AI intelligent marketing based on multi-source data selling point generation method

Abstract

The application discloses an AI intelligent marketing selling point generation method based on multi-source data, relates to the technical field of big data analysis and application, and comprises the following steps: acquiring a uniform time matrix in the multi-source data access stage, constructing extreme value suppression prior based on the uniform time matrix, establishing a collection synchronization baseline, and realizing full-link time alignment through the collection synchronization baseline; acquiring extreme value diagnosis fingerprints under the constraint of full-link time alignment, generating threshold trajectories based on quantile residual spectrum and kurtosis skewness, and taking the threshold trajectories as dynamic benchmarks to capture and calibrate sudden extreme values. The application constructs time sequence consistency basis, accurately identifies and corrects data extreme values, ensures content authenticity and reliability in combination with causal correction and semantic screening, realizes dynamic closed-loop optimization of content generation and delivery through closed-loop feedback and phase control, and effectively improves the authenticity, stability and compliance of AI marketing content.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data analysis application, in particular to an AI intelligent marketing based on multi-source data selling point generation method. BACKGROUND

[0002] "AI intelligent marketing based on multi-source data selling point generation" refers to using artificial intelligence technology to fuse and analyze multi-dimensional data from different sources (such as user portrait, consumer behavior, market feedback, competitor information, social media comments, product attributes, etc.), automatically identify key advantage points that can impress target users, and through natural language generation or knowledge-driven content construction methods, convert these advantages into clear, accurate, and marketing value selling point descriptions. Its core lies in the dynamic interaction of multi-source data and the reasoning ability of deep learning models, avoiding the inefficiency and one-sidedness of traditional marketing relying on manual experience to screen selling points, realizing the automation, personalization and real-time of selling point generation, and thus improving the pertinence and conversion rate of marketing content.

[0003] The prior art has the following disadvantages:

[0004] In the prior art, the selling point generation process of AI intelligent marketing generally relies on real-time input of multi-source data and obtains results through feature calculation. However, when external market environment appears sudden events or user behavior changes dramatically, input data often presents sharp fluctuations. In this case, extreme value signals are easy to enter the feature calculation link without effective screening, and due to the lack of dynamic abnormality suppression and robustness processing mechanism, such extreme values may be mistakenly amplified in the weight calculation process, resulting in reverse exaggeration phenomenon of selling point generation results. Specifically, short-term sales surge is misjudged as long-term best-selling trend, and transient negative information peak is extended to overall failure evaluation, resulting in significant deviation of generated content from actual situation. Once the above deviation is spread in large-scale delivery scenarios, it is easy to be tracked and verified at user side and regulatory side, thereby causing trust crisis and possibly bringing legal compliance risks such as false advertising identification.

[0005] The above information disclosed in the background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0006] The purpose of the present application is to provide an AI intelligent marketing based on multi-source data selling point generation method to solve the problems in the background technology.

[0007] In order to achieve the above purpose, the present application provides the following technical scheme: an AI intelligent marketing based on multi-source data selling point generation method, comprising the following steps:

[0008] A uniform time matrix is acquired in the multi-source data access stage, an extreme value suppression prior is constructed based on the uniform time matrix, and a collection synchronization baseline is established, time alignment of the whole link is realized through the collection synchronization baseline, so as to provide stable timing consistency;

[0009] An extreme value diagnostic fingerprint is acquired under the constraint of the whole link time alignment, a threshold trajectory is generated based on the quantile residual spectrum and kurtosis skewness, and the threshold trajectory is taken as a dynamic reference, so as to capture and calibrate the burst extreme value;

[0010] A causal de-biased view is acquired under the constraint of the threshold trajectory, a tool variable method is used to perform timing backtracking on the promotion intervention, the backtracking result is corrected in combination with the extreme value diagnostic fingerprint, and a real trend mapping is generated;

[0011] A semantic consistency constraint is acquired under the guidance of the real trend mapping, a cross-domain contrast coding scorer is constructed, and the selling point description is screened in combination with the causal de-biased view, so that the generated semantic output is consistent with the real trend;

[0012] A closed-loop feedback device is acquired under the semantic consistency constraint, small-sample empirical flow is injected into the weight updating link of the real trend mapping, and dominant index re-weighting is performed, so as to generate a stable weight field, so as to ensure the balance and persistence of the output;

[0013] A phase programmable launch control mechanism is acquired under the driving of the stable weight field, a self-triggering dynamic closed loop is formed through the spectrum return traction, time difference micro-injection and gain trajectory return strategy, and the selling point generation is continuously optimized in the dynamic closed loop, and the reverse exaggeration phenomenon is suppressed.

[0014] Preferably, the full-link time alignment step is as follows:

[0015] Raw data is collected from multiple data sources, a set of time fields of multi-source heterogeneous marketing related data is extracted, and is uniformly converted into coordinated universal time format, including time fields in user purchase behavior data, marketing activity execution data, commodity dimension information, user comment content, social platform content and market feedback data, to form a uniform time sequence with a minimum interval of milliseconds;

[0016] On the basis of the uniform time sequence, the behavior density change of each source data in the same time period is extracted, an extreme value suppression prior is constructed, and a risk label is generated by identifying the sudden increase interval through residual spectrum analysis;

[0017] Based on the uniform time sequence and the risk label, a collection synchronization baseline is established, and time sequence rearrangement and behavior node closure are realized through sliding window analysis and behavior chain relationship matrix;

[0018] According to the collection of synchronous baseline, the full-link time alignment process is constructed, the time stamp legality, behavior node rationality and risk label association of all incremental data are compared, and the data standardization and semantic integration are realized.

[0019] Preferably, under the constraint of full-link time alignment, the extreme value diagnostic fingerprint is obtained, and the threshold trajectory is generated based on the quantile residual spectrum and kurtosis skewness as follows:

[0020] On the basis of unified time series, the marketing behavior events in the extreme value suppression prior risk label section are extracted, and the behavior density atlas containing event type, trigger mode and data source identification is constructed;

[0021] Based on the behavior density atlas, the quantile residual, kurtosis and skewness statistical characteristics of the time series behavior data are extracted, and the extreme value diagnostic fingerprint containing event type, intensity value, behavior directionality and time position is constructed;

[0022] Based on the extreme value diagnostic fingerprint, the confidence interval boundary of behavior density is calculated through high and low behavior comparison window, and the threshold trajectory dynamically changing with time is connected;

[0023] Under the action of the threshold trajectory, the behavior density is compared and the burst extreme value node is identified, and the behavior node is classified and the dynamic label is outputted in combination with the event link diagram and the marketing activity strategy table.

