AI intelligent marketing data-driven placement strategy generation method

Abstract

The application discloses an AI intelligent marketing data-driven delivery strategy generation method, relates to the technical field of artificial intelligence and big data analysis, and comprises the following steps: establishing an opinion monitoring matrix, gathering multiple high-frequency indexes, constructing a unified time baseline and generating a multi-time-domain steady-state baseline, obtaining a short window fluctuation intensity sequence and an emotion gradient sequence; under the constraint of the short window fluctuation intensity sequence and the emotion gradient sequence, performing multi-scale spectrum decomposition and cross-channel time delay alignment, reconstructing a phase trajectory and identifying a phase acceleration peak, extracting a phase transition early warning section, and obtaining a causal verification candidate window. Through multi-dimensional index modeling and phase analysis, the application accurately identifies opinion mutation and behavior chain rupture, combines causal verification and counterfactual deduction to evaluate jump risk, and introduces weight inversion and dynamic backfilling to realize strategy correction, thereby significantly improving the prediction stability, robustness and efficiency of marketing strategies in extreme scenarios, and enhancing the coherence and risk resistance of brand communication.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and big data analytics, specifically to a method for generating AI-driven marketing data-driven campaign strategies. Background Technology

[0002] "AI-driven intelligent marketing data-driven campaign generation" refers to the use of artificial intelligence technology to deeply analyze and model multi-source marketing data (such as user behavior data, product sales data, social media interaction data, and advertising performance data). This automatically identifies key elements such as target user characteristics, content preferences, campaign timing, media selection, and geographic distribution, thereby dynamically generating targeted and high-conversion advertising or content delivery strategies. This method not only adapts to rapid changes in market demands and user behavior but also enables continuous strategy optimization and closed-loop feedback, improving the accuracy of ad targeting, reducing customer acquisition costs, and enhancing brand influence. Its core lies in the deep integration of artificial intelligence with marketing business processes, achieving intelligent management across the entire "data-cognition-decision-execution" chain.

[0003] The existing technology has the following shortcomings:

[0004] In existing technologies, AI-powered intelligent marketing campaigns typically rely on the continuous collection and modeling of user behavior data, communication interaction data, and real-time feedback data to dynamically generate campaign strategies that adapt to market changes. However, during sudden public opinion crises, data performance often exhibits short-term, high-frequency, and drastic fluctuations. Existing predictive mechanisms struggle to accurately identify these non-linear trends within a very short timeframe. This leads to predictive jumps in the model during critical campaign windows, where judgments about user attention and sentiment can reverse within minutes, causing abnormal fluctuations in strategy output. Such anomalies not only trigger frequent switching of ad placements, resulting in significant waste of budget and resources, but also disrupt the overall campaign rhythm, reduce the continuity of the communication chain, and even amplify the spread of negative public opinion, ultimately severely impacting marketing effectiveness and brand stability.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an AI-powered intelligent marketing data-driven strategy generation method to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent marketing data-driven campaign generation method, comprising the following steps:

[0008] Establish a public opinion trigger monitoring matrix, aggregate multiple high-frequency indicators, construct a unified time baseline and generate multi-time domain steady-state baselines, and obtain short-window fluctuation intensity sequence and sentiment gradient sequence;

[0009] Under the constraints of short-window fluctuation intensity sequence and sentiment gradient sequence, multi-scale spectral decomposition is performed and cross-channel time delay alignment is carried out to reconstruct phase trajectory and identify phase acceleration peaks, extract phase transition warning segments, and obtain causal verification candidate windows.

[0010] Within the causal verification candidate window, perform consistency verification of the reach-interaction-conversion causal chain, calculate the stability of the causal direction, and mark the time segment where the causal direction is reversed as a risk segment to obtain the inference input;

[0011] Counterfactual stable trajectories are generated for risk segments and compared and deduced. The relative divergence between the real trajectory and the counterfactual trajectory is calculated, and the phase transition threshold characteristics are statistically analyzed to obtain the jump intensity index and the phase transition threshold rate.

[0012] The jump intensity index, phase transition threshold rate, causal direction stability and short window fluctuation intensity sequence are weighted and fused to generate window-level predicted jump judgment, and output the predicted jump event label, dominant feature, severity level and intervention time.

[0013] Based on the window-level prediction jump judgment results, a weighted inversion mechanism is introduced to compensate for abnormal data in risk segments and dynamically replenish it under the multi-time domain steady-state baseline. This ensures that the delivery strategy output maintains a balance between historical trends and local corrections, resulting in a continuously stable delivery strategy.

[0014] Preferably, the steps for obtaining the short-window fluctuation intensity sequence and the sentiment gradient sequence are as follows:

[0015] Four types of user behavior data were collected: search volume, comment volume, forward volume, and click volume. The raw indicator sequences were obtained by keyword query records, comment area statistics, forward record aggregation, and click event counting, respectively.

[0016] The index series is processed with unified timestamps, and missing values ​​are filled with linear interpolation. The maximum and minimum normalization is used to scale the data to a uniform scale to complete the data fusion.

[0017] By analyzing the average lag time of different indicators in historical events, a delay correction function is constructed, the original data sequence is time-shifted to generate a unified time baseline, and the data is divided into statistical structured indicators with equal time windows.

[0018] Based on a unified time baseline, three types of time-domain steady-state baselines—short-term, medium-term, and long-term—were constructed. The standard deviation of the rate of change, the mean and standard deviation, and the moving average and standard deviation threshold were calculated for each. Short-window fluctuation intensity sequences and sentiment gradient sequences were extracted as inputs for subsequent analysis.

[0019] Preferably, the steps for obtaining the causality check candidate window are as follows:

[0020] Continuous wavelet transform is performed based on short-window fluctuation intensity sequence and emotion gradient sequence to extract spectral components at seven scales, construct frequency dominant spectrum, and mark candidate points of jump signals with phase synchronization error of less than three seconds.

[0021] Based on the candidate point set of jump signals, the cross mutual information between search volume, comment volume, forward volume and click volume is calculated to complete the time delay alignment of the time series and generate a composite behavior spectrum structure.

[0022] Based on the composite behavioral spectrum structure, the dominant frequency component is selected, the instantaneous phase and phase acceleration of four types of indicators are calculated, and the phase transition activation segment is extracted by combining the emotional gradient acceleration characteristics.

[0023] For each phase transition activation segment, the integral of the fluctuation intensity, the slope of the emotion gradient, and the amplitude of the phase acceleration are calculated, fused to generate a score, and the transition segment with the highest score is selected as the candidate window for causal verification.

[0024] Preferably, the time segment where the causal direction is reversed is marked as a risk segment, and the inference input steps are as follows:

[0025] Within the phase transition early warning segment, a contact behavior sequence, an interaction behavior sequence, and a conversion behavior sequence are constructed. User contact, participation, and outcome behavior data are collected, and the sampling frequency is uniformly set to five seconds to complete the normalization processing of the behavior chain time series.

[0026] Based on behavioral time series, lag analysis is performed to calculate the optimal lag time points for reaching the two stages of interaction and interaction to transformation, and the causal path delay structure is confirmed when the correlation meets the stable interval condition.

