An AI intelligent marketing promotion strategy generation method based on word-of-mouth trends

By constructing a net trend sequence and fitting a logistic growth function, a precise marketing and promotion strategy is generated, solving the problems of misjudgment of category fluctuations and waste of resources in existing technologies, and realizing precise marketing of target products.

CN123071058APending Publication Date: 2026-11-10MICRO-ZHIJU INTELLIGENT TECHNOLOGY (NANTONG) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611408739.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-11
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing marketing strategy generation methods cannot effectively isolate the overall fluctuations of leaf categories, cannot accurately determine the evaluation objects of different sample sizes, and lack quantitative means for the life cycle of word-of-mouth elements, resulting in deviations in the direction of ad placement and waste of resources.

Method used

By collecting comments from multiple sources and extracting attribute opinion pairs, a net trend sequence is constructed. Multi-scale slope estimation and logistic growth function fitting are used to generate campaign execution parameters, including audience packages, keywords, and bidding strategies.

Benefits of technology

It enables accurate determination of the excess change of target products relative to the category as a whole, avoids misdirection of resources, and improves the accuracy of marketing strategies and the efficiency of resource utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071058A_ABST
    Figure CN123071058A_ABST
Patent Text Reader

Abstract

This invention discloses an AI-powered intelligent marketing strategy generation method based on word-of-mouth trends, belonging to the field of e-commerce marketing data processing technology. The method first obtains a target product review set and a category-matching product review set through multi-source review collection. Then, it extracts attribute-viewpoint binary pairs and labels them with sentiment polarity to obtain a set of attribute-viewpoint binary pairs with polarity tags and timestamps. Next, it obtains a net trend sequence after removing overall fluctuations in leaf categories through time window segmentation, window mention intensity calculation, comparison sequence alignment, and net trend sequence construction. Finally, it obtains a candidate list through multi-scale slope estimation, self-fluctuation amplitude normalization, and inflection point candidate labeling. Finally, it obtains a list of rising-period elements through rising curve fitting, element confidence grading, and elimination of decaying elements. This invention models the propagation process of word-of-mouth elements as a growth process with a capacity limit and outputs the remaining rising window duration, making it suitable for generating marketing strategies for e-commerce products and consumer brands.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of e-commerce marketing data processing technology, specifically to an AI-powered intelligent marketing promotion strategy generation method based on word-of-mouth trends. Background Technology

[0002] E-commerce product and consumer brand marketing relies on user reviews in the comment sections. These sections aggregate spontaneous expressions from buyers regarding product attributes, user experience, and shortcomings. The combination of attribute words and opinion words in the comment text constitutes a user's evaluation of a specific aspect of the product. User ratings and comment posting times record the intensity and timing of these evaluations. Advertising platforms offer category-based audience tag libraries, keyword bidding interfaces, and time-based bidding interfaces, allowing operators to set keywords, audience segments, bids, and ad placement periods.

[0003] Current marketing promotion strategies are primarily based on comment sentiment statistics: First, comments on the target product are segmented and categorized by sentiment; then, positive and negative mentions across various attributes are tallied, and a popularity ranking of selling points is generated by sorting them in descending order of mention volume. Operations staff then select several selling points from this ranking, manually write keywords for ad placement, and define audience segments and ad placement periods based on experience. The ad placement period and bid levels are determined by operations staff based on current popularity data. Some methods further incorporate moving averages or exponential smoothing, smoothing the mention volume sequence and using the month-on-month change in the smoothed value as a trend indicator.

[0004] This method of using the absolute number of mentions in the target product's own reviews as the basis for popularity will produce distorted results when the overall leaf category fluctuates. Seasonal changes in leaf categories, changes in platform traffic allocation strategies, and the impact of industry-wide public events all affect the target product and all similar products within that leaf category simultaneously. The changes in mentions caused by these three factors are indistinguishable from the changes in mentions caused by changes in the target product's own selling points in a single sequence. For example, in the sun-protective clothing category, the number of mentions related to breathability generally increases during the summer. This increase is a seasonal change in the category, and the popularity ranking will interpret it as an increase in the target product's selling points, thus concentrating resources on that selling point. However, the target product has not gained any advantage over similar products in that selling point.

[0005] Regarding the generation of marketing and promotion strategies driven by word-of-mouth data, existing methods have the following shortcomings: First, popularity statistics are based on the absolute number of mentions in the target product's own reviews, failing to isolate the overall fluctuations of the subcategories. Seasonal changes, platform traffic allocation strategy changes, and industry-wide public events are often misinterpreted as changes in the product's selling points, leading to a shift in the targeting direction. Second, the judgment is based on the current popularity value or the absolute value of a single-scale slope. The sample size of reviews across different attribute dimensions can span two orders of magnitude, making a uniform threshold unsuitable. When using a uniform threshold across all categories, normal jagged fluctuations in low-frequency attribute dimensions are judged as an upward trend, while using standard deviation normalization for all samples flattens the true trend of high-frequency attribute dimensions due to fluctuations in low-frequency attribute dimensions. Third... The lack of quantitative methods for measuring the lifecycle of word-of-mouth elements means that the output of moving averages and exponential smoothing are smoothed values ​​of historical data, which can only describe the trend up to the present moment and cannot indicate how many more time windows the trend can continue in the future. The setting of the campaign cycle relies on the experience and judgment of the operators, which can lead to two types of deviations: continuing to bid high after the selling point has reached saturation and prematurely reducing the budget when the selling point still has a long period of growth. Fourth, the keywords for the campaign are manually selected by the operators from the comment text. Manual selection tends to favor mature selling points that have already achieved scale and is not sensitive to word-of-mouth elements in the early stages of growth. In addition, both positive and negative comments are converted into keywords for the campaign. The traffic bought after negative comments enter the bidding does not match the actual capabilities of the product, and the decline in conversion data interferes with the next round of popularity statistics.

[0006] Therefore, a method is needed that can isolate the overall fluctuations of leaf categories, place evaluation objects of different sample sizes on the same judgment scale, quantify the remaining room for improvement of word-of-mouth elements, and generate execution parameters based on this, so that the basis for generating marketing and promotion strategies can be changed from the current popularity level to how long that popularity can be maintained. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an AI-powered intelligent marketing and promotion strategy generation method based on word-of-mouth trends.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an AI-powered intelligent marketing promotion strategy generation method based on word-of-mouth trends, the method comprising: S1: Obtain the target product review set and the category-matching product review set through multi-source review collection. Then, extract the attribute opinion pairs to parse each review text into attribute opinion pairs. Finally, attach a polarity label to each attribute opinion pair through sentiment polarity labeling to obtain a set of attribute opinion pairs with polarity labels and timestamps. S2: Taking the set of attribute viewpoint pairs obtained in step S1 as input, the set of attribute viewpoint pairs is first divided into observation window sequences by time window partitioning. Then, the target product mention intensity sequence and the category comparison mention intensity sequence are calculated by window mention intensity. The sparse window and unique elements are aligned by the comparison sequence. Finally, the net trend sequence is constructed by subtracting the logarithmic change of the category comparison mention intensity sequence from the logarithmic change of the target product mention intensity sequence. S3: Using the net trend sequence obtained in step S2 as input, first perform least squares line fitting on the short window, medium window and long window scales through multi-scale slope estimation to obtain the slope of the short window, the slope of the medium window and the slope of the long window. Then, through the normalization of its own fluctuation amplitude, use the standard deviation of the historical slope sample of the attribute viewpoint binary as the normalization denominator to obtain the rise degree of the short window, the rise degree of the medium window and the rise degree of the long window. Finally, through the inflection point candidate label, perform the same sign test, the magnitude increase test and the sparse window test on the three rise degrees in turn. The attribute viewpoint binary that passes all three tests is written into the candidate list, and the attribute viewpoint binary that fails any one test is transferred to the regular word-of-mouth element pool. S4: Using the candidate list obtained in step S3 as input, first, by fitting the rising curve, the cumulative net trend sequence is fitted with the logistic growth function to obtain the saturation scale, intrinsic growth rate and the midpoint of the curve. Then, the propagation stage is determined by the propagation stage, and the remaining rising window duration is calculated by the remaining rising window. Subsequently, the confidence level is obtained by the confidence level of the elements. Finally, the unqualified inflection point candidate elements are eliminated by the decaying elements, and the rising period element list is output. S5: Using the list of rising elements obtained in step S4 as input, first obtain four layers of audience packages—test audience package, main audience package, follower audience package, and maintenance audience package—through element audience mapping. Then, generate the target keywords and negative keywords through keywords and negative keywords. Through joint allocation of bid and budget, calculate the bid coefficient and channel budget allocation ratio by combining the remaining rising window duration and the short window rising degree. Finally, obtain the time-based bid multiplier through the placement of placement time periods and jointly output the placement execution parameter set.

[0009] This invention provides an AI-powered intelligent marketing promotion strategy generation method based on word-of-mouth trends. It has the following beneficial effects: (1) This invention places two sets of review data of the target product and the top-selling similar products in the leaf category under the same observation window system. The net trend sequence is constructed by subtracting the logarithmic changes of the two mention intensity sequences. The three common factors of seasonal changes in the leaf category, changes in platform traffic allocation strategies, and the impact of industry-wide public events are eliminated from the word-of-mouth changes of the target product. This overcomes the shortcomings of the existing technology, which uses the absolute mention volume of the target product's own reviews as the basis for popularity and misinterprets the overall fluctuation of the category as a change in the product's selling points. The subsequent judgment is only based on the excess change of the target product relative to the overall leaf category. At the same time, the mention intensity is taken as the relative mention rate. The overall surge in the total number of reviews brought about by the promotion activities is canceled out in the ratio calculation. The net trend sequence only retains the changes in the word-of-mouth structure.