[0024] Preferably, the real trend mapping generation steps are as follows:

[0025] Based on the threshold trajectory, the extreme value events coinciding with the promotion behavior are screened, and the promotion mode, channel source, user touch path and target population label are extracted;

[0026] Around the intervention type extreme segment, a tool variable set is constructed, and time series backtracking analysis is carried out to obtain the behavior trend difference before and after the promotion behavior;

[0027] The tool variable backtracking result and the extreme value diagnostic fingerprint are bidirectionally compared, the baseline backpropagation method is used to strip the promotion intervention net effect, and the non-intervention trend path is reconstructed;

[0028] Based on the corrected net effect behavior sequence, a trend mapping atlas is constructed, the response time difference, change consistency and intensity correlation between behavior variables are marked, and a real trend mapping for semantic optimization is formed.

[0029] Preferably, the steps of constructing semantic consistency constraint and screening selling point description are as follows:

[0030] Based on the real trend mapping, a semantic consistency constraint structure is constructed, the trend changes of behavior variables are one by one corresponding to language expression, and language boundary standards are established;

[0031] A cross-domain contrast coding scorer is constructed based on the semantic consistency constraint structure to extract language elements from the selling point description text, match and cross-score feedback, and output a five-dimensional score table containing semantic intensity values and trend support scores.

[0032] According to the score results, all marketing language is screened, and sentences that do not conform to the data trend are rewritten to ensure that all language outputs meet the three semantic constraint conditions of reasonable structure, trend support, and sentiment evidence.

[0033] Preferably, the semantic rewriting process is based on the offset results output by the cross-domain contrast coding scorer. Only when the semantic expression offset exceeds the preset threshold and the trend variable does not meet the minimum change amplitude of the modifier, the automatic replacement operation of the corresponding selling point description is triggered.

[0034] Preferably, the stable weight field generation step is as follows:

[0035] A semantic trend mapping alignment index table is constructed to establish a mapping between each expression in the marketing text and the corresponding behavior variable in the real trend graph.

[0036] Collect user behavior data associated with the mapped expression as a small sample empirical flow, and bind the timestamp and behavior path;

[0037] Inject the small sample empirical flow into the trend graph weight revision link, and perform weighted adjustment on the variable path according to the behavior feedback intensity;

[0038] According to the weight revision results, identify the dominant variable, and perform re-weighting operation on the dominant variable and its related variables;

[0039] Generate a stable weight field from the revised trend graph structure, and embed it into the subsequent text generation task as the main weight input basis.

[0040] Preferably, the self-triggered dynamic closed loop is formed by the frequency spectrum return traction, time difference micro-injection, and gain trajectory return strategy, and the selling point generation is continuously optimized in the dynamic closed loop to suppress the reverse exaggeration phenomenon. The steps are as follows:

[0041] After the stable weight field is constructed, the periodic response characteristics of each behavior variable in the time dimension are extracted to construct a phase behavior window.

[0042] Perform frequency spectrum return traction operation based on the phase behavior window, and design a staggered and staggered delivery rhythm model according to the user behavior frequency structure;

[0043] After the frequency spectrum return traction is completed, introduce the time difference micro-injection mechanism, and bind the phase point and delay offset value for the content according to the semantic intensity and user response delay characteristics;

[0044] The gain trajectory folding strategy is performed after the time difference micro-injection, the trend variable path priority is adjusted according to the feedback result of the delivery behavior, and the self-correction of the content generation direction is realized.

[0045] In the above technical solution, the present application provides technical effects and advantages:

[0046] The present application avoids the disturbance of asynchronous data on model judgment by constructing data timing consistency basis through unified time scale matrix and acquisition synchronization baseline; realizes accurate identification and correction of sudden data fluctuation through extreme value diagnosis fingerprint and threshold trajectory dynamic capture mechanism; effectively distinguishes incidental behavior and real trend by combining with the backtracking and correction of promotion intervention based on causal unbiased view; makes the generated content strictly comply with the actual trend by means of cross-domain comparison coding scorer to implement semantic consistency constraint, improves the authenticity and credibility of semantic expression; realizes small sample empirical feedback and dominant index reweighting through closed loop feedback device, dynamically optimizes semantic generation weight distribution, and guarantees the stability and continuous evolution of output structure; finally, a dynamic regulation and control system based on user behavior frequency and content response rhythm is formed by means of phase programmable delivery mechanism, realizing adaptive closed loop control from content generation to content delivery. In summary, the method effectively improves the authenticity, controllability and compliance of AI intelligent marketing content, and reduces the probability of occurrence of false propaganda and trust risk. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0048] Figure 1 The method flowchart of the present application AI intelligent marketing based on multi-source data selling point generation method. DETAILED DESCRIPTION

[0049] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the inventive aspects to those skilled in the art.

[0050] The present application provides an AI intelligent marketing based on multi-source data selling point generation method as shown in Figure 1 The present application provides an AI intelligent marketing based on multi-source data selling point generation method as shown in

[0051] A uniform time matrix is obtained in the multi-source data access stage, an extreme value suppression prior is constructed on the basis of the uniform time matrix, a collection synchronization baseline is established, and full-link time alignment is achieved through the collection synchronization baseline to ensure stable time sequence consistency for subsequent dynamic feature analysis;

[0052] To solve the problem of distorted selling point generation results caused by inconsistent time of multi-source data in the AI intelligent marketing scenario, a method for realizing full-link time alignment is proposed, which specifically includes the following steps:

[0053] Raw data is collected from multiple sources and time standardization operations are performed. The data sources include: (1) user purchase behavior data, which contains order time, payment time, delivery time, and evaluation time; (2) marketing activity execution data, including activity start time, end time, and phased push time points; (3) product dimension information, including listing time, price adjustment time, and inventory change time; (4) user comment content, including comment publishing time, like response time, and reply interaction time; (5) social platform content, including publishing time, forwarding time, comment time, and emotion change record time; (6) market feedback data, including third-party e-commerce platform sales fluctuation publishing time and competitor publishing event time. While collecting the above data, the corresponding original time stamp of each data is extracted and uniformly converted to coordinated universal time (UTC) format with a precision of milliseconds. Subsequently, the time distribution of each type of data is checked, and the formatted time markers are inserted into a unified main time sequence. This sequence has a minimum interval of milliseconds and runs through all data sources, forming a uniform time matrix. This matrix arranges the distribution of each type of data on the time axis in the horizontal direction and preserves the data content and its source label in the vertical direction, forming a basic structure that can be used for subsequent alignment processing.

