Intelligent screening method for high conversion potential keywords based on commodity attributes and hot-selling data

Abstract

The application discloses a high-conversion-potential keyword intelligent screening method based on commodity attributes and hot-selling data, relates to the technical field of keyword screening, and comprises the following steps: obtaining historical conversion data of candidate keywords; correcting the historical conversion data according to external environment parameters; obtaining the use coverage ratio of the candidate keywords in competitive commodity; comprehensively scoring the candidate keywords; tracking the change trend of the use coverage ratio in a recent time window; and selecting the candidate keywords whose adjusted comprehensive scores reach a preset threshold value. The application solves the problem of insufficient signal-to-noise ratio of the input signal of the evaluation module, solves the problem that the system cannot identify the degree of competition monopoly, solves the problem that the prediction deviation of the static model continuously expands after the competition pattern changes, realizes accurate matching with the specific attribute positioning of the target commodity, meets the keyword selection demand of different advertisement budgets, and enables the screening model to continuously evolve and improve the precision in actual application.

Description

Technical Field

[0001] This invention relates to the field of keyword screening technology, and in particular to an intelligent screening method for high-conversion-potential keywords based on product attributes and best-selling data. Background Technology

[0002] In e-commerce search advertising, keywords serve as the core bridge connecting users' search intent and product display pages. The quality of keyword selection directly impacts the advertising system's exposure, click-through rate, and conversion efficiency. However, existing keyword selection technologies suffer from the following main shortcomings:

[0003] First, historical conversion data is severely affected by external factors. External environmental factors such as product discounts, platform promotions, and seasonal changes can cause the original conversion data to be mixed with a large number of artificially high conversions that are not due to the attributes of the keywords themselves. Existing technologies lack effective data correction mechanisms and cannot separate external interference, which makes it easy for advertising systems to allocate budgets to keywords that are driven up by short-term factors but have insufficient real conversion capabilities, resulting in a waste of resources.

[0004] Secondly, it ignores the impact of the degree of monopoly of competitors on the difficulty of conversion. Existing technologies focus more on the semantic relevance of keywords and products, but do not quantify the coverage ratio of keyword usage in competitors and the degree of concentration of related traffic. When keywords have been widely used by top competitors and traffic is highly concentrated, new advertising campaigns are difficult to obtain effective exposure, and the marginal return on advertising resources is extremely low. Existing competitor analysis tools also do not provide quantitative analysis of traffic concentration.

[0005] Third, the screening results lack dynamic tracking and adaptive adjustment capabilities. Existing prediction models mostly output scores in a static way, which cannot track the temporal trend of keyword coverage ratio among competitors. Keywords with continuously rising competition may quickly become red oceans, while keywords with declining popularity may present opportunity windows. Static scores without trend perception cannot provide the advertising system with forward-looking keyword value judgments. Fourth, the screening granularity is coarse and is not grouped according to product attribute dimensions. The conversion contribution of the same keyword varies significantly under different product categories, functional selling points, applicable scenarios, and specifications. Existing technology does not group and correct historical conversion data according to the above attribute dimensions, resulting in screening results that cannot accurately match the product needs of specific attribute positioning.

[0006] In summary, the existing technology has the following technical defects: First, historical conversion data is affected by external environmental factors such as discounts, major promotions, and seasonality, making it impossible for the system to separate external noise signals from the inherent conversion characteristics of keywords, resulting in insufficient signal-to-noise ratio of the input signal for the evaluation module. Second, there is a lack of joint quantitative analysis of the breadth of keyword coverage and traffic concentration in the competitor space, making the evaluation model unable to perceive the degree of competitive monopoly faced by keywords, causing the advertising system to allocate budgets to low-marginal-return keywords already locked in by leading competitors. Third, the evaluation model is a static, time-invariant structure, unable to perceive the evolution trend of competitor coverage ratios over time, and the system lacks adaptive correction capabilities after changes in the competitive landscape, leading to a continuous increase in prediction bias. Fourth, keyword evaluation is not grouped according to product attribute dimensions, and the differences in conversion contribution of the same keyword in different categories, functions, and scenarios are ignored by the system, resulting in the screening results failing to accurately match the product placement needs of specific attribute positioning. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0008] In view of the problems existing in the above-mentioned intelligent screening methods for high conversion potential keywords based on product attributes and best-selling data, this invention is proposed.

[0009] Therefore, the purpose of this invention is to provide an intelligent keyword screening method based on product attributes and best-selling data. This method addresses the problems of existing technologies where historical conversion data is coupled with interference from external environmental factors such as discounts, promotions, and seasonality, leading to insufficient signal-to-noise ratio in the input signal of the evaluation module due to the inability of the system to separate external noise signals from the inherent conversion characteristics of keywords; a lack of joint quantitative analysis of keyword coverage breadth and traffic concentration in the competitor space, resulting in the evaluation model's inability to perceive the degree of competitive monopoly faced by keywords, causing the advertising system to allocate budgets to low-marginal-return keywords already locked in by leading competitors; a static, time-invariant evaluation model that cannot perceive the evolution trend of competitor coverage ratios over time, resulting in a lack of adaptive correction capability after changes in the competitive landscape and a continuously expanding prediction bias; and the failure to group keyword evaluation according to product attributes, causing the system to ignore the differences in conversion contribution of the same keyword across different categories, functions, and scenarios, resulting in screening results that cannot accurately match the product placement needs of specific attribute positioning.

[0010] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0011] In a first aspect, embodiments of the present invention provide an intelligent screening method for high-conversion-potential keywords based on product attributes and best-selling data. This method includes: acquiring historical conversion data of candidate keywords and simultaneously acquiring external environmental parameters within the corresponding time period; correcting the historical conversion data according to the external environmental parameters; calculating the pure conversion contribution value of the candidate keyword based on the corrected conversion data; acquiring the usage coverage ratio of the candidate keyword among competing products and the degree of concentrated occupation of the candidate keyword's associated traffic among various competing products; comprehensively scoring the candidate keyword based on the pure conversion contribution value, the usage coverage ratio, and the degree of concentrated occupation; tracking the changing trend of the usage coverage ratio within a recent time window and adjusting the comprehensive score according to the changing trend; and identifying and outputting candidate keywords whose adjusted comprehensive scores reach a preset threshold as high-conversion-potential keywords.

[0012] As a preferred embodiment of the intelligent keyword selection method for high conversion potential based on product attributes and best-selling data described in this invention, the historical conversion data is grouped and statistically analyzed according to the product attribute dimension; the product attributes include product category, functional selling points, applicable scenarios, and specification parameters; the external environment parameters include product discount level, promotional activity duration, seasonal change factor, and platform promotion identifier; the competing products refer to the set of products that belong to the same category as the target product and rank in the top N in terms of transaction volume within a preset time period, where N is a preset positive integer.

[0013] As a preferred embodiment of the intelligent keyword selection method for high conversion potential based on product attributes and best-selling data according to the present invention, the step of correcting the historical conversion data according to external environmental parameters and calculating the pure conversion contribution value of the candidate keyword based on the corrected conversion data is as follows: extracting normal conversion period data unaffected by external environmental factors from the historical conversion data, and calculating a baseline conversion value based on the normal conversion period data; identifying the interfered periods in the historical conversion data according to the external environmental parameters, comparing the actual conversion value of the interfered periods with the baseline conversion value, and calculating the interference deviation caused by external environmental interference; removing the interference deviation from the total value of the historical conversion data to obtain the corrected conversion data; and calculating the pure conversion contribution value of the candidate keyword based on the corrected conversion data.

