Type price-based hot competitive product mining method for AI intelligent marketing

Abstract

The invention discloses a category price-based hot competitive product mining method for AI intelligent marketing, and relates to the technical field of marketing, and the method comprises the following steps: building a unified time base line, extracting a continuous sales volume curve of commodities under a target category, calculating the instantaneous slope of sales volume change in the sales volume curve, and calculating the instantaneous slope of the sales volume change; based on the sudden change interval of the instantaneous slope, determining a candidate region of short-period sales volume increase; and constructing a causal factor set in the candidate region, and mapping the channel expansion track and the promotion track to a sales volume curve point by point. According to the method, through the joint constraint of the causal residual and the user behavior density tensor, the sales volume anomaly weight index is generated and written back to the competitive product portrait, false sales volume signals are effectively filtered, and the accuracy of competitive product identification and marketing decision is improved. And meanwhile, short-time Fourier transform is introduced to decompose a sales volume curve, high-frequency abnormity is dynamically weakened by using a spectral domain attenuation coefficient, and a hot sell weight is ensured to truly reflect a long-term and periodic trend, so that the reliability and the analysis precision of a competitive product portrait are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marketing, in particular to an AI intelligent marketing hot-selling competitive product mining method based on category price. BACKGROUND

[0002] In AI intelligent marketing, hot-selling competitive product mining based on category price refers to using big data and intelligent algorithms to deeply analyze the market sales of different price intervals under a specific category, so as to identify competitive products with high sales, fast growth, good user evaluation and potential market threat in the same sub-interval. This process is not just a simple screening of the sales ranking, but a combination of category attributes, price band distribution, user preference characteristics and dynamic market trends to build a competitive product portrait and mine the most representative contrast object of the target user's real purchase selection. Through this way, enterprises can accurately lock in direct competitors, discover potential substitutes, analyze their marketing strategies and product advantages, and provide data support for their own pricing, promotion, product selection and differentiation positioning, thereby improving the pertinence and foresight of marketing decisions.

[0003] The prior art has the following disadvantages: In the prior art, some products may show abnormal sales growth in a short period due to single behavior or sudden channel tilt, but such abnormal patterns are often not effectively identified and filtered, resulting in a false hot-selling weight amplification effect in the competitive product portrait construction process. Once such virtual expansion is used as the basis for subsequent analysis and strategy development, it will cause deviation in competitive product identification, misjudging objects that are not truly market leaders as core competitors, and then directly causing errors in pricing strategy adjustment, leading to resource imbalance and market response delay and other serious consequences.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an AI intelligent marketing hot-selling competitive product mining method based on category price to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical solution: an AI intelligent marketing hot-selling competitive product mining method based on category price, comprising the following steps: A unified time baseline is established and the continuous sales curve of the goods under the target category is extracted, the instantaneous slope of the sales change in the sales curve is calculated, and the candidate area of short-period sales growth is determined based on the mutation interval of the instantaneous slope; A set of causal factors is constructed within the candidate region, and the channel expansion trajectory and promotion trajectory are mapped point by point to the sales curve. The causal residuals are calculated based on the mapping results to characterize the degree of deviation between sales growth and causal factors. Based on the causal residual, a cross-channel consistency spectrum is generated, and the phase difference and synchronicity indicators of sales growth in each channel are projected onto the causal residual space. The projection results are then combined with the causal residual to form an abnormal signal. Under the constraint of abnormal signals, a user behavior density tensor is constructed to model the density of the purchase time interval and geographical distribution of target users. The density peak is compared with the abnormal signals one by one to enhance the reliability of sales anomaly identification. Under the joint constraints of user behavior density tensor and causal residual, an abnormal sales weight index is generated. Based on the abnormal sales weight index, false sales growth and normal sales trends are dynamically separated. The abnormal sales weight index is then written back to the competitor profile to complete the correction of the best-selling weight. Driven by the abnormal sales weight index, the sales curve is transformed into a time spectrum signal. The sales changes are decomposed into low-frequency trends, mid-frequency fluctuations and high-frequency anomalies through short-time Fourier transform. The low-frequency trends and mid-frequency fluctuations are included in the hot-selling weight, and the high-frequency anomalies are dynamically weakened through the spectral domain attenuation coefficient to reduce the interference of false sales bursts on the competitor profile.

[0007] Preferably, the identification of short-term sales growth candidate regions based on sales data of products within the target category includes the following steps: Establish a unified time baseline and extract continuous sales curves on this time baseline. Perform time mapping, interpolation, and moving average processing on sales data from different sources to construct complete and continuous sales curves. The instantaneous slope of sales change is calculated on the continuous sales curve. The instantaneous slope curve is obtained through difference operation. The sales curve is smoothed before calculation to reduce noise interference, thereby identifying the abrupt change characteristics of the slope curve. Candidate regions are determined based on the abrupt change points of the instantaneous slope curve. Abrupt change points are marked by threshold detection, and the time range before and after the abrupt change points is expanded to cover the entire growth process. If the interval between adjacent abrupt change points is insufficient, they are merged according to the set threshold. The starting sales, peak sales, ending sales and duration of the region are recorded.

[0008] Preferably, constructing a set of causal factors and calculating causal residuals within the candidate region includes the following steps: External influencing factors related to sales changes are collected within the candidate region, and a set of causal factors is constructed, including channel expansion trajectory and promotion trajectory. The causal factors are then standardized and time-aligned to form a time series on a unified time baseline. The causal factor is mapped to the sales curve in the candidate area point by point, the value of the causal factor is extracted at each time point and paired with the sales value, and the weight is set according to the historical contribution, so as to establish the corresponding relationship between the causal factor and the sales change; The causal residual is calculated based on the corresponding relationship, the sales are predicted by the causal factor, and the residual is obtained by comparing the actual sales, and the residual is superimposed with the sales slope mutation interval in the time dimension to identify the abnormal growth area that cannot be explained by the causal factor.

