligent delivery target penetration decision-making method for precision marketing. BACKGROUND

[0002] In the existing intelligent advertising delivery system, a prediction model is usually used to predict the trend of key indicators such as return on investment (ROI) and conversion rate, so that the system can adjust the budget and bid in advance. However, under sudden traffic fluctuations or special marketing nodes, the prediction model driven by historical data is often distorted, resulting in an overestimation of future ROI. When the system increases the delivery based on the inaccurate prediction results, it will quickly increase the budget consumption and bid level. At the same time, the risk control module protects according to the real-time monitoring of multiple thresholds (such as excessive consumption speed, substandard conversion effect, insufficient account balance, etc.). Under the condition of rapid budget growth, these thresholds are easily triggered at the same time. Therefore, the over-optimistic prediction layer and the over-tight risk control layer form a contradiction, the former pushes the system to expand the delivery, and the latter forces to freeze or suspend the plan. The superposition of the two directly leads to the dramatic fluctuation of the delivery strategy, which is manifested as frequent switching between "increasing delivery - forced suspension" in a short period of time, which not only destroys the continuity and convergence of the delivery, but also causes the dramatic fluctuation of ROI and the waste of resources, and even may miss the key traffic window. SUMMARY

[0003] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present application provide an intelligent delivery target penetration decision-making method for precision marketing to solve the problems raised in the background art.

[0004] To achieve the above object, the present application provides the following technical scheme: The intelligent delivery target penetration decision-making method for precision marketing comprises the following steps: Through a cross-platform interface, behavior log collection and third-party user tag data, delivery effect, traffic, material performance and account health indicators are obtained, and the obtained data are time-aligned, format-unified and missing value-completed to generate a standardized indicator set; A multi-level threshold check is applied to the standardized indicator set to identify abnormal states, and a dynamic correlation between indicators is constructed according to the abnormal markers to form an indicator correlation information set; Based on the indicator correlation information set and historical time series data, in combination with periodic characteristics and external factors, short-term trends of indicators are predicted, and prediction inaccuracy and sudden fluctuation information and risk control trigger conflict information are obtained to generate a delivery strategy shock index; According to the index prediction trend, index correlation information and the shock index of the delivery strategy, dynamically adjust the bid of each channel, select high-quality material reuse or generate new material, and automatically configure the A / B test scheme, while adjusting the proportion of traffic in each channel; Combined with user behavior characteristics, platform traffic distribution and interest labels, execute competition penetration, scene penetration, crowd penetration and material penetration, and return the delivery effect data to update the standardized index set.

[0005] In a preferred embodiment, a multi-level threshold check is applied to the standardized index set to identify abnormal states, and a dynamic correlation between indexes is constructed according to the abnormal markers to form an index correlation information set, as follows: Perform multi-scale time series decomposition on the standardized index set to generate a time series component set; Calculate the adaptive threshold for the time series component set, and output the threshold set and threshold confidence; Apply a multi-strategy detector to the threshold set in parallel and reference the standardized index set for detection, and output the original abnormal marker set; Perform confidence fusion and multi-level grading on the original abnormal marker set to generate a multi-level abnormal marker set; Perform time series aggregation and denoising on the multi-level abnormal marker set to generate a cleaned abnormal time period set and an abnormal co-occurrence statistical table; Calculate the conditional synergy and time lag correlation coefficients for the cleaned abnormal time period set and the time-aligned standardized index set, and output the initial correlation candidate set: Perform causal directionality verification and counterfactual disturbance test on the initial correlation candidate set to generate a weighted directed correlation edge list; Visualize the weighted directed correlation edge list, smooth the dynamic weights, and calculate the abnormal propagation ability to output the final index correlation information set.