[0054] On the basis of the unified time matrix, an extreme value suppression prior is constructed. Specifically, taking the unified time matrix as the coordinate axis, the behavior density of each source data in the same time period is analyzed, and the data surge interval is extracted. For example, if the user order volume surges, the advertisement click volume rises sharply, the commodity inventory drops synchronously, and the mention volume of the commodity on the social platform rises sharply in the period from 18:00 to 20:00 on a certain day, and there is no normal marketing activity arranged in this period, the system judges that this time period is caused by an abnormal event. In this case, combined with the statistical indicators such as the mean, median, range, and standard deviation of the historical behavior density in the same period, a residual spectrum is constructed to identify abnormal deviation values, and a risk classification is generated according to the deviation degree. For the risk area identified above, a “extreme value risk label” is marked in the unified time matrix, and a warning coefficient is set. For example, when the sales growth rate is more than 3 times the average value in the same period, the inventory consumption speed is more than twice, and the user comment keywords appear “emergency”, “sold out”, “rush to buy” and other high-intensity emotional words, the section will be marked as a first-level risk area. The extreme value suppression prior is used to guide the processing priority, participation degree and sampling strategy adjustment of the data in subsequent processing.

[0055] On the basis of the unified time matrix and the extreme value suppression prior label, a synchronous baseline is established. This process first performs point-to-point comparison of data points to check whether there is a delay backfill phenomenon in the timestamp. For example, a user completes an order at 10:00 am, but the platform actually records the time as 10:07, the comment publishing time is 11:00, but the social platform comment of the same user appears at 10:30, which can be initially judged as a data writing delay. At this time, sliding window analysis is performed on all source data, and the time sequence is adjusted by using the pair-by-pair rearrangement strategy. Then, the key behavior nodes are extracted according to the data event content, including the time nodes of each behavior in the “exposure→click→add→order→comment” link, and the relationship matrix of the behavior chain is constructed to map the sequence of the behavior nodes. The nodes that do not conform to the time sequence logic are compressed or extrapolated in time to ensure that each data chain is closed and coherent in the time dimension. On this basis, according to the synchronization accuracy of each data source, the network upload period, the event response time, and other factors, a minimum co-integration multiple time window is established as the time framework of the synchronous baseline, so that the data from different sources needs to be aligned and mapped before participating in subsequent processing.

[0056] Based on the acquisition of synchronous baseline and extreme value suppression priori label, the full link time alignment process is constructed. In this process, all subsequent incremental data must first be mapped and compared with the main time series to test its timestamp legality, behavior node rationality and risk label relevance. The specific operation includes: extracting the timestamp field of the new data, finding the adjacent data items in the unified timestamp matrix, calculating the time difference and behavior correlation degree; if it falls into the extreme risk label interval, adjust its sampling frequency to reduce its dominant influence on the overall eigenvalue; if it is outside the time window set by the acquisition synchronization baseline, it will be delayed or low-weighted. This process ensures that all data has been standardized in time and integrated in semantics before entering the AI intelligent marketing engine.

[0057] Under the constraint of full link time alignment, the extreme value diagnosis fingerprint is obtained, the threshold trajectory is generated based on the joint determination method of quantile residual spectrum and kurtosis skewness, the threshold trajectory is used as the dynamic reference for extreme value recognition, and the dynamic reference is used to realize the accurate capture and calibration of the sudden extreme value;

[0058] To solve the problem of distorted sell point recognition results caused by data sudden fluctuations in AI intelligent marketing, based on full link time alignment, by constructing extreme value diagnosis fingerprint and establishing threshold trajectory, accurate recognition and behavior calibration of abnormal extreme value are realized. This method specifically includes the following steps:

[0059] Taking the unified time series as a reference, all marketing behavior events in the extreme value suppression priori first risk label segment are extracted, and a high-precision behavior density atlas is constructed. The data includes user-generated browsing behavior, clicking behavior, adding to shopping cart, submitting order, completing payment, submitting evaluation, applying for after-sales, etc. in the marketing path; also includes platform behaviors such as advertisement pushing time, promotion activity node, price adjustment time, inventory change time; also includes external dissemination behaviors such as keyword mention time, forwarding time, liking time, comment publishing time in social platforms. All the above events are mapped to the unified time axis with millisecond as the time granularity, a sliding time window is constructed, the behavior density of each type of event is normalized, and a multi-source behavior density total graph is constructed in a superimposed manner. The behavior density atlas needs to retain the event type label, behavior triggering method and data source identification in each time slice to support the subsequent semantic reasoning process.

[0060] Based on the behavior density atlas, the statistical characteristics of the time series behavior data on the time axis are extracted, and the extreme value diagnostic fingerprint is constructed. The specific method is: for the data in each time slice, the quantile residual of the segment in the overall time sequence is calculated, that is, the deviation degree from the behavior mean and median in the same time period; then combined with the kurtosis of the behavior segment, whether the behavior has a sharp concentration trend is judged; at the same time, combined with the skewness, whether the behavior density deviates, that is, whether there is an excessive forward or backward of the behavior in time. Joint analysis of the three types of characteristic parameters forms a multi-dimensional extreme behavior feature structure, and marks the high-risk behavior nodes with significantly higher concentration than the benchmark. Each high-risk node has structured information such as event type, intensity value, behavior directionality, and time position, which is part of the extreme value diagnostic fingerprint.