[0009] Preferably, the cross-channel consistency spectrum based on the causal residual comprises the following steps: The sales curve of each channel in the candidate area is extracted, and the local growth rate is calculated by a sliding window, the growth rate is converted into a phase signal, and the phase difference sequence is obtained by taking the reference curve as a reference; The synchronization index between channels is calculated according to the phase difference sequence, the cross-channel consistency matrix is obtained through correlation analysis and coherence function, and the consistency matrix is matched with the causal residual point by point in the time dimension to generate a consistency spectrum; The consistency spectrum is normalized and superimposed with the causal residual curve in the time dimension to form a joint abnormal signal, and the signal amplitude is enhanced when the causal residual deviates positively and the consistency is low, so as to mark the abnormal growth area.

[0010] Preferably, when the consistency spectrum is normalized, the numerical interval of the consistency spectrum is adjusted to be consistent with the numerical interval of the causal residual, and a weighting coefficient is set in the superposition process, the causal residual is given a high weight, and the consistency spectrum is given a low weight, so as to enhance the identification accuracy of the abnormal signal.

[0011] Preferably, the user behavior density tensor is constructed under the constraint of the abnormal signal, comprising the following steps: The user order time is extracted from the transaction data of the candidate area under the constraint of the abnormal signal, the time interval of continuous orders is calculated, and the time interval density distribution is formed by kernel density estimation; The geographic location information corresponding to the transaction data is extracted, the order location is mapped to a two-dimensional space, and the geographic density distribution is obtained by density clustering; The time interval density distribution and the geographic density distribution are combined to construct the user behavior density tensor, and the density peak value in the tensor space is compared with the abnormal signal one by one, so as to enhance the reliability of the sales abnormality identification.

[0012] Preferably, the sales abnormality weight index is generated under the joint constraint of the user behavior density tensor and the causal residual, comprising the following steps: Based on the user behavior density tensor and the causal residual, a feature vector is extracted, a weight factor set is constructed, and it is projected to a unified numerical space by feature standardization; The sales anomaly weight index is generated according to the weight factor set, the abnormal signal is amplified through weighted superposition and an exponential function, and the output result is normalized to limit the value range; The sales anomaly weight index is used for dynamic peeling of the candidate region sales curve, the abnormal growth part is weakened through a decay coefficient, and the normal trend part is retained; The sales anomaly weight index is written back to the competitor portrait, and the original hot sales weight is corrected to avoid deviation in competitor identification caused by false sales expansion.

[0013] Preferably, the steps of converting the sales curve into a time spectrum signal under the driving of the sales anomaly weight index include the following steps: The candidate region sales curve is adjusted under the driving of the sales anomaly weight index, and is divided into adjacent time windows to form a frequency spectrum analysis input signal; The short-time Fourier transform is applied to the time window to obtain the frequency spectrum distribution of low-frequency trend, medium-frequency fluctuation and high-frequency anomaly; The low-frequency trend and the medium-frequency fluctuation are directly included in the hot sales weight, and the amplitude attenuation is used in the high-frequency anomaly range to weaken the abnormal signal; The low-frequency trend and the medium-frequency fluctuation are combined and updated to the hot sales weight of the competitor portrait, and the attenuated high-frequency anomaly is written back to the competitor portrait as a monitoring index.

[0014] In the above technical solution, the technical effects and advantages provided by the present application are as follows: The present application can accurately identify abnormal sales growth caused by brushing or channel tilt in a short period by constructing a joint constraint mechanism based on causal residual and user behavior density tensor, and effectively filter false sales signals. By dynamically calculating the sales anomaly weight index and writing it back to the competitor portrait, the amplification effect of false sales on the hot sales weight is effectively avoided, thereby significantly improving the accuracy of competitor identification and the reliability of marketing decisions, and solving the problem that the competitor portrait is easily distorted by abnormal data in the prior art.

[0015] The present application decomposes the sales curve into three types of signals, i.e. low-frequency trend, medium-frequency fluctuation and high-frequency anomaly, by introducing short-time Fourier transform, and dynamically weakens the high-frequency anomaly by combining the spectral domain decay coefficient, effectively reducing the interference of abnormal sales in the frequency domain. This method not only enhances the identification ability of the real market trend, but also makes the hot sales weight more truly reflect the competitiveness of the goods in the long-term and periodic dimensions, and comprehensively improves the adaptability and analysis accuracy of the competitor portrait in the multi-frequency behavior mode. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description only represent some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0017] Figure 1 The method flowchart of the AI intelligent marketing based on category price hot-selling competitive product mining method of the present application. DETAILED DESCRIPTION

[0018] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the gist of the example implementations to those skilled in the art.

[0019] The present application provides an AI intelligent marketing based on category price hot-selling competitive product mining method as shown in Figure 1 The method flowchart of the AI intelligent marketing based on category price hot-selling competitive product mining method of the present application. S101, establishing a unified time baseline and extracting the continuous sales curve of the goods under the target category, calculating the instantaneous slope of the sales change in the sales curve, and determining the candidate area of short-period sales growth based on the mutation interval of the instantaneous slope; The candidate area identification of short-period sales growth for the sales data of the goods under the target category includes the following steps: First, a unified time baseline is established and a continuous sales curve is extracted on this baseline. To this end, the sales data from different sources is first imported into the same time sequence structure, all sales data is uniformly mapped to a continuous time axis in units of minutes, hours or days, and the time stamps of different source data are format-converted to have uniform time accuracy. Subsequently, for the missing sales data points on the time axis, piecewise linear interpolation is used to fill in the missing sales data points, so that each time point on the time axis has a corresponding sales value. For some channel data that is too densely sampled, the sliding window average method is used for down-sampling to make it consistent with the time interval of the unified time baseline. Next, the sales data of all channels are added up or kept independent by channel to construct a continuous sales curve under the unified time baseline. The sales curve obtained in this way can completely cover the sales change trajectory of the target category goods in a given time period, and there is no breakpoint, jump point or abnormal fluctuation caused by inconsistent sampling granularity. Through the continuous sales curve, it can be ensured that the subsequent steps are performed under a unified time reference framework, avoiding the situation that the results are not comparable due to differences in data sources.