[0014] As a preferred embodiment of the intelligent keyword screening method for high conversion potential based on product attributes and best-selling data described in this invention, the normal conversion period unaffected by external environmental factors refers to a continuous period of time without product discount activities, platform promotional activities, or seasonal transition periods.

[0015] As a preferred embodiment of the intelligent keyword screening method for high conversion potential based on product attributes and best-selling data described in this invention, wherein: the coverage ratio of candidate keywords in competitor products refers to the ratio of the number of competitor products that use the candidate keywords to the total number of competitor products; the degree of concentration of candidate keyword-related traffic among competitor products refers to the distribution of the share of related traffic obtained by each competitor product using the candidate keywords to the total related traffic, and the more concentrated the share distribution is in a few competitor products, the higher the degree of concentration.

[0016] As a preferred embodiment of the intelligent keyword selection method for high conversion potential based on product attributes and best-selling data described in this invention, the following steps are taken: The candidate keywords are comprehensively scored based on their pure conversion contribution value, usage coverage ratio, and concentration level. Specifically, the pure conversion contribution value is used as the baseline score; an additional score is added based on the usage coverage ratio, with the increment of the additional score decreasing as the usage coverage ratio increases; and the concentration level is used to deduct from the above results, with a greater deduction for higher concentration levels.

[0017] As a preferred embodiment of the intelligent keyword screening method for high conversion potential based on product attributes and best-selling data described in this invention, the method of tracking the trend of usage coverage ratio within a recent time window and adjusting the comprehensive score according to the trend is as follows: The recent time window is divided into multiple consecutive sub-periods, and the usage coverage ratio of the candidate keywords in competing products within each sub-period is statistically analyzed to obtain time series data of the coverage ratio; based on the time series data, the difference in coverage ratio between adjacent sub-periods is calculated; if multiple consecutive differences are positive and the average of each positive value exceeds a first preset change threshold, the trend is determined to be continuously rising. In this case, the attenuation range is determined based on the average value of the consecutive positive values. The larger the mean, the greater the attenuation. The overall score is multiplied by an attenuation coefficient, which decreases as the attenuation increases. If multiple consecutive differences are negative and the average of these negative values ​​is below a second preset change threshold, the trend is determined to be a continuous decline. In this case, no attenuation adjustment is applied to the overall score, and the gain is determined based on the average of the consecutive negative values. The larger the absolute value of the negative value, the greater the gain. The overall score is multiplied by a gain coefficient, which increases as the gain increases. If the average of multiple consecutive differences does not exceed the first preset change threshold and is not lower than the second preset change threshold, the trend is determined to be without significant change. No adjustment is applied to the overall score, and the overall score remains unchanged.

[0018] As a preferred embodiment of the intelligent keyword screening method for high conversion potential based on product attributes and best-selling data described in this invention, the step of determining and outputting candidate keywords whose adjusted comprehensive scores reach a preset threshold as high conversion potential keywords specifically involves: pre-setting two different levels of scoring thresholds, including a first threshold and a second threshold, wherein the first threshold is higher than the second threshold; comparing the adjusted comprehensive scores with the first threshold, and marking candidate keywords that reach or exceed the first threshold as first-level high conversion potential keywords; marking candidate keywords whose adjusted comprehensive scores are between the second threshold and the first threshold as second-level high conversion potential keywords; marking candidate keywords whose adjusted comprehensive scores are lower than the second threshold as invalid keywords and not included in the output list; and outputting the marked candidate keywords and their corresponding comprehensive scores in descending order of level.

[0019] As a preferred embodiment of the intelligent keyword selection method for high conversion potential based on product attributes and best-selling data described in this invention, the method involves: after determining and outputting candidate keywords whose comprehensive scores after attenuation adjustment reach a preset threshold as high conversion potential keywords, obtaining the actual conversion performance data of each candidate keyword determined as a high conversion potential keyword after output; adjusting the attenuation adjustment range according to the deviation between the actual conversion performance data and the comprehensive score; if the actual conversion performance data is higher than the predicted value corresponding to the comprehensive score, reducing the attenuation range under the same trend in subsequent cycles; if the actual conversion performance data is lower than the predicted value corresponding to the comprehensive score, increasing the attenuation range under the same trend in subsequent cycles.

[0020] Secondly, to further address the aforementioned technical problems, the present invention provides an intelligent screening system for high-conversion-potential keywords based on product attributes and best-selling data. This system includes: a data acquisition module for acquiring historical conversion data of candidate keywords and simultaneously acquiring external environmental parameters within the corresponding time period; a purity value calculation module for correcting historical conversion data based on external environmental parameters and calculating the pure conversion contribution value of candidate keywords based on the corrected conversion data; a competitor analysis module for acquiring the usage coverage ratio of candidate keywords among competitor products and the degree of concentration of traffic associated with the candidate keywords among various competitors; a comprehensive scoring module for comprehensively scoring candidate keywords based on their pure conversion contribution value, usage coverage ratio, and concentration degree; a trend adjustment module for tracking the changing trend of usage coverage ratio within a recent time window and adjusting the comprehensive score based on the changing trend; and a result output module for identifying and outputting candidate keywords whose adjusted comprehensive scores reach a preset threshold as high-conversion-potential keywords.

[0021] The beneficial effects of this invention are as follows: By acquiring external environmental parameters and eliminating their interference with historical conversion data, this invention enables the screening results to reflect the true matching value between keywords and product attributes rather than the abnormally high conversion rates, thus solving the problem of insufficient signal-to-noise ratio in the input signal of the evaluation module; by simultaneously quantifying the usage coverage ratio and traffic concentration of keywords among competitors, it guides sellers to avoid keywords that have been highly monopolized by leading competitors, thus solving the problem of the system's inability to identify the degree of competitive monopoly; by tracking the temporal trend of coverage ratio changes and applying attenuation when it is continuously rising and gain when it is continuously falling, the screening results have a forward-looking perspective in the time dimension, solving the problem of the static model's prediction deviation continuously expanding after changes in the competitive landscape; by grouping, correcting, and scoring historical conversion data according to attributes such as product category, functional selling points, applicable scenarios, and specifications, it achieves accurate matching with the specific attribute positioning of the target product, solving the problem of the system ignoring the differences in conversion contribution of keywords under different attribute dimensions; by preset multi-level scoring thresholds for graded output and by performing closed-loop feedback correction on the adjustment range based on the actual conversion performance after keyword placement, it not only meets the keyword selection needs of different advertising budgets but also enables the screening model to continuously evolve and improve accuracy in practical applications. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0023] Figure 1 This is a flowchart illustrating the implementation of the present invention in Example 1.

[0024] Figure 2 This is a schematic diagram of the dynamic adjustment and hierarchical output logic of the keyword comprehensive score in Example 1. Detailed Implementation

[0025] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Example 1

[0026] Reference Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a method for intelligently filtering high-conversion-potential keywords based on product attributes and best-selling data, including the following steps:

[0027] S1: Obtain historical conversion data for candidate keywords and simultaneously obtain external environmental parameters for the corresponding time period.

[0028] Specifically, historical conversion data includes the search volume, click volume, and corresponding number of completed orders for each candidate keyword within a preset historical period. Historical conversion data is statistically analyzed on a daily basis, with the preset historical period being the most recent 30 to 90 days.

[0029] Furthermore, the aforementioned historical conversion data is grouped and statistically analyzed according to product attribute dimensions:

[0030] Among them, product attributes include product category, functional selling points, applicable scenarios, and specifications;

[0031] External environmental parameters include product discount levels, promotional duration, seasonal factors, and platform promotion indicators;

[0032] Specifically, the historical conversion data and external environmental parameters mentioned above were obtained by calling the official API interface of the e-commerce platform. It should be noted that the seasonal change factor in the external environmental parameters was extracted by time series decomposition of the historical conversion data.