[0027] Based on the hysteresis structure, a causal direction stability sequence is constructed, the reverse behavior relationship state is marked, and combined with the short window fluctuation intensity and the change of sentiment gradient, the causal direction reversal segment is identified.

[0028] The time window that meets the conditions of causal reversal, fluctuation intensity exceeding the threshold, and dramatic shift in emotional gradient is marked as a risk segment. Its start and end time and abnormal characteristics of behavioral chain response are extracted to form the input for subsequent inference.

[0029] Preferably, when marking risk segments, they are identified as high-confidence risk segments only when the causal direction stability is less than -0.5, the duration is not less than 30 seconds, and the period includes at least one complete dynamic process of the reach, interaction, and conversion behavior chain. The amplitude of the lag structure fluctuation and the degree of correlation destruction are then extracted as the basis for intervention variables in subsequent inferences.

[0030] Preferably, the steps for obtaining the jump intensity index and phase transition threshold rate are as follows:

[0031] Based on risk segments, three types of counterfactual behavioral trajectories are constructed: frozen baseline trajectory, emotion-neutral trajectory, and noise-suppressed trajectory, which respectively simulate the evolution path of user behavior under the condition of no sudden change events;

[0032] The actual behavior trajectory is compared with the three types of counterfactual behavior trajectories hour by hour within a sliding time window. The total behavior offset and phase perturbation intensity are calculated respectively to generate a joint perturbation value sequence.

[0033] The jump intensity index and phase transition threshold rate are calculated based on the joint perturbation value sequence to quantify the abnormal jump intensity and phase stability perturbation characteristics of risk segments.

[0034] Preferably, the window-level prediction jump determination generation steps are as follows:

[0035] Normalization preprocessing is performed on the jump intensity index, phase transition threshold rate, causal direction stability and short window fluctuation intensity sequence to ensure consistent dimensions and construct a unified feature vector sequence;

[0036] The indicators are weighted and fused with fixed weights, and a moving average is performed every five seconds to generate a smooth score sequence.

[0037] Based on the smoothing score, a jump score threshold is set, and time window segments that continuously exceed the threshold are identified to determine whether they constitute a jump event and to determine the start and end time of the event.

[0038] Analyze the scoring structure within the transition event segment to determine the transition type label, dominant behavioral characteristic subband, and severity level;

[0039] Based on the local peak or plateau start point within the scoring sequence, the optimal intervention time for the transition event is output, and a complete event structure is formed for strategy intervention invocation.

[0040] Preferably, a weighted inversion mechanism is introduced based on the window-level prediction jump judgment result to compensate for abnormal data in risk segments, and dynamic replenishment is performed under multi-time domain steady-state baselines. The steps are as follows:

[0041] Based on the window-level prediction jump judgment results, identify the sudden increase and decrease behavior points and phase acceleration direction reversal points within the jump event segment, and construct an abnormal data point set;

[0042] Based on the jump score, different levels of negative weight coefficients are set, and negative weight compensation is applied to abnormal data points. By reducing the behavioral participation value, the interference intensity of the behavior on the strategy output is reduced.

[0043] The system invokes multiple time-domain steady-state baselines, selects short-period, medium-period, or long-period baselines based on the jump type of the outlier, and uses linear interpolation to dynamically replenish the outlier behavior data.

[0044] After the backfilling is completed, a transition fusion zone is constructed at the boundary of the jump segment. The behavior curve is smoothly connected by a linear slope adjustment method, and the final delivery strategy sequence containing local correction and global stability is output.

[0045] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0046] This invention constructs a multi-time-domain steady-state baseline and a public opinion trigger monitoring matrix to achieve unified modeling and sequential expression of multi-dimensional indicators such as search volume, comment volume, repost volume, and click volume. It further introduces multi-scale spectral decomposition, cross-channel time delay alignment, and phase trajectory reconstruction mechanisms to accurately identify hidden phase transition characteristics and potential behavioral chain breakpoints during sudden public opinion changes. Simultaneously, by combining causal direction stability assessment and counterfactual trajectory comparison and deduction, the degree of jump risk is quantified, improving the structured understanding of strategy disturbance trends. In the strategy output stage, a weight inversion and dynamic replenishment mechanism is further introduced to weaken and steadily repair abnormal data within risk segments, achieving closed-loop control across the entire chain from behavior identification to strategy correction. Overall, this invention not only significantly improves the predictive accuracy and stability of marketing strategies in extreme scenarios but also enhances the robustness and adaptability of the model in the face of nonlinear disturbances, thereby significantly reducing budget waste, improving campaign efficiency, and enhancing the consistency and risk resistance of brand communication. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0048] Figure 1 This is a flowchart of the AI-powered intelligent marketing data-driven campaign generation method of the present invention. Detailed Implementation

[0049] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0050] This invention provides, for example Figure 1 The AI-powered intelligent marketing data-driven campaign generation method shown includes the following steps:

[0051] A public opinion trigger monitoring matrix was established, which aggregated four high-frequency indicators: search volume, comment volume, forward volume and click volume. A unified time baseline was constructed and a multi-time domain steady-state baseline was generated to obtain the short-window fluctuation intensity sequence and the sentiment gradient sequence simultaneously.

[0052] To achieve dynamic response to marketing data during sudden public opinion events and to support subsequent prediction of abrupt changes and strategy stability, a public opinion trigger monitoring construction method based on high-frequency indicator fusion and multi-time domain processing is proposed, which includes the following steps:

[0053] Four representative user behavior data categories—search volume, comment volume, forwarding volume, and click volume—were selected as raw input metrics. Search volume was collected from keyword query records of multiple mainstream search engines. Specifically, a set of keywords highly relevant to the target brand, product, and event was defined, and the number of search requests for each keyword in this set was extracted every 60 seconds and aggregated into a total search volume value. Comment volume was collected from user comment sections of short video platforms, e-commerce platforms, and news clients. Specifically, a mapping relationship was constructed between content distribution paths and comment aggregation paths, and the number of user comments was captured from under a specified content ID, with the number of new comments counted every 30 seconds. Forwarding volume originated from the usage of content forwarding functions on social media platforms. By monitoring the forwarding records of specific content links or hashtags among different users, forwarding volume was aggregated by timestamp to form a forwarding time series. Click volume refers to user click behavior on advertisements, recommended content, and activity pages. The collection method is to count click events triggered by the user in real time and accumulate them every 10 seconds. The above four types of indicators are difficult to directly fuse in their original state due to differences in sampling frequency, units, and platform attributes. Therefore, this step requires first performing data alignment to convert all timestamps to a unified Beijing time format and setting a 5-second basic sampling granularity. Missing points in each data point are then filled using linear interpolation. Subsequently, the min-max normalization method is used to scale each type of data to the [0,1] interval. Specifically, each value in each sequence is subtracted from its minimum value and then divided by the range of the sequence, ensuring that all data have the same scale basis, thus providing comparability for subsequent fusion operations.