[0020] After obtaining the continuous sales curve, the instantaneous slope of sales changes is calculated to capture the dynamic growth characteristics of sales over time. To do this, the continuous sales curve is differenced between two adjacent time points to obtain the change in sales, which is then divided by the time interval between the two adjacent time points to obtain the instantaneous slope for that interval. To improve the stability and noise resistance of the calculation, the continuous sales curve can be first Gaussian smoothed to reduce outliers in the instantaneous slope caused by random fluctuations. During the calculation, the instantaneous slope at each time point is saved as a new time series, forming a slope curve that corresponds one-to-one with the sales curve. When the sales curve experiences a sharp rise within a short period, the corresponding instantaneous slope value will show a significant positive peak, and the difference between the slope and the adjacent time point will widen significantly. Mathematically, this phenomenon is represented by a sudden change in the slope curve, indicating an abnormally rapid increase in the sales growth rate. By calculating the instantaneous slope point by point and constructing slope curves, key moments in the sales curve that may represent abnormal growth can be clearly identified.

[0021] After obtaining the instantaneous slope curve, candidate regions for short-cycle sales growth are constructed based on abrupt change points. Specifically, the instantaneous slope curve is first scanned to detect points exceeding a preset threshold. This threshold can be calculated based on the long-term average slope and standard deviation of historical sales curves, for example, set as the average plus three times the standard deviation. When the instantaneous slope at a certain time point is greater than this threshold, that time point is marked as a candidate abrupt change point. Next, the candidate abrupt change points are extended forward and backward by a certain time range, for example, by 2 to 4 sampling intervals each, to cover the initial and declining phases of sales growth. If the interval between adjacent abrupt change points is less than a set time threshold, these abrupt change points are merged into a continuous candidate region to avoid incorrectly splitting the same growth process into multiple regions. For each candidate region, the initial sales, peak sales, and ending sales of the sales curve within that region, as well as the corresponding duration and average slope, need to be recorded simultaneously for use in subsequent causal factor matching and anomaly identification. The resulting candidate regions are not just single instantaneous slope abrupt changes, but cover the entire process of sales from a stable state to a sharp increase and then back to stability. This ensures that subsequent steps can be analyzed based on the complete growth trajectory without missing important contextual information.

[0022] By implementing the above steps, a unified time baseline is first established to ensure the comparability and continuity of data over time. Secondly, the dynamic characteristics of sales change rates are captured through point-by-point calculation of instantaneous slopes. Finally, continuous candidate regions are formed through abrupt change point expansion and regional aggregation. This process enables the accurate identification of short-cycle abnormal growth in the sales curve. This method not only avoids misjudgments of false signals caused by short-term promotions or fraudulent orders in existing technologies, but also ensures the integrity and robustness of candidate region identification.

[0023] S102, constructing a causal factor set in the candidate region, mapping the channel expansion trajectory and the promotion trajectory to the sales curve point by point, and calculating a causal residual according to the mapping result to represent the deviation degree between the sales growth and the causal factor; In order to make a causal explanation of the authenticity of the sales growth in the candidate region, a causal factor set is constructed in the candidate region, and a causal residual is calculated by mapping the channel expansion trajectory and the promotion trajectory to the sales curve point by point, so as to represent the deviation degree between the sales growth and the causal factor. The process includes the following steps: In the candidate region that has been determined, external influencing factors related to sales changes are collected, and a causal factor set is constructed. The content of the causal factor set mainly includes two types of data, channel expansion trajectory and promotion trajectory. The channel expansion trajectory refers to the change of the target product in the sales channel within the time range of the candidate region, such as whether a new online e-commerce channel is added, whether a new distribution point is opened, or whether the exposure position of platform recommendation is increased. The promotion trajectory refers to various promotion activities implemented for the target product within the time range of the candidate region, such as full reduction, discount, coupon issuance, time-limited flash sale, live broadcast promotion, and cross-category joint promotion. In order to make the causal factor set fully cover the external variables related to sales, standardization and time alignment operations need to be performed on different data sources, so that all causal factors can be expressed on the same time baseline as the candidate region. Each causal factor is represented as a time series, where each time point corresponds to a numerical value, which reflects the intensity of the factor at that time. For example, whether the channel expansion trajectory adds a distribution node at a certain time can be represented by a binary number, and the promotion intensity can be represented by the discount ratio or the promotion budget size.

[0024] The above-mentioned causal factors are mapped to the sales curve in the candidate region point by point to establish the corresponding relationship between the causal factors and the sales change. Specifically, at each time point in the candidate region, the values of all causal factors corresponding to the time point are extracted and paired with the sales value at the time point to form a causal factor and sales mapping at the time point level. In order to ensure the accuracy of the mapping, the importance of different causal factors can be considered by weighting, for example, different weights are given to the channel expansion factor and the promotion factor, and the weight value can be determined according to the statistical analysis result of the contribution of the causal factor to the sales in the historical data. Through this point-by-point mapping operation, it can be determined whether the fluctuation on the sales curve corresponds to a certain channel expansion event or promotion activity in each time segment of the candidate region. If the sales change and the causal factor change appear synchronously, it means that the sales growth has the possibility of causal explanation; if the sales change appears suddenly without the change of the causal factor, it means that there is an additional abnormal factor.

[0025] After mapping causal factors to the sales curve point by point, it is necessary to calculate the causal residual based on the mapping results to measure the degree of deviation between sales growth and the causal factors. Specifically, a sales prediction function based on causal factors is established. This function uses the values ​​of channel expansion trajectory and promotion trajectory to predict the theoretically expected sales increase at each time point, and then compares the predicted sales value with the actual sales value. The difference between the two is the causal residual, and the magnitude of the residual directly reflects the part of sales change that cannot be explained by known causal factors. When the residual shows a significant positive deviation at consecutive time points, it means that the sales growth is much higher than the level that promotional activities and channel expansion can explain, which is very likely to be due to interference from fraudulent order behavior or abnormal channel traffic tilt. When the residual shows a negative deviation at multiple time points, it may indicate that the promotion or channel expansion has not achieved the expected effect, and the sales have not grown as theoretically predicted. By calculating the causal residual, we can not only identify which parts of the sales curve belong to the explainable normal growth and which parts belong to the unexplained abnormal growth, but also provide a basis for subsequent amplification of abnormal signals and multi-dimensional cross-validation.