[0033] Specifically, the seasonal change factor in the aforementioned external environmental parameters is obtained through the following methods:

[0034] The STL (Seasonal-Trend decomposition using LOESS) time series decomposition method is used to decompose the historical conversion data of candidate keywords into three additive components: trend component, seasonal component, and residual component.

[0035] STL decomposition uses LOESS (locally weighted regression) to iteratively smooth the estimation of the three components. In specific implementation, the key parameters can be set according to the needs, such as a seasonal smoothing window of 13 weeks, a trend smoothing window of 19 or 21 days, a low-pass filter window of 7 days, a seasonal polynomial degree of 0, a trend polynomial degree of 1, and enabling robust estimation to resist outlier interference.

[0036] After the decomposition is completed, the average seasonal component of each period is extracted according to a fixed period (e.g., 7 days as a period), and the absolute rate of change of the average seasonal component of adjacent periods is calculated.

[0037] When the rate of change exceeds the preset seasonal change threshold, the corresponding date interval is marked as the seasonal change critical period, and the rate of change is used as the seasonal change factor for that period.

[0038] Furthermore, the above-mentioned seasonal change threshold is obtained as follows:

[0039] Obtain the numerical sequence of the seasonal component variation rate between all adjacent periods, and calculate the arithmetic mean and standard deviation of the sequence;

[0040] The mean plus the standard deviation is used as the seasonal change threshold.

[0041] For example, if a sports brand's flagship store plans to promote a "lightweight cushioned running shoe," the historical data period is set to the most recent 90 days (April 1, 2026 to June 30, 2026), the data granularity is daily, and the product attributes are grouped according to four dimensions: "product category (sports shoes / running shoes)," "features and selling points (lightweight / cushioning)," "applicable scenarios (road running / marathon)," and "specifications (size / shoe weight)." The search term report for "lightweight cushioned running shoes" in the past 90 days is obtained from the e-commerce platform's backend, and "lightweight running shoes," "cushioned running shoes," "lightweight cushioned running shoes for men," "road running shoes," and "marathon running shoes" are extracted. Historical conversion data for candidate keywords (including daily search volume, click volume, and number of completed orders) are grouped and statistically analyzed according to four attribute dimensions: product category, functional selling points, applicable scenarios, and specifications. At the same time, the daily product discount, promotion start and end time, and platform promotion indicators (e.g., no major promotion in April, "520" promotion in May, and "618" promotion in June) are obtained synchronously through the platform API interface. The historical conversion data is then decomposed into time series to extract seasonal transition factors (April-May is the transition period from spring to summer, and June is the peak summer sales season), providing complete external environmental parameter inputs for subsequent data correction.

[0042] S2: Correct historical conversion data based on external environmental parameters, and calculate the pure conversion contribution value of candidate keywords based on the corrected conversion data;

[0043] Specifically, historical conversion data is corrected based on external environmental parameters, and the clean conversion contribution value of candidate keywords is calculated based on the corrected conversion data. The details are as follows:

[0044] Extract normal conversion period data from historical conversion data that is unaffected by external environmental factors, and calculate the baseline conversion value based on the normal conversion period data. The specific formula is as follows:

[0045] ;

[0046] In the formula, The baseline conversion value serves as the baseline conversion level after weighted averaging. The first during the normal conversion period The conversion value for each time unit is extracted from historical conversion data by time unit, providing conversion data samples for each time point within a normal period. For the first The weighting coefficient for each time unit is calculated using the formula: The calculation shows that the contribution of each time unit in the calculation of the benchmark value is controlled, and the closer to the current time, the greater the weight, making the benchmark value more timely. The time unit is the sequence number within the normal conversion period, used to determine the position of each data point in the time series, thereby calculating the corresponding weight coefficient; This represents the total number of time units included within the normal conversion period, calculated based on the start and end dates of the normal conversion period. The decay rate constant is a preset hyperparameter, which is obtained by fitting historical data or manually setting, and its value ranges from 0.02 to 0.08, controlling the decay rate of the weights over time. The larger the value, the stronger the influence of recent data, and the more sensitive the benchmark value is to recent changes. The smaller the value, the more evenly weighted the various time units, and the smoother the baseline value.

[0047] Based on external environmental parameters, identify the interference periods in historical conversion data, compare the actual conversion values ​​of the interference periods with the baseline conversion values, and calculate the interference deviation caused by external environmental interference.

[0048] The total historical conversion data is removed to eliminate interference and deviation, thereby eliminating conversion fluctuations caused by changes in the external environment and obtaining corrected conversion data. The pure conversion contribution value of candidate keywords is calculated based on the corrected conversion data. The normal conversion period unaffected by external environmental factors refers to a continuous period of time without product discount activities, platform promotion activities, or seasonal transitions.

[0049] It should be noted that the above serial numbers The furthest historical moment corresponding to the normal conversion period, Corresponding to the most recent moment (i.e., the current moment) within the normal conversion period; Serial Number As time progresses, the value increases, and the closer to the current moment... The larger the value, the higher its corresponding weight. The larger the value, the greater the contribution of recent data to the baseline transformation value.

[0050] Specifically, the aforementioned decay rate constant The value ranges from 0.02 to 0.08, and its determination is based on the following: To control the forgetting rate of historical samples in the exponentially weighted moving average, taking daily granular data as an example, =0.02 corresponds to an effective memory period of approximately 50 days. =0.08 corresponds to an effective memory period of approximately 12.5 days, which covers the typical conversion period of most product categories in e-commerce scenarios;

[0051] Different scenarios The selection criteria are as follows: for short-cycle categories such as fast-moving consumer goods and apparel, The value is relatively large, ranging from 0.06 to 0.08; for long-cycle categories such as home appliances and digital products, The values ​​are relatively small, ranging from 0.02 to 0.04; when using weekly granular data, The value range is adjusted accordingly to 0.1 to 0.3.

[0052] It should be noted that the aforementioned normal conversion periods unaffected by external environmental factors are allowed to exist as "quasi-normal periods" in actual e-commerce scenarios, and the determination rules are as follows:

[0053] Mark the interference status of historical transformation data by day, labeling them as interfered days or non-interfered days;

[0054] Using a preset step size to slide and scan historical conversion data, find continuous date intervals that meet the following conditions as quasi-normal periods: the proportion of the number of days marked as "interference days" in the interval to the total number of days in the interval does not exceed the preset tolerance threshold.

[0055] The tolerance threshold is set according to the promotion intensity of the target product category: 20% for frequently promoted categories such as FMCG and apparel; 15% for moderately promoted categories such as home appliances and digital products; and 10% for low-frequency promoted categories such as industrial products and niche categories. For example, within a continuous 14-day period, a maximum of 2 days are allowed to be interfered with, and the remaining 12 days are allowed to be non-interference days. This period can be considered as a quasi-normal period that can be used as a benchmark for conversion value calculation. The minimum continuous length of the quasi-normal period is no less than 14 days (daily granular data), which can be reduced to no less than 7 days when the historical conversion data period is short.

[0056] If multiple consecutive intervals that meet the conditions are scanned, the longest one is selected as the quasi-normal time period; if the lengths are the same, the one closest to the current time is selected.

[0057] If no continuous intervals that meet the tolerance threshold can be found, the tolerance threshold will be gradually widened by a preset step size (increasing by 5 percentage points each time) until a quasi-normal period that meets the criteria is selected.