[0054] To address the time delay differences in user behavior triggers across different platforms, a time delay correction function was first constructed by analyzing the average lag time of peak indicators on each platform during similar historical events (such as brand crises, new product launches, and sudden negative trending topics). This function was then used to shift all high-frequency indicator sequences, aligning the peak segments of multiple indicators as closely as possible along the time axis. Next, a unified time segmentation strategy was established, dividing the complete observation period into equally spaced time windows (e.g., every 5 seconds). For each window, the average, extreme values, fluctuation amplitude, and growth rate of four types of indicators were statistically analyzed, forming a time-structured observation matrix. To further enhance the stability of the unified baseline, a dynamic window smoothing strategy was introduced. The values ​​of each 5-second window were weighted and averaged between the preceding and following windows, with weights calculated using a Gaussian distribution kernel to ensure that abrupt changes were not overly smoothed and distorted. The resulting unified time baseline is a fixed-resolution, cross-platform corrected, and structurally complete indicator time matrix, which can be used for subsequent high-precision modeling.

[0055] The "time delay correction function" is a function method used to adjust the differences in the response of data from different sources on the time axis. Its purpose is to align the peak performance of high-frequency indicators (such as search volume, comment volume, repost volume and click volume) from different platforms or channels after the outbreak of an event to a unified time reference system, thereby constructing an accurate unified time baseline.

[0056] In practice, users' responses to the same event on different platforms such as search engines, social media, and video platforms naturally exhibit time lags. For example, a user clicking on trending content on a news platform might happen within the first minute of the event, while related comments might appear several minutes later, and forwarding might lag by 10 minutes or even longer. Without time alignment, the peak values ​​from different data sources will be distributed at different points on the timeline, distorting the unified time baseline and affecting the accuracy of subsequent identification of behavioral fluctuations and jumps.

[0057] Therefore, the core function of introducing a time delay correction function is to automatically identify and adjust the lag of various indicators relative to the baseline time based on historical patterns, thereby achieving peak time alignment. This type of function can be implemented using the following common function forms:

[0058] Linear time offset function: By calculating the average peak latency of each metric in historical events (e.g., an average lag of 180 seconds for comments and 300 seconds for reposts), a fixed offset value is constructed, and the timestamp of each metric is shifted forward by the corresponding number of seconds. This method is simple and efficient, and suitable for scenarios with stable lag patterns.

[0059] The sliding cross-correlation function (SCF) calculates the time offset corresponding to the maximum correlation between two time series using a sliding window approach, finding the optimal alignment point between the two indicators. This method can dynamically adapt to the lag relationship at different times and is suitable for complex public opinion events.

[0060] Dynamic Time Warping (DTW) uses a dynamic programming algorithm to non-linearly align two time series, making their overall shapes as similar as possible. This method is suitable for aligning indicators with inconsistent fluctuation patterns but consistent overall trends, offering high accuracy but also significant computational cost.

[0061] Exponential decay adjustment function: For metrics with severe lag, an exponential decay model is used for weighted forward shifting to simulate the response curve of user behavior and achieve flexible time alignment. This method is suitable for data types with slow responses, such as comment and interactive data.

[0062] In summary, the role of the time delay correction function in this embodiment is to uniformly align high-frequency data from multiple platforms with different response delays, so that various indicators form a consistent structure on a unified time baseline, improve the accuracy of data fusion and the sensitivity of behavior fluctuation detection, and is a key preliminary step for constructing a multidimensional steady-state baseline and subsequent jump identification.

[0063] A multi-time-domain steady-state baseline is constructed on top of a unified time baseline to identify stable behavioral trends at different time scales. Specifically, three time-domain branches are constructed: a short-term baseline, a medium-term baseline, and a long-term baseline. The short-term baseline uses a 5-minute observation window with a 60-second step, employing the standard deviation of the rate of change of search volume, comment volume, repost volume, and click volume within the sliding window as a reference value to measure the stability of the indicators. The medium-term baseline uses a 30-minute window with a 5-minute step, using the mean and standard deviation of similar indicators as the upper and lower boundaries of the stable interval. The long-term baseline uses a 3-hour window with a 30-minute step, employing a moving average method to calculate the long-term trend of each indicator, and setting three times the standard deviation as a fluctuation warning line. Each type of time-domain baseline maintains a baseline state record set that is updated over time. This record set stores stability indicator values, deviation rates, critical approach frequencies, etc., over a past period, used to determine whether the current behavior is transitioning from a steady state to a jump state. Among the three baselines, the short-term baseline primarily responds to sudden outbreaks of public opinion, the medium-term baseline responds to situations of sustained high public attention, and the long-term baseline identifies periodic changes in public opinion and distinguishes them from normal fluctuations. By operating these three types of time domains in parallel, the system possesses a comprehensive reference system when facing public opinion events of varying intensities and rhythms.

[0064] Short-window fluctuation intensity sequences are extracted based on a unified time baseline and multi-time-domain steady-state baselines, and sentiment gradient sequences are generated simultaneously as core dynamic inputs for subsequent modeling and prediction. The short-window fluctuation intensity sequence is constructed as follows: The change amplitude (i.e., maximum value minus minimum value) of four high-frequency indicators is calculated every 5 seconds within the corresponding window, and compared with their average level in the corresponding time-domain steady-state baseline to obtain the relative offset rate of each window. This offset rate is then weighted and summarized to obtain the fluctuation intensity value at the current time point. The sentiment gradient sequence is generated as follows: Each comment text undergoes Chinese word segmentation and stop word removal, and a sentiment dictionary scoring model is introduced to label the sentiment tendency score of each sentence. The trend of sentiment score changes within each window is statistically analyzed, and its first derivative is calculated as the sentiment gradient value, reflecting the speed at which the current user's sentiment evolves towards a positive or negative direction. Finally, the short-window fluctuation intensity values ​​and sentiment gradient values ​​are aligned on the same time axis to construct a bivariate sequence containing behavioral intensity and sentiment change trends, providing accurate multi-dimensional input for subsequent phase trajectory recognition and risk segment prediction.

[0065] Under the constraints of short-window fluctuation intensity sequence and sentiment gradient sequence, multi-scale spectral decomposition is performed and cross-channel time delay alignment is carried out to reconstruct phase trajectory, identify phase acceleration peaks and extract phase transition warning segments to obtain causal verification candidate windows;

[0066] To accurately identify abrupt trends and construct an analysis window for subsequent causal verification, multi-scale spectral expansion, time alignment, phase trajectory reconstruction, and warning segment extraction are performed based on the generated short-window fluctuation intensity sequence and sentiment gradient sequence, as detailed below:

[0067] Under the dual constraints of short-window fluctuation intensity sequence and sentiment gradient sequence, continuous wavelet transform is performed at each time point to complete multi-scale spectral decomposition. In the implementation, the short-window fluctuation intensity sequence is derived from the weighted average of the changes in four high-frequency indicators—search volume, comment volume, repost volume, and click volume—under a five-second sliding window. The sentiment gradient sequence is derived from the evolution rate of the positive and negative sentiment trends in user comments at a five-second granularity. Both sequences are aligned to a unified baseline in time, and a third-order moving average is used to denoise abnormal peaks. Subsequently, Morlet mother wavelet expansion is applied to both sequences to extract a set of spectral coefficients covering seven scales, with the scale range covering a variation period from 10 seconds to 180 seconds, ensuring coverage of the entire frequency band of sudden public opinion response rhythms. Low-scale components are used to capture sudden short-term strong oscillations, while high-scale components are used to characterize background rhythm fluctuations. The position of the frequency component with the largest amplitude and its corresponding time point are extracted at each scale to form a frequency-dominant spectrum. The short-window fluctuation intensity spectrum and the emotion gradient spectrum are compared point by point. Points with amplitude direction consistency exceeding 90% and phase synchronization error less than three seconds are marked as potential jump signal candidate points and proceed to the next stage of processing.