[0010] (2) This invention uses the standard deviation of the historical slope sample of the attribute viewpoint binary as the normalized denominator, and converts the slope of the three scales into short window rise degree, medium window rise degree and long window rise degree, so that the attribute viewpoint binary with the sample size of the comments spanning two orders of magnitude is on the same judgment scale. The inflection point candidate elements are jointly judged by the same sign relationship and the order of magnitude increase relationship of the three rise degrees, which makes up for the shortcomings of the existing technology that uses a unified threshold for the whole category or the normalization of the standard deviation of the whole sample, and the normal sawtooth jitter of low frequency attribute dimensions is judged as an upward trend. On this basis, the cumulative net trend sequence is fitted with a logistic growth function, and the three process parameters of saturation scale, intrinsic growth rate and midpoint of the curve are output. The remaining rise window duration is calculated from the three process parameters, and the basis for setting the placement cycle is changed from the current popularity to how long the popularity can be maintained. This avoids two types of deviations: continuing to maintain a high bid after the selling point is saturated and prematurely shrinking the budget when the selling point still has a long rise period.

[0011] (3) The present invention constructs the bidding coefficient in the form of the product of the remaining rising window duration and the rising degree of the short window. When either of the two factors approaches zero, the overall bidding coefficient approaches zero, so that rising period elements with short remaining windows or small rising amplitudes do not occupy high bidding space. The direction of the bidding coefficient is opposite to the direction of higher bids in conventional placements where higher heat values ​​result in higher bids. At the same time, the rising period element list is divided into positive element groups and negative element groups according to polarity labels. The positive element group follows the placement layer path and enters the four-layer audience package, while the negative element group follows the defense layer path and is transformed into negative keywords and defense audience packages. This replaces the existing technology's practice of converting both positive and negative expressions into placement keywords and negative expressions into bidding. In addition, the comment posting time is restored to the distribution of product usage time after a six-hour cyclical shift correction. The placement period is aligned with the product usage time rather than the comment writing time, so that the allocation basis of the time-sharing bidding multiplier is consistent with the user's actual usage scenario. Attached Figure Description

[0012] Figure 1Flowchart of the overall solution of this invention; Figure 2 Flowchart of the net trend sequence construction of this invention; Figure 3 Flowchart of the candidate inflection point marker of this invention; Figure 4 Flowchart for determining the rising phase elements of this invention; Figure 5 The flowchart for generating execution parameters in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] like Figure 1-5 As shown in the figure, this invention provides a method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends. The method includes: S1: Reputation data collection and attribute opinion binary labeling; By collecting reviews from multiple sources, we obtain a set of reviews for the target product and a set of reviews for products in the same category. Then, by extracting attribute-viewpoint pairs, we parse each review text into attribute-viewpoint pairs consisting of attribute words and viewpoint words. Finally, we attach a polarity label to each attribute-viewpoint pair using sentiment polarity labeling, resulting in a set of attribute-viewpoint pairs with polarity labels and timestamps.

[0015] The multi-source review collection first retrieves all reviews from the e-commerce platform's product review interface based on the target product's identifier. The retrieval scope covers the review text, user rating, review posting time, and follow-up review text and posting time. Then, it retrieves the leaf category number to which the target product belongs, and within that leaf category, sorts the top 50 product identifiers by sales volume over the past 30 days in descending order. The same retrieval operation is performed on each of these 50 product identifiers, and the retrieval results are merged into a category-matched product review set.

[0016] In this embodiment, it should be specifically noted that the number of category comparison products is fifty. The basis for this is to ensure that the total number of comments in the category comparison product comment set is on the same order of magnitude as the total number of comments in the target product. If the total number of comments in the comparison set is too small, the category comparison mention intensity calculated in step S2 will fluctuate drastically, and this fluctuation will be directly transmitted to the net trend sequence. During the collection process, the comment posting time is standardized to a unified time zone, and the cross-platform time format is converted into an integer sequence in days. The appended comment text is concatenated with the main comment text as a complete comment for subsequent processing. The two sets are written into separate data tables. The data table fields include the comment unique number, product identifier, product source marker, leaf category number, comment text, user rating, comment posting time, and source interface identifier. The product source marker has only two values: target product and category comparison product. This marker is used when distinguishing between the two sequences in step S2. The source interface identifier records whether the comment comes from the e-commerce platform's internal comment section interface, the social platform's product recommendation content comment section interface, or the Q&A section interface. This identifier is used when calculating the placement time arrangement by channel in step S5.

[0017] The specific steps for extracting the attribute viewpoint tuples are as follows: S101, Sentence and Word Segmentation: Segment and word segment the comment text, and remove stop words, emoticons and pure punctuation segments; S102. Paired Structure Location: Using dependency parsing, locate the paired structures of attribute words and opinion words in the sentence segmentation results. The paired structures include two types: the collocation relationship between the attributive and the headword, and the collocation relationship between the subject and the predicate. The attribute words come from a pre-constructed category attribute word list, and the opinion words are directly extracted by dependency parsing from the modification position or predicate position of the attribute words. S103. Binary number generation: Concatenate the strings of attribute words and opinion words and take the hash value to generate binary numbers.

[0018] It should be further explained that the category attribute term list is manually obtained by statistically analyzing the word frequency of all product titles and details pages under the leaf category. The term list entries are stored hierarchically by attribute name, with levels including material, specifications, performance, appearance, and smell. Opinion terms do not rely on a pre-defined term list; they are directly extracted through dependency parsing. The extracted opinion terms retain the original comment text form and are not subject to synonym merging, which would flatten expressions with different semantic strengths, affecting the estimation of fluctuation range in step S3. After a successful pairing of attribute terms and opinion terms, an attribute-opinion tuple is formed. The record fields of the attribute-opinion tuple include the tuple number, attribute term, opinion term, unique comment number, and the character offset position of the sentence in the comment text. Multiple attribute-opinion tuples can be generated from the same comment, and all attribute-opinion tuples generated from the same comment share the user rating and comment publication time of that comment. The tuple number is the primary key connecting the five steps.

[0019] The sentiment polarity labeling takes the semantic vector of the opinion word for each attribute opinion tuple, calculates the cosine similarity with the center vector of the positive sentiment dictionary and the center vector of the negative sentiment dictionary, and takes the side with the larger similarity value as the polarity label. When the difference between the similarity values ​​of the two sides is less than a set threshold, the polarity label is set to neutral.

[0020] In this embodiment, it should be specifically noted that the threshold is set to 0.1. This value is based on the fact that samples with a difference less than 0.1 have a lower consistency rate in polarity determination during manual sampling than 60%. These samples contribute mainly to the trend sequence as noise and are marked as neutral, thus not proceeding to the verification stage in step S3. The annotation results are written into the annotation table along with the attribute / opinion pairs. The annotation table fields include pair number, product identifier, product source marker, attribute word, opinion word, polarity tag, comment publication time, and comment unique number. After annotation, the missing percentage of polarity tags is calculated. If the missing percentage is greater than 5%, the process returns to the attribute / opinion pair extraction stage for re-execution; if the missing percentage is less than 5%, the missing samples are directly removed. The attribute / opinion pair set is stored separately according to two sources: the target product and the category comparison product, for use in step S2 to divide the observation window and calculation window to mention intensity.

[0021] Step S2: Category comparison baseline stripping and net trend sequence construction; Using the set of attribute viewpoint pairs obtained in step S1 as input, the set of attribute viewpoint pairs is first divided into observation window sequences by time window partitioning. Then, the target product mention intensity sequence and the category comparison mention intensity sequence are obtained by calculating the window mention intensity. The sparse window and unique elements are then aligned by the comparison sequence. Finally, the net trend sequence is constructed by subtracting the logarithmic change of the category comparison mention intensity sequence from the logarithmic change of the target product mention intensity sequence.

[0022] The time window division sets the total backtracking length to eighty-four days and the window length to seven days, resulting in twelve consecutive observation windows. The observation window number t is taken from one to twelve, with the number twelve corresponding to the window at the current time.

[0023] It should be further explained that the seven-day window length is based on the large fluctuations in comment data on a daily scale. The number of mentions of attribute-opinion pairs on a single day is often zero. The dense occurrence of zero values ​​renders subsequent slope estimation meaningless. The seven-day window absorbs the differences in shopping and delivery rhythms within a week, ensuring sufficient comment samples within each observation window. The total backtracking length of eighty-four days is based on the requirement in step S4 to fit the cumulative net trend sequence with a logistic growth function. The fitting requires the curve segment to cover the complete shape from the initial level to the current level. Differentiation of the twelve observation windows yields eleven net trend values, and eleven data points are the minimum acceptable number of points for a nonlinear three-parameter fitting. Window segmentation uses the zero point of the current calendar day as the right boundary for backtracking to the left. The left-closed and right-open intervals of each observation window are explicitly recorded in the data table to avoid duplicate counting of comments across windows.

[0024] The window mention intensity calculation involves counting the number of comments containing attribute viewpoint tuple e within each observation window t, then dividing this number by the total number of comments within that observation window. The resulting ratio is denoted as the mention intensity. Mention intensity is calculated separately for each product source tag, resulting in a target product mention intensity sequence m and a category-matching mention intensity sequence c. Both sequences use the observation window number as the horizontal axis and the tuple number as the index, with a sequence length of twelve. Attribute viewpoint tuples with neutral polarity tags are not included in the mention intensity calculation. Attribute viewpoint tuples with positive polarity tags and those with negative polarity tags have naturally separate tuple numbers due to the different viewpoint terms; the two polarity sequences are constructed independently and do not cancel each other out.

[0025] In this embodiment, it's important to clarify that the mention intensity is calculated using relative mention rate, and the absolute number of comments is not included in the calculation. The method of determining this value is based on two principles. First, the total number of comments for the target product and its category counterpart can differ by two orders of magnitude, making direct comparison of absolute numbers between the two sequences impossible. Second, promotional activities can lead to a surge in the overall number of comments; during major promotions, the total number of comments for the target product can increase to three times its usual amount. Absolute numbers interpret this increase in total comments as a change in reputation factors, while relative mention rate cancels out the overall effect, retaining only structural changes.