[0061] On the basis of the extreme value diagnostic fingerprint, a dynamically adjustable threshold trajectory is constructed, which is used as a benchmark for subsequent extreme value judgment. The specific steps are: taking each extreme behavior fingerprint as the center, expanding a specified length of time period forward and backward, sampling normal behavior data of the same type to construct a high-low behavior comparison window. In each window, the cumulative probability distribution of behavior density is calculated, and the upper and lower limits between the 25th and 75th percentiles are extracted as the fluctuation interval to form the confidence interval. With the upper and lower boundaries of the confidence interval as the boundary, the behavior intensity in the continuous time period is marked point by point, and the complete threshold trajectory is obtained by connecting the points. The trajectory presents dynamic changes on the entire unified time axis, which can automatically reflect the boundary range of high-density behavior. Compare the actual behavior density with the trajectory. If the behavior density in the continuous time slice exceeds the upper limit of the threshold and matches the extreme value diagnostic characteristics, it is determined that the behavior node is a sudden extreme value.

[0062] Under the action of the threshold trajectory, the behavior capture and attribution of the sudden extreme value are completed. The specific method includes: extracting the identified extreme behavior node, conducting linkage analysis with other behavior nodes in time, constructing an event link graph, and identifying the causal order relationship between behaviors. Further combined with the marketing activity strategy table, it is judged whether the extreme value occurs before or after the promotion intervention or the input of sudden market information. If it is closely coupled with external thrust events, it is classified as a non-structural extreme value, and its weight suppression strategy in subsequent feature modeling is set; if the extreme behavior shows continuity and is derived from the user's natural behavior trajectory, for example, a large number of positive feedbacks due to product attribute optimization or natural word-of-mouth fermentation, it is defined as a structural extreme value, and its full behavior intensity information is retained. Finally, all the identified extreme nodes will be re-embedded in the unified time sequence and output as dynamic labels, providing reliable behavior basis and boundary control reference for subsequent causal de-biasing analysis and semantic output.

[0063] Under the constraint of threshold trajectory, the causal de-biased view is obtained, the tool variable method is used to perform time series backtracking on the promotion intervention, the time series backtracking result is combined with the extreme value diagnostic fingerprint for correction, and the true trend mapping is generated based on the correction to provide reliable basis for subsequent semantic level selling point optimization;

[0064] To avoid the misleading influence of promotion behavior on data trends, on the basis of dynamic threshold trajectory identification, a causal de-biased view needs to be constructed to realize the root identification of extreme value events and the true trend modeling, which includes the following steps:

[0065] Under the limitation of dynamic threshold trajectory, all time periods that have been marked as extreme value events are identified, and the part highly coinciding with the promotion behavior is screened and structured. This operation is based on the timestamp corresponding mechanism, and the time range, behavior variable type, peak intensity and residual characteristics of each extreme value record are checked one by one to find out whether there is a coincidence relationship with the preset promotion activity time window. For example, within 14:00-18:00, if the order quantity, browsing volume and comment volume of a product all break through their respective dynamic threshold upper limit, and this time period coincides with the activity period of the product participating in “limited time discount 30%”, it can be determined that the extreme value event is affected by the promotion behavior. After screening, the promotion background information of the intervention extreme value behavior is extracted, including the promotion method (such as price discount, gift stacking), channel source (such as in-site notification, social media push), user touch path (such as active search, passive exposure) and target user group label (such as new user, high-frequency repeat purchase user). This process provides a quantifiable intervention variable basis for the next step of establishing a causal analysis structure.

[0066] Around the screened intervention type extreme value segment, a tool variable set is constructed and time series backtracking analysis is carried out to identify the behavior structure difference before and after the promotion activity. First, the time period before the execution of the promotion behavior is selected, and auxiliary variables with relevant behavior characteristics to the current extreme value variable but not directly affected by the promotion are extracted, such as user daily activity level, natural browsing trend of the product, stability of advertisement exposure, historical daily sales median, average of previous conversion rate, etc. Then, the behavior trend sequence before and after the intervention is constructed, taking the promotion trigger time point as the dividing line, the stable behavior before the intervention as the control group, and the behavior after the intervention as the experimental group. Time stamp is used as the basis for window division, and the change slope, amplitude and directionality shift of the behavior variable are calculated in each window, and the difference characteristics are compared with the control group tool variables. For example, when the sales volume of a product doubles during the promotion period compared with before the promotion, while the browsing volume growth does not exceed one time of the standard deviation of historical fluctuations, it shows that the increase of sales volume is mainly driven by external factors rather than internal trend. Through backtracking comparison, without introducing prediction model, the specific shift strength of marketing behavior on behavior variable is identified.

[0067] Based on the tool variable backtracking results and the extreme value diagnostic fingerprint content, the causal net effect sequence after the offset stripping is constructed. First, the event information matching the current variable in the extreme value diagnostic fingerprint is called, including the start and end time of the event, the fluctuation intensity, kurtosis, skewness, and the resonance degree of the collaborative variable. Then the behavior offset trend estimated by the tool variable is compared with the actual fluctuation amplitude in the extreme value fingerprint value by value, to judge whether there is a fluctuation overlap phenomenon caused by non-promotion related factors. If the natural trend reflected by the tool variable is significantly lower than the fluctuation level of the extreme value fingerprint, it means that there is a significant intervention effect of the behavior variable. The offset is quantitatively corrected, and the baseline backtracking method is used to reconstruct the trend path without intervention: that is, the predicted trajectory without the influence of promotion intervention is extended from the natural trend line before promotion, and the difference between the actual observation value during the promotion period and the predicted trajectory is removed in the form of difference. For example, if the actual sales in the next three hours exceed the sum of the predicted values by 30%, the difference can be used as the net effect of the promotion intervention, and the difference is removed in the trend data. Finally, the purified data sequence after removing the interference of the promotion behavior is constructed as the input of the subsequent trend mapping.