[0006] In a preferred embodiment, based on the index correlation information set and historical time series data, combined with periodic characteristics and external factors, predict the short-term trend of the index, as follows: Merge the index correlation information set, historical time series data, periodic characteristics and external factors according to a unified time reference, perform data integrity check, missing marker and delay completion, and form an enhanced time series data set with source tracing; Perform multi-scale time series decomposition on the enhanced time series data set to extract long-term trends, periodic components and short-term residuals, and apply abnormal window mask and noise suppression to the residuals, and output the decomposed component set; Take the decomposed component set and the index correlation information set as input, construct cross-index lag features, short-term residual features, periodic embedding features and external event encodings according to the correlation edge weight and lag candidate list, and form a feature matrix for modeling; The feature matrix is supplied to the predictor set for parallel inference and rolling adaptation, outputting an original prediction value sequence and each predictor error distribution; With the original prediction value sequence and each predictor error distribution as input, backtest residual calibration, short-term deviation correction and confidence interval estimation are performed, and a calibrated short-term prediction package is output.

[0007] In a preferred embodiment, the prediction misalignment and sudden fluctuation information includes the following indicators: investment return rate, conversion rate, consumption speed, click rate and material performance deviation.

[0008] In a preferred embodiment, the risk control trigger conflict information includes the following indicators: budget consumption, insufficient balance, and event sequence with multiple thresholds triggered simultaneously.

[0009] In a preferred embodiment, the delivery strategy shock index is obtained by weighted summation of each indicator in the prediction misalignment and sudden fluctuation information and the risk control trigger conflict information.

[0010] Technical effects and advantages of the present application: 1. The present application introduces the delivery strategy shock index in the precise marketing scenario, explicitly quantifies the conflict information between the prediction layer and the risk control layer, and integrates it into the dynamic decision-making, so that the system can balance the forward-looking and robustness when facing sudden traffic fluctuations and special marketing nodes, thereby avoiding over-expansion due to prediction misalignment and frequent freezing due to risk tightening.

[0011] 2. The present application constructs a unified processing of standardized indicator set and dynamic association of multi-level threshold. The system can not only identify potential abnormalities comprehensively, but also enhance the adaptability to complex scenarios in trend prediction by combining periodic characteristics and external events; with the constraint of delivery strategy shock index on delivery behavior, dynamic bidding, material optimization and traffic distribution can be smoothly adjusted in fluctuating environment, ensuring the controllability of budget consumption and the continuity of delivery effect. At the same time, combined with the multi-dimensional penetration strategy of competition, scene, crowd and material, the system can more accurately reach the high-value user group and improve the material matching degree, so that the investment return rate and conversion rate remain stable and increase, not only significantly reducing the shock and resource waste in the delivery process, but also improving the cross-platform collaborative optimization ability and risk prevention and control level, so that the efficient use of marketing resources and the continuous improvement of return rate can still be realized in an uncertain environment. BRIEF DESCRIPTION OF DRAWINGS

[0012] In order to facilitate understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings; Figure 1 The flowchart of the embodiment method of the present application. DETAILED DESCRIPTION