[0026] The alignment process addresses two types of non-alignable scenarios. The first is sparsity in the observation window dimension. During alignment, the total number of valid reviews in the category-matching product review set within each observation window is counted. Valid reviews are defined as those that have completed attribute opinion tuple extraction and have non-neutral polarity labels. Observation windows with a total number of valid reviews less than a set lower limit (500 reviews) are marked as sparse windows. The category-matching mention intensity of sparse windows is filled by linear interpolation between two adjacent non-sparse windows. The interpolation endpoints are the pair of non-sparse windows whose window indices are closest to the sparse window, and the interpolation weights are evenly distributed according to the window indices. The second type is missing items in the tuple dimension. If a certain attribute opinion tuple appears only in the target product review set, and the number of mentions in the category-matching product review set across all twelve observation windows is zero, this attribute opinion tuple is marked as a unique element.

[0027] It should be further explained that the category comparison mention intensity of unique elements is smoothed by a constant across all observation windows and is not involved in the subtraction operation in the construction of the net trend sequence. Instead, only the logarithmic change of the target product mention intensity sequence is taken as the net trend value of the element, and the "whether it is a unique element" field is set to "yes" in the output. Unique elements will follow a separate associative word expansion path in step S5, and this is marked in advance to allow subsequent steps to treat them differently.

[0028] The net trend sequence is constructed by using each attribute view tuple e and each observation window number t from two to twelve, and calculating the net trend value according to the following formula: ; The d e,t Let m be the net trend value of the attribute-view pair e at observation window index t; e,t Intensity of mention of the target product; c e,t The intensity of mention in category comparison; ε is a smoothing constant.

[0029] The net trend value structurally comprises two layers of processing. The first layer is the logarithmic ratio, which expresses the multiple change in mention intensity between adjacent observation windows. This multiple change puts binary pairs with different magnitudes on a comparable scale. The logarithmic ratio is the same for mention intensity increasing from 0.1% to 0.2% and from 0.1% to 0.2%, allowing direct comparison of the magnitude of change in the two binary pairs. When calculating the difference directly, the latter's value will be 100 times that of the former. The true changes in less popular binary pairs will be completely masked by the daily fluctuations of popular binary pairs. The second layer is the subtraction of the logarithmic ratios of the two sequences. The category-related mention intensity sequence carries three common factors: seasonal changes in leaf categories, changes in platform traffic allocation strategies, and the impact of industry-wide public events. The target product mention intensity sequence also carries these three common factors. After subtraction, the common factors are eliminated, and the remaining part is the excess change of the target product relative to the overall leaf category. A net trend value greater than zero indicates that the target product's mentions on that attribute view tuple are growing faster than the category as a whole, a net trend value less than zero indicates that it is growing slower than the category as a whole, and a net trend value fluctuating around zero indicates that it is in sync with the category as a whole.

[0030] After the net trend sequence is constructed, it is output line by line according to the attribute viewpoint tuples. Each sequence is eleven in length and includes three auxiliary fields. The first auxiliary field is the total number of reviews for the target product window, recording the sequence of valid reviews for the target product within the twelve observation windows. The second auxiliary field is the total number of valid reviews for the category comparison window, recording the sequence of valid reviews for the category comparison product within the twelve observation windows. The third auxiliary field indicates whether it is a unique feature, with a value of yes or no. These three auxiliary fields are read in the sparse window test in step S3 and the feature confidence grading in step S4. The net trend sequence, auxiliary fields, smoothing constant, and tuple number are combined and written into the net trend data table, which is the only input for step S3.

[0031] Step S3: Multi-scale slope analysis and inflection point candidate markers; Using the net trend sequence obtained in step S2 as input, the least squares line fitting is first performed on the short window, medium window and long window scales through multi-scale slope estimation to obtain the slope of the short window, the slope of the medium window and the slope of the long window. Then, by normalizing the fluctuation amplitude, the standard deviation of the historical slope sample of the attribute viewpoint binary is used as the normalization denominator to obtain the rise degree of the short window, the rise degree of the medium window and the rise degree of the long window. Finally, through the inflection point candidate label, the same sign test, the magnitude increase test and the sparse window test are performed on the three rise degrees in sequence. The attribute viewpoint binary that passes all three tests is written into the candidate list, and the attribute viewpoint binary that fails any one test is transferred to the regular word-of-mouth element pool.

[0032] The multi-scale slope estimation performs least-squares linear fitting on each net trend sequence at three window scales. The short window covers the three net trend values ​​corresponding to observation window numbers 10, 11, and 12; the medium window covers the six net trend values ​​corresponding to observation window numbers 7 to 12; and the long window covers the nine net trend values ​​corresponding to observation window numbers 4 to 12. The fitting uses the observation window number as the independent variable and the net trend value as the dependent variable. The physical meaning of the slope is the change in net trend value with each additional observation window.

[0033] It's important to further clarify that the three windows share the same right endpoint. The reason for this shared endpoint is that all three slopes need to describe the rate of change under the same current state. If the right endpoints were staggered, the three slopes would mix information from different periods, rendering the magnitude-increasing test results meaningless. The reason for using three scales instead of a single scale is that the slope given by a single window length cannot distinguish between two completely different scenarios for investment decisions. The first scenario is that the net trend sequence steadily rises over a long period, and the trend of the last three observation windows is simply a continuation of this long-term upward trend. The slopes of the short, medium, and long windows are all close in value and positive. The second scenario is that the net trend sequence fluctuates close to zero over a long period, and an isolated peak appears in the last three observation windows. In this case, the short window slope is very large, and the long window slope is close to zero. While the short window slopes can have the same value in both scenarios, their investment value differs significantly. The first scenario represents a continuous upward trend that can be further increased, while the second scenario represents an isolated fluctuation that will subside within a few weeks. After juxtaposing the three scales, the two cases are distinguished by the ratio of the long window slope to the short window slope. The ratio in the first case is close to one, while the ratio in the second case is much greater than one.

[0034] The self-variance normalization process first performs a moving least squares fit with a window length of three on all eleven values ​​of the net trend sequence, with a sliding step size of one, resulting in nine historical slope samples. The standard deviation of these nine historical slope samples is then calculated, and the standard deviation is denoted as the self-dispersion amplitude of the attribute viewpoint binary. Subsequently, the short window slope is divided by its self-dispersion amplitude to obtain the short window rise, the medium window slope is divided by its self-dispersion amplitude to obtain the medium window rise, and the long window slope is divided by its self-dispersion amplitude to obtain the long window rise. The self-dispersion amplitude is calculated for each binary pair, and no two binary pairs share the same value.

[0035] This embodiment requires further explanation regarding the necessity of normalizing the fluctuation amplitude, stemming from the differences in the net trend sequence morphology between different binary pairs. The net trend sequence of high-frequency binary pairs is supported by sufficient samples across eleven observation windows, exhibiting a smooth sequence shape and small dispersion amplitude. Conversely, the net trend sequence of low-frequency binary pairs has only single-digit sample sizes in some observation windows, resulting in a jagged sequence shape and large dispersion amplitude. When skipping normalization and directly comparing the absolute values ​​of the slopes, the slope generated by a normal jagged jitter in a low-frequency binary pair can exceed the true trend slope of the high-frequency binary pair, filling the candidate list. However, the sample size of comments from low-frequency binary pairs is insufficient to support the deployment decision. After dividing by its own dispersion amplitude, the rise rate becomes the multiple of the current slope relative to the historical fluctuation level of the binary pair itself. This multiple can be directly compared between different binary pairs, placing high-frequency and low-frequency binary pairs on the same judgment scale. This approach, which uses the historical fluctuation range of the object itself as the normalized denominator, originates from the process capability determination in statistical process control. It is not common in e-commerce reputation trend analysis, which usually uses a uniform fixed threshold for the entire category or the standard deviation of all binary groups as the normalized denominator. Both approaches will result in inconsistent judgment scales between low-frequency binary groups and high-frequency binary groups.

[0036] The inflection point candidate markers perform three checks on the three degrees of ascent in sequence. The execution order of the three checks is arranged from low to high computational cost. If the previous check fails, the next check will not be executed.

[0037] The first item is the same sign test, which requires all three conditions to be met simultaneously: the rise of the short window is greater than zero, the rise of the medium window is greater than zero, and the rise of the long window is greater than zero. If any one of the rises is less than or equal to zero, the test fails. The same sign test excludes attribute pairs where the net trend sequence is in a downward process or an alternating positive and negative oscillation process.

[0038] The second test is the magnitude increase test. The criteria are that both the short window rise rate must be greater than the product of the medium window rise rate and the tolerance coefficient, and the medium window rise rate must be greater than the product of the long window rise rate and the tolerance coefficient, with the tolerance coefficient set to 0.9. The magnitude increase test identifies attribute-viewpoint pairs where the rate of change itself is increasing. Three increases forming an increasing sequence indicate that the net trend sequence curve is curving upwards, corresponding to the segment before and after the inflection point of the logistic growth curve. Attribute-viewpoint pairs where all three increases are positive but the increasing relationship does not hold are in a decelerating upward segment, having already passed the inflection point and moved towards saturation; therefore, this test fails.

[0039] It should be further explained that the tolerance coefficient is set to 0.9 because there is an overlap between the windows of the three scales. The three observation windows of the short window are simultaneously covered by the medium and long windows. This overlap leads to a structural correlation between the three slopes. A completely strict increasing relationship would have too low a pass rate in actual data.