[0068] Based on the corrected net effect behavior sequence, a real trend mapping of multi-variable synchronous evolution is constructed. In this process, all behavior variables are first processed by a unified time window slicing to ensure that the values of each variable at the same time point are comparable. Then, trend smoothing operation is performed on each type of variable to remove small high-frequency fluctuations and retain the trend trajectory on the medium scale. Then the causal order mapping relationship between different variables is established, such as advertisement display affecting browsing volume, browsing volume promoting add-to-cart behavior, add-to-cart behavior affecting order conversion, order conversion driving comment quantity and emotion change, etc. Based on this variable influence path, a trend chain is constructed to mark the response time difference, change direction consistency and strength correlation between different variables in each time period, forming a clear structure trend map. In this map, each node represents the real state of a specific variable at a specific time, and each edge represents the confirmed behavior influence path between variables based on causal debiasing. For example, when a stable corresponding relationship between the peak value of an advertisement click volume and the subsequent order number increase within 5 minutes is confirmed after promotion stripping, this path will be retained as a high-confidence link in the trend mapping graph. The final trend mapping result not only removes the offset misleading introduced by marketing means, but also retains the collaborative change pattern of multi-dimensional behavior variables in the natural evolution state, providing a stable and high-confidence data foundation for accurately identifying the "essence selling point" in subsequent semantic expression.

[0069] Under the guidance of real trend mapping, semantic consistency constraints are obtained, a cross-domain contrastive coding scorer is constructed, and the selling point description is screened item by item using the cross-domain contrastive coding scorer combined with the causal unbiased view, so as to ensure that the generated semantic output is consistent with the real trend, thereby eliminating the reverse exaggeration phenomenon caused by extreme value amplification.

[0070] To avoid semantic deviation and exaggerated description in the natural language generation process misleading users, on the basis of the constructed real trend mapping, a selling point semantic screening and strengthening mechanism oriented to data consistency constraints is proposed, which includes the following steps:

[0071] On the basis of real trend mapping, a semantic consistency constraint structure is constructed to strictly correspond the trend changes of behavior variables to language expression. In real trend mapping, the change direction, amplitude, change frequency, change period and mutual response relationship between variables of each behavior variable under the unified time line have been recorded. Specifically, for the payment amount variable, the percentage of growth or decline in the continuous time period is extracted; for the order quantity variable, the average rate of change and the abnormal deviation interval within a day are calculated; for the comment quantity variable, the sentiment tendency category is marked and the positive comment growth rate in the continuous time period is extracted; for the page browsing variable, the high-frequency access time point, page bounce rate and stay time are counted; for the product add-to-cart quantity, the synchronization ratio between it and the conversion behavior is calculated. In the process of constructing semantic constraints, the above data trends need to be semantically mapped and classified. For the target description semantic "rapid growth", it needs to correspond to the browsing volume or order quantity that increases by more than 30% within 20 minutes and the trend slope continuously rises for at least three time windows; for the target semantic "user feedback is enthusiastic", it needs to correspond to the total number of comments rising more than the average level of the previous week, the proportion of positive sentiment words exceeding 70%, and involving more than three dimensional keywords (such as "satisfactory", "cost-effective", "recommended"). In this way, a data-driven language boundary standard is established, and all subsequent generated selling point language structures must run within this boundary. If "explosion", "record-breaking" or "continuous rise" type of words are used, there must be at least two variables rising simultaneously, and the increase amplitude lasts for more than a certain interval.

[0072] Based on the semantic constraint structure, a cross-domain contrast coding scorer is constructed to score the semantic strength and trend actuality of the generated selling point language for each sentence. The scorer is divided into three parts: language element extraction, trend mapping matching, and cross-scoring feedback. In the language element extraction stage, the marketing selling point text is disassembled into five types of language units, namely main statement subject (such as product name), action description (such as "growth", "good comment"), degree modification (such as "rapid", "large"), time modification (such as "short time", "continuous"), and emotional modifier (such as "enthusiastic", "unprecedented"). In the trend mapping matching stage, each text unit is mapped to the corresponding relationship in the trend variable set. For example, "order growth" is mapped to the payment amount and order quantity variables, and "user feedback" is mapped to the comment quantity and comment emotion variables, respectively. The trend direction and slope offset in the mapping time period are read. By setting threshold rules, for example, "growth" must correspond to a continuous growth rate of more than 5%, "continuous" must reflect that the index has not fallen back for three consecutive time periods, and "explosive" must be accompanied by the historical maximum amplitude or behavior peak position. In the cross-scoring feedback stage, the scorer outputs a five-dimensional score table, including the strength value of semantic expression, the support value of corresponding data trend, the expression and fact offset, the trend synchronization score, and the emotional intensity rationality score. If the offset exceeds the allowed threshold (such as using "explosion" but the trend amplitude is only 8%), the sentence is marked as "potential exaggeration risk" and a downgrade suggestion is given, such as replacing "explosive growth" with "rapid rise".

[0073] After completing the cross-domain scoring, all candidate selling point languages are screened item by item, and semantic rewriting operations are performed when sentences deviating from the actual trend are found. In the screening process, each piece of marketing language is taken as a unit, and the trend source of each keyword in the sentence is confirmed by reverse mapping in combination with the causal bias-free view and the real trend atlas. If an emotional expression is found to be unsupported by behavioral data, the risk level of the expression is recorded, and the expression is replaced according to the rewriting suggestion provided by the scorer. For example, "sales hit a new high" must be replaced with "sales are significantly higher than the recent average" if it does not coincide with the historical maximum sales peak; "users widely recognized" must be replaced with "users have positive feedback" if the number of comments or covered users is insufficient. This rewriting mechanism uses trend variables as constraints and scoring mechanisms as feedback sources to achieve a closed-loop check between language expression and data facts. Each language output must meet three constraint conditions: (1) structural reasonableness, i.e., the semantic clarity of subject-predicate-object is not jumping; (2) trend support, i.e., all modifiers are traceable in data; (3) emotional evidence, i.e., positive and negative polarity language must have comment sentiment or behavioral response data support.