[0013] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0014] Embodiment: Figure 1 The intelligent delivery target penetration decision method for precision marketing of the present application is given, including the following steps: Through a cross-platform interface, behavior log collection and third-party user tag data, delivery effect, traffic, material performance and account health indicators are obtained, and the obtained data is time-aligned, format-unified and missing value-completed to generate a standardized indicator set; A multi-level threshold check is applied to the standardized indicator set to identify abnormal states, and a dynamic correlation between indicators is constructed according to the abnormal markers to form an indicator correlation information set; Based on the indicator correlation information set and historical time series data, combined with periodic characteristics and external factors, a short-term trend of the indicators is predicted, and prediction inaccuracy and sudden fluctuation information and risk control trigger conflict information are obtained to generate a delivery strategy shock index; According to the indicator prediction trend, the indicator correlation information and the delivery strategy shock index, the bid of each channel is dynamically adjusted, the high-quality material is reused or new material is generated, and the A / B test scheme is automatically configured, and the traffic proportion of each channel is adjusted; Combined with user behavior characteristics, platform traffic distribution and interest labels, competitive penetration, scene penetration, crowd penetration and material penetration delivery are performed, and delivery effect data is backflowed to update the standardized indicator set; Through a cross-platform interface, behavior log collection and third-party user tag data, delivery effect, traffic, material performance and account health indicators are obtained, and the obtained data is time-aligned, format-unified and missing value-completed to generate a standardized indicator set, specifically as follows: Pull the original event stream: pull the original event stream through platform interface query, real-time log collector and third-party data export interface, including delivery effect, traffic, material performance and account health indicators, to generate an original event stream with a source timestamp; The delivery effect indicators include: return on investment, conversion cost, conversion rate, interaction rate; The traffic indicators include: exposure, click volume, consumption speed; The material performance indicators include: material competitiveness score, completion rate, interaction rate; The account health indicators include: budget utilization rate, balance sufficiency, planned delivery amount; Buffer and label event stream: write the original event stream to a low-latency buffer layer and add acquisition metadata (data source identification, receiving time, original timestamp, access batch ID) to each event to generate a buffered event stream with metadata; Unified time reference and alignment of buffered event stream: time standardization and event reordering of the buffered event stream according to a unified time reference (e.g. UTC or specified time zone according to advertising activities), handling cross-source clock drift and time zone differences, generating a time-aligned event sequence; Mapping and deduplicating time-aligned event sequence: performing identity mapping and deduplication on the time-aligned event sequence, unifying cross-platform identities through user ID mapping tables, device fingerprints, or probabilistic matching, and merging duplicate events according to the same request ID or time window rules to generate a deduplicated and uniformly identified event record set; Correlation attribution window and division of attribution event record set: marking event attribution according to business attribution rules (such as click attribution window, view attribution window, attribution priority) for the deduplicated and uniformly identified event record set to form an event record set with attribution attributes; Supplementing delayed attribution and merging late events: applying late data supplement strategy (based on water level / watermark judgment, incremental backfilling, and confidence decay) to the event record set with attribution attributes to merge delayed arrival conversion or cost events into corresponding attribution records to generate a complete event set with enhanced integrity; Standardizing fields and semantic mapping of complete event set with enhanced integrity: performing field mapping, type conversion, and enumeration normalization (e.g. mapping "impression / cnt" from different platforms to a unified "exposure" field) according to a unified semantic dictionary for the complete event set with enhanced integrity to generate a semantically consistent intermediate indicator set; Quality check and intelligent correction of intermediate indicator set: performing quality checks (range inspection, anomaly detection, cross-source verification) on the intermediate indicator set and correcting abnormal or missing items according to priority strategies: first, trace back and re-pull, second, use recent window interpolation or regression filling based on similar channels, and finally, mark confidence and retain correction logs to generate a standardized indicator set with confidence and traceability records; Packaging and writing features into storage and releasing standardized indicator set for downstream calls: writing the standardized indicator set with confidence and traceability records into a queryable feature storage (supporting time window retrieval and incremental update), updating metadata directory and version number, and outputting a standardized indicator set that can be directly called; Applying multi-level threshold checks to identify abnormal states and constructing dynamic correlation relationships between indicators based on abnormal markers to form an indicator correlation information set, as follows: Multi-scale time series decomposition on the standardized indicator set to generate a time series component set (trend, seasonality, residual): Through multi-scale decomposition, the standardized indicator set is decomposed into long-term trend, periodic component and short-term residual, providing clear objects for subsequent threshold setting and noise suppression; Adaptive threshold calculation on the time series component set, outputting a threshold set and threshold