[0040] The third step is the sparse window test. The criteria are that the number of sparse windows marked in the three observation windows covered by the short window is zero, and the minimum number of reviews for the target product across the three observation windows covered by the short window is greater than one hundred. The sparse window test eliminates false acceleration signals driven by interpolation of the control sequence. After linear interpolation in step S2, the category control mention intensity sequence becomes overly smooth in the sparse section. The logarithmic change in the smooth section is close to zero, and the net trend value is entirely determined by the logarithmic change of the target product itself. Normal fluctuations in the target product are amplified into an accelerated upward trend by the difference operation. These false signals must be intercepted in the candidate stage; proceeding to step S4 for fitting would consume computational resources and produce incorrect lifecycle prediction results.

[0041] Attribute-viewpoint binary pairs that pass all three tests are marked as inflection point candidate elements and added to the candidate list. The candidate list fields include binary pair number, attribute term, viewpoint term, polarity label, short-window rise degree, medium-window rise degree, long-window rise degree, self-dispersion magnitude, whether it is a unique element, and smoothing constant. The short-window rise degree is read in the bid coefficient calculation in step S5, and whether it is a unique element is read in the element confidence level classification in step S4 and the associated word expansion in step S5.

[0042] Any attribute / opinion tuple that fails any test is moved to the regular word-of-mouth element pool. Attribute / opinion tuples in the regular word-of-mouth element pool still enter the keyword and negative word generation stage in step S5, but are only maintained as existing keywords and do not participate in the adjustment of bidding coefficients and channel budget allocation ratios.

[0043] In this embodiment, it should be specifically explained that the purpose of traffic splitting is to ensure that changes in delivery parameters are driven only by attribute viewpoint tuples that are in the rising phase. When all tuples participate in parameter adjustment at the same time, the parameter fluctuations between two adjacent executions are too large, the delivery system needs to relearn frequently, and the conversion data during the learning period is unavailable, which will further interfere with the judgment input of the next round.

[0044] Step S4: Determining the factors in the upward phase and predicting the life cycle; Using the candidate list obtained in step S3 as input, the cumulative net trend sequence is first fitted with a logistic growth function through rising curve fitting to obtain the saturation scale, intrinsic growth rate and midpoint of the curve. Then, the propagation stage is determined by the propagation stage determination, and the remaining rising window duration is calculated by the remaining rising window. Subsequently, the confidence level is obtained by the confidence level of the elements. Finally, unqualified inflection point candidate elements are eliminated by the decaying elements, and the rising period element list is output.

[0045] The first step in fitting the upward curve is to perform a cumulative summation of the net trend sequence. The eleven values ​​of the net trend sequence describe the change between adjacent observation windows. The sequence has a pulse-like shape, and pulse-like sequences cannot be directly used for growth curve fitting. The cumulative summation starts from observation window number two and adds up the net trend values ​​for each observation window to obtain the cumulative net trend sequence. The sequence length is eleven, and the horizontal axis represents observation window numbers two to twelve. The cumulative net trend sequence describes the total relative change accumulated from the start of the retrospective period to the current observation window, and the sequence has an S-shaped growth profile.

[0046] It should be further noted that the cumulative net trend sequence may take negative values ​​in some observation windows; the negative value segments come from observation windows with negative net trend values. Before fitting, the cumulative net trend sequence is shifted to its minimum value, moving it to a smoothing constant position one unit above zero. The shift amount is recorded in the data table. The shift operation does not change the shape of the curve; the intrinsic growth rate and the midpoint time of the curve have the same values ​​before and after the shift. The absolute value of the saturation scale changes with the shift amount. The saturation scale is only used subsequently to calculate the completion ratio; the bias introduced by the shift is canceled out in the ratio calculation, and the result of the propagation stage determination is not affected by the shift.

[0047] After the minimum value shift is completed, the cumulative net trend sequence is fitted using the logistic growth function, which has the following form: ; The Se (t) K represents the cumulative net trend value of candidate inflection point element e at observation window index t; e For saturation scale; r e Internal growth rate; t 0,e The time is the midpoint of the curve; the value of t ranges from two to twelve.

[0048] In this embodiment, it's important to note that the selection of the logistic growth function is based on the diffusion structure of word-of-mouth elements within the user group. The diffusion speed of a newly emerging product's selling point, from being mentioned by a few early adopters to being mentioned by the majority of buyers, is limited by the size of the potential user base who haven't yet purchased, used, or written reviews. This potential user base gradually diminishes as diffusion progresses, with the growth rate initially accelerating and then slowing down. The cumulative mentions eventually approach the upper limit of the natural mention ratio of this selling point among all buyers; this upper limit corresponds to the saturation point. The overall pattern of this process is consistent with the logistic growth curve. In e-commerce marketing, word-of-mouth time series processing typically employs two methods: moving average and exponential smoothing. Both methods output smoothed values ​​of historical data, which only describe the trend up to the current moment and cannot indicate how many more observation windows this trend can continue in the future. The output of the logistic fitting includes three process parameters: saturation point, intrinsic growth rate, and the midpoint of the curve. These three parameters together determine the shape of the entire curve in the future, and the remaining upward window duration can be directly calculated from these three parameters. This step differs from conventional word-of-mouth popularity statistics by modeling word-of-mouth elements as a propagation process with a capacity limit and outputting process parameters.

[0049] The fitting employs nonlinear least squares, with the initial values ​​for the three parameters set as follows: The initial value for the saturation scale is twice the maximum value of the cumulative net trend sequence. This is based on the premise that the maximum value of the cumulative net trend sequence is necessarily less than the saturation scale, and setting it to twice ensures that the initial value and the true value fall on the same side, preventing iterations from crossing the curve from below. The initial value for the intrinsic growth rate is 0.3, corresponding to the common speed at which the main upward phase is completed in approximately ten observation windows. The initial value for the curve midpoint time is taken as the middle window number of the backtracking period, with a value of seven. The upper limit for fitting iterations is set to 500, and the convergence criterion is that the relative change in the sum of squared residuals is less than 10 to the power of -6. The fitting performs joint optimization on the three parameters, with parameter boundary constraints that the saturation scale is greater than the maximum value of the cumulative net trend sequence, the intrinsic growth rate is greater than zero, and the curve midpoint time falls within the observation window number range of 2 to 14. The upper bound for the curve midpoint time is widened to 14 because the curve midpoint of inflection point candidate elements may fall after the current time, and falling after this time is precisely the source of a positive remaining upward window duration.

[0050] The propagation stage determination is based on the ratio of the current cumulative net trend value to the saturation scale. This ratio is denoted as the completion degree, and the current moment corresponds to observation window number twelve. When the completion degree is less than 0.25, the propagation stage is determined to be the budding stage; when the completion degree is greater than or equal to 0.25 and less than or equal to 0.75, the propagation stage is determined to be the diffusion stage; and when the completion degree is greater than 0.75, the propagation stage is determined to be the near-saturation stage.

[0051] It needs further explanation that the values ​​of the three dividing points correspond to three characteristic positions on the logistic growth curve. A completion level of 0.25 marks the point where the curve begins to clearly deviate from the horizontal segment. Before this position, the absolute value of the accumulated amount is relatively small, making it difficult to achieve scale with further investment. The attribute opinion tuple corresponding to this position is only mentioned by a few early buyers. A completion level of 0.75 corresponds to the point where the growth rate begins to decline significantly after the midpoint of the curve. After this position, the remaining space only accounts for a quarter of the saturated scale, and the marginal output of further investment decreases rapidly. The midpoint of the curve corresponds to a completion level of 0.5. The growth rate at this position is the maximum value of the entire curve, and the attribute opinion tuple in the middle of the diffusion segment has the largest increment per unit time. The determination of the propagation stage determines the audience package stratification and the stratification adjustment multiple of the bidding coefficient in step S5. The propagation stage and confidence level are used together to segment the audience package. When segmenting with a single dimension, high-confidence elements in the nascent stage will be pushed to the main position, while the absolute amount in the nascent stage is insufficient to support the main investment.

[0052] The remaining rising window is calculated by taking the midpoint of the curve, the intrinsic growth rate, and the saturation determination ratio. The calculation method is to add the midpoint of the curve to the product of the reciprocal of the intrinsic growth rate and the logarithmic probability of the saturation determination ratio, and then subtract the observation window number corresponding to the current moment. The saturation determination ratio is set to 0.95. The result is recorded as the remaining rising window duration. The unit of the remaining rising window duration is the number of observation windows. When converted to days, it is multiplied by the window length of seven days.

[0053] In this embodiment, it's important to clarify that a saturation determination ratio of 0.95% means that the upward trend of a candidate element at the inflection point is considered to have ended when the cumulative net trend value reaches 95% of the saturation level. When the saturation determination ratio is 100%, the logistic growth curve approaches the saturation level at a progressively slower rate, and the observation window occupied by the last 5% becomes significantly longer. Including this tail in the delivery window will cause the bidding coefficient to remain in a high range for an extended period. If the saturation determination ratio is too low, the upward trend is prematurely determined to have ended, and elements with remaining growth potential are prematurely budgeted. 0.95% is a compromise between the two directions; under this value, the remaining upward window duration falls between zero and eight observation windows, which aligns with the actual lifecycle length of e-commerce product selling points.

[0054] It needs further clarification that a remaining upward window duration of less than zero indicates that the current time has passed the saturation judgment position according to the fitted results, and this inflection point candidate element is eliminated in the decay element elimination stage. A remaining upward window duration greater than zero and less than two indicates that the remaining space is limited, and this inflection point candidate element is processed as a maintenance audience package after entering step S5. A remaining upward window duration greater than two indicates entry into the main audience package, follower audience package, or trial audience package, with the specific allocation determined by the dissemination stage and confidence level. The remaining upward window duration provides a quantitative value indicating how long a selling point can continue to be promoted. Conventional word-of-mouth advertising only considers the current level of popularity, which cannot answer the question of how long it can be maintained. The setting of the advertising cycle relies on the experience and judgment of the operations personnel. Experience-based judgment can lead to two types of biases: one is continuing to maintain high bids after the selling point has reached saturation, resulting in a decrease in conversion rate and sunk costs for the traffic purchased at high bids after the selling point reaches saturation; the other is prematurely reducing the budget due to short-term data fluctuations when the selling point still has a long upward period, missing the main incremental range. The remaining upward window duration transforms the setting of the placement cycle from empirical judgment to direct calculation of fitting parameters. The calculation result is inversely related to the bidding coefficient in step S5. The smaller the value of the remaining upward window duration, the lower the bidding coefficient is suppressed.