[0074] Under the action of semantic consistency constraints, a closed-loop feedback device is obtained, small sample empirical flow is injected into the weight update link of real trend mapping, and dominant indicator reweighting is performed under the guidance of semantic consistency constraints. A stable weight field is generated through reweighting to ensure the balance and persistence of global output;

[0075] To ensure that the marketing content generation achieves feedback closed loop between trend dominance and semantic driving, and further enhances the stability and long-term performance consistency of the generated results, a closed-loop feedback and weight stabilization mechanism is proposed, which includes the following steps:

[0076] Under the action of semantic consistency constraint mechanism, marketing selling point texts that have passed screening and scoring are selected, and corresponding behavior variable paths are extracted in combination with real trend mapping structure to construct a semantic trend mapping alignment index table. The index table establishes a one-to-one correspondence between the core expression items in the text and the trend variables. For example, "the order volume continues to rise" corresponds to the behavior variable "order quantity per unit time", and "user review heat continues to rise" corresponds to the variables "review number growth rate" and "positive sentiment rate". Through the variable path, starting time, maximum slope position, growth duration, feedback response delay time and other information marked in the trend graph, the semantic expression is structured and positioned on the trend data node. This process not only confirms whether the current text maintains real trend consistency, but also serves as a trigger entry for the subsequent feedback mechanism, providing a basic mapping structure for small sample feedback access and weight structure revision.

[0077] After the semantic trend index structure is established, the small sample empirical flow collection phase is entered. The empirical flow refers to the reaction behavior data associated with the generated text produced by real users in a short period of time. The collected indicators include: page dwell time after the marketing text is browsed, whether the user's add-to-cart behavior occurs after reading, the difference between click-through rate and bounce rate, whether it is actively forwarded, the distribution density of like number among different user groups, and the difference in the ratio of positive emotion words to neutral or negative words in the comment content. Each item of behavior data must be bound to a specific timestamp and user behavior path, and classified according to the source channel, such as e-commerce home page browsing path, search path, and content recommendation path. Subsequently, these behavior indicators are mapped back to the corresponding variable structure in the semantic trend index table, for example, the feedback path of the "payment conversion rate improvement" statement should collect whether the order conversion is successful, the conversion time comparison, and the sustained behavior performance after the payment success page. This small sample flow is not used for overall model training, but as an important signal source for feedback weight adjustment, reflecting whether the current language expression accurately triggers the user's expected behavior.

[0078] After obtaining the structured small sample feedback data, it is injected into the trend mapping weight correction link to build a "feedback-variable" linkage structure. Each type of feedback behavior is weighted and quantified according to its behavior intensity, change rate and consistency. For example, when a certain selling point description obtains a significantly higher click rate, comment heat and repeat purchase behavior number within 24 hours than similar texts, and its trend mapping path is "order growth + commodity evaluation heat improvement", it can be judged that the trend variable combination has real conversion value. At this time, the basic weight of the "order growth" path in the original trend map is increased from 0.78 to 0.89, and if the "commodity evaluation heat improvement" path performs generally in the feedback data, its initial weight is retained. If there is a deviation between user actual behavior feedback and selling point semantic expression, such as an increase in click-out rate, a negative word set in the comment, the weight of the related variable path in the trend map will be adjusted downward. For example, if the original setting is 0.81, it can be adjusted to 0.66 according to the deviation amplitude. The entire weight revision process is established in the trend mapping graph structure, and the sliding time window mechanism is used to avoid short-term extreme feedback from causing fluctuations in long-term trend weights, achieving dynamic adjustment and control of stable weight updates.

[0079] After completing the trend map weight revision, relying on the semantic consistency constraint mechanism, the updated behavior variable set is executed for dominant variable identification and re-weighting processing. The identification criteria include: first, the variable is called frequently in multiple selling point texts; second, the variable shows strong response aggregation in small sample feedback behavior data; third, the variable and other behavior variables show obvious synchronous change relationship on the trend time axis. Through these judgment criteria, the variables with "content response center" characteristics in the current marketing period are identified as dominant variables. For example, during a certain promotion, "add-to-cart behavior proportion" and "order conversion cycle shortening" show synchronous increase for three consecutive days, and are called by multiple high-score marketing texts, so they are considered as dominant behavior variables. On this basis, re-weighting processing is performed to further increase the weight of the dominant variable to 1.0, forming a weight anchor point in the trend map. At the same time, the auxiliary behavior variables related to it are maintained in the range of 0.85-0.9, constructing a trend guiding main line structure, so that it is easier to focus on key behavior changes in subsequent marketing text generation, and reduce the content style divergence or conversion rate decline caused by non-key variables.

[0080] After the dominant variables are reweighted, a stable weight field is generated based on the full-quantity trend map structure. This stable weight field is a continuous variable-weight space structure that has been corrected through multiple rounds of feedback, variable selection, and behavior response induction. Its characteristics are: the weight of the same variable in multiple time windows changes within a threshold, for example, no more than ±5%; the weight of the dominant variable remains in the peak interval and no longer fluctuates dramatically with non-structural feedback; the marginal variable is dynamically exited or marginalized according to the actual feedback result. The establishment of the stable weight field marks the completion of the dynamic adaptation of the trend map among semantic content, real feedback, and behavior path. In the subsequent semantic generation, copy screening, and trend matching processes, it can be used as a main weight reference input. The structure of the weight field will be embedded in the data input process of future content generation tasks, driving the generation engine to preferentially select the selling point expression structure associated with high-weight variables, thereby improving content conversion rate, user experience, and expression stability. Through continuous iteration, this weight field can also absorb new rounds of feedback information, forming a self-updating, self-strengthening, and self-adapting ability. It is the central mechanism for the integration of intelligent generation of marketing semantics and real user behavior feedback.

[0081] Under the drive of the stable weight field, a phase-programmable delivery control mechanism is obtained. By introducing frequency spectrum return traction, time difference micro-injection, and gain trajectory return strategy, the delivery control process forms a self-triggering dynamic closed loop, and continuously optimizes the selling point generation in the dynamic closed loop, thereby suppressing the reverse exaggeration phenomenon in the whole link operation process.

[0082] To realize the rhythm optimization and precise control of the marketing content generation process under the guidance of real trend data, based on the stable weight field constructed, a phase-programmable delivery control mechanism is further proposed. This mechanism completes the delivery closed loop through time sequence deduction, dynamic return, and feedback linkage, which includes the following steps.