confidence: Based on quantiles, seasonal periods after decomposition and rolling baseline, short-term and long-term thresholds are generated respectively, and confidence for each threshold is calculated (confidence is obtained by weighted sum of standard deviation and sample size of historical indicators); Parallel application of multi-strategy detectors to the threshold set and reference to the standardized indicator set for detection, outputting the original anomaly label set and each detector score: Parallel use of robust Z-score (based on MAD), EWMA / CUSUM control chart, change point detection (such as abrupt change point detection), and prediction residual detection (prediction model residual exceeding limit) and other multi-strategy detectors to score and output the original anomaly label (including detector source and score) under the constraint of the threshold set; Confidence fusion and multi-level classification on the original anomaly label set to generate a multi-level anomaly label set (including start / end time, severity, detector voting and confidence): Weighted voting and confidence fusion strategy is used for each detector output to produce mild / moderate / severe three-level anomaly labels, and detector consistency, duration and confidence curve are recorded for each anomaly label; Time series aggregation and denoising on the multi-level anomaly label set to generate a cleaned anomaly time period set and anomaly co-occurrence statistical table: Close anomaly events are merged according to the pre-set merging window, and isolated short-time noise events are removed, and the co-occurrence number and co-occurrence sequence pattern of different indicators in the same time window are counted, and the cleaned anomaly time period set and anomaly co-occurrence statistical table are output; Calculation of conditional synergy and time lag correlation coefficient on the cleaned anomaly time period set and time-aligned standardized indicator set to output an initial correlation candidate set (including time lag, correlation directionality candidate, correlation strength): The logic of obtaining the conditional synergy is as follows: Select a probability density estimation strategy for continuous variables (for example: equidistant histogram estimation, kernel density estimation or Gaussian assumption parameter estimation), and estimate the marginal and joint probability density of indicators A, B and condition set Z on the sample set; Based on the estimated probability density, the following entropy terms (text description is "entropy") are calculated: Conditional entropy of condition set Z; Joint entropy of the joint distribution of indicators A and condition set Z; Joint entropy of the joint distribution of indicators B and condition set Z; Calculate the joint entropy of the triad of indicator A, indicator B and the set of conditions Z; (Each of the above calculations is discretized or continuous entropy approximated with the selected density estimation method for sample probability, and then summed / integrated to obtain the numerical value according to the entropy definition) The above entropy is numerically synthesized according to the "combination of the derivative of conditional mutual information": that is, the numerical expression of the conditional synergy quantity is obtained by the sum and difference of the corresponding entropy terms (the numerical value is the conditional synergy quantity by the derivative combination of the entropy term as the intermediate quantity); The acquisition logic of the time lag correlation coefficient is as follows: Set the lag scanning range (the maximum time length of the positive lag and the negative lag range, for example, ±N time steps), and determine the lag step; For each pair of indicators A and B and for each lag value τ, the time lag correlation value is calculated according to the following steps: a. Shift the time series of indicator A and the time series of indicator B on the time axis according to τ (if τ is positive, shift the B sequence backward relative to A; if τ is negative, shift forward); b. After shifting, only the overlapping time window part of the two sequences is retained, and the de-meaning process is performed on the overlapping section (the sample mean of each overlapping section is subtracted respectively); c. Calculate the sample covariance of the overlapping section, that is, the sample mean of the product of the de-meaned two sequences of the overlapping section; d. Calculate the sample standard deviation of each overlapping section; e. Divide the covariance by the product of the two sample standard deviations to obtain the normalized correlation coefficient value under the lag τ; Repeat the calculation steps of the time lag correlation value for each lag τ on all abnormal time periods and aggregate the corresponding lag correlation coefficients according to the predetermined aggregation strategy (for example, take the average or weighted average, and the weight can be based on the severity of the anomaly or the length of the segment); Perform causal direction verification and counterfactual disturbance test on the initial association candidate set to generate a weighted directed association edge list (including direction, lag, weight and confidence): take causal determination means (such as Granger type determination or constraint test based on time series prediction residual) for the candidate association, and implement small counterfactual disturbance simulation in the controlled history window to verify the downstream response, and combine the results of the two to give each candidate association directionality, lag value, weight and confidence; The weighted directed association edge list is graphed, the dynamic weight is smoothed, and the abnormal propagation ability is calculated, and the final index association information set (including nodes, weighted directed edges, lags, positive and negative relationships, confidence, abnormal propagation scores, and the latest triggering time) is output: write the weighted edges into the time series graph database, use exponential smoothing to integrate historical weights and this time's abnormal weight update, and perform influence propagation simulation on the graph to calculate the abnormal propagation score of each edge, and finally form the index association information set that can be referenced for decision-making; Based on the index association information set and historical time series data, combined with periodic characteristics and external factors, the short-term trend of the