[0058] For days with tolerable interference within the quasi-normal period, data smoothing is performed by linear interpolation or by replacing the median of adjacent non-interference days. This replacement value is only used for the calculation of the baseline transformation value and does not change the original historical transformation data.

[0059] When the amount of historical conversion data is severely insufficient (e.g., the total number of days is less than 30 days), the global median or global average is used as an alternative estimate of the baseline conversion value.

[0060] Specifically, the system identifies the period of interference based on external environmental parameters: based on the platform's promotional logo, the discount level of the product, the duration of the promotional activity, and the seasonal change factor, the period of the platform's promotion, the date when the discount level exceeds the preset threshold, the start and end interval of the promotional activity, and the critical period of the seasonal change are marked as the period of interference, respectively; when multiple parameters indicate the same date, they are combined and marked as one period of interference.

[0061] Specifically, obtaining the actual conversion value for each day within the period of interference includes:

[0062] By using the official API interface, obtain the number of completed orders and the corresponding clicks or search volume for candidate keywords on each affected date.

[0063] Divide the number of completed orders by the number of clicks to get the actual click-through rate on the affected days; or divide the number of completed orders by the number of searches to get the actual search conversion rate on the affected days.

[0064] Use the actual click-through rate or actual search conversion rate as the actual conversion value for the affected date.

[0065] Furthermore, the specific formula for the above-mentioned interference deviation is as follows:

[0066] ;

[0067] In the formula, For the first The daily interference deviation quantifies the degree of deviation caused by external environmental factors on the daily transformation. Positive values ​​indicate a positive improvement, while negative values ​​indicate a suppressive effect. It is used for subsequent weighted summation to obtain the total interference deviation. For the first The actual conversion value of the day provides the actual conversion performance of the disturbed day and is the basic input for calculating the daily disturbance deviation. The baseline conversion value serves as a reference under interference-free conditions, used to determine the degree to which the actual value deviates from the normal level.

[0068] Specifically, the daily interference deviations of each effective interfered day within the aforementioned interference period are directly summed to obtain the total interference deviation.

[0069] Preferably, when click volume data is available, click conversion rate should be used as the actual conversion value; when click volume data is difficult to obtain, search conversion rate should be used as a substitute.

[0070] Furthermore, after calculating the total interference deviation, the total interference deviation is subtracted from the total value of the historical conversion data to obtain the corrected conversion data. This corrected conversion data eliminates the conversion fluctuations caused by external environmental factors and reflects the true conversion level of the keywords under conditions without external interference. It is used to calculate the pure conversion contribution value of candidate keywords in subsequent calculations.

[0071] Specifically, the net conversion contribution value of candidate keywords is calculated based on the corrected conversion data, as detailed below:

[0072] Obtain corrected conversion data after external interference removal. Each conversion event in this data has been freed from interference from external environmental factors and reflects the conversion performance of keywords under normal market conditions.

[0073] According to the preset attribution rules, each conversion event in the corrected conversion data is attributed to the corresponding candidate keyword. The total number of conversion events attributed to the candidate keyword in the corrected conversion data is counted, and this statistical value is the pure conversion contribution value of the candidate keyword.

[0074] It should be noted that the basic logic of the above attribution rules is: if a user enters a product page by searching for a certain keyword and eventually completes a transaction, then the transaction event is attributed to that keyword. When a user searches for multiple keywords before making a transaction, different attribution methods can be selected according to actual needs, including: attributing the transaction to the keyword the user last searched, or attributing the transaction equally to the multiple keywords the user searched before making a transaction.

[0075] Furthermore, the selection of the above attribution methods follows these rules:

[0076] For categories with low average order value, such as fast-moving consumer goods, apparel, and daily necessities, the user decision-making path is relatively short. The last click attribution method is adopted to attribute the transaction to the keyword searched by the user last time.

[0077] For high-priced categories such as home appliances, digital products, and furniture, where the user decision-making path is relatively long, a linear attribution approach is adopted, which weights the transaction based on the proportion of the number of times the user searched for each keyword before making the transaction.

[0078] The system defaults to the last click attribution method and allows users to customize the configuration based on their actual business understanding.

[0079] When linear attribution is used, if a user has searched for multiple keywords before making a purchase, each keyword will receive a corresponding attribution weight based on the proportion of its search count to the total number of searches. When the same keyword is searched multiple times, its search count will be accumulated and included in the total number of searches for that keyword, and the attribution weight will be determined based on the accumulated count.

[0080] When there are conversion events in the corrected conversion data that cannot be attributed to any candidate keyword, these conversion events are marked as invalid attribution events and are not included in the pure conversion contribution value statistics of any candidate keyword.

[0081] Specifically, in this embodiment, the pure conversion contribution value is calculated as follows: the total number of conversion events attributed to the candidate keyword in the statistically corrected conversion data. The attribution adopts the last click attribution model or the linear attribution model. For example, for keyword A, it generated 100 transactions in the corrected 30-day conversion data, so its pure conversion contribution value is recorded as 100. This value is used to characterize the actual conversion capability driven independently by the keyword after excluding external environmental interference.

[0082] For example, from 90 days of historical conversion data, continuous normal periods without any promotional activities, major sales events, or seasonal transitions are selected. The daily click-through rate (CTR) of "lightweight cushioning running shoes" within these periods (e.g., an average of 2.0%) is used as the base conversion rate, weighted by exponential decay based on time proximity. Then, disrupted periods are identified based on external environmental parameters—June 1st to June 18th (during the "618" sales event) is marked as a disrupted period, and the actual daily CTR within this period is obtained. (e.g., 6.0% on June 18th), subtract the baseline conversion value of 2.0% each day to obtain the daily interference deviation (e.g., +4.0% on June 18th), and directly sum the daily interference deviations of all affected dates to obtain the total interference deviation (assumed to be 8.5%); after removing this total interference deviation from the total value of the original conversion data, count the total number of corrected transaction events attributed to the candidate keyword. For example, if "lightweight cushioning running shoes" is attributed to 100 transactions in the corrected data, then the pure conversion contribution value of this keyword is recorded as 100.

[0083] S3: Obtain the coverage ratio of candidate keywords in competitor products, and the degree of concentration of traffic associated with the candidate keyword among various competitors.

[0084] Specifically, the aforementioned competing products refer to the set of products that belong to the same category as the target product and rank in the top N in terms of transaction volume within a preset time period, where N is a preset positive integer.

[0085] Among them, the granularity of the same category adopts the leaf category, which is the finest category level in the platform's classification system; when the number of products under the leaf category is insufficient, it is traced back to the next higher level category to supplement, and the tracing level does not exceed two levels.

[0086] The preset time period is consistent with the preset historical period of historical conversion data, and the default time is the same as that in step S1, to ensure that the competitive analysis results are consistent with the timeliness of historical conversion data;

[0087] The value of N is determined based on the market concentration of the target product's category: for top-tier concentrated categories (such as mobile phones and digital products), N is 10 to 20; for dispersed categories (such as clothing and home furnishings), N is 30 to 50; for long-tail categories (such as handicrafts and niche collectibles), N is 50 to 100; when the category type is not specified, N defaults to 30.

[0088] Furthermore, the coverage ratio of the aforementioned candidate keywords in competitor products refers to the ratio of the number of competitor products that used the candidate keywords to the total number of competitor products.

[0089] The degree of concentration of traffic associated with candidate keywords among competing products refers to the distribution of the share of traffic obtained by each competing product using candidate keywords relative to the total associated traffic. The more concentrated the share distribution is among a few competing products, the higher the degree of concentration.

[0090] Specifically, the coverage ratio of the aforementioned candidate keywords in competitor products refers to the appearance of candidate keywords in the title of competitor products or in the keyword list of paid search ads placed by competitor products.