[0083] After the stable weight field is completed, the periodic response characteristics of each behavior variable in the time dimension are obtained, and a phase behavior window is established based on the characteristics. In the stable weight field, behavior variables such as “commodity browsing volume growth rate”, “add-to-cart conversion rate”, “user positive evaluation frequency” and the like have been corrected by weighting for multiple feedback cycles, forming a stable weight and clear trend expression structure. The behavior intensity change of each variable in the past multiple time cycles will be extracted as the original time series data, and then a periodic pattern recognition operation is performed to identify the peak points and valley points in 24 hours, 7 days and the like. Taking “commodity browsing volume growth rate” as an example, if analysis finds that the variable has obvious concentrated rising behavior trend from 12:30 to 13:30 in the afternoon and from 20:00 to 21:00 in the evening, it is determined that the variable has a response active phase in the two time periods. These high response periods are defined as “phase behavior windows”, and the phase windows of all variables are uniformly constructed into a time—behavior variable two-axis structure table, which will serve as a reference atlas for the starting rhythm of content placement.

[0084] On the basis of the phase behavior window, a frequency spectrum return traction operation is performed to adjust the actual placement rhythm of the marketing content. The operation takes the identified high-frequency phase window as the center, and extends the content pushing time before and after it, so that the content can also obtain effective exposure in the upper and lower bands near the high-activity time period, thereby dispersing the placement peak concentration and reducing the user information load. In this process, the user behavior frequency sequence is converted into a frequency domain expression, the main behavior concentration period and its harmonic structure are identified, and a traction trajectory is constructed according to the user response curve. For example, when the “add-to-cart conversion rate” concentration of the user in the period from 20:00 to 21:00 is significantly improved, but there are weak frequency activation fluctuations at 19:45 and 21:15, part of the content can be displayed in advance at 19:45 and part of the content can be displayed in advance at 21:15 through return traction, forming a time sequence strategy of staggered placement. At the same time, the amplitude and step length of the traction are adaptively determined according to the frequency spectrum energy distribution, and the design of the traction rhythm needs to keep the content placement density balanced, the user perception rhythm reasonable, and the up-and-down carousel logic coherent, so as to finally build a placement rhythm model that fits the user rhythm.

[0085] After completing the spectrum return traction, combined with semantic intensity and user response delay characteristics, time difference fine-tuning mechanism is injected to realize fine-grained content accurate delivery. According to the complexity of content semantic structure, emotional guidance intensity and data driven weight, all selling point contents are semantically graded, and the adaptability of time difference is matched for different level contents. For example, for the description content emphasizing brand trust, service response and other emotional stickiness, according to the historical user response delay characteristics, it is set to early exposure, so as to carry out cognitive pre-embedding before the user forms a clear purchase decision; while for the content emphasizing price advantage, inventory shortage and other contents with instant conversion orientation, it is set to peak point synchronization or delayed exposure, so as to trigger user behavior response at the decision critical point. In specific implementation, each content sentence is bound with a delivery phase point and delay offset value, for example, the content of "the number of orders continues to rise" is delayed for 12 minutes, so as to match the appearance of price fluctuation path peak, realize the accurate overlap of content and trend peak, and maximize the conversion possibility. The time difference fine injection strategy controls the triggering time of semantic content while establishing the time sequence coupling relationship between content semantic intensity and trend variable, which significantly improves the content effectiveness and behavior response matching degree.

[0086] After completing the rhythm traction and time difference fine injection, the gain trajectory return strategy is executed to realize the content generation direction regulation closed loop based on the delivery result. The strategy takes the user behavior performance after delivery as the evaluation basis, extracts behavior data such as content click rate, comment density, conversion rate and stay time, and classifies and statistics according to the content-variable path mapping structure. For the content with excellent behavior performance after delivery, trace its path source in the stable weight field, extract its dominant trend variable, and correspondingly improve the path weight priority in the next round of content generation; for the content with weak performance, locate its trend variable path, and judge whether it is caused by time mismatch, insufficient semantic expression or trend weight deviation, if there is deviation, execute the return operation, that is, in the next round of content generation, turn to the coordinated variable path of the variable to reconstruct the expression structure. For example, if the original content is generated based on "evaluation preference rate improvement", but there are more negative comments, the next round of generation will turn to new content generation based on "customer service response time" or "logistics delivery accuracy" and other coordinated variables. The gain return not only regulates the generation direction of semantic path, but also realizes the self-correction mechanism based on user behavior data in content strategy. Finally, under the linkage of frequency spectrum rhythm control, semantic level injection and behavior feedback, a dynamic programmable, rhythm self-evolution, expression controllable delivery mechanism is built, which ensures that the marketing content realizes continuous optimization and behavior closed loop under the guidance of complex trend variable.

[0087] The application avoids the disturbance of asynchronous data on model judgment by constructing data time sequence consistency basis through unified time scale matrix and collection synchronization baseline; realizes accurate identification and correction of sudden data fluctuation through extreme value diagnosis fingerprint and threshold trajectory dynamic capture mechanism; effectively distinguishes accidental behavior and real trend by combining with promotion intervention backtracking and correction of causal debiasing view; makes generated content strictly comply with trend actuality by implementing semantic consistency constraint with the aid of cross-domain comparison coding score; guarantees stability and continuous evolution of output structure by realizing small sample empirical feedback and dominant index reweighting through closed loop feedback device; finally, forms dynamic regulation and control system based on user behavior frequency and content response rhythm with the aid of phase programmable delivery mechanism, realizes adaptive closed loop control from content generation to content delivery. In conclusion, the method effectively improves the authenticity, controllability and compliance of AI intelligent marketing content, and reduces the probability of occurrence of false propaganda and trust risk.

[0088] The above only describes certain exemplary embodiments of the application by way of illustration, and it is needless to say that the described embodiments can be modified in various ways without departing from the spirit and scope of the application for those skilled in the art. Therefore, the above figures and description are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the application.