index is predicted, as follows: Merge the index association information set, historical time series data, periodic characteristics (hour / day / week encoding), and external factors (holiday, promotion plan, market event identification, etc.) according to a unified time reference, perform data integrity check, missing mark, and delay completion, and form an enhanced time series data set with source tracing and confidence annotation; Perform multi-scale time series decomposition on the enhanced time series data set, extract long-term trends, periodic components, and short-term residuals respectively, and apply abnormal window mask and noise suppression to the residuals, and output the decomposed component set (trend / season / residual) and its time window label; Take the decomposed component set and the index association information set as input, construct cross-index lag features, short-term residual features, periodic embedding features, and external event encodings according to the associated edge weight and lag candidate list; form a feature matrix for modeling and attach feature confidence; Provide the feature matrix to the predictor set for parallel inference and rolling adaptation: adopt complementary short-term pattern predictors (local trend extrapolator, residual predictor, time series learner, and tree regressor, etc.) to train / online update on the rolling window, and dynamically adjust the weights of each predictor based on the performance of the latest validation window, output the original prediction value sequence and the error distribution of each predictor; Take the original prediction value sequence and the error distribution of each predictor as input, perform backtest residual calibration, short-term bias correction, and confidence interval estimation, and output the calibrated short-term prediction package; Obtain prediction inaccuracy and sudden fluctuation information and risk control trigger conflict information, and generate a launch strategy shock index; The prediction inaccuracy and sudden fluctuation information includes the following indicators: investment return rate, conversion rate, consumption speed, click-through rate, and material performance deviation; The investment return rate is a ratio of the economic benefit generated by the advertisement launch to the input cost; by counting the total revenue brought by the advertisement, then counting the corresponding advertisement expenditure, and then normalizing the ratio of revenue to expenditure to obtain the final investment return rate; The conversion rate is used to measure the proportion of target behaviors (such as purchase, registration or download) generated after the user is reached by the advertisement; by counting the number of users who complete the target behavior within a fixed time period, and dividing it by the total number of users reached by the advertisement, the conversion rate is obtained; The consumption speed is used to measure the use speed and consumption efficiency of the advertising budget, by counting the actual consumption amount of the advertising budget within a fixed time period, and comparing it with the historical consumption average, the consumption speed is obtained; The click rate is used to measure the interest and interaction degree of users to the advertisement, by counting the number of times the advertisement is clicked within a fixed time period, and dividing it by the number of times the advertisement is exposed, the click rate is obtained; The material performance deviation is used to measure the deviation of the comprehensive performance of the advertisement material on each index relative to the historical or expected baseline, by first extracting the core effect index (completion rate, interaction rate, click rate, etc.) of each material, and then comparing the core effect index of each material with the historical average, the material performance deviation value is obtained and accumulated; The risk control trigger conflict information includes the following indicators: budget consumption, insufficient balance, and event sequence triggered by multiple thresholds at the same time; The budget consumption is used to measure the use progress and expenditure scale of the advertising budget within the delivery period, by counting the advertising expenses consumed within a fixed delivery period, and comparing it with the budget amount allocated for that period, the budget consumption ratio is obtained, which is used as the representative indicator of budget consumption; The insufficient balance is used to measure the adequacy of the remaining budget of the account relative to the subsequent delivery demand, by real-time statistics of the available balance of the account, and comparing it with the future estimated delivery demand, when the balance is less than the demand, it is marked as insufficient balance, and the ratio of available balance to delivery demand is calculated as the representative indicator of insufficient balance; The event sequence triggered by multiple thresholds at the same time is used to measure the situation that multiple risk thresholds are triggered at the same time within a time window during the delivery process, by counting the threshold trigger events of two or more thresholds triggered within the same window within a continuous time window, and forming an event sequence by recording the continuous occurrence of the trigger, and counting the total number of threshold trigger events in the event sequence as the representative indicator of the event sequence triggered by multiple thresholds at the same time; The delivery strategy shock index is obtained by weighted summation of each indicator in the prediction inaccuracy and sudden fluctuation information and the risk control trigger conflict information; According to the index prediction trend, the index correlation information and the delivery strategy shock index, dynamically adjust the bid of each channel, select high-quality material reuse or generate new material, and automatically configure A / B test scheme, while adjusting the proportion of traffic of each channel, as follows: Receiving the calibrated short-term prediction package, the indicator correlation information set, the launch strategy shock index, the current account budget and launch constraints, the real-time performance snapshot of each channel and the historical bidding backtracking data, and merging to generate a decision context package with confidence and version number; Taking the decision context package as input, generating candidate