[0055] The confidence grading of the elements assigns a score to each inflection point candidate element based on three indicators, with the scores from the three indicators added together to obtain the total score. The first indicator is the relative residual, calculated by dividing the sum of squared residuals by the number of samples in the cumulative net trend sequence to obtain the mean squared residual. This mean squared residual is then divided by the maximum value of the cumulative net trend sequence to obtain the relative residual. A relative residual less than 0.15 scores 3 points, greater than or equal to 0.15 but less than or equal to 0.3 scores 2 points, and greater than 0.3 scores 1 point. The second indicator is the total number of reviews during the backtesting period, calculated by the number of reviews containing the attribute's opinion binary within an 84-day backtesting period for the target product. A number greater than 50 scores 3 points, greater than or equal to 20 but less than or equal to 50 scores 2 points, and less than 20 scores 1 point. The third indicator is whether it is a unique element. A "no" value for the "unique element" field scores 3 points, and a "yes" value scores 2 points. A total score of 9 points is Grade A, 7 or 8 points is Grade B, 5 or 6 points is Grade C, and a total score less than 5 points is Grade D.

[0056] It needs further explanation that dividing the relative residual by the maximum value of the cumulative net trend sequence converts the residual into a dimensionless quantity, allowing inflection point candidate elements of different magnitudes to be compared under the same set of thresholds. The total number of comments during the retrospective period directly determines the estimation accuracy of the mention intensity. When the number of comments is low, both the denominator and numerator of the mention intensity are small, meaning that the comment behavior of a single user can change the value of the mention intensity. The basis for reducing the score of unique elements by one point is that unique elements do not have category-matched mention intensity sequences participating in the difference calculation. The net trend sequence consists only of the logarithmic change of the target product itself, lacking the correction layer of overall fluctuation of the leaf category. Therefore, the robustness of the judgment result is weaker than that of inflection point candidate elements with comparisons participating in the difference calculation. The four levels of grading correspond to different stratification adjustment multiples and matching methods in step S5.

[0057] The decay factor excludes four types of inflection point candidate elements. The first type is fitting failure, with three criteria: nonlinear least squares fails to converge within 500 iterations; the intrinsic growth rate after convergence is less than or equal to zero; and the saturation size after convergence is less than the maximum value of the cumulative net trend sequence. An intrinsic growth rate less than or equal to zero indicates the fitted curve is in a downward trend, contradicting the premise that the inflection point candidate element passed the order-of-magnitude test in step S3. This type of result usually comes from the candidate stage test being misled by noise and then exposed during the fitting stage. A saturation size less than the maximum value of the cumulative net trend sequence indicates the curve cannot accommodate the existing data, and the fitting result is unusable. The second type is a remaining rising window duration less than zero. The third type is a confidence level of D. The fourth type is a negative polarity label and a completion rate greater than 0.75.

[0058] In this embodiment, it should be specifically noted that the fourth category of elimination is based on the fact that negative attribute opinion tuples nearing saturation indicate that negative word-of-mouth has been fully spread, and actions on the advertising side cannot change the spread at this stage. This element is then transferred to the product and customer service processing flow, and no more resources are allocated to it in the advertising parameter generation stage. Inflection point candidate elements with negative polarity tags and a completion rate of less than or equal to 0.75 are not eliminated. These elements enter step S5 and proceed through the negative keyword and defensive audience segmentation branches. The remaining inflection point candidate elements after elimination constitute the rising phase element list. The fields of the rising phase element list include tuple number, attribute word, opinion word, polarity tag, dissemination stage, completion rate, remaining rising window duration, confidence level, saturation scale, intrinsic growth rate, midpoint of the curve, short window rising degree, and whether it is a unique element. The rising phase element list is the only element input for step S5, while the regular word-of-mouth element pool serves as an auxiliary input for step S5 for maintaining existing keywords.

[0059] Step S5: Generate deployment execution parameters; Using the list of rising elements obtained in step S4 as input, first obtain four layers of audience packages—test audience package, main audience package, follower audience package, and maintenance audience package—through element audience mapping. Then, generate ad keywords and negative keywords through keywords and negative keywords. Through joint allocation of bid and budget, calculate the bid coefficient and channel budget allocation ratio by combining the remaining rising window duration and the short window rising degree. Finally, obtain the time-based bid multiplier through the placement of ad placement periods and jointly output the ad placement execution parameter set.

[0060] The element-based population mapping first divides the rising element list into positive and negative element groups based on polarity labels. The positive element group follows a layered deployment path, while the negative element group follows a layered defense path. The two groups share the same rising element list but are processed completely separately after this step.

[0061] For each attribute-opinion binary in the positive element group, the attribute word and opinion word strings are retrieved and searched in two levels in the category audience tag library of the advertising platform. The category audience tag library is an open tag system of the advertising platform, including three categories: category interest tags, behavioral preference tags, and spending power tags. The first level of search recalls candidate tags by attribute words, and the second level of search filters by opinion words within the candidate tag range. Tags that are matched in both levels of search are considered direct related tags for that binary. If the number of directly related tags is less than three, expanded recall is initiated. Expanded recall includes the hypernyms and hyponyms of the attribute words in the search scope. The hypernyms and hyponyms come from the hierarchical structure of the category attribute word table. Tags matched in expanded recall are considered indirect related tags. In the subsequent budget weight calculation, the indirect related tags take half the weight of the directly related tags.

[0062] It should be further explained that the two-level retrieval order is set based on the fact that the recall scope of attribute words is much larger than that of opinion words. First, the candidate set is narrowed down by attribute words, and then refined by opinion words, resulting in lower retrieval costs than the reverse order. Audience groups are segmented according to two dimensions: propagation stage and confidence level. For binary groups in the nascent stage of propagation, their directly and indirectly related tags form an exploratory audience group. For binary groups in the diffusion stage with a confidence level of A, their tags form a main audience group. For binary groups in the diffusion stage with a confidence level of B or C, their tags form a follower audience group. For binary groups in the near-saturation stage of propagation, their tags form a maintenance audience group. Tag deduplication is performed among the four levels of audience groups. When the same tag appears in multiple layers, the highest level is retained. The hierarchy is arranged as: main audience group, follower audience group, exploratory audience group, and maintenance audience group. The reason for deduplication is that counting the same tag multiple times will distort the scale weight in the budget weight calculation. The scale weight will be obtained from the audience scale query interface of the advertising platform, and the interface only counts duplicate tags once.

[0063] This embodiment requires further explanation. The negative element group's binary tuples are retrieved by attribute words to identify the "purchased" audience tags, resulting in a "purchased defense audience package." This package is used to push usage guidance content and after-sales access to users who have completed their purchases. The semantic vectors of the negative element group's binary tuples are taken from the opinion words, and their cosine similarity is calculated with the semantic vectors of the positive opinion words of the competitor's selling points. A similarity greater than 0.8 indicates that the opinion words are consistent. For binary tuples with this consistent similarity, competitor audience tags are retrieved by attribute words, resulting in a competitor defense audience package. This package is used to strengthen the target product's description of corresponding attributes within the competitor's audience. The positive opinion words of the competitor's selling points are extracted from the category-matching product review set in the same manner as steps S1 to S3, with the extraction scope limited to attribute opinion binary tuples with positive polarity tags and a completion rate greater than 0.5.

[0064] The keyword and negative word generation process combines the binary pairs of positive element groups into a phrase according to the original text order of attribute words and opinion words, and then combines them into another phrase in reverse order with opinion words first and attribute words last. Both phrases are written into the keyword table. The matching method is set according to the audience package to which the binary pair belongs. Broad match is used for the keywords corresponding to the trial audience package, phrase match is used for the keywords corresponding to the main audience package and the follower audience package, and exact match is used for the keywords corresponding to the keywords corresponding to the maintenance audience package.

[0065] It's important to further clarify that the matching method settings are consistent with the purpose of each audience segment. The purpose of the exploratory audience segment is to observe real search behavior; broad match maximizes the recall range; the main audience segment and follower audience segment bear the main conversion task; phrase match strikes a balance between recall range and precision; the maintenance audience segment only retains core keywords that have been verified; and exact match limits traffic to verified expressions. For binary pairs where the "unique element" field is set to "yes," the target keywords undergo an additional keyword expansion. This expansion involves inputting a phrase combining attribute and opinion terms into the e-commerce platform's search box association interface. The top ten words with the smallest edit distance from this phrase in the returned association list are then added to the target keyword table, with the matching method set to broad match. The basis for expanding association keywords for unique elements is that if the unique element does not appear in the category-matching product review set, it indicates that the selling point has not yet formed an industry-standard expression within the leaf category. The words used by users in actual searches may differ from the expressions in reviews. The association keyword list reflects the actual search terms used by users.

[0066] Negative element groups are combined into negative keywords in the original text order of attribute words and opinion words, and written into the negative keyword list. The negative keyword list is divided into two levels according to confidence level: binary pairs with a confidence level of A are written as exact negations, and binary pairs with a confidence level of B or C are written as phrase negations. The negative keyword list is attached to the main audience package and the follower audience package; the test audience package does not have the negative keyword list attached.

[0067] In this embodiment, it's important to note that the reason for not attaching a negative keyword list to the test audience is that the purpose of the test audience includes observing the actual traffic performance of negative keywords. Prematurely blocking these keywords would result in the loss of this judgment basis. Furthermore, the traffic share of the test audience is small, and the ineffective consumption caused by negative keywords is within a controllable range. The binary pairs in the regular word-of-mouth element pool are combined with attribute words and opinion words and then written into the existing keyword table. The matching method for the existing keyword table remains unchanged from the previous round of campaigning, and the bid remains unchanged from the benchmark bid of the campaign platform.