Claims

1. An AI-powered intelligent marketing method for generating selling points based on multi-source data, characterized in that: Includes the following steps: In the multi-source data access phase, a unified time-stamped matrix is ​​obtained. Based on the unified time-stamped matrix, an extreme value suppression prior is constructed and a data collection synchronization baseline is established. The data collection synchronization baseline is used to achieve end-to-end time alignment. Specifically, this includes: collecting raw data from multiple data sources, extracting the time field set of multi-source heterogeneous marketing-related data, and uniformly converting it into Coordinated Universal Time (UTC) format to form a unified time series with milliseconds as the minimum interval. Based on the unified time series, the behavioral density changes of data from each source within the same time period are extracted, an extreme value suppression prior is constructed, and sudden increase intervals are identified and risk labels are generated through residual spectrum analysis. Under the constraint of end-to-end time alignment, extreme value diagnostic fingerprints are obtained. Threshold trajectories are generated based on quantile residual spectra and kurtosis skewness. The threshold trajectory is used as a dynamic benchmark to capture and calibrate sudden extreme values. Specifically, this includes: extracting marketing behavior events within the extreme value suppression prior first-level risk label segment based on a unified time series, constructing a behavior density map, extracting the quantile residuals, kurtosis and skewness statistical features of time series behavior data based on the behavior density map, and constructing extreme value diagnostic fingerprints. Under threshold trajectory constraints, a causal debiased view is obtained. The instrumental variable method is used to perform time-series backtracking on promotional interventions. The backtracking results are jointly corrected with extreme value diagnostic fingerprints to generate a true trend mapping. Under the guidance of real trend mapping, semantic consistency constraints are obtained, a cross-domain comparison encoding scorer is constructed, and the selling point description is screened by combining causal debiasing view to ensure that the generated semantic output is consistent with the real trend. Under semantic consistency constraints, a closed-loop feedback loop is obtained, a small sample empirical stream is injected into the weight update stage of the real trend mapping, and the dominant indicator is reweighted to generate a stable weight field. Driven by a stable weight field, a phase-programmable delivery control mechanism is acquired. This mechanism forms a self-triggered dynamic closed loop through spectrum backtracking, time-difference micro-injection, and gain trajectory backtracking strategies. Within this dynamic closed loop, selling point generation is continuously optimized, and reverse exaggeration is suppressed. Specifically, after the stable weight field is constructed, the periodic response characteristics of each behavioral variable within the time dimension are extracted to construct a phase behavior window. Based on this window, a spectrum backtracking operation is executed to adjust the actual delivery rhythm of marketing content. The spectrum backtracking operation extends the content push time before and after the identified high-frequency phase window. After spectrum backtracking, a time-difference micro-injection mechanism is introduced. This mechanism binds phase points and delay offset values ​​to the content based on semantic strength and user response latency characteristics. After time-difference micro-injection, a gain trajectory backtracking strategy is executed. This strategy adjusts the path priority of trend variables based on the delivery behavior feedback results, achieving self-correction of the content generation direction.

2. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 1, characterized in that, Achieving end-to-end time alignment also includes: Based on a unified time series and risk labels, a synchronous baseline for data collection is established, and time sequence rearrangement and behavior node closure are achieved through sliding window analysis and behavior chain relationship matrix. Based on the data collection and synchronization baseline, a full-link time alignment process is constructed. All incremental data are compared for timestamp validity, behavioral node rationality, and risk label correlation to achieve data standardization and semantic integration.

3. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 2, characterized in that, Obtaining extreme value diagnostic fingerprints under the constraint of end-to-end time alignment, and generating threshold trajectories based on quantile residual spectra and kurtosis skewness, also includes: Based on extreme value diagnostic fingerprints, the confidence interval boundary of behavior density is calculated through a high-low behavior comparison window, and then connected to form a threshold trajectory that changes dynamically over time. Under the influence of threshold trajectories, behavioral density is compared and sudden extreme value nodes are identified. The behavioral nodes are attributed and classified by combining the event link diagram and the marketing campaign strategy table, and dynamic labels are output.

4. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 3, characterized in that, The steps for generating a true trend map are as follows: Based on the extreme events that overlap with the threshold trajectory screening and promotional behavior, we extract the promotional methods, channel sources, user reach paths and target audience tags; A set of instrumental variables was constructed around the extreme value segment of the intervention, and a time-series backtesting analysis was conducted to obtain the differences in behavioral trends before and after the promotional behavior; The instrumental variable backtracking results are compared bidirectionally with the extreme value diagnostic fingerprint. The baseline backtracking method is used to remove the net effect of the promotion intervention and reconstruct the uninterventional trend path. That is, starting from the natural trend line before the promotion, the predicted trajectory without the influence of the promotion intervention is extended, and the difference between the actual observed value and the predicted trajectory during the promotion period is removed by the difference method. A trend mapping map is constructed based on the corrected net effect behavior sequence, marking the response time difference, consistency of change and intensity of correlation between behavioral variables, forming a true trend mapping for semantic optimization.

5. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 4, characterized in that, The steps for constructing semantic consistency constraints and screening selling point descriptions are as follows: A semantic consistency constraint structure is constructed based on real trend mapping, which maps the trend changes of behavioral variables to language expressions one by one, and establishes language boundary standards. Based on the semantic consistency constraint structure, a cross-domain comparison coding scorer is constructed to extract language elements, match trends, and provide cross-scoring feedback on the selling point description text, and output a five-dimensional score table. Based on the scoring results, all marketing language was screened, and sentences whose expressions did not conform to the data trends were semantically rewritten to ensure that all language outputs met three semantic constraints: reasonable structure (i.e., clear and non-jumping subject-verb-object semantics), trend support (i.e., all modifiers are traceable in the data), and verifiable sentiment (i.e., positive and negative polarity language must be supported by commentary sentiment or behavioral response data).

6. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 5, characterized in that, The semantic rewriting process is based on the offset results output by the cross-domain comparison encoding scorer. The automatic replacement operation of the corresponding selling point description is triggered only when the semantic expression offset exceeds the preset threshold and the trend variable does not meet the minimum change range of the modifier.

7. The AI-powered intelligent marketing method for generating selling points based on multi-source data according to claim 5, characterized in that, The steps for generating a stable weight field are as follows: Construct a semantic trend mapping alignment index table to map each expression in the marketing text to the corresponding behavioral variable in the real trend map; Collect user behavior data associated with the mapping expression as a small sample empirical stream, and bind timestamps and behavior paths; Inject small-sample empirical data into the trend graph weight revision stage, and perform weighted adjustments on variable paths based on the intensity of behavioral feedback. Based on the weight revision results, the dominant variable is identified, and a reweighting operation is performed on the dominant variable and its related variables; The revised trend graph structure generates a stable weight field, which is then embedded into subsequent text generation tasks as the basis for sovereign re-input.

Images