actions for each advertising plan and channel, including: differentiated bidding adjustment instructions (percentage increase / decrease, target bidding interval and minimum modification granularity), material operation instructions (high-quality material reuse, new material suggestion, variant generation request), traffic distribution instructions (channel proportion adjustment, time period redistribution) and candidate A / B experiment configuration (shunt ratio, grouping key, early stop rule); output the executable candidate action set with metadata (priority, expected impact, minimum execution unit); Taking the candidate action set and historical bidding backtracking data, real-time traffic prediction and launch strategy shock index as input, performing rapid simulation in a low-cost shadow environment (including shadow auction replay, budget burning estimation and conversion link response simulation), obtaining the estimated revenue, budget consumption curve and trigger risk threshold probability of each candidate action, and then weighting and summing to obtain the candidate action score, outputting the candidate action score table with estimated effect and risk score; Taking the candidate action score table as input, filtering actions according to multi-objective strategies (such as ROI stability priority, short-term conversion priority, and long-term value parity) and anti-shock strategies (using lag suppression, momentum buffer and launch strategy shock index driven change amplitude upper limit), generating execution plans by stage and time sequence (start time, duration, rate limit), outputting executable action plans with rollback conditions and observation points; Taking the executable action plan as input, issuing bidding change instructions, material launch commands and traffic distribution commands through channel interfaces according to the plan, and synchronously creating micro-experiment (A / B) configurations: shunt key, grouping ratio, early stop and statistical efficacy threshold, confusion control and sample size requirement; at the same time, recording the version, reason and expected indicators of each change for audit and traceability, outputting execution change event stream and experiment ID; Combining user behavior characteristics, platform traffic distribution and interest labels to perform competitive penetration, scene penetration, crowd penetration and material penetration, and returning the launch effect data to update the standardized indicator set, as follows: Merging processed user behavior characteristics (access frequency, recent path, conversion probability, value estimation, etc.), platform traffic distribution (time period traffic curve, bidding depth, winning price distribution), interest label set, and current creative library, budget constraints and decision signals (prediction trend, indicator correlation information, launch strategy shock index) to generate a launch execution context package with timestamp, confidence and version number; With the input of the execution context package, the candidate actions are generated for competition penetration, scene penetration, crowd penetration and material penetration: Competition penetration candidates (such as time period premium, automatic bidding upper / lower limit, pre-emptive bidding window); Scene penetration candidates (such as specified time period weighted delivery, scene triggered centralized over-delivery plan); Crowd penetration candidates (such as interest label combination package, similarity expansion rule and exclusion list); Material penetration candidates (such as format priority mapping, creative variant priority and personalized template); The candidate action list with priority, expected impact estimation and minimum execution unit is obtained; With the input of the candidate action list, the upper limit of execution frequency, budget rate limit, privacy protection check (edge hash and homomorphic matching rule), anti-fraud check and delivery strategy shock index compliance detection are performed, and the non-compliant / high-risk actions are removed or downgraded to generate a purified executable action list and a safety label; With the input of the executable action list, the creative template is mapped into a platform compatible creative package according to the display specification and format requirement of each platform, the personalized placeholder is replaced and the variant (including downgraded and fallback version) is generated, and the routing rule (target crowd, priority time period, frequency upper limit, A / B grouping key) of each creative package is generated, and the creative package to be delivered and the routing table are output; With the input of the creative package to be delivered and the routing table, the bidding instruction, routing allocation and creative uploading are executed through the channel interface; during the execution process, atomic level event stream (including display / bid winning / click / conversion event, bid, win rate, actual consumption, timestamp) is collected and recorded, and real-time execution event stream containing complete metadata is output; With the input of the real-time execution event stream, the short-term performance indicators (return on investment, conversion rate, click rate, consumption speed, win rate and material complete play / interaction rate) are calculated according to the experiment ID and shunt key, and the incremental and causal backtest is done through the reserved control group / time grouping, the risk trigger event log is supplemented, and the performance summary set with confidence and attribution window label is output; With the input of the performance summary set, first, the late attribution merging and confidence adjustment (time window backfilling is done for the delayed conversion and the confidence is lowered) are performed, then the adjusted indicators are written into the standardized indicator storage, and the user behavior characteristics, material performance record, platform traffic sample and risk trigger count are updated; The present application introduces the delivery strategy shock index in the precise marketing scene, explicitly quantifies the conflict information between the prediction layer and the risk control layer and integrates it into the dynamic decision, so that the system can balance the forward-looking and robustness when facing sudden traffic fluctuations and special marketing nodes, thereby avoiding over-expansion caused by prediction error and frequent freezing caused by risk tightening.