[0068] The joint allocation of bids and budgets first calculates the bid coefficient for each factor during the upward phase using the following formula: ; The β e β0 is the bidding coefficient for factor e during the rising period; β0 is the base bidding coefficient, input by the operations side based on the target product's gross profit margin and target production ratio; T e The remaining rising window duration is the output of step S4; τ is the window duration scaling constant, which takes values ​​from four observation windows; g e κ is the short-window rise degree output in step S3; κ is the rise degree half-saturation constant, which takes the median of the short-window rise degree of all rising period elements in this leaf category.

[0069] The first factor in the bidding coefficient maps the remaining rising window duration to a value between zero and one. As the remaining rising window duration approaches zero, this factor approaches zero, and the bidding coefficient is pushed towards a lower value range. Once the remaining rising window duration exceeds three times the window duration scaling constant, this factor approaches one and ceases to amplify further.

[0070] It's important to further explain that the first factor's technical meaning is that the fitting results show selling points nearing saturation should not receive high bids. Traffic acquired through high bids will see a decrease in conversion rates after the selling point saturates, making it impossible to recover the investment cost. This factor contradicts the conventional logic of higher bids for higher popularity. Conventional logic reads the current popularity value, while this factor reads how long the current popularity can be maintained. These two values ​​will yield opposite conclusions in the later stages of a selling point's lifecycle. The second factor in the bidding coefficient maps the short-window rise rate to between zero and one. When the short-window rise rate is equal to the half-saturation constant, this factor is half; when the short-window rise rate is much greater than the half-saturation constant, this factor approaches one; and when the short-window rise rate approaches zero, this factor approaches zero. The second factor's function is to limit the bidding space occupied by elements with large remaining rise window durations but small rise rates. These elements have flat and long curves with limited absolute increases. The half-saturation constant for the rise degree is taken as the median, not a constant. The median is recalculated for each leaf category in each round of execution. The magnitude of the rise degree varies greatly between categories. A fixed constant would cause all elements of some categories to fall into the high-saturation zone of the factor, and all elements of some categories to fall into the low-saturation zone, causing the factor to lose its distinguishing ability. The two factors are multiplied. The multiplication form makes the overall bidding coefficient approach zero when either factor approaches zero. Both conditions must be met simultaneously to obtain a high bid. Elements with long remaining windows and weak rising momentum, as well as elements with strong rising momentum and short remaining windows, are suppressed.

[0071] The bid multiplier is adjusted twice based on audience segmentation. The adjustment multiplier for the main audience segment is 1.2, for the follower audience segment it is 1.0, for the test audience segment it is 0.6, and for the maintenance audience segment it is 0.4. The final bid multiplier is obtained by multiplying the segment adjustment multiplier by the bid multiplier, and this final bid multiplier is applied to the benchmark bid on the advertising platform.

[0072] In this embodiment, it should be specifically explained that the values ​​of the four tiered adjustment multiples are based on the following: the main audience group is responsible for the main conversion task and needs to grab traffic; the follower audience group maintains the baseline; the test audience group is for the purpose of small-scale verification and does not need to grab traffic; and the elements of the maintenance audience group have entered the later stage of the life cycle and only need to maintain basic exposure.

[0073] Channel budget allocation is conducted under the constraint of a total budget, which is input by the operations team. The budget weight for each audience segment is the sum of the final bid coefficients of all binary groups within that segment, multiplied by the segment's scale weight. The scale weight is defined as the proportion of reachable users covered by the associated tags of that segment to the total reachable users of all four audience segments. The reachable user count is read from the audience scale query interface of the advertising platform, and this reading is performed after tag deduplication. The channel budget allocation ratio is obtained after the budget weights are normalized.

[0074] The channel budget allocation ratio is subject to upper and lower limits. The upper limit for the budget allocation of a single audience segment is set at 60%, and the lower limit is set at 5%. The upper limit constraint prevents a single top-tier element in the diffusion segment from monopolizing the budget. Monopolization would cause the campaign to lose element diversity, and if the fitting result of that element deviates, there will be no buffer for the entire round of campaigning. The lower limit constraint ensures that there is always traffic entering the test audience segment. New comments generated by the test traffic are the collection targets for step S1 in the next round. When the test budget is reduced to zero, the new element discovery channel is interrupted, and the plan can only make repeated adjustments to existing elements in subsequent rounds. After the constraint is triggered, the budget is renormalized proportionally. Renormalization is performed for a maximum of three rounds. After three rounds, if there are still out-of-bounds entries, they are locked according to the boundary value, and the remaining budget is allocated among the other audience segments according to budget weight.

[0075] The placement time arrangement is based on the posting time of all comments corresponding to the binary pairs in the rising period element list, and aggregated by hour to obtain the 24-hour time period density distribution. The time period density is defined as the proportion of the number of comments posted in a certain hour to the total number of related comments.

[0076] Before use, the time-period density distribution is cyclically shifted and corrected. The shift amount for the cyclical shift correction is six hours, moving the time-period density value of each hour to the position six hours before that hour, and wrapping back to the end of the day for the position that crosses the zero point.

[0077] It's important to further clarify that the cyclical shift correction is based on a systematic offset between review posting time and actual product usage time. Users typically write reviews after using a product within a certain timeframe, with the peak of review posting time lagging behind the peak of product usage time. Shifting the overall time-period density distribution forward by six hours restores the distribution of product usage time. When the ad placement time aligns with product usage time, the conversion path from ad viewing to purchase and usage is shorter. The time-period density distribution after cyclical shift correction is arranged in descending order of value: the first eight hours are designated as the core placement period, hours nine to sixteen as secondary placement periods, and the remaining eight hours as low placement periods. The time-period bid multiplier for the core placement period is 1.3, for the secondary placement period it's 1.0, and for the low placement period it's 0.5. The time-period bid multiplier, along with the final bid coefficient and channel budget allocation ratio, constitutes the time-period placement plan.

[0078] Time-based density distribution is calculated separately for each channel. The target product's review sources include three channels: e-commerce platform in-site review sections, social media product recommendation content review sections, and Q&A sections. These three channels have different user activity time distributions. Mixed calculations will result in the time-based density distribution being dominated by the channel with the largest number of reviews. The channel with the largest number of reviews is typically the e-commerce platform's in-site review section, where reviews are posted primarily in the evening after receiving the goods. Social media product recommendation content review sections are posted primarily at midday and late at night. After channel-specific calculations, each channel obtains an independent 24-hour time-based density distribution, an independent core ad placement time, and an independent time-based bid multiplier, which are combined with the channel's budget allocation ratio for the output. The channel classification is based on the unique review ID recorded during step S1's data collection of the source interface identifier.

[0079] The final output fields of the campaign execution parameter set include audience segment number, associated tag list, keyword list, negative keyword list, match method, bid coefficient, final bid coefficient, channel budget allocation ratio, core campaign period, secondary campaign period, low campaign period, time-based bid multiplier, and source channel identifier. The parameter set is written to the campaign interface directory as a structured file and is read and distributed by the campaign platform's batch upload interface. The parameter set is also written to the archive table for this round of execution. The archive table is indexed by the execution batch number, and the parameter set from the previous round is read in the next execution round for maintaining existing keywords and comparing traffic from test audience segments.

[0080] Steps S1 to S5 are executed cyclically within each campaign cycle, forming a closed loop of reputation data collection, category baseline stripping, inflection point candidate marking, rise phase element determination, and campaign execution parameter generation. New comments generated by the previous round of trial audience packages enter the multi-source comment collection scope of the next round of step S1.

[0081] Example 2: Scenarios where the cooling sensation of sun-protective clothing drives sales. This embodiment of the invention provides an AI-powered intelligent marketing promotion strategy generation method based on word-of-mouth trends.

[0082] The processing procedure is as follows: The target product is a women's sun-protective clothing item, belonging to the leaf category "sun-protective clothing." The attribute opinion tuple is taken as the combination of fabric and cooling sensation, denoted as 'e', ​​and the polarity label is positive. Step S1 involves multi-source review collection, pulling a set of 3,600 valid reviews for the target product, and pulling a set of 68,000 valid reviews for the top 50 products in the leaf category by sales volume in descending order over the past 30 days. For attribute opinion tuple extraction, if tuple 'e' matches 180 reviews in the target product review set, the "whether it is a unique element" field is set to "no".

[0083] Step S2 divides the time window into twelve observation windows, each seven days long, with a total backtracking period of eighty-four days. The current moment corresponds to observation window number twelve. The window mention intensity is calculated to obtain the target product mention intensity sequence and the category comparison mention intensity sequence for tuple e. The mean window mention intensity of tuple e during the backtracking period is 0.050, and the smoothing constant is taken as one percent of this mean, with a value of 0.0005. The net trend sequence is constructed by calculating eleven net trend values. The net trend values ​​corresponding to observation window numbers two through twelve are, respectively, -0.030, 0.012, -0.018, 0.035, 0.028, 0.072, 0.045, 0.128, 0.098, 0.205, and 0.248.

[0084] Step S3's multi-scale slope estimation yields a short-window slope of 0.0750, a medium-window slope of 0.0380, and a long-window slope of 0.0295. All three slopes are positive and form an increasing sequence. To normalize the self-dispersion amplitude, a moving least squares fit with a window length of three and a sliding step size of one is performed on the eleven net trend values, resulting in nine historical slope samples. The standard deviation of these nine historical slope samples is 0.01984, which represents the self-dispersion amplitude of the binary tuple e. Dividing each of the three slopes by its self-dispersion amplitude yields a short-window rise of 3.780, a medium-window rise of 1.915, and a long-window rise of 1.487. In the inflection point candidate label identity test, all three degrees of ascent are greater than zero, so this item passes. In the order of magnitude increase test, 3.780 is greater than the product of 1.915 and the tolerance coefficient 0.9, which is 1.7235, and 1.915 is greater than the product of 1.487 and the tolerance coefficient 0.9, which is 1.3383, so both conditions are met, and this item passes. In the sparse window test, the observation windows numbered 10, 11, and 12 covered by the short window are not marked as sparse windows, and the minimum total number of comments for the target product window across the three observation windows is 280, which is greater than 100, so this item passes. All three tests pass, and the tuple e is added to the candidate list.