[0026] Causal residuals are used as a preliminary measure of anomaly signals in candidate regions, and then combined with sales curves for enhanced analysis. This process involves comparing the causal residuals and sales curves point-by-point over time, superimposing areas of sudden increases in the residuals with instantaneous slope abrupt changes in the sales curve. If the causal residuals deviate continuously and significantly within the sales slope abrupt change range, the candidate region can be marked as an anomalous growth region, indicating that the sales growth cannot be explained by normal channels or promotions, exhibiting obvious characteristics of spurious inflation. Simultaneously, statistical characteristics of the residuals, such as mean, variance, kurtosis, and skewness, can be calculated for each candidate region to further quantify the strength of the anomaly signal. These statistical characteristics will continue to be used in subsequent cross-channel consistency spectrum construction and user behavior density tensor modeling, forming a progressively validating chain of anomaly detection. In this way, causal residuals are not merely a single measure of deviation, but become a core indicator for analyzing the authenticity of sales within candidate regions.

[0027] In summary, through the above steps, a complete set of causal factors within the candidate region is constructed, and the channel expansion trajectory and promotion trajectory can be mapped point by point to the sales curve. Furthermore, quantitative characterization of the abnormal portion of sales growth is achieved through causal residual calculation. This implementation method, through residual analysis, establishes a mechanism for distinguishing between the explainable and non-explainable portions of sales growth, laying a data and methodological foundation for subsequent amplification of abnormal signals, cross-channel consistency analysis, and user behavior verification.

[0028] S103, generating a cross-channel consistency spectrum based on the causal residual, projecting the phase difference and synchronism indicators of the sales growth of each channel to the causal residual space, and superimposing the projection result and the causal residual to form an abnormal signal; In order to more accurately identify the abnormal part in the sales growth that cannot be explained by the causal factors, cross-channel consistency analysis is introduced on the basis of the causal residual. By generating a cross-channel consistency spectrum, the phase difference and synchronism indicators of the sales growth of each channel are projected to the causal residual space and superimposed with the causal residual to form a clear abnormal signal. The process includes the following steps: The sales curves of each sales channel in the candidate area are independently extracted, and the phase characteristics of the growth trajectory are calculated so as to be combined with the causal residual. In the specific implementation process, for the sales curve of each channel, the local growth rate curve is first calculated by means of sliding window, and then the growth rate curve is converted into a phase signal in the form of time series. The phase signal refers to the offset of the sales growth fluctuation of a channel in the time axis relative to the reference time point in the sales growth process. In order to ensure the comparability of the phase characteristics of each channel, a unified reference curve needs to be selected as the reference, for example, the weighted average curve of all channel sales or the channel curve most representative in the long-term historical data can be selected as the reference. By calculating the offset in the time dimension between each channel curve and the reference curve, a phase difference sequence can be obtained. The phase difference sequence can reveal the leading or lagging relationship of the sales growth of different channels in time, and is an important basis for judging whether there is synchronization between channels.

[0029] After obtaining the phase difference sequence of each channel, the synchronization indicators between channels are further calculated, and these indicators are associated with the causal residual. Specifically, first, the correlation coefficient between the sales curves of different channels is calculated by the method of cross-correlation analysis, and the correlation intensity in different time periods within the candidate region is extracted. The higher the correlation intensity, the more synchronized the sales changes of different channels in that time period; the lower the correlation intensity, the more obvious the difference in the sales growth performance of different channels. In order to avoid the limitations of a single correlation indicator, the coherence function can also be introduced to calculate the synchronization in the frequency domain of the sales curves of different channels and obtain the synchronization spectrum distribution across channels. Then, these synchronization indicators are combined with the phase difference sequence obtained in the previous step to form a cross-channel consistency matrix. Each element of the consistency matrix represents the phase difference and synchronization level between a pair of channels at a certain time point. Next, the consistency matrix is projected into the causal residual space, i.e. each element in the matrix is matched with the causal residual at the same time point, and a consistency spectrum is generated by vectorization and superposition. In this process, if the causal residual value is large and the cross-channel consistency is low, it means that the sales growth lacks consistent support across multiple channels and is likely to be caused by abnormal behavior of a single channel; on the contrary, if the causal residual value is small and the cross-channel consistency is high, it means that the sales growth has multi-channel support and belongs to a more normal market phenomenon.

[0030] After completing the generation of the cross-channel consistency spectrum, it is combined with the causal residual to form a clearer abnormal signal. In the specific implementation process, first, the consistency spectrum is normalized to make its numerical interval consistent with that of the causal residual, so as to avoid the bias caused by different dimensions on the superposition result. Then, the normalized cross-channel consistency spectrum and the causal residual curve are superimposed point by point in the time dimension to obtain a joint abnormal signal curve. On this curve, when the causal residual shows a persistent positive deviation and the consistency spectrum shows that the channels are out of sync or the phase difference is too large, the amplitude of the joint signal will be significantly enhanced, forming a prominent abnormal peak. These abnormal peaks clearly mark the unreasonable part of the sales growth in the candidate region. Even if the growth shows explosive growth in a single channel, it will still be identified as an anomaly in the joint signal due to the lack of cross-channel synchronization support. In order to further improve the robustness of the abnormal signal, a weighting coefficient can be introduced in the superposition process, such as assigning a higher weight to the causal residual and a relatively lower weight to the consistency spectrum, to emphasize the impact of insufficient causal explanation. The final joint abnormal signal not only retains the abnormal explanation ability of the causal residual, but also introduces cross-channel consistency as an auxiliary verification means, thereby significantly reducing the risk of misjudgment caused by accidental promotion or local traffic tilt.