[0068] Based on the obtained potential transition signal candidate point set, cross-channel time delay alignment processing is required to eliminate response time differences between data from different platforms. In this step, time series cross-mutual information matrices are constructed for the four types of indicators. Taking the search volume time series as the benchmark, the mutual information score curves of its interaction with comment volume, forward volume, and click volume within a sliding window are calculated, and the time lag value corresponding to the peak point is identified. For example, if the maximum mutual information between search volume and click volume occurs at a lag of 20 seconds, the entire click volume time series is shifted forward by 20 seconds. The above adjustment process processes non-integer time points through linear interpolation and fills in missing values ​​at the boundaries. For the sentiment gradient series, comment volume is used as a reference, and the same time delay alignment operation is performed based on the overlap rate between the two on the transition candidate point set. After all time series are aligned, they are recombined in chronological order on a unified time axis to form a single aligned composite behavioral spectrum structure. Each time point contains the alignment fluctuation intensity of the four high-frequency indicators (search, comment, forward, and click) and the corresponding sentiment change rate at that time point.

[0069] Based on the composite behavioral spectrum structure, the phase evolution path of the behavioral trajectory is reconstructed, and the phase acceleration peak region is identified. The phase trajectory is calculated as follows: First, the dominant frequency component in the spectrum of each type of indicator is selected, and an analytical signal is constructed from the dominant frequency waveform using Hilbert transform to extract instantaneous phase information. For each time point, the phase values ​​of the four types of indicators are calculated, and their weighted average is used to construct a unified phase sequence. The weights are set based on the average change intensity of each indicator over the past ten minutes. The obtained phase sequence is further differentiated to obtain the phase velocity sequence, and then the second-order difference is performed to obtain the phase acceleration sequence. Among all time points, points with phase acceleration values ​​greater than twice the historical mean of the sequence are identified as "candidate peak points". For periods with three or more consecutive peak points, they are judged as phase transition activation segments. To avoid misidentifying isolated peaks, this step introduces an emotional gradient acceleration judgment constraint, that is, if the emotional gradient value in the peak segment increases by more than 0.3 (emotional value units) between two adjacent time points, it is confirmed as a real transition segment, and the start and end times of the transition are recorded.

[0070] Based on the identified phase transition activation segments, a causal verification candidate window is constructed. This window is generated based on the fusion of three indicators: fluctuation intensity integral value, emotion gradient slope, and phase acceleration amplitude. For each transition activation segment, firstly, the cumulative sum of fluctuation intensity values ​​within the range of 30 seconds before the transition begins and 60 seconds after it ends is calculated as an indicator of the total fluctuation; then, the linear fitting slope of the emotion gradient over the same time period is calculated, representing the stability of the emotion change direction; finally, the maximum value of the phase acceleration peak is extracted. These three indicators are standardized and weighted, with weight ratios of 0.4, 0.3, and 0.3 respectively. The top five transition activation segments with the highest fusion scores are selected as the final causal verification candidate windows, and their start and end times are precisely recorded to ensure sufficient behavioral span and a complete contextual event environment for subsequent analysis of the reach, interaction, and conversion chains.

[0071] Within the phase transition early warning segment, a consistency check of the reach-interaction-transformation causal chain is carried out. Based on the lead-lag structure, the stability of the causal direction is calculated. Time segments in which the causal direction reverses are marked as risk segments, and the inference input is obtained.

[0072] To identify specific risks associated with abnormal responses in the marketing chain during abrupt events, this step establishes a behavioral sequence causal direction judgment mechanism based on the extracted phase transition warning segments, verifies the consistency of user responses across the stages of reach, interaction, and conversion, and constructs a reversal segment identification mechanism, as detailed below:

[0073] A complete user behavior chain sequence is constructed within the phase transition warning segment for subsequent causal direction verification. User behavior data is divided into three time series: reach behavior sequence, interaction behavior sequence, and conversion behavior sequence. The reach behavior sequence refers to the total number of times a user visits a page, advertisement, or search results page within each five-second time window within the warning segment; the interaction behavior sequence includes the number of user participation behaviors such as clicks, likes, and comments within the corresponding time window; and the conversion behavior sequence refers to the total number of outcome actions such as purchases, order submissions, and contact information entry within the time window. The three types of sequence data are sampled from user access logs under a unified tracking path, with a sampling frequency of five seconds to ensure a one-to-one correspondence between the sequences. Each type of sequence is normalized to ensure consistency in numerical scale for subsequent time series comparisons. During construction, the analysis window length for each time point is set to 120 seconds, specifically including two directions: 60 seconds forward and 60 seconds backward, to capture the temporal response delay characteristics of the behavior chain.

[0074] Based on the constructed three types of behavioral time series, causal path delay analysis was performed on the two stages of reach-to-interaction and interaction-to-conversion to identify the optimal lag relationship. Specifically, using the reach behavior sequence as a baseline, each time point was shifted backward by 0 to 60 seconds. The Pearson correlation coefficient between the two sequences was calculated at each lag step, comparing it with the trend of the interaction behavior sequence within the same time period, and the lag time point corresponding to the maximum correlation coefficient was recorded. Subsequently, using the interaction behavior sequence as a baseline, the same method was used to match it with the conversion behavior sequence to obtain the optimal lag time point for the second stage. The two lag time points represent the maximum response delay of reach-driven interaction and interaction-driven conversion, respectively. In each stage, the correlation coefficient within a range of ±5 seconds around the lag peak was considered a stable interval. If the mean correlation within this interval is not lower than 0.6, the lag structure is considered to have high causal consistency. If the correlation is lower than 0.3 or negative, the causal relationship in this stage is preliminarily determined to be disordered, requiring further verification.

[0075] Based on the results of lag structure identification, a causal direction stability sequence is established to quantify the temporal distribution of causal direction reversal phenomena in the behavioral chain. The method is as follows: For each five-second time window, the causal direction state is marked as positive, negative, or without a clear trend, depending on whether the reach precedes the interaction and whether the interaction precedes the conversion. Specifically, if the reach peak is earlier than the interaction peak, and the interaction peak is earlier than the conversion peak, it is marked as positive; if the order is disrupted, i.e., the fluctuation signal of interaction or conversion appears before its preceding behavior, it is marked as negative; if there is no significant correlation, it is marked as null. In each 120-second sliding window, the ratio of the number of occurrences of positive state to the number of occurrences of negative state is counted as the causal direction stability value of that window, ranging from [-1, 1]. If the number of negative states in a window exceeds twice that of positive states, it is determined to be a causal direction reversal segment. To eliminate occasional reversals caused by data errors, this implementation method also sets the following conditions: when the causal direction reverses, the short-window fluctuation intensity must be observed to be more than twice its mean within the same time window, and the emotional gradient direction must undergo a rapid shift with an amplitude greater than 0.4. Only when all three conditions are met simultaneously can a high confidence risk characteristic be constituted.