[0091] The title is detected by obtaining competitor product title text through the e-commerce platform's front-end search or product detail page API.

[0092] The advertising keywords are detected by obtaining a list of advertising keywords for competitor products through the e-commerce platform's advertising backend API or third-party advertising insight tools;

[0093] If a candidate keyword appears in at least one place in the title or ad keyword list of a competitor's product, it is determined that the competitor's product uses that candidate keyword.

[0094] Specifically, the aforementioned degree of concentrated occupancy is quantified by calculating the Herfindahl-Hirschman Index (HHI), using the following formula:

[0095] ;

[0096] In the formula, The Herfindahl-Hirschman Index has a value range of (0,1). The larger the value, the more concentrated the traffic. It quantifies the degree of concentration of keyword-related traffic among competitors and is used as a basis for deduction in the subsequent comprehensive score. For the first The percentage of related traffic received by each competitor's product from this keyword relative to the total related traffic for that keyword is used to reflect traffic distribution. The more uneven the distribution, the higher the probability of traffic loss. The higher; This represents the total number of competing products that used this keyword.

[0097] Specifically, the above-mentioned related traffic is obtained by: obtaining the search exposure or click volume of each competitor's product through each candidate keyword within a preset time period through the back-end data interface of the e-commerce platform or third-party data services, and using the exposure or click volume as the related traffic of the competitor's product under that keyword.

[0098] When it is impossible to obtain the precise search exposure or click volume of each competitor's product under candidate keywords through the e-commerce platform's backend data interface or third-party data services, the following alternative method is used to calculate the associated traffic share of each competitor:

[0099] Alternative Solution 1: Obtain the search ranking position of each competitor's product under the target keyword. Based on the preset empirical mapping relationship between search ranking and click-through rate, map the ranking position of each competitor's product to the estimated traffic share. The higher the ranking of the product, the higher the estimated traffic share it will receive.

[0100] Alternative Option 2: Utilize the exclusive data analysis tools provided by e-commerce platforms to obtain the search exposure and click volume of the brand's own products under target keywords. Combine this with publicly available competitor ranking information to make proportional estimates and obtain the estimated related traffic for each competitor.

[0101] For example, the top 50 products in terms of sales volume in the past 30 days under the "sports shoes / running shoes" category were selected as the competitor product set. For the candidate keyword "lightweight cushioning running shoes", it was found that 35 out of the 50 competitors used words related to "lightweight" or "cushioning" in their titles or advertising keywords, which is a coverage rate of 70%. Further, the search exposure traffic obtained by each of these 35 competitors through this keyword was obtained, and the sum of squares of the share of traffic of each competitor in the total traffic (HHI) was calculated. The calculated HHI value was 0.65, indicating that the traffic of this keyword is highly concentrated in a few top competitors, and the degree of concentration is high.

[0102] S4: Based on the pure conversion contribution value, usage coverage ratio, and concentration degree, the candidate keywords are comprehensively scored.

[0103] Specifically, the candidate keywords are comprehensively scored based on their contribution to net conversion, coverage ratio, and concentration level, as detailed below:

[0104] The contribution value of pure conversion is used as the benchmark score.

[0105] Based on the baseline score, additional points are added according to the coverage ratio, and the increment of additional points decreases as the coverage ratio increases.

[0106] The above results will be deducted based on the degree of concentrated occupancy; the higher the degree of concentrated occupancy, the greater the deduction.

[0107] Furthermore, the higher the contribution value of the pure conversion, the higher the score; the higher the concentration of occupation, the lower the score; and the improvement of the score by the coverage ratio decreases as the coverage ratio increases.

[0108] Specifically, the formula for calculating the above comprehensive score is as follows:

[0109] ;

[0110] In the formula, The overall score for keywords is used to measure their conversion potential and serves as the basis for final ranking and threshold screening. The contribution value to pure conversion serves as a benchmark score, reflecting the keyword's true conversion capability. To use the coverage ratio, additional points are added. The wider the coverage, the more points are added, but the marginal return decreases. The degree of concentration of occupancy is used for deduction; the higher the concentration, the greater the deduction. The coverage ratio weighting coefficient controls the positive influence of the coverage ratio on the score, with a value ranging from 0.2 to 0.6, preferably 0.4. The concentration penalty coefficient controls the negative penalty intensity of concentration on the score, with a value ranging from 0.3 to 0.8, preferably 0.5.

[0111] The comprehensive scoring in the above embodiment uses a multiplicative approach to create an interactive effect between the pure conversion contribution value, the coverage ratio bonus, and the concentration deduction. When the concentration level is high, the coverage ratio bonus effect is weakened simultaneously; when the concentration level is low, the coverage ratio bonus effect is fully preserved. This interactive effect is consistent with the actual business logic of e-commerce keyword placement, that is, in a highly monopolistic market environment, the coverage ratio bonus effect itself should be weakened, rather than maintaining an independent and constant contribution.

[0112] Among them, the logarithmic function in the above comprehensive scoring formula The selection criteria are as follows: the logarithmic function provides a smooth marginal diminishing effect, and the score increment when the coverage ratio increases from 0% to 20% is greater than the increment when it increases from 60% to 80%, which is in line with the business rule that the value of early coverage information is higher than that of later coverage information in the keyword competition landscape; at the same time, the increment of the logarithmic function when the coverage ratio approaches 0 is a limited value, avoiding the oversensitivity problem caused by the divergence of the derivative of the square root or power function in the low coverage area.

[0113] Furthermore, the aforementioned coverage ratio weighting coefficient and concentration penalty coefficient The method for determining the parameters is as follows: historical conversion data is divided into training set and validation set. A grid search is performed within the preset parameter candidate range, and the set of parameters that minimizes the average error between the comprehensive score on the validation set and the actual conversion performance is selected as the optimal parameters. When the amount of historical data is insufficient to support the above cross-validation, the initial value is selected from the recommended value table according to the competitive characteristics of the category of the target product, and dynamically corrected according to the actual prediction error during the system operation.

[0114] It should be noted that the above comprehensive score The unit of measurement is "equivalent number of transactions". Its physical meaning is: the effective conversion contribution of the candidate keyword after excluding external environmental interference and comprehensively considering the breadth of competitor coverage and the degree of traffic concentration. For example, if the comprehensive score of a keyword is 80, it means that its equivalent conversion contribution is equivalent to 80 transactions.

[0115] For example, the pure conversion contribution value of 100 calculated by S2 is used as the baseline score. Since the usage coverage of "lightweight cushioning running shoes" among competitors is 70%, an additional score is added on the basis of the baseline score. However, 70% coverage is already a high level, so the increase in the additional score slows down significantly (the marginal diminishing effect of the logarithmic function). At the same time, since the HHI index calculated by S3 is 0.65, it indicates that the traffic is highly concentrated in a few top competitors, so the comprehensive score of this keyword is significantly reduced. The comprehensive score of this keyword is calculated by a non-linear composite scoring formula (assumed to be 78 points). This score is used for subsequent dynamic adjustment and threshold determination.

[0116] S5: Track the trend of usage coverage ratio changes within a recent time window and adjust the overall score accordingly.

[0117] Specifically, the above tracking uses the trend of changes in coverage ratio within a recent time window, and adjusts the overall score based on this trend, as detailed below:

[0118] The recent time window is divided into multiple consecutive sub-periods, and the coverage ratio of candidate keywords in competitor products within each sub-period is statistically analyzed to obtain time series data of the coverage ratio.