[0031] Through the layer-by-layer advancement of the above steps, firstly, the phase characteristics are extracted from the channel sales curve to ensure that the differences in time between different channels can be quantified; secondly, the synchronicity index is calculated and projected into the causal residual space to realize the organic combination of causal explanation and cross-channel consistency; finally, the abnormal signal is generated by joint superposition, so that the identification of abnormal sales growth is more intuitive and accurate. This embodiment combines causal residuals and cross-channel consistency analysis for the first time, considering whether the sales growth can be explained by external factors and the coordination of different channels in the growth process, thereby significantly improving the reliability and accuracy of abnormal identification under double constraints. This method can not only deal with false brushing behavior in a single channel, but also identify false outbreaks caused by channel traffic tilt.

[0032] In S104, under the constraint of the abnormal signal, a user behavior density tensor is constructed to model the density of the purchase time interval and the geographical distribution of the target user, and the density peaks are compared with the abnormal signals one by one to enhance the reliability of abnormal sales identification. In order to further improve the reliability of abnormal sales identification, under the constraint of the generated abnormal signal, the user behavior dimension is introduced for cross verification. By constructing a user behavior density tensor, the behavior characteristics of the target user in the purchase time interval and the geographical distribution are converted into a density model, and the density peaks are compared with the abnormal signals one by one, so as to realize multi-dimensional abnormal identification. The process includes the following steps: Under the constraint of the obtained abnormal signal, the purchase behavior information of the target user is extracted from the transaction data in the candidate area, and the density distribution of the purchase time interval is constructed based on this. Specifically, for all valid orders in the candidate area, the unique identifier of the user and the corresponding order time are extracted one by one, and then the time interval between two consecutive orders is calculated for each user to form a time interval sequence. Under normal circumstances, the time interval distribution of different users has strong randomness and diversity, and the density distribution presents a relatively smooth shape; while in the presence of abnormalities, the time interval distribution will show significant concentration, such as a large number of orders completed by a small number of users in a short period of time, or a large number of orders repeatedly appearing in a very short time interval. In order to characterize this concentration feature, the time interval sequence is mapped to a unified time interval coordinate axis, and the probability density function is calculated on this axis by kernel density estimation method, thereby forming the time interval density distribution. This density distribution can intuitively reflect the concentration and abnormality of user purchase behavior in the time dimension, and is an important part of the subsequent construction of the density tensor.

[0033] While obtaining the distribution of purchase time interval density, the purchase distribution of the target user in the geographical space is modeled in density. Specifically, for all orders in the candidate region, the geographical position information corresponding to the order is extracted, such as the longitude and latitude coordinates corresponding to the delivery address, and these coordinate points are mapped into a two-dimensional geographical space. Under normal circumstances, the geographical distribution should be consistent with the market coverage of the target category of goods, which is uniformly distributed in multiple cities and regions; and in abnormal circumstances, the orders are often concentrated in a small number of geographical locations, even in the same city or the same community. In order to depict this abnormal concentration feature, the geographical coordinate points need to be input into a spatial clustering algorithm, such as a density-based clustering method, to identify regions with highly concentrated distribution in space and calculate the order quantity density in each clustering region. Further, a geographical density distribution map can be constructed on a two-dimensional plane to visually display the concentration degree of purchase behavior. In this way, if it is found that the sales growth in the candidate region is accompanied by abnormal geographical concentration, it can be mutually confirmed with the abnormal signal, thereby improving the credibility of abnormal identification.

[0034] After obtaining the purchase time interval density distribution and the geographical density distribution, the two are combined to construct a user behavior density tensor, and the tensor is compared with the abnormal signal one by one to enhance the reliability of identification. Specifically, the user behavior density tensor is represented in three dimensions, one of which is the time interval density, the other is the geographical density, and the third is the time series. By modeling the user behavior in the candidate region in a three-dimensional tensor space, the concentration features in the time and space dimensions can be captured at the same time, and the evolution trend over time can be observed. In the tensor, if there are peak regions with too dense time interval and too concentrated geographical distribution in a certain time period, and the abnormal signal is highly coincident in time, it means that the sales growth in this time period is most likely caused by abnormal transaction behavior. In order to ensure the rigor of matching, the strength of the density peak in the tensor space needs to be calculated and correlated with the amplitude of the abnormal signal. If the correlation coefficient is significantly higher than the set threshold, it is confirmed that the sales growth belongs to the abnormal signal; if the correlation is low, it means that the abnormal signal may be caused by other unobserved factors, which needs to be further verified in the subsequent link. Through this one-by-one comparison, the user behavior density tensor not only serves as an auxiliary verification means, but also as an enhancer of the abnormal signal strength, so that the abnormal identification is no longer dependent on the sales curve and causal factor analysis, but integrates the behavior characteristics of the user dimension, thereby realizing multi-dimensional cross-validation.

[0035] Through the organic combination of the above steps, firstly, the abnormal concentration of user ordering behavior in the time dimension is identified through the purchase time interval distribution, secondly, the abnormal aggregation of orders in the spatial dimension is revealed through the geographical density distribution, and finally, the joint verification of time and space dimensions is realized by constructing the user behavior density tensor and comparing it with the abnormal signal one by one. The specific embodiment couples the time interval features and geographical distribution features of user behavior into the same density tensor, and performs superimposed analysis with the abnormal signal under the constraint of causal residual error, thereby forming a multi-level verification framework across the sales dimension and the user behavior dimension. It can not only effectively exclude normal growth caused by promotion activities or channel adjustment, but also accurately identify false explosions caused by brushing or abnormal channel traffic.

[0036] S105, under the joint constraint of the user behavior density tensor and the causal residual error, a sales abnormal weight index is generated, and false sales growth and normal sales trend are dynamically separated according to the sales abnormal weight index. The sales abnormal weight index is written back to the competitor portrait to complete the correction of the hot sales weight; In order to realize the separation of false sales growth and the correction of the hot sales weight in the competitor portrait, a sales abnormal weight index is generated under the joint constraint of the user behavior density tensor and the causal residual error, and the sales abnormal weight index is used to distinguish false sales growth and normal sales trend. Finally, the sales abnormal weight index is written back to the competitor portrait. The process includes the following steps: Under the constraint of the user behavior density tensor and the causal residual error that has been obtained, a weight factor set for measuring the sales authenticity is constructed. Specifically, the user behavior density tensor provides the concentration index of user ordering behavior in the time interval and geographical distribution dimensions, and the causal residual error reflects the part of the sales growth that cannot be explained by channel expansion and promotion. In this embodiment, the peak intensity, peak duration and peak occurrence frequency in the user behavior density tensor are extracted as a group of behavior feature vectors, and the mean, variance and abnormal deviation amplitude in the causal residual error sequence are extracted as a group of residual feature vectors. Then, the behavior feature vectors and the residual feature vectors are projected into the same numerical space through the feature standardization method, so that they can be calculated comprehensively in the same dimension. The core of this process is to provide quantitative input factors for the generation of the sales abnormal weight index by jointly modeling the user behavior and the causal residual error.