[0076] Windows meeting the above conditions are marked as risk segments, and their boundary times are defined as input time intervals for subsequent counterfactual inference stages. The starting point of a risk segment is set at the time point when the stability of the causal direction is first observed to drop below -0.5, and the ending point is the time point when the stability recovers to above zero and remains for three consecutive time windows. To ensure the integrity of the behavioral characteristics of the identified segments, this implementation also sets a minimum duration of no less than 30 seconds, and it must contain at least one complete reach-interaction-conversion behavioral chain dynamic process. After completing the above screening, all time periods that meet the conditions are summarized into a risk segment set, and their start and end times, lag structure fluctuation amplitude, correlation disruption degree, and abnormal intensity of behavioral chain response are recorded in the form of time tags. This information will be used to construct boundary conditions, set intervention variables, and assess jump intensity in subsequent counterfactual trajectory inference, providing contextual basis for the strategy correction mechanism.

[0077] Counterfactual stable trajectories are generated for risk segments. Parallel extrapolation is performed using frozen baselines, sentiment-neutral substitutions, and noise-suppressed probes. The relative divergence between the true trajectory and the counterfactual trajectory is calculated. During the extrapolation process, the number and proportion of phase transitions exceeding the threshold are statistically analyzed to obtain the jump intensity index and the phase transition threshold rate.

[0078] To accurately assess the anomalous perturbations caused by causal reversal fragments to marketing behavior paths, a stability quantification method based on counterfactual construction and behavioral comparison is proposed, based on identified risk fragments. The method includes the following steps:

[0079] For the identified risk segments, multiple counterfactual stable trajectories are constructed to simulate the natural evolution of user behavior in the absence of sudden abrupt events. The counterfactual trajectory construction is based on the original behavior sequence and incorporates three types of intervention mechanisms for trajectory replacement and generation. The first type of intervention mechanism is baseline freezing replacement, which involves extracting the average values ​​of outreach behavior, interaction behavior, and conversion behavior within the first 60 seconds of the risk segment and constructing a stable input sequence of the same length as the risk segment. This sequence does not contain any high-frequency mutations or abrupt trend changes and only reflects basic traffic and natural behavior expectations. The second type of intervention mechanism is emotion-neutral replacement, which involves cleaning and segmenting all user comment data within the risk segment, identifying extreme emotional words (such as "angry," "bad review," "disgusting," etc.), and replacing them with neutral words with zero semantic strength (such as "received," "okay," "see," etc.). Then, the emotion gradient sequence is recalculated, and the results are embedded into the original behavior sequence to simulate the user's behavioral feedback path in a non-excited emotional state. The third type of intervention mechanism is noise suppression probe injection, which involves identifying outlier outreach behavior data points within the risk segment and replacing them with corresponding position data from the 30-second stable behavior segment with the smallest fluctuations over the past three hours to remove local jumps caused by single-point anomalies. All three types of intervention trajectories are generated with the same time granularity as the original trajectory, producing a behavioral sample every five seconds, forming three parallel versions of counterfactual behavioral trajectories, which are used for subsequent comparative analysis.

[0080] The actual behavioral trajectory is compared hourly with three counterfactual behavioral trajectories, and multidimensional dynamic offset analysis is performed using phase information to assess the differences and intensity of abrupt changes. In this step, a sliding time window mechanism is used to construct the comparison window, with a window length of 30 seconds and a step size of 5 seconds to ensure full coverage of the risk segment and sufficient temporal resolution. Within each time window, the numerical differences in reach behavior, interaction behavior, and conversion behavior between the actual trajectory and the frozen baseline trajectory, the emotion-neutral trajectory, and the noise-suppressed trajectory are calculated. The difference index is defined as the sum of the squares of the differences between the corresponding time points of the same behavioral dimension between the actual and counterfactual trajectories, and normalized to a offset intensity score. The average offset intensity score of the three counterfactual trajectories is taken as the total behavioral offset within that time window. Simultaneously, combined with the phase trajectory calculated in the previous stage, the phase change trend within the current time window is differentially processed to obtain the phase change rate within the current time window, which is compared with the phase trajectory derived from the counterfactual trajectory to obtain the absolute value of the phase difference, serving as the phase perturbation intensity index for the current time window. The total behavioral offset is weighted and superimposed with the phase perturbation intensity to construct a "joint perturbation value," which represents the overall deviation between the true trajectory and the counterfactual trajectory within the current window, and is used for the next step of indicator extraction.

[0081] After completing the multi-window comparative analysis, two core stability assessment indicators corresponding to the risk segment are extracted: the jump intensity index and the phase transition threshold rate. The jump intensity index is defined as the weighted average of the joint perturbation values ​​of all time windows within the entire risk segment. The weights are calculated based on the standard deviation of behavioral fluctuations in the true trajectory within each time window; the more severe the fluctuation, the higher the weight, to enhance the impact of the jump segment on the overall assessment result. The jump intensity index ranges from 0 to 1; the higher the value, the greater the deviation of the true trajectory from the counterfactual trajectory, and the more obvious the jump characteristics. The extraction process of the phase transition threshold rate is as follows: all phase transition events are identified in the sliding time window, specifically those whose phase change rate exceeds three times the historical average of the sequence within two consecutive time points, and whose corresponding behavioral deviation value is higher than twice its average. The total number of time windows meeting the conditions is divided by the total number of sliding windows to obtain the transition window ratio, which is the phase transition threshold rate. If the jump intensity index is higher than 0.7 and the phase transition threshold rate is higher than 0.4, the risk segment is marked as a high-risk jump segment and the policy intervention mechanism will be triggered first in the next stage. All segment evaluation results, including parameters such as jump intensity index, phase transition threshold rate, highest behavior offset time point, and longest continuous perturbation duration, will be recorded and input into the subsequent policy control module to support dynamic weight adjustment and policy path repair.

[0082] The jump intensity index, phase transition threshold rate, causal direction stability and short window fluctuation intensity sequence are weighted and fused to generate window-level predicted jump judgment, and output the predicted jump event label, dominant feature sub-band, severity level and intervention time point.

[0083] To accurately identify strategy jump trends at the window level and assist the strategy system in dynamic adjustment, a weighted fusion analysis is performed on a unified time baseline based on the jump intensity index, phase transition threshold rate, causal direction stability, and short-window fluctuation intensity sequence obtained from the simulation results. This ultimately generates jump determination results and outputs event labels, dominant feature subbands, severity levels, and intervention points. Specifically, the following steps are included:

[0084] Four indicators—jump intensity index, phase transition threshold rate, causal direction stability, and short-window fluctuation intensity sequence—are preprocessed to ensure comparability and consistency in the dimensions and distribution ranges of the data. The original value of the jump intensity index is the behavioral-phase joint offset score of the true trajectory and the three types of counterfactual trajectories, with a fixed value range between 0 and 1. The phase transition threshold rate represents the proportion of windows with drastic phase changes per unit time, also with a value between 0 and 1. The causal direction stability, obtained through behavioral chain consistency analysis in the previous stage, originally ranged from -1 to 1. This step uses a linear transformation to map it to the 0-1 range, giving it a positive contribution characteristic in the weighted fusion. The short-window fluctuation intensity sequence is the result of normalizing the fluctuation amplitudes of four types of behaviors—search, comment, forward, and click—at a five-second granularity. Its mean is set to 0.5, the maximum value to 1, and the minimum value to 0, ensuring consistent response in the fusion calculation. After processing, the four indicators constitute a complete feature vector sequence at each time point, laying the foundation for subsequent scoring calculations.