[0119] Calculate the difference in coverage ratio between adjacent sub-time periods based on time series data;

[0120] If multiple consecutive differences are positive and the average of these positive values ​​exceeds a first preset change threshold, the trend is determined to be continuously rising. In this case, the attenuation rate is determined based on the average of the consecutive positive values; the larger the average, the greater the attenuation rate. The comprehensive score is then multiplied by an attenuation coefficient, which decreases as the attenuation rate increases. The specific formula for the attenuation coefficient is as follows:

[0121] ;

[0122] In the formula, This is the attenuation coefficient, with a value range of (0,1); The average of multiple consecutive positive differences (the average increase in the coverage ratio of adjacent sub-periods) reflects the slope of the increase in keyword competition intensity. The larger the value, the faster the market becomes a red ocean. This is the attenuation intensity factor (preset positive number), with a value range of 10~30, preferably 20, which controls the sensitivity of the attenuation amplitude to the rise rate. The larger the value, the stronger the decay for the same increase.

[0123] If multiple consecutive differences are negative and the average of these negative values ​​is below the second preset change threshold, the trend is determined to be a continuous decline. In this case, no attenuation adjustment is applied to the overall score, and the gain amplitude is determined based on the average of the consecutive negative values. The larger the absolute value of the negative value, the larger the gain amplitude. The overall score is multiplied by a gain coefficient, which increases with the increase in gain amplitude. The specific formula for the gain coefficient is as follows:

[0124] ;

[0125] In the formula, This is the gain coefficient, with a value range of (1.0, 1.5). The average of the absolute values ​​of multiple consecutive negative differences (i.e., the average of the rate of decline) reflects the speed at which the keyword competition intensity decreases; the larger the value, the more obvious the trend towards a blue ocean market. This is the gain intensity factor (a preset positive number), with a value ranging from 0.5 to 2, preferably 1.0, which controls the sensitivity of the gain amplitude to the descent rate. The larger the value, the stronger the gain for the same decrease.

[0126] If the average of multiple consecutive differences does not exceed the first preset change threshold and is not lower than the second preset change threshold, the change trend is determined to be no significant change, and no adjustment is made to the overall score, maintaining the overall score unchanged.

[0127] Specifically, the rules for determining the number of sub-periods are as follows: the number of sub-periods is dynamically determined based on the total length of the recent time window, each sub-period contains at least 2 days of data, and the total number of sub-periods is an integer between 5 and 14; when the length of the time window changes, the number of sub-periods is adjusted accordingly to ensure that each sub-period has sufficient data and that the total number of sub-periods meets the minimum requirements for trend determination.

[0128] Furthermore, the aforementioned attenuation coefficient The value range is set to between 0.5 and 1.0 (excluding 1.0). When calculated according to the formula... If the calculated value is lower than 0.5, it will be automatically truncated to 0.5.

[0129] Gain coefficient The value range is set to between 1.0 and 1.5 (excluding 1.0). When calculated according to the formula... If the calculated value is higher than 1.5, it will be automatically truncated to 1.5;

[0130] The above-mentioned cutoff protection mechanism is based on the following: In keyword selection scenarios, even if a keyword is in a rapid red ocean process, its comprehensive score should not be suppressed indefinitely, otherwise a large number of keywords will be directly judged as invalid, affecting the diversity of the candidate keyword pool of the advertising system; similarly, even if a keyword is in a rapid blue ocean process, its comprehensive score should not be unreasonably amplified, otherwise the advertising budget will be overly tilted towards keywords with high uncertainty, affecting the overall delivery efficiency.

[0131] The 50% lower bound and 50% upper bound are default values ​​set based on the conservative and stable principles of e-commerce advertising. The system allows users to customize and adjust them according to their own risk preferences.

[0132] When the attenuation coefficient or gain coefficient is truncated, the system automatically records the truncation event and outputs the corresponding warning or prompt information.

[0133] It should be noted that the number of consecutive positive differences mentioned above is either 3 or 5.

[0134] Among them, "multiple consecutive positive differences" means that at least 3 or 5 consecutive differences are required to determine the trend, that is, at least 4 or 6 sub-periods are required.

[0135] When the recent time window length is insufficient to divide the minimum number of sub-periods required to meet the judgment criteria, the following approach is adopted: First, try to shorten the data granularity of each sub-period (e.g., from 2 days to 1 day) to increase the number of sub-periods; if the minimum number of sub-periods still cannot be met, slide the time window forward from the current time as the endpoint, expanding by 1 day each time, until the window length is sufficient to divide the required number of sub-periods.

[0136] When the time window is extended to the starting point of historical data but still cannot meet the minimum sub-period requirement, the judgment condition will be temporarily adjusted from "multiple consecutive" to "two consecutive", and the preset change threshold will be increased accordingly (e.g., from one standard deviation to 1.5 standard deviation) to maintain the reliability of trend judgment.

[0137] Specifically, the first preset change threshold and the second preset change threshold are determined as follows:

[0138] The time series data of the coverage ratio is smoothed by moving average to filter out short-term random fluctuations and obtain a smoothed series.

[0139] Calculate the difference between adjacent sub-time periods in the smoothed sequence to obtain the difference sequence, and calculate the mean, standard deviation and standard error of the difference sequence;

[0140] The first preset change threshold is set as the product of the mean difference plus the critical value of the statistical distribution and the standard error; the second preset change threshold is set as the product of the mean difference minus the critical value of the statistical distribution and the standard error.

[0141] When the cumulative confidence level of multiple consecutive positive differences in the difference sequence exceeds the preset confidence level threshold, the trend is determined to be continuously rising; when the cumulative confidence level of multiple consecutive negative differences exceeds the preset confidence level threshold, the trend is determined to be continuously falling.

[0142] The pre-set reliability threshold corresponds to a significance level of 5%. When the number of sub-periods is insufficient to support statistical inference, the threshold is set to the standard deviation of the smoothed sequence difference with an additional 50% margin as a simplified judgment rule.

[0143] For example, the recent 14-day period (June 17th to June 30th) was divided into 7 sub-periods (each sub-period is 2 days). The usage coverage ratio of "lightweight cushioning running shoes" among competitors was statistically analyzed in each sub-period to obtain time series data. The coverage ratio difference between adjacent sub-periods was calculated. It was found that four consecutive differences were positive (+0.6%, +0.8%, +0.7%, +0.9%), and the average of each positive value was +0.75% / sub-period, which exceeded the preset change threshold (the standard deviation of time series data is 0.5% / sub-period). It was determined that the keyword was in the process of becoming a red ocean. In step S4, the keyword had a usage coverage ratio of 55% and an HHI index of 0.5, resulting in a comprehensive score of 85 points. After multiplying by the decay coefficient of 0.83 in step S5, the adjusted score dropped to about 70 points, still within the second level (60~80 points). It was finally determined to be a second-level high conversion potential keyword and recommended as a medium priority.

[0144] S6: Select candidate keywords whose adjusted overall scores reach the preset threshold and identify them as high-conversion-potential keywords and output them.

[0145] Specifically, candidate keywords whose adjusted overall scores reach a preset threshold are identified as high-conversion-potential keywords and output, as detailed below:

[0146] Two different scoring thresholds are preset, including a first threshold and a second threshold, with the first threshold being higher than the second threshold.

[0147] The adjusted overall score is compared with the first threshold, and candidate keywords that reach or exceed the first threshold are marked as first-level high conversion potential keywords.

[0148] Candidate keywords whose adjusted overall scores fall between the second and first thresholds are marked as second-level high-conversion-potential keywords.

[0149] Candidate keywords whose adjusted overall score is below the second threshold will be marked as invalid keywords and will not be included in the output list.

[0150] The marked candidate keywords and their corresponding comprehensive scores are output in descending order of level.