[0037] After obtaining the set of weighting factors, a sales anomaly weighting index is generated based on this set. Specifically, an exponential function is first defined to weight and superimpose user behavior features and residual features according to preset weight parameters, amplifying the contribution of abnormal signals through exponentialization. For example, when the peak of user behavior density and the peak of causal residuals highly overlap in time, their superposition result will be non-linearly amplified by the exponential function, resulting in a significantly high value in the sales anomaly weighting index. To ensure the stability of the calculation results, the output range of the exponential function can be normalized, so that the value of the sales anomaly weighting index is distributed between 0 and 1. Values ​​close to 0 indicate that sales growth can be interpreted as a normal trend, while values ​​close to 1 indicate that sales growth is highly likely to be spurious inflation. In this way, the sales anomaly weighting index achieves a unified expression of complex multidimensional data, allowing anomaly identification results to be presented in an intuitive and comparable numerical form.

[0038] After generating the abnormal sales weight index, this index is used to dynamically peel off the sales curves within the candidate region, thereby distinguishing between false growth and normal trends. Specifically, the original sales curves in the candidate region are decomposed into two parts: one part represents the normal sales trend, and the other part represents abnormal sales growth. During the calculation process, for time points where the abnormal sales weight index is higher than a set threshold, the corresponding sales increment is classified as the abnormal growth part, and its contribution is weakened by a decay coefficient; while for time points where the abnormal sales weight index is lower than the threshold, the sales increment is completely retained as part of the normal trend. Furthermore, a sliding window method can be used to smooth the abnormal sales weight index to avoid excessive impact of fluctuations at a single time point on the overall peeling result. Ultimately, the sales curve after dynamic peeling accurately reflects the true sales growth trend of the target product within the candidate region, while the peeled-off abnormal part clearly identifies possible order-brushing behavior or false inflation caused by channel traffic skew. This process not only improves the authenticity of sales data but also provides a reliable basis for subsequent competitor profiling.

[0039] After the stripping of false sales growth, the sales anomaly weight index is written back to the competitive product portrait to correct the hot sales weight. Specifically, the competitive product portrait is a multi-dimensional market performance characterization tool, and the hot sales weight is one of the core indicators. In this embodiment, the sales anomaly weight index is used as a correction factor to dynamically adjust the original hot sales weight. For example, in the original competitive product portrait, if a certain product has a high hot sales weight due to short-term sales surge, and its corresponding sales anomaly weight index is close to 1, then through the index correction formula, the hot sales weight of the product is adjusted to a reasonable range, thereby avoiding the weight expansion caused by false sales. At the same time, for products with low sales anomaly weight index, their hot sales weight can remain unchanged or be adjusted appropriately to highlight their real market competitiveness. By writing the corrected hot sales weight back to the competitive product portrait, the final competitive product portrait can more truly and comprehensively reflect the market pattern, helping enterprises make more accurate decisions in competitive product identification, pricing strategy and resource investment.

[0040] Through the above steps, first, the user behavior density tensor and the feature vector of the causal residual are extracted and a weight factor set is established, second, the sales anomaly weight index that uniformly expresses the degree of abnormality is generated based on the set, third, the index is used to dynamically strip the sales curve in the candidate area to distinguish between false growth and normal trend, and finally the index is written back to the competitive product portrait to correct the hot sales weight. This specific embodiment not only establishes a joint constraint mechanism across behavior dimensions and causal dimensions, but also realizes the quantitative differentiation and correction application of false sales through the methods of indexing and dynamic stripping, thereby significantly improving the accuracy of the competitive product portrait.

[0041] S106, under the driving of the sales anomaly weight index, the sales curve is converted into a time spectrum signal, the sales change is decomposed into low-frequency trend, medium-frequency fluctuation and high-frequency anomaly through short-time Fourier transform, and the low-frequency trend and medium-frequency fluctuation are included in the hot sales weight, and the high-frequency anomaly is dynamically weakened by the spectral domain attenuation coefficient to reduce the interference of false sales explosion on the competitive product portrait; In order to completely reduce the interference of false sales explosion on the competitive product portrait, under the driving of the sales anomaly weight index, the sales curve is converted into a time spectrum signal, and the sales change is decomposed into low-frequency trend, medium-frequency fluctuation and high-frequency anomaly through short-time Fourier transform, and finally the high-frequency anomaly part is dynamically weakened by the spectral domain attenuation coefficient, thereby realizing the accurate correction of the hot sales weight. The process includes the following steps: Under the driving of the obtained sales anomaly weight index, the sales curve of the candidate region is converted into a time signal that can be subjected to spectral analysis. Specifically, the sales anomaly weight index serves as a dynamic factor to adjust the sales curve at each time point, so that the value of the sales curve at a time point with a high anomaly weight is appropriately weakened, and the value of the sales curve at a time point with a low anomaly weight remains original. In this way, the converted time signal can already embed preliminary information of anomaly suppression in the time dimension before entering the frequency domain analysis. Next, the weighted sales curve is subjected to discretization processing to divide it into a plurality of adjacent time windows, and the length of each time window is set according to the time span of the candidate region, for example, 5 days, 7 days or 10 days can be selected as a window. Each window will serve as the input unit of the short-time Fourier transform to ensure that the dynamic characteristics of the sales curve in different time scales can be captured. The core of this processing process is to ensure that the time signal processed in the spectral analysis stage already has the prior constraint of distinguishing normal trend and abnormal fluctuation through the guidance of the sales anomaly weight index.