[0085] A fusion strategy is set and window-level scoring is performed to form the basis for judging jump risk. In this embodiment, a fixed weight strategy is used to complete the fusion of multi-dimensional indicators: the jump intensity index is assigned a weight of 0.35, reflecting the overall trajectory deviation trend; the phase transition threshold rate is assigned a weight of 0.30, representing the frequency of trajectory structure disturbances; the causal direction stability is assigned a weight of 0.20, representing the coherence of the causal chain; and the short-window fluctuation intensity is assigned a weight of 0.15, used to measure the intensity of micro-level behavior. At each time point, the four indicator values ​​are multiplied by their corresponding weights and then summed to obtain the jump score at that time point. To reduce the impact of occasional anomalies, a three-window moving average is used to process the jump score sequence, that is, the score at each time point is the average of its own score and the scores of the window before and after it, as the smoothed score result. The smoothed score value range is also between 0 and 1, with a larger value indicating a stronger probability of jump.

[0086] Abrupt event identification is performed based on the smoothed scoring sequence, and the existence of the event is confirmed in a window-level structure. This step sets a threshold of 0.65 for abrupt scoring. When the scores in three or more consecutive time windows are not lower than this threshold, the current time segment is confirmed as an abrupt event segment. If only one or two isolated windows have scores higher than the threshold, it is judged as low-significance fluctuation and does not constitute an abrupt event. After confirming an abrupt event, its start time (the first window exceeding the threshold) and end time (the last high-score window after the score is first continuously lower than the threshold for three windows) are recorded. Each abrupt event segment must have a minimum duration of 15 seconds, i.e., at least three consecutive windows, to ensure that the event has sufficient behavioral persistence. The event segment time information output in this step serves as a key basis for subsequent label generation and strategy intervention time point identification.

[0087] For confirmed jump event segments, content classification, behavioral feature localization, and severity determination are performed to generate structured event descriptions. Jump event type labels are determined by ranking the percentage of scoring components: if the jump intensity index has the highest percentage within the event segment, it is labeled "Behavioral Deviation Type"; if the phase transition threshold rate is dominant, it is labeled "Trajectory Abrupt Change Type"; if the causal direction stability is the lowest, it is labeled "Link Break Type"; if the short-window fluctuation intensity is the largest, it is labeled "Severe Behavioral Type". The dominant feature sub-band identification method is as follows: within the event segment, sequentially check which of the three data types—reach behavior, interaction behavior, and conversion behavior—experienced the highest fluctuation value within 5 seconds before the jump score surge; the corresponding behavioral dimension is the dominant feature. The severity level is set based on the highest score of the event segment: a score between 0.65 and 0.75 is a mild jump, a score between 0.75 and 0.85 is a moderate jump, and a score above 0.85 is a high jump. Combining the duration and dominant features, the output results can be further refined into structural labels such as "moderate-interactive jump" or "high-link break jump".

[0088] Based on the trend of the scoring curve of the abrupt change event, the optimal time point for strategic intervention is identified, and a complete event structure is output for subsequent use. The intervention time point is the point when the score within the event segment first reaches its local maximum value. If there is a plateau in the scoring curve, the starting point of the plateau is used as the intervention suggestion point, indicating that intervention should be carried out before the risk of abrupt change is maximized. The final output includes seven elements: start time of the abrupt change event, end time, peak score, intervention time point, event type label, dominant behavior type, and severity level. Each element corresponds to a clear value or identification rule, leaving no room for speculation, ensuring that the event description has a structural foundation that can be directly used for strategy reasoning and intervention scheduling.

[0089] Based on the window-level prediction jump judgment results, a weight inversion mechanism is introduced to apply negative weight compensation to abnormal data in risk segments, and dynamic replenishment is carried out with the support of multi-time domain steady-state baselines, so that the delivery strategy output maintains a balance between historical trends and local corrections, and obtains a continuous and stable delivery strategy.

[0090] To ensure the continuity of the strategy deployment path, reduce misjudgment intervention, and achieve trend-based balance control after identifying abrupt changes, a weight inversion and dynamic replenishment mechanism is constructed based on the window-level prediction abrupt change judgment results. This enables stable processing of risk segments and generates a final deployment strategy that combines behavioral continuity and local correction capabilities. The specific steps include the following:

[0091] Based on the jump event labels, score peaks, time period start and end points, and dominant feature sub-band information output in the window-level prediction jump judgment results, abnormal data in risk segments are identified. The identification process is based on behavioral time-series data, using a sliding window check at a time granularity of five seconds to filter out behavioral points with sharp increases or decreases in the jump event segment. The specific judgment criteria are as follows: if the increase or decrease of a behavior in the current time window exceeds three times its average change in the previous hour, and the behavior type is exactly consistent with the dominant feature sub-band of the jump event, it is marked as a "critical anomaly"; if the phase trajectory acceleration direction in the same time window is reversed compared to the previous time window, and the behavioral fluctuation amplitude exceeds twice the average level, it is marked as a "structural disturbance point". Anomalies of the two categories are merged to form a set of abnormal data points to be processed, which serves as the starting object for subsequent strategy interventions.

[0092] After identifying outlier data points, a compensation process based on negative weight adjustment is performed on these data points to offset the drastic disturbances they cause to the strategy calculation process. This step proposes a weight inversion mechanism, which assigns negative weights to behavioral data before it participates in the strategy model, making its impact value lower than the actual observed value, or even showing a weakening trend. The operation is as follows: using jump scores as a guide, if the score of a certain outlier data point in its time window exceeds 0.85, the corresponding negative weight is set to -1; if the score is between 0.75 and 0.85, the negative weight is set to -0.6; if the score is between 0.65 and 0.75, the negative weight is set to -0.3. This negative weight does not directly modify the original behavior value, but rather scales the strategy participation coefficient of that behavior value. For example, if the original value of a certain click behavior is 220 times, if its negative weight is -0.6, only 88 times will be counted when participating in the strategy evaluation, i.e., 220 × (1 - 0.6), significantly reducing its disturbance intensity. In this way, the dominance of outliers in decision input is systematically reduced, avoiding unreasonable and drastic adjustments to the strategy at fluctuation points.