[0015] The application constructs the unified processing of the standardized index set and the dynamic association of the multi-level threshold, the system can not only comprehensively identify potential abnormalities, but also enhance the adaptability to complex scenarios in trend prediction by combining periodic characteristics and external events; with the help of the constraint of the launch strategy shock index on the launch behavior, dynamic bidding, material optimization and traffic distribution can be smoothly adjusted in the fluctuating environment, ensuring the controllability of budget consumption and the continuity of launch effect. At the same time, combined with the multi-dimensional penetration strategy of competition, scene, crowd and material, the system can more accurately reach the high-value user group and improve the material matching degree, so that the return on investment and conversion rate remain stable in growth, not only significantly reducing the shock and resource waste in the launch process, but also improving the cross-platform collaborative optimization ability and risk prevention and control level, so as to realize the efficient use of marketing resources and the continuous improvement of return on investment in the uncertain environment.

[0016] The above formulas are dimensionless numerical calculations, the formula is obtained by collecting a large amount of data to simulate the recent real situation, and the preset parameters in the formula are set by the person skilled in the art according to the actual situation.

[0017] It should be understood that the size of the sequence number of each process described above in various embodiments of the present application does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0018] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An intelligent targeted and penetrating decision-making method for precision marketing, characterized by: Includes the following steps: By acquiring campaign performance, traffic, creative performance, and account health metrics through cross-platform interfaces, behavioral log collection, and third-party user tag data, the acquired data is time-aligned, formatted, and missing values ​​are filled in to generate a standardized set of metrics. Multi-level threshold checks are applied to the standardized indicator set to identify abnormal states, and dynamic correlations between indicators are constructed based on the abnormality markers to form an indicator correlation information set; Based on the indicator correlation information set and historical time series data, combined with periodic characteristics and external factors, the short-term trend of the indicator is predicted, and information on prediction inaccuracies, sudden fluctuations and risk control trigger conflicts is obtained to generate the oscillation index of the investment strategy. Based on the trend prediction of indicators, the correlation information of indicators and the oscillation index of the advertising strategy, the bids of each channel are dynamically adjusted, high-quality creatives are selected for reuse or new creatives are generated, and A / B testing plans are automatically configured, while the traffic ratio of each channel is adjusted. By combining user behavior characteristics, platform traffic distribution, and interest tags, competitive penetration, scenario penetration, audience penetration, and creative penetration are executed, and the data on the performance of the campaign is fed back to update the standardized indicator set.