[0042] After obtaining the weighted time signal, the short-time Fourier transform is applied to each time window, so as to convert the sales curve in the time domain into the frequency spectrum distribution in the frequency domain. The specific steps of the short-time Fourier transform include: first, selecting a suitable window function, such as a Hamming window or a Gaussian window, to avoid the influence of boundary effects on the transformation result; then, performing windowing processing on the sales signal in each time window, and calculating the amplitude and phase of the sales signal in different frequency components to obtain the frequency spectrum diagram of the time window. By performing the short-time Fourier transform on all time windows, the characteristics of sales change can be observed in both time and frequency dimensions. In the frequency spectrum diagram, the low-frequency part corresponds to the long-term trend of sales, such as stable growth lasting for weeks or months; the medium-frequency part corresponds to periodic fluctuations, such as sales fluctuations caused by weekly promotional activities or holiday promotions; and the high-frequency part corresponds to short-term sharp fluctuations, which are often related to brushing behavior, sudden traffic tilt or abnormal promotional activities. Therefore, the short-time Fourier transform can not only reveal the overall trend of the sales curve, but also finely distinguish the growth patterns in different frequency ranges, providing a clear decomposition basis for subsequent anomaly weakening processing.

[0043] After completing the spectral decomposition of the sales curve, the low-frequency trend, medium-frequency fluctuation and high-frequency anomaly are classified and processed to play different roles in the hot-selling weight correction. Specifically, the low-frequency trend is regarded as the core part of sales growth and directly included in the hot-selling weight calculation; the medium-frequency fluctuation is regarded as market regular fluctuation and should also be retained and included in the hot-selling weight to ensure that the competitive product portrait can accurately reflect the influence of periodic factors on the market; and the high-frequency anomaly part needs to be further weakened in combination with the sales anomaly weight index. In this embodiment, first, the frequency range of the high-frequency component is identified in the frequency spectrum, and the corresponding amplitude distribution is extracted, and then the sales anomaly weight index is used as an adjustment factor to attenuate the amplitudes. Specifically, the sales anomaly weight index is multiplied point by point with the high-frequency amplitude to form a spectral domain adjustment vector. In the time window with high abnormal weight, the high-frequency amplitude is significantly weakened, and in the time window with low abnormal weight, the high-frequency amplitude remains relatively complete. The spectral domain adjustment based on the sales anomaly weight index dynamically weakens the high-frequency anomaly component without affecting the authenticity of the low-frequency and medium-frequency components.

[0044] After completing the dynamic weakening of the high-frequency anomaly, the low-frequency trend and the medium-frequency fluctuation are combined back into the hot-selling weight, and the processed frequency spectrum result is written back to the competitive product portrait, thereby completing the final correction. Specifically, the low-frequency and medium-frequency amplitude parts are re-integrated to form the corrected hot-selling weight value, which is updated to the corresponding hot-selling dimension index in the competitive product portrait. At the same time, the weakened high-frequency anomaly part is retained as an additional monitoring index to mark the potential risk of false sales explosion. In this way, the corrected competitive product portrait not only accurately reflects the long-term trend and periodic fluctuation of the goods in the market, but also timely warns of abnormal situations to avoid misjudgment of false sales as real competitiveness. The final competitive product portrait has dual advantages: on the one hand, the hot-selling weight is more accurate and reliable, avoiding interference from false explosions; on the other hand, the risk monitoring is more sensitive, providing forward-looking support for the enterprise to develop pricing, promotion and product selection strategies.

[0045] Through the implementation of the above steps, first, the sales curve is weighted and divided into time windows under the drive of the sales anomaly weight index, second, the time-frequency spectrum is decomposed through short-time Fourier transform, third, the low-frequency, medium-frequency and high-frequency are effectively distinguished through classification and anomaly weakening, and finally the correction result is written back to the competitive product portrait to complete the adjustment of the hot-selling weight, which not only strips false growth in the time domain, but also dynamically weakens high-frequency anomalies through spectral domain attenuation coefficients in the frequency domain, thereby constructing a two-dimensional correction mechanism across the time domain and the frequency domain, greatly improving the authenticity and reliability of the competitive product portrait.

[0046] The application can accurately identify abnormal sales growth in a short period caused by brushing or channel tilt through constructing a joint constraint mechanism based on causal residual and user behavior density tensor, and effectively filter false sales signals. By dynamically calculating the sales anomaly weight index and writing it back to the competitor portrait, the application effectively avoids the amplification effect of false sales on hot sales weight, thereby significantly improving the accuracy of competitor identification and the reliability of marketing decisions, solving the problem that the competitor portrait is easily distorted by abnormal data in the prior art.

[0047] The application further introduces short-time Fourier transform to decompose the sales curve into three types of signals, i.e. low-frequency trend, medium-frequency fluctuation and high-frequency anomaly, and dynamically weakens the high-frequency anomaly in combination with the spectral domain attenuation coefficient, effectively reducing the interference of abnormal sales in the frequency domain. This method not only enhances the ability to identify real market trends, but also enables the hot sales weight to more truly reflect the competitiveness of the goods in the long-term and periodic dimensions, and comprehensively improves the adaptability and analysis accuracy of the competitor portrait in the multi-frequency behavior mode.

[0048] The above only describes certain exemplary embodiments of the application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the application. Therefore, the above figures and description are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the application.