[0093] To address the structural gaps in the strategy input caused by negative weighting compensation, a multi-time-domain steady-state baseline is introduced as a reference template. Dynamic partitioning and backfilling are then performed to maintain the structural integrity and trend continuity of the delivery path. The multi-time-domain steady-state baseline is constructed from historical behavioral data, including a short-cycle baseline (based on the average value of the previous 5-minute window), a medium-cycle baseline (based on the sliding trend of the previous 30 minutes), and a long-cycle baseline (based on the global average trend of the past 3 hours). These three time-domain baselines correspond to high volatility, stable transition, and global reference scenarios, respectively. During the backfilling process, the current anomaly is first assessed to determine the type of abrupt change: if the abrupt change score is accompanied by high-frequency fluctuations, the short-cycle baseline is used; if the score remains stable but the structural trend shifts, the medium-cycle baseline is used; if the score fluctuations are small but the impact spans a long period, the long-cycle baseline is used. After determining the baseline type, steady-state values ​​from two adjacent time windows are selected around the anomaly, and linear interpolation is performed to construct new behavioral values ​​as replacement data for the anomaly. By performing the backfilling operation, not only are the input gaps caused by negative weight processing filled, but the behavioral sequence is also ensured to have trend continuity over time, avoiding local pruning or misalignment.

[0094] After completing negative weight compensation and baseline restoration, the overall fusion output of the strategy path is executed, ensuring the final strategy curve maintains stable progress across the entire time axis. This step employs a strategy repair fusion mechanism to smooth out potential boundary faults between the restored and original segments. Specifically, a 10-second "transition fusion band" is inserted before and after the start and end points of each jump event segment. Behavioral data within the fusion band is connected to adjacent segments via linear slope adjustment, ensuring a natural transition in slope and amplitude of the behavioral change curve. Furthermore, when generating the final strategy output, the correction mechanism is applied only to segments that experience jumps; all non-jump segments retain their original behavioral values ​​and strategy input values. The entire strategy output includes a continuous time behavioral sequence, the start and end points of the jump correction segment, statistical differences in behavioral values ​​before and after correction, and the ratio of the impact of behavioral correction on the volatility of the strategy curve. Through full-process control, the strategy deployment path can respond quickly to sudden public opinion interference without deviating from its original rhythm, truly achieving a dual balance of "short-term response + long-term consistency."

[0095] This invention constructs a multi-time-domain steady-state baseline and a public opinion trigger monitoring matrix to achieve unified modeling and sequential expression of multi-dimensional indicators such as search volume, comment volume, repost volume, and click volume. It further introduces multi-scale spectral decomposition, cross-channel time delay alignment, and phase trajectory reconstruction mechanisms to accurately identify hidden phase transition characteristics and potential behavioral chain breakpoints during sudden public opinion changes. Simultaneously, by combining causal direction stability assessment and counterfactual trajectory comparison and deduction, the degree of jump risk is quantified, improving the structured understanding of strategy disturbance trends. In the strategy output stage, a weight inversion and dynamic replenishment mechanism is further introduced to weaken and steadily repair abnormal data within risk segments, achieving closed-loop control across the entire chain from behavior identification to strategy correction. Overall, this invention not only significantly improves the predictive accuracy and stability of marketing strategies in extreme scenarios but also enhances the robustness and adaptability of the model in the face of nonlinear disturbances, thereby significantly reducing budget waste, improving campaign efficiency, and enhancing the consistency and risk resistance of brand communication.

[0096] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An AI-powered intelligent marketing data-driven strategy generation method, characterized in that, Includes the following steps: Establish a public opinion trigger monitoring matrix, aggregate multiple high-frequency indicators, construct a unified time baseline and generate multi-time domain steady-state baselines, and obtain short-window fluctuation intensity sequence and sentiment gradient sequence; The method for constructing the short-window fluctuation intensity sequence is as follows: The change amplitude of four types of high-frequency indicators is calculated every 5 seconds within the corresponding window, and compared with their average level in the corresponding time-domain steady-state baseline to obtain the relative offset rate of each window. The offset rates are then weighted and summarized to obtain the fluctuation intensity value at the current time point. The method for generating the sentiment gradient sequence is as follows: Each comment text is processed by Chinese word segmentation and stop word removal, and a sentiment dictionary scoring model is introduced to label the sentiment tendency score of each sentence. The trend of sentiment score changes within each window is statistically analyzed, and its first derivative is calculated as the sentiment gradient value, reflecting the speed at which the current user's sentiment evolves towards a positive or negative direction. Under the constraints of short-window fluctuation intensity sequence and sentiment gradient sequence, multi-scale spectral decomposition is performed and cross-channel time delay alignment is carried out to reconstruct phase trajectory and identify phase acceleration peaks, extract phase transition warning segments, and obtain causal verification candidate windows. The generation of causal verification candidate windows is based on the fusion of three indicators: fluctuation intensity integral value, emotion gradient slope, and phase acceleration amplitude. For each transition activation segment, the cumulative sum of fluctuation intensity values ​​within the range from 30 seconds before the start of the transition to 60 seconds after the end is first calculated as an indicator of the total fluctuation. Then, the linear fitting slope of the emotion gradient in the same time period is calculated to represent the stability of the direction of emotion change. Finally, the maximum value of the phase acceleration peak is extracted. After standardizing these three indicators, they are weighted and fused with weight ratios of 0.4, 0.3, and 0.3 respectively. The top five transition activation segments with the highest fusion scores are selected as the final causal verification candidate windows, and their start and end times are accurately recorded to ensure sufficient behavioral span and complete contextual event environment when performing subsequent analysis of the reach, interaction, and conversion chain. Within the causal verification candidate window, perform consistency verification of the reach-interaction-conversion causal chain, calculate the stability of the causal direction, and mark the time segment where the causal direction is reversed as a risk segment to obtain the inference input; The consistency verification of the causal chain of reach-interaction-conversion refers to: constructing a complete user behavior chain sequence within the phase transition warning segment; dividing the data into reach behavior sequence, interaction behavior sequence, and conversion behavior sequence, all derived from user access log sampling under a unified tracking path, with a sampling frequency maintained at five seconds and normalized; using a 120-second analysis window, performing causal path delay analysis in two stages: reach-to-interaction and interaction-to-conversion; shifting the baseline sequence backward by 0 to 60 seconds to calculate the Pearson correlation coefficient and recording the optimal lag time point; if the lag peak is ±5 seconds... A correlation mean of 0.6 or higher is considered highly consistent causally, while a mean of 0.3 or a negative value is initially considered disordered causality. Based on this, each five-second window is marked as positive, negative, or null. Within a 120-second sliding window, the stability of the causal direction is defined by the ratio of the number of occurrences of positive and negative states. When the negative state exceeds twice the number of the positive state and simultaneously the short window fluctuation intensity exceeds twice its mean and the emotional gradient direction undergoes a rapid shift with an amplitude greater than 0.4, the time segment in which the causal direction reverses is marked as a risk segment and used as input for counterfactual inference. Counterfactual stable trajectories are generated for risk segments and compared and deduced. The relative divergence between the real trajectory and the counterfactual trajectory is calculated, and the phase transition threshold characteristics are statistically analyzed to obtain the jump intensity index and the phase transition threshold rate. The jump intensity index is defined as the weighted average of the joint disturbance values ​​of all time windows within the entire risk segment. The weights are calculated based on the standard deviation of the behavioral fluctuations in the actual trajectory within each time window. The more violent the fluctuations, the higher the weight, so as to enhance the impact of the jump segment on the overall assessment results. The jump intensity index ranges from 0 to 1. The higher the value, the greater the deviation of the true trajectory from the counterfactual trajectory, and the more obvious the jump characteristics. The extraction process of the phase transition threshold rate is as follows: Identify all phase transition events in the sliding time window, specifically those whose phase change rate exceeds three times the historical average of the sequence within two consecutive time points, and whose corresponding behavior offset value is more than twice its average; divide the total number of time windows that meet the conditions by the total number of sliding windows to obtain the transition window ratio, which is the phase transition threshold rate. A window-level prediction jump determination is generated by weighted fusion of jump intensity index, phase transition threshold rate, causal direction stability and short window fluctuation intensity sequence. Based on the window-level prediction jump judgment results, a weighted inversion mechanism is introduced to compensate for abnormal data in risk segments and dynamically replenish it under the multi-time domain steady-state baseline. This ensures that the delivery strategy output maintains a balance between historical trends and local corrections, resulting in a continuously stable delivery strategy.