[0085] Step S4, the rising curve fitting, first involves accumulating and summing the eleven net trend values ​​to obtain the cumulative net trend sequence. The cumulative net trend values ​​corresponding to observation window numbers two through twelve are, respectively, -0.030, -0.018, -0.036, -0.001, 0.027, 0.099, 0.144, 0.272, 0.370, 0.575, and 0.823. The minimum value shifting method shifts the minimum value of the cumulative net trend sequence from -0.036 to 0.0005, a shift of 0.0365. The value of the shifted cumulative net trend sequence at observation window number twelve is 0.8595. The logistic growth function is then fitted using nonlinear least squares to determine the saturation scale. K e The internal growth rate is 2.34%. eThe time t is 0.479, the midpoint of the curve. 0,e The value is 13.15. The saturation scale of 2.34 is greater than the maximum value of the cumulative net trend sequence of 0.8595. The intrinsic growth rate of 0.479 is greater than zero. The midpoint of the curve at 13.15 falls within the observation window interval from number 2 to 14. All three parameters satisfy the boundary constraints, and the fit converges within the upper limit of 500 iterations. The propagation stage is determined by the completion degree, which is the ratio of the cumulative net trend value of 0.8595 at the current time after translation to the saturation scale of 2.34. The value is 0.367, falling between 0.25 and 0.75. The propagation stage is determined to be the diffusion stage.

[0086] The remaining upward window is calculated using a saturation judgment ratio of 0.95. The logarithmic probability of 0.95 is 2.9444. The reciprocal of the intrinsic growth rate of 0.479 is 2.0877. The product of the two is 6.1471. Adding 6.1471 to the midpoint of the curve at 13:15 and then subtracting the observation window number 12 corresponding to the current time, we get the remaining upward window duration as 7.30 observation windows, which is equivalent to 7.30 multiplied by 7 days, equaling 51.1 days. In the factor confidence rating, the sum of squared residuals is 0.00244, the sample size is 11, the mean squared residual is 0.000222, the maximum value of the cumulative net trend sequence after shift is 0.8595, and the relative residual is 0.000258, which is less than 0.15. The first item scores 3 points; the total number of comments in the retrospective period is 180, which is greater than 50. The second item scores 3 points; the field for whether it is a unique factor is no. The third item scores 3 points; the total score is 9 points, and the confidence rating is A. In the decay factor exclusion, the fit has converged and the intrinsic growth rate is greater than zero, the saturation size is greater than the maximum value of the cumulative net trend sequence, the remaining rising window duration is 7.30, which is greater than zero, the confidence rating is A, the polarity label is positive, all four exclusion conditions are not met, and the binary tuple 'e' is added to the rising period factor list.

[0087] In step S5, during the element-audience mapping, the polarity label of binary tuple e is positive, so it is classified into the positive element group. The attribute word "fabric" retrieves twelve candidate tags in the first-level search of the category audience tag library, and the opinion word "coolness" hits four tags in the second-level search. The number of directly associated tags is four, no less than three, and no expanded recall is initiated. The propagation stage of binary tuple e is the diffusion stage, and the confidence level is A. The four directly associated tags form the main audience package. In the generation of keywords and negative words, both the phrase "fabric coolness" combined in the original text order and the phrase "coolness fabric" combined in reverse order are written into the keyword table. The matching method is set to phrase matching according to the main audience package, and the value of the "unique element" field is "no". No additional related words are added.

[0088] In the joint allocation of bids and budgets, the base bid coefficient β0 is input by the operations side based on the target product's gross profit margin and target input-output ratio, and is set to 2.00. The window duration scaling constant τ is set to four observation windows, and the remaining rising window duration T is... e The value is 7.30, with the first factor being 1 minus the natural constant raised to the power of negative 7.30, resulting in a value of 0.8386; the short window rise angle g. e The median of the short-window rise rate of all rising elements in this leaf category, i.e., the rise rate half-saturation constant κ, is 3.780. The second factor is 3.780 divided by the sum of 3.780 and 1.80, which is 0.6774. Calculated using the bidding coefficient, the bidding coefficient is 2.00 multiplied by 0.8386 and then multiplied by 0.6774, resulting in a value of 1.1363. The tiered adjustment multiple for the main user group package is 1.2, and the final bidding coefficient is 1.1363 multiplied by 1.2, resulting in a value of 1.36.

[0089] In this round of execution, the four-tiered audience packages are as follows: the main audience package contains three binary groups with a final bid coefficient sum of 2.35 and a size weight of 0.42; the follower audience package contains five binary groups with a final bid coefficient sum of 1.86 and a size weight of 0.31; the test audience package contains four binary groups with a final bid coefficient sum of 0.92 and a size weight of 0.18; and the maintenance audience package contains six binary groups with a final bid coefficient sum of 1.05 and a size weight of 0.09. The budget weights for the four tiers are 0.9870, 0.5766, 0.1656, and 0.0945, respectively, totaling 1.8237. After normalization, the channel budget allocation ratios are 54.1%, 31.6%, 9.1%, and 5.2%, respectively. All four tier percentages fall between the lower limit of 5% and the upper limit of 60%, thus not triggering constraints and requiring no re-normalization. When the total budget is 30,000 yuan, the allocation amounts for the four levels are 16,236 yuan, 9,485 yuan, 2,724 yuan and 1,555 yuan respectively.

[0090] In the distribution of ad placement times, the 24-hour density distribution of the comment section channel on the e-commerce platform reached its maximum at 10 PM and 9 PM, at 8.6% and 8.4% respectively, and its minimum at 4 AM and 3 AM, at 0.9% and 1.1% respectively. A cyclical shift correction moved the hourly density value to the position six hours prior to that hour; the value for 10 PM was moved to 4 PM, and the value for 9 PM was moved to 3 PM. After the shift, the peak values ​​appeared at 4 PM and 3 PM, at 8.6% and 8.4% respectively. The shifted time-period density distribution is arranged in descending order of value. The first eight hours are the core delivery periods, with values ​​of 5:00, 6:00, 7:00, 12:00, 13:00, 14:00, 15:00, and 16:00, and a time-period bid multiplier of 1.3. The ninth to sixteenth hours are the secondary delivery periods, with values ​​of 2:00, 3:00, 4:00, 8:00, 10:00, 11:00, 17:00, and 19:00, and a time-period bid multiplier of 1.0. The remaining eight hours are the low delivery periods, with values ​​of 1:00, 9:00, 18:00, 20:00, 21:00, 22:00, 23:00, and 0:00, and a time-period bid multiplier of 0.5.

[0091] This process illustrates that each step is sequentially connected under real data: the eleven values ​​of the net trend sequence are calculated from the net trend value from the two mention strength sequences; the short window rise is calculated by dividing the short window slope by its own dispersion; the short window rise is used both for the magnitude increase test in step S3 and as the bid coefficient g. e The remaining rising window duration is used both for the decay factor exclusion in step S4 and as the T value of the bid coefficient. e The bid coefficient calculation, dissemination stage, and confidence level jointly determine the audience segmentation, which in turn determines the segmentation adjustment multiplier and matching method. Data is passed along the binary number 'e' through these five steps while maintaining consistent values. Peak usage of sun-protective clothing occurs during daytime outdoor hours, while review writing peaks in the evening after receipt. Cyclic shifting and correction restores the ad placement period from the evening review writing peak to the daytime product usage peak, with the core ad placement period falling between 12:00 and 16:00, consistent with the actual usage scenario of sun-protective clothing.

[0092] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for generating AI-powered intelligent marketing and promotion strategies based on word-of-mouth trends, characterized in that, The method includes: S1: Obtain the target product review set and the category-matching product review set through multi-source review collection. Then, extract the attribute opinion pairs to parse each review text into attribute opinion pairs. Finally, attach a polarity label to each attribute opinion pair through sentiment polarity labeling to obtain a set of attribute opinion pairs with polarity labels and timestamps. S2: Taking the set of attribute viewpoint pairs obtained in step S1 as input, the set of attribute viewpoint pairs is first divided into observation window sequences by time window partitioning. Then, the target product mention intensity sequence and the category comparison mention intensity sequence are calculated by window mention intensity. The sparse window and unique elements are aligned by the comparison sequence. Finally, the net trend sequence is constructed by subtracting the logarithmic change of the category comparison mention intensity sequence from the logarithmic change of the target product mention intensity sequence. S3: Using the net trend sequence obtained in step S2 as input, first perform least squares line fitting on the short window, medium window and long window scales through multi-scale slope estimation to obtain the slope of the short window, the slope of the medium window and the slope of the long window. Then, through the normalization of its own fluctuation amplitude, use the standard deviation of the historical slope sample of the attribute viewpoint binary as the normalization denominator to obtain the rise degree of the short window, the rise degree of the medium window and the rise degree of the long window. Finally, through the inflection point candidate label, perform the same sign test, the magnitude increase test and the sparse window test on the three rise degrees in turn. The attribute viewpoint binary that passes all three tests is written into the candidate list, and the attribute viewpoint binary that fails any one test is transferred to the regular word-of-mouth element pool. S4: Using the candidate list obtained in step S3 as input, first, by fitting the rising curve, the cumulative net trend sequence is fitted with the logistic growth function to obtain the saturation scale, intrinsic growth rate and the midpoint of the curve. Then, the propagation stage is determined by the propagation stage, and the remaining rising window duration is calculated by the remaining rising window. Subsequently, the confidence level is obtained by the confidence level of the elements. Finally, the unqualified inflection point candidate elements are eliminated by the decaying elements, and the rising period element list is output. S5: Using the list of rising elements obtained in step S4 as input, first obtain four layers of audience packages—test audience package, main audience package, follower audience package, and maintenance audience package—through element audience mapping. Then, generate the target keywords and negative keywords through keywords and negative keywords. Through joint allocation of bid and budget, calculate the bid coefficient and channel budget allocation ratio by combining the remaining rising window duration and the short window rising degree. Finally, obtain the time-based bid multiplier through the placement of placement time periods and jointly output the placement execution parameter set.

2. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The multi-source review collection includes two sets of reviews: the first set is the target product review set, which retrieves all reviews based on the target product's identifier; the second set is the category-matching product review set, which retrieves the top 50 product identifiers from the leaf category to which the target product belongs, sorted by sales volume in descending order over the past 30 days, and then merges the reviews for each. Both sets of reviews include the review text, user rating, review publication time, follow-up review text, follow-up review publication time, and the leaf category number to which the product belongs. The multi-source review collection interface is categorized into three types based on the source: e-commerce platform in-site review area interface, social media platform product recommendation content review area interface, and Q&A area interface. The source interface identifier is recorded along with the unique review number for use in step S5, where the placement time is calculated separately for each channel. The two sets are distinguished by a product source marker, which is either the target product or the category-matching product.

3. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The specific steps for extracting the attribute viewpoint tuples are as follows: S101, Sentence and Word Segmentation: Segment and word segment the comment text, remove stop words, emoticons and pure punctuation segments, and then concatenate the appended comment text with the main comment text as a complete comment for processing; S102. Paired Structure Location: Dependency parsing is used to locate the paired structures of attribute words and opinion words in the sentence segmentation results. The paired structures include two types: the collocation relationship between the attributive and the headword, and the collocation relationship between the subject and the predicate. Attribute words are taken from a pre-constructed category attribute word list. Opinion words are directly extracted by dependency parsing from the modification position or predicate position of the attribute words and the original form of the comment is preserved without synonym merging. S103. Generation of binary tuple numbers: After concatenating the strings of attribute words and opinion words, the hash value is taken to generate a binary tuple number. The binary tuple number is the primary key concatenated between steps S1 and S5. All attribute opinion binary tuples generated from the same comment share the user rating and comment publication time of the comment.

4. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The window mention intensity calculation is performed within each observation window by counting the ratio of the number of comments containing a certain attribute opinion binary to the total number of comments within that observation window. The resulting ratio is recorded as the mention intensity. The mention intensity is calculated separately according to the product source tag. Attribute opinion binary with neutral polarity tags are not included in the mention intensity calculation. Attribute opinion binary with positive polarity tags and attribute opinion binary with negative polarity tags are used to construct independent target product mention intensity sequences and category-matching mention intensity sequences, respectively.

5. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 4, characterized in that, The identification result verification performs three consistency checks on the decomposition results of each control cycle: the first is the phase angle range check, which requires the leading phase angle to be within the range of greater than 5° and less than 70°; the second is the numerical continuity check, which requires the absolute value of the difference between the equivalent parallel resistance of the load in this cycle and the equivalent parallel resistance of the load that passed the check in the previous cycle to not exceed 50% of the value that passed the check in the previous cycle; the third is the positive susceptance check, which requires the susceptance of the coil branch to be positive. When all three checks pass, the equivalent parallel resistance of the load, the equivalent inductance of the coil, and the resonant frequency of the load in this cycle are written into the identification result register. If any one check fails, the result of this cycle is discarded, the identification result register retains the value that passed the check in the previous cycle and increments the check failure count by one. When the check failure count reaches 5 consecutive times, an identification anomaly flag is output.

6. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The multi-scale slope estimation uses a short window covering the three most recent observation windows, a medium window covering the six most recent observation windows, and a long window covering the nine most recent observation windows. All three windows share the same right endpoint, which is the observation window at the current time. The fitting process uses the observation window number as the independent variable and the net trend value as the dependent variable. The self-variance normalization first performs a sliding least squares fit on the net trend sequence with a window length of three and a sliding step size of one, obtaining nine historical slope samples. The standard deviation of these nine historical slope samples is taken as the self-dispersion amplitude of the attribute viewpoint binary. Then, the short-window slope, medium-window slope, and long-window slope are each divided by their self-dispersion amplitude. The three tests for the inflection point candidate markers are performed in ascending order of computational cost; if the previous test fails, the next test is not performed. The criteria for the same sign test are that the rise degree of the short window is greater than zero, the rise degree of the medium window is greater than zero, and the rise degree of the long window is greater than zero, all three conditions must be met simultaneously. The criteria for the order of magnitude increase test are that the rise degree of the short window is greater than the product of the rise degree of the medium window and the tolerance coefficient, and the rise degree of the medium window is greater than the product of the rise degree of the long window and the tolerance coefficient, both of which must be met simultaneously. The tolerance coefficient is set to 0.

9. The criterion for the sparse window test is that the number of sparse windows marked in the three observation windows covered by the short window is zero, and the minimum number of comments in the target product window across the three observation windows covered by the short window is greater than one hundred.

7. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The propagation stage determination is calculated by the ratio of the current cumulative net trend value to the saturation scale and recorded as the completion degree. When the completion degree is less than 0.25, the propagation stage is determined to be the nascent stage; when the completion degree is greater than or equal to 0.25 and less than or equal to 0.75, the propagation stage is determined to be the diffusion stage; when the completion degree is greater than 0.75, the propagation stage is determined to be the near-saturation stage. The remaining rising window duration is calculated by taking the midpoint of the curve, the intrinsic growth rate, and the saturation determination ratio. The calculation method is to add the midpoint of the curve to the reciprocal of the intrinsic growth rate and the logarithmic probability of the saturation determination ratio, and then subtract the observation window number corresponding to the current moment. The saturation determination ratio is set to 0.

95. The unit of the remaining rising window duration is the number of observation windows, which is multiplied by the window length when converted to days.

8. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The confidence level grading of each inflection point candidate element scores it using three indicators, and the scores are summed to obtain the total score. The first indicator is the relative residual, which is calculated by dividing the sum of squared residuals by the number of samples in the cumulative net trend sequence to obtain the mean squared residual, and then dividing the mean squared residual by the maximum value of the cumulative net trend sequence to obtain the relative residual. A relative residual less than 0.15 is scored as 3 points, greater than or equal to 0.15 and less than or equal to 0.3 is scored as 2 points, and greater than 0.3 is scored as 1 point. The second indicator is the total number of comments during the retrospective period, which is the number of comments containing the opinion binary of this attribute for the target product during the retrospective period. If the number of comments is greater than five, the total number of comments is calculated as follows: Ten points are awarded as three points, twenty or more points and fifty or less points are awarded as two points, and less than twenty points are awarded as one point. The third indicator is whether it is a unique element. If the value of the "unique element" field is "no", it is awarded as three points, and if the value is "yes", it is awarded as two points. A total score of nine points is Grade A, a total score of seven or eight points is Grade B, a total score of five or six points is Grade C, and a total score of less than five points is Grade D. The attenuation element excludes four types of inflection point candidate elements: the first type is fitting failure, the second type is the remaining rising window duration is less than zero, the third type is the confidence level is Grade D, and the fourth type is the polarity label is negative and the completion degree is greater than 0.

75.

9. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The element-based audience mapping first divides the rising element list into positive and negative element groups based on polarity tags. For each attribute-opinion binary in the positive element group, the attribute words and opinion words are retrieved from the category-based audience tag library of the advertising platform and subjected to a two-level search. The first-level search recalls candidate tags based on attribute words, and the second-level search filters within the candidate tag range based on opinion words. Tags matched by both levels of search are considered directly associated tags for that binary. If the number of directly associated tags is less than three, an expanded recall is initiated, which includes the hypernyms and hyponyms of the attribute words in the search scope. Tags matched by the expanded recall are considered indirectly associated tags. The audience package is stratified according to... The propagation stage is divided into two dimensions: the propagation phase and the confidence level. The direct and indirect related tags of the binary groups in the budding stage form the test audience package. The tags of the binary groups in the diffusion stage with a confidence level of A form the main audience package. The tags of the binary groups in the diffusion stage with a confidence level of B or C form the follower audience package. The tags of the binary groups in the near-saturation stage form the maintenance audience package. The four audience packages are deduplicated according to the hierarchical order of the main audience package, follower audience package, test audience package and maintenance audience package. When the same tag appears in multiple layers, the highest level layer is retained.

10. The method for generating AI-powered intelligent marketing promotion strategies based on word-of-mouth trends according to claim 1, characterized in that, The generation of attribute-opinion binary pairs for positive element groups using keywords and negative words is performed by combining attribute words and opinion words in their original text order into a phrase, and then combining them again in reverse order with opinion words first and attribute words last. Both phrases are written into the keyword table. The matching method is set according to the audience package to which the binary pair belongs. Broad match is used for keywords corresponding to the trial audience package, phrase match is used for keywords corresponding to the main audience package and the follower audience package, and exact match is used for keywords corresponding to the maintenance audience package. For binary pairs with a unique element field value of "yes", additional related word expansion is performed. For attribute-opinion binary pairs for negative element groups, negative keywords are combined in their original text order with attribute words and opinion words and written into the negative word table. Binary pairs with a confidence level of A are written into exact negative, and binary pairs with a confidence level of B or C are written into phrase negative. The negative word table is attached to the main audience package and the follower audience package, but not to the trial audience package. The binary pairs in the regular word-of-mouth element pool are combined with attribute words and opinion words and then written into the stock keyword table. The matching method and bidding of the stock keyword table remain unchanged from the previous round of campaigning.