[0151] Specifically, the adjusted comprehensive score of all candidate keywords is calculated, and the top 20% percentile is taken as the first threshold, and the top 40% percentile is taken as the second threshold, which is applicable to conventional keyword selection scenarios.

[0152] Furthermore, both the first threshold and the second threshold mentioned above are screening thresholds based on "equivalent number of transactions". The first threshold represents the minimum number of equivalent transactions required to determine a keyword as a high-value keyword, and the second threshold represents the minimum number of equivalent transactions required to determine a keyword as a medium-value keyword.

[0153] The specific values ​​of the first and second thresholds mentioned above are related to the actual traffic volume of the target product: for products with larger traffic volume, the overall comprehensive score of their keywords is generally higher, and the threshold is increased accordingly; for products with smaller traffic volume, the threshold is decreased accordingly. The threshold is set by default using the quantile method—the comprehensive score of all candidate keywords is counted, and the top quantile is taken as the first and second thresholds, so that the thresholds are automatically adapted to the actual traffic volume of the product.

[0154] It should be noted that after identifying and outputting candidate keywords whose comprehensive scores after attenuation adjustment reach a preset threshold as high conversion potential keywords, the actual conversion performance data of each candidate keyword identified as high conversion potential keyword is obtained after the output.

[0155] The magnitude of the attenuation adjustment is corrected based on the degree of deviation between the actual conversion performance data and the overall score. If the actual conversion performance data is higher than the predicted value corresponding to the overall score, the attenuation magnitude under the same trend is reduced; if the actual conversion performance data is lower than the predicted value corresponding to the overall score, the attenuation magnitude under the same trend is increased.

[0156] Specifically, the method for correcting the attenuation adjustment range is as follows:

[0157] Set an observation period (preferably 7 days or 14 days) to obtain the actual click-through rate or actual transaction conversion rate of each output keyword within that period.

[0158] The actual conversion performance data is compared with the corresponding comprehensive score prediction value at the time of output, and the deviation rate is calculated.

[0159] When the deviation rate is greater than zero, it is determined that the actual performance is better than expected. The original attenuation coefficient is multiplied by an adjustment factor greater than 1, so that the attenuation amplitude is reduced under the same trend of change.

[0160] When the deviation rate is less than zero, it is determined that the actual performance is worse than expected. The original attenuation coefficient is multiplied by an adjustment factor less than 1, so that the attenuation amplitude increases under the same trend of change.

[0161] When the absolute value of the deviation rate is within the preset small fluctuation threshold range, the original attenuation coefficient remains unchanged.

[0162] The degree to which the adjustment factor deviates from 1.0 is positively correlated with the absolute value of the deviation rate.

[0163] Furthermore, the above deviation rate is calculated as follows: obtain the actual conversion rate of the candidate keyword within the observation period, subtract the predicted comprehensive score at the time of output, and then divide by the predicted value. The resulting ratio is the deviation rate. A positive deviation rate indicates that the actual performance is better than expected, and a negative deviation rate indicates that the actual performance is worse than expected.

[0164] It should be noted that the above actual conversion rates are click-through rates or transaction conversion rates obtained by e-commerce platform data tools, reflecting the real conversion effect of the keyword in the actual advertising environment; the comprehensive score prediction value is the comprehensive score corresponding to the keyword at the time of output, serving as a reference benchmark for judging the expected level.

[0165] Specifically, the correspondence between the above-mentioned comprehensive score prediction value and the predicted conversion rate is established by fitting historical campaign data: collect the comprehensive scores of the campaigned keywords and their corresponding actual conversion rates, and establish the mapping relationship between the comprehensive score and the predicted conversion rate through linear regression or nonlinear fitting methods; when the target product lacks sufficient historical campaign data, the average click-through conversion rate of the target product's category is used as a benchmark, and the deviation of the current keyword comprehensive score from the category's average comprehensive score is combined to make proportional adjustments. The degree of amplification or reduction of the predicted conversion rate due to the deviation is controlled by a preset sensitivity coefficient.

[0166] Furthermore, the rules for determining the values ​​of the aforementioned adjustment factors are as follows: they are set in tiers based on the absolute value of the deviation rate; no adjustment is made when the deviation rate is within a reasonable fluctuation range; moderate adjustment is made when the deviation rate exceeds the reasonable range but is within acceptable limits; and a larger adjustment is made when the deviation rate exceeds the acceptable limits. The direction of the adjustment factor deviating from 1.0 is consistent with the sign of the deviation rate, and the degree of deviation is positively correlated with the absolute value of the deviation rate. The magnitude of a single correction does not exceed the preset upper limit. When the direction of the deviation rate reverses, the correction is paused once, and the adjustment is made again after the trend is confirmed in the next observation period, in order to avoid excessive correction at the trend reversal point.

[0167] Specifically, the aforementioned correction period is determined based on the product category conversion cycle: for example, 7 days for short-cycle categories and 14 days for long-cycle categories, with 14 days as the default. When the number of clicks or completed orders for keywords falls below the preset threshold within the observation period, the data confidence level for that period is deemed insufficient, and the observation period is automatically extended until sufficient confidence data is collected. The system convergence criterion is that the absolute value of the deviation rate after multiple consecutive corrections is lower than the preset stability threshold. If, after convergence, the absolute value of the deviation rate exceeds the preset warning threshold multiple times, the correction mechanism is re-triggered.

[0168] For example, the overall score of 70 points after dynamic adjustment by S5 is compared with two preset level thresholds (first threshold 80 points, second threshold 60 points). Since 70 points falls between the second threshold of 60 points and the first threshold of 80 points, the candidate keyword "lightweight cushioning running shoes" is determined to meet the high conversion potential screening criteria, marked as a second-level high conversion potential keyword, and included in the medium-priority recommendation list. The system continues to execute steps S1 to S5 on other candidate keywords, ultimately filtering out those with an overall score reaching the first threshold (e.g., "road running shoes" score of 85 points) and the second threshold (e.g., "cushioning running shoes for men"). Keywords with scores of 72 and 70 (e.g., "lightweight cushioning running shoes") were categorized and output. Taking "road running shoes" as an example, the overall score for this keyword was 85 (corresponding to a predicted conversion rate of 3.4%). After the campaign was launched, the operations staff set a 14-day observation period and obtained the actual click-through conversion rate of 4.08% through the e-commerce platform's backend. The calculated deviation rate was +20%, indicating that the actual performance was better than expected. The original attenuation coefficient was multiplied by an adjustment factor of 1.1 to obtain a corrected attenuation coefficient, which reduced the attenuation amplitude under the same upward trend. The system parameters were continuously optimized based on feedback, and the screening accuracy was continuously improved.

[0169] In summary, this invention addresses the problem of insufficient signal-to-noise ratio in the evaluation module by acquiring external environmental parameters and eliminating their interference with historical conversion data, ensuring that the screening results reflect the true matching value between keywords and product attributes rather than the artificially high conversion rates. By simultaneously quantifying the coverage ratio and traffic concentration of keywords among competitors, it guides sellers to avoid keywords highly monopolized by leading competitors, solving the system's inability to identify the degree of competitive monopoly. Through an adaptive adjustment method that tracks the temporal trend of coverage ratio changes and applies attenuation when it is continuously rising and gain when it is continuously falling, the screening results have a forward-looking perspective in the time dimension, solving the problem of continuously expanding prediction bias in static models after changes in the competitive landscape. By grouping, correcting, and scoring historical conversion data according to attributes such as product category, functional selling points, applicable scenarios, and specifications, it achieves accurate matching with the specific attribute positioning of the target product, solving the problem of the system ignoring the differences in conversion contribution of keywords under different attribute dimensions. By pre-setting multi-level scoring thresholds for tiered output and using closed-loop feedback correction based on the actual conversion performance after keyword deployment, it not only meets the keyword selection needs of different advertising budgets but also allows the screening model to continuously evolve and improve accuracy in practical applications.