2. The AI-powered intelligent marketing data-driven campaign generation method according to claim 1, characterized in that, The steps for obtaining the short-window fluctuation intensity sequence and sentiment gradient sequence are as follows: Four types of user behavior data were collected: search volume, comment volume, forward volume, and click volume. The raw indicator sequences were obtained by keyword query records, comment area statistics, forward record aggregation, and click event counting, respectively. The index series is processed with unified timestamps, and missing values ​​are filled with linear interpolation. The maximum and minimum normalization is used to scale the data to a uniform scale to complete the data fusion. By analyzing the average lag time of different indicators in historical events, a delay correction function is constructed, the original data sequence is time-shifted to generate a unified time baseline, and the data is divided into statistical structured indicators with equal time windows. Based on a unified time baseline, three types of time-domain steady-state baselines—short-term, medium-term, and long-term—were constructed. The standard deviation of the rate of change, the mean and standard deviation, and the moving average and standard deviation threshold were calculated for each. Short-window fluctuation intensity sequences and sentiment gradient sequences were extracted as inputs for subsequent analysis.

3. The AI-powered intelligent marketing data-driven campaign generation method according to claim 2, characterized in that, The steps for obtaining the candidate window for causality verification are as follows: Continuous wavelet transform is performed based on short-window fluctuation intensity sequence and emotion gradient sequence to extract spectral components at seven scales, construct frequency dominant spectrum, and mark candidate points of jump signals with phase synchronization error of less than three seconds. Based on the candidate point set of jump signals, cross mutual information is calculated to complete the time delay alignment of the time series and generate a composite behavioral spectrum structure. Based on the composite behavioral spectrum structure, the dominant frequency component is selected, the instantaneous phase and phase acceleration of four types of indicators are calculated, and the phase transition activation segment is extracted by combining the emotional gradient acceleration characteristics. For each phase transition activation segment, the integral of the fluctuation intensity, the slope of the emotion gradient, and the amplitude of the phase acceleration are calculated, fused to generate a score, and the transition segment with the highest score is selected as the candidate window for causal verification.

4. The AI-powered intelligent marketing data-driven campaign generation method according to claim 3, characterized in that, The time segment where the causal direction is reversed is marked as a risk segment, and the following steps are taken to obtain the inference input: Within the phase transition early warning segment, a contact behavior sequence, an interaction behavior sequence, and a conversion behavior sequence are constructed. User contact, participation, and outcome behavior data are collected, and the sampling frequency is uniformly set to five seconds to complete the normalization processing of the behavior chain time series. Based on behavioral time series, lag analysis is performed to calculate the optimal lag time points for reaching the two stages of interaction and interaction to transformation, and the causal path delay structure is confirmed when the correlation meets the stable interval condition. Based on the hysteresis structure, a causal direction stability sequence is constructed, the reverse behavior relationship state is marked, and combined with the short window fluctuation intensity and the change of sentiment gradient, the causal direction reversal segment is identified. The time window that meets the conditions of causal reversal, fluctuation intensity exceeding the threshold, and dramatic shift in emotional gradient is marked as a risk segment. Its start and end time and abnormal characteristics of behavioral chain response are extracted to form the input for subsequent inference.

5. The AI-powered intelligent marketing data-driven campaign generation method according to claim 4, characterized in that, When labeling risk segments, they are identified as high-confidence risk segments only when the causal direction stability is less than -0.5, the duration is not less than 30 seconds, and the period includes at least one complete dynamic process of the reach, interaction, and conversion behavior chain. The amplitude of the lag structure fluctuation and the degree of correlation destruction are then extracted as the basis for intervention variables in subsequent inferences.

6. The AI-powered intelligent marketing data-driven campaign generation method according to claim 4, characterized in that, The steps for obtaining the jump intensity index and phase transition threshold rate are as follows: Based on risk segments, three types of counterfactual behavioral trajectories are constructed: frozen baseline trajectory, emotion-neutral trajectory, and noise-suppressed trajectory, which respectively simulate the evolution path of user behavior under the condition of no sudden change events; The actual behavior trajectory is compared with the three types of counterfactual behavior trajectories hour by hour within a sliding time window. The total behavior offset and phase perturbation intensity are calculated respectively to generate a joint perturbation value sequence. The jump intensity index and phase transition threshold rate are calculated based on the joint perturbation value sequence to quantify the abnormal jump intensity and phase stability perturbation characteristics of risk segments.

7. The AI-powered intelligent marketing data-driven campaign generation method according to claim 6, characterized in that, The steps for generating window-level prediction jump determination are as follows: Normalization preprocessing is performed on the jump intensity index, phase transition threshold rate, causal direction stability and short window fluctuation intensity sequence to ensure consistent dimensions and construct a unified feature vector sequence; The indicators are weighted and fused with fixed weights, and a moving average is performed every five seconds to generate a smooth score sequence. Based on the smoothing score, a jump score threshold is set, and time window segments that continuously exceed the threshold are identified to determine whether they constitute a jump event and to determine the start and end time of the event. Analyze the scoring structure within the transition event segment to determine the transition type label, dominant behavioral characteristic subband, and severity level; Based on the local peak or plateau start point within the scoring sequence, the optimal intervention time for the transition event is output, and a complete event structure is formed for strategy intervention invocation.

8. The AI-powered intelligent marketing data-driven campaign generation method according to claim 7, characterized in that, Based on the window-level prediction jump judgment results, a weighted inversion mechanism is introduced to compensate for abnormal data in risk segments, and dynamic replenishment is performed under multi-time domain steady-state baselines. The steps are as follows: Based on the window-level prediction jump judgment results, identify the sudden increase and decrease behavior points and phase acceleration direction reversal points within the jump event segment, and construct an abnormal data point set; Based on the jump score, different levels of negative weight coefficients are set, and negative weight compensation is applied to abnormal data points. By reducing the behavioral participation value, the interference intensity of the behavior on the strategy output is reduced. The system invokes multiple time-domain steady-state baselines, selects short-period, medium-period, or long-period baselines based on the jump type of the outlier, and uses linear interpolation to dynamically replenish the outlier behavior data. After the backfilling is completed, a transition fusion zone is constructed at the boundary of the jump segment. The behavior curve is smoothly connected by a linear slope adjustment method, and the final delivery strategy sequence containing local correction and global stability is output.

Images