Claims

1.An AI intelligent marketing method for mining hot-selling competitive products based on category price, characterized in that, The method comprises the following steps: A unified time baseline is established, and a continuous sales curve of a commodity under a target category is extracted. A momentary slope of sales change is calculated in the sales curve, and a candidate area of short-period sales growth is determined based on a mutation interval of the momentary slope. A causal factor set is constructed in the candidate area. Channel expansion trajectories and promotion trajectories are mapped to the sales curve point by point, and causal residuals are calculated according to the mapping results. A cross-channel consistency spectrum is generated based on the causal residuals. The phase difference and synchronism index of sales growth of each channel are projected into the causal residual space, and the projection results are superimposed with the causal residuals to form abnormal signals. A user behavior density tensor is constructed under the constraint of the abnormal signals. The purchase time interval and geographical distribution of the target user are modeled in density, and the density peaks are compared with the abnormal signals one by one. A sales anomaly weight index is generated under the joint constraint of the user behavior density tensor and the causal residuals. The sales anomaly weight index is used to dynamically separate false sales growth from normal sales trends. The sales anomaly weight index is written back to the competitor portrait to correct the hot sales weight. Under the driving of the sales anomaly weight index, the sales curve is converted into a time spectrum signal. The sales change is decomposed into low-frequency trends, medium-frequency fluctuations and high-frequency anomalies through short-time Fourier transform. The low-frequency trends and medium-frequency fluctuations are included in the hot sales weight, and the high-frequency anomalies are dynamically weakened by the spectral domain attenuation coefficient. 2.The AI intelligent marketing method for mining hot-selling competitive products based on category price according to claim 1, characterized in that, The sales data of the commodity under the target category is used to identify the candidate area of short-period sales growth, which comprises the following steps: A unified time baseline is established, and a continuous sales curve is extracted on the time baseline. The sales data from different sources are time-mapped, interpolated, filled and smoothed to build a complete and continuous sales curve. The momentary slope of sales change is calculated on the continuous sales curve. The momentary slope curve is obtained through difference operation. The sales curve is smoothed before calculation to identify the mutation characteristics of the slope curve. The candidate area is determined based on the mutation points of the momentary slope curve. The mutation points are marked through threshold detection, and the time range before and after the mutation points is expanded to cover the whole growth process. If the interval between adjacent mutation points is less than a certain threshold, they are merged. 3.The AI intelligent marketing method for mining hot-selling competitive products based on category price according to claim 1, characterized in that, The causal factor set is constructed in the candidate area, and the causal residuals are calculated, which comprises the following steps: External influence factors related to sales change are collected in the candidate area, and a causal factor set is constructed. The causal factor set includes channel expansion trajectories and promotion trajectories. The causal factors are standardized and time-aligned to form a time series on the unified time baseline. The causal factors are mapped to the sales curve in the candidate area point by point. The values of the causal factors are extracted at each time point and paired with the sales data. The weights are set according to the historical contribution to establish the correspondence between the causal factors and the sales change. The causal residuals are calculated based on the correspondence. The sales are predicted by the causal factors, and the residuals are obtained by comparing the actual sales. The residuals are superimposed with the sales slope mutation interval in the time dimension to identify the abnormal growth area that cannot be explained by the causal factors. 4.The AI intelligent marketing method for mining hot-selling competitive products based on category price according to claim 3, characterized in that, The cross-channel consistency spectrum is generated based on the causal residuals, which comprises the following steps: The sales curves of each channel in the candidate region are extracted, and the local growth rate is calculated through a sliding window. The growth rate is converted into a phase signal, and the phase difference sequence is obtained by taking the reference curve as the reference; The synchronism index between channels is calculated according to the phase difference sequence, and the cross-channel consistency matrix is obtained through correlation analysis and coherence function. The consistency matrix is matched with the causal residual in the time dimension point by point to generate a consistency spectrum; The consistency spectrum is normalized and superimposed with the causal residual curve in the time dimension to form a joint abnormal signal. The signal amplitude is enhanced when the causal residual deviates positively and the consistency is low to mark the abnormal growth region. 5.The AI intelligent marketing method for mining hot-selling competitive products based on category price according to claim 4, characterized in that, When normalizing the consistency spectrum, the numerical interval of the consistency spectrum is adjusted to be consistent with the numerical interval of the causal residual, and a weighting coefficient is set in the superposition process. The causal residual is given a high weight, and the consistency spectrum is given a low weight. 6.The AI intelligent marketing method based on the hot-selling competitive product mining of the category price according to claim 4, characterized in that, The steps of constructing the user behavior density tensor under the constraint of the abnormal signal include: Under the constraint of the abnormal signal, the user order time is extracted from the transaction data of the candidate region. The time interval of consecutive orders is calculated and the time interval density distribution is formed by kernel density estimation; The geographic location information corresponding to the transaction data is extracted, the order location is mapped to a two-dimensional space, and the geographic density distribution is obtained by density clustering; The time interval density distribution and the geographic density distribution are combined to construct the user behavior density tensor, and the density peak value in the tensor space is compared with the abnormal signal one by one. 7.The AI intelligent marketing method for hot-selling competitive product mining based on category price according to claim 1, characterized in that, The steps of generating the sales abnormal weight index under the joint constraint of the user behavior density tensor and the causal residual include: Based on the user behavior density tensor and the causal residual, a feature vector is extracted, a weight factor set is constructed, and it is projected to a unified numerical space through feature standardization; The sales abnormal weight index is generated according to the weight factor set. The abnormal signal is amplified by weighted superposition and exponential function. The output result is normalized to limit the value range; Under the driving of the sales abnormal weight index, the sales curve of the candidate region is dynamically stripped. The abnormal growth part is weakened by the attenuation coefficient, and the normal trend part is retained; The sales abnormal weight index is written back to the competitor portrait to correct the original hot sales weight, avoiding the deviation of competitor identification caused by false sales expansion. 8.The AI intelligent marketing method based on the hot-selling competitive product mining of the category price according to claim 7, characterized in that, The steps of converting the sales curve into a time frequency spectrum signal under the driving of the sales abnormal weight index include: Under the driving of the sales abnormal weight index, the sales curve of the candidate region is weighted and adjusted, and is divided into adjacent time windows to form a frequency spectrum analysis input signal; Short-time Fourier transform is applied to the time window to obtain the frequency spectrum distribution of low-frequency trend, medium-frequency fluctuation and high-frequency anomaly; The low-frequency trend and medium-frequency fluctuation are directly included in the hot sales weight, and the amplitude attenuation is performed on the high-frequency anomaly range by using the sales abnormal weight index to weaken the abnormal signal; The low-frequency trend and medium-frequency fluctuation are combined and updated to the hot sales weight of the competitor portrait, and the attenuated high-frequency anomaly is written back to the competitor portrait as a monitoring indicator.

Images