2. The intelligent targeted penetration decision-making method for precision marketing according to claim 1, characterized in that: Multi-level threshold checks are applied to the standardized indicator set to identify abnormal states, and dynamic correlations between indicators are constructed based on the anomaly markers to form an indicator correlation information set, as detailed below: The standardized index set is decomposed into a multi-scale time series to generate a set of time series components. Calculate adaptive thresholds for the time series component set, and output the threshold set and threshold confidence. A multi-strategy detector is applied in parallel to the threshold set and a standardized metric set is used for detection, outputting the original anomaly label set; Confidence fusion and multi-level classification are performed on the original anomaly label set to generate a multi-level anomaly label set; Perform temporal aggregation and denoising on the multi-level anomaly marker set to generate a cleaned set of anomaly time periods and an anomaly co-occurrence statistics table; For the cleaned set of abnormal time periods and the standardized index set aligned with time, calculate the conditional coherence and time-delay correlation coefficients, and output the initial candidate set of associations: Perform causal directionality verification and counterfactual perturbation test on the initial candidate set of associations to generate a weighted list of directed association edges; The weighted directed edge list is graph-based, dynamically weighted and smoothed, and anomaly propagation capabilities are calculated to output the final set of index association information.

3. The intelligent targeted penetration decision-making method for precision marketing according to claim 2, characterized in that: Based on the indicator correlation information set and historical time series data, combined with periodic characteristics and external factors, the short-term trend of the indicator is predicted as follows: The indicator association information set, historical time series data, periodic features and external factors are merged according to a unified time benchmark, and data integrity verification, missing data marking and delay completion are performed to form an enhanced time series dataset with source traceability; The enhanced time series dataset is decomposed into multiple scales to extract long-term trends, periodic components and short-term residuals. Anomaly window masking and noise suppression are applied to the residuals, and the decomposed component set is output. Using the decomposed component set and the indicator association information set as input, propagation-weighted cross-indicator lag features, short-term residual features, periodic embedding features and external event codes are constructed according to the association edge weights and lag candidate lists; thus forming a feature matrix for modeling. The feature matrix is ​​fed into the predictor set for parallel inference and rolling adaptation, and the original predicted value sequence and the error distribution of each predictor are output. Using the original predicted value sequence and the error distribution of each predictor as input, backtesting residual calibration, short-term bias correction and confidence interval estimation are performed, and the calibrated short-term prediction package is output.

4. The intelligent targeted penetration decision-making method for precision marketing according to claim 1, characterized in that: The information on prediction inaccuracies and sudden fluctuations includes the following indicators: return on investment, conversion rate, consumption rate, click-through rate, and deviation in creative performance.

5. The intelligent targeted penetration decision-making method for precision marketing according to claim 1, characterized in that: The risk control trigger conflict information includes the following indicators: budget consumption, insufficient balance, and event sequences triggered by multiple thresholds simultaneously.

6. The intelligent targeted penetration decision-making method for precision marketing according to claim 1, characterized in that: The oscillation index of the deployment strategy is obtained by weighted summation of each indicator in the information of prediction inaccuracy, sudden fluctuations, and risk control-triggered conflict.

Images