[0170] Example 2, an embodiment of the present invention, provides an intelligent screening system for high-conversion-potential keywords based on product attributes and best-selling data. The system includes: a data acquisition module for acquiring historical conversion data of candidate keywords and simultaneously acquiring external environmental parameters within the corresponding time period; a purity value calculation module for correcting historical conversion data based on external environmental parameters and calculating the pure conversion contribution value of candidate keywords based on the corrected conversion data; a competitor analysis module for acquiring the usage coverage ratio of candidate keywords among competitor products and the degree of concentration of traffic associated with the candidate keywords among various competitors; a comprehensive scoring module for comprehensively scoring candidate keywords based on their pure conversion contribution value, usage coverage ratio, and concentration degree; a trend adjustment module for tracking the changing trend of usage coverage ratio within a recent time window and adjusting the comprehensive score based on the changing trend; and a result output module for identifying and outputting candidate keywords whose adjusted comprehensive scores reach a preset threshold as high-conversion-potential keywords.

[0171] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. 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 be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligently selecting high-conversion-potential keywords based on product attributes and best-selling data, characterized by: include: Obtain historical conversion data for candidate keywords and simultaneously acquire external environmental parameters for the corresponding time period; The historical conversion data is corrected based on the external environment parameters, and the pure conversion contribution value of the candidate keywords is calculated based on the corrected conversion data. Obtain the coverage ratio of the candidate keywords in competing products, and the degree of concentration of traffic associated with the candidate keywords among various competing products; The candidate keywords are comprehensively scored based on the pure conversion contribution value, the usage coverage ratio, and the degree of concentrated occupation. Track the trend of the usage coverage ratio within a recent time window, and adjust the overall score based on the trend; Candidate keywords whose adjusted overall scores reach the preset threshold are identified as high-conversion-potential keywords and output.

2. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The historical conversion data is grouped and statistically analyzed according to product attribute dimensions; The product attributes include product category, functional selling points, applicable scenarios, and specifications. The external environmental parameters include product discount levels, promotional activity duration, seasonal change factors, and platform promotion indicators. The competing products refer to the set of products that belong to the same category as the target product and rank in the top N in terms of sales volume within a preset time period, where N is a preset positive integer.

3. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The process involves correcting the historical conversion data based on external environmental parameters, and then calculating the pure conversion contribution value of the candidate keywords based on the corrected conversion data. The specific details are as follows: Extract normal conversion period data that is not affected by external environmental factors from the historical conversion data, and calculate the baseline conversion value based on the normal conversion period data; Based on the external environment parameters, identify the interference period in the historical conversion data, compare the actual conversion value of the interference period with the benchmark conversion value, and calculate the interference deviation caused by the external environment interference. The interference deviation is removed from the total value of the historical conversion data to obtain the corrected conversion data; The clean conversion contribution value of the candidate keywords is calculated based on the corrected conversion data.

4. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 3, characterized in that: The normal conversion period unaffected by external environmental factors refers to a continuous period of time during which there are no product discount activities, no platform promotional activities, and no seasonal transition critical periods.

5. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The coverage ratio of the candidate keywords in competing products refers to the ratio of the number of competing products that used the candidate keywords to the total number of competing products. The degree of concentration of traffic associated with candidate keywords among competing products refers to the distribution of the share of traffic obtained by each competing product using the candidate keywords relative to the total associated traffic. The more concentrated the share distribution is among a few competing products, the higher the degree of concentration.

6. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The candidate keywords are comprehensively scored based on their pure conversion contribution value, usage coverage ratio, and concentration degree, as detailed below: The purity conversion contribution value is used as the benchmark score; Based on the baseline score, additional points are added according to the usage coverage ratio, and the increment of the additional points decreases as the usage coverage ratio increases. The above results are deducted based on the degree of concentrated occupancy; the higher the degree of concentrated occupancy, the greater the deduction.

7. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The tracking uses the trend of coverage ratio changes within a recent time window, and adjusts the overall score based on this trend, as detailed below: The recent time window is divided into multiple consecutive sub-periods, and the usage coverage ratio of the candidate keywords in competing products within each sub-period is statistically analyzed to obtain time series data of the coverage ratio. Based on the time series data, calculate the difference in coverage ratio between adjacent sub-time periods; If multiple consecutive differences are positive and the average of each positive value exceeds the first preset change threshold, then the change trend is determined to be continuously rising. At this time, the attenuation amplitude is determined according to the average value of the consecutive positive values. The larger the average value, the larger the attenuation amplitude. The comprehensive score is multiplied by the attenuation coefficient, and the attenuation coefficient decreases as the attenuation amplitude increases. If multiple consecutive differences are negative and the average of each negative value is lower than the second preset change threshold, then the change trend is determined to be a continuous decline. In this case, no attenuation adjustment is applied to the comprehensive score, and the gain amplitude is determined according to the average value of the consecutive negative values. The larger the absolute value of the negative value, the larger the gain amplitude. The comprehensive score is multiplied by the gain coefficient, and the gain coefficient increases with the increase of the gain amplitude. If the average of the consecutive differences does not exceed the first preset change threshold and is not lower than the second preset change threshold, then the change trend is determined to be without significant change, and no adjustment is made to the comprehensive score, maintaining the comprehensive score unchanged.

8. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 1, characterized in that: The process involves identifying and outputting candidate keywords whose adjusted overall scores reach a preset threshold as high-conversion-potential keywords, as detailed below: Two different levels of scoring thresholds are preset, including a first threshold and a second threshold, wherein the first threshold is higher than the second threshold; The adjusted comprehensive score is compared with the first threshold, and candidate keywords that reach or exceed the first threshold are marked as first-level high conversion potential keywords. Candidate keywords whose adjusted comprehensive scores fall between the second threshold and the first threshold are marked as second-level high-conversion-potential keywords. Candidate keywords whose adjusted overall score is lower than the second threshold are marked as invalid keywords and are not included in the output list; The marked candidate keywords and their corresponding comprehensive scores are output in descending order of level.

9. The intelligent keyword selection method for high conversion potential based on product attributes and best-selling data as described in claim 7, characterized in that: After identifying and outputting candidate keywords whose comprehensive scores after attenuation adjustment reach a preset threshold as high conversion potential keywords, the actual conversion performance data of each candidate keyword identified as high conversion potential keyword is obtained after output. Based on the degree of deviation between the actual conversion performance data and the comprehensive score, the magnitude of the attenuation adjustment is corrected. If the actual conversion performance data is higher than the predicted value corresponding to the comprehensive score, the attenuation magnitude under the same trend is reduced; if the actual conversion performance data is lower than the predicted value corresponding to the comprehensive score, the attenuation magnitude under the same trend is increased.

10. A high-conversion-potential keyword intelligent screening system based on product attributes and best-selling data, based on the high-conversion-potential keyword intelligent screening method based on product attributes and best-selling data as described in any one of claims 1 to 9, characterized in that: include, The data acquisition module is used to acquire historical conversion data of candidate keywords and simultaneously acquire external environmental parameters within the corresponding time period; The purity value calculation module is used to correct historical conversion data based on external environmental parameters, and calculate the purity conversion contribution value of candidate keywords based on the corrected conversion data. The competitor analysis module is used to obtain the coverage ratio of candidate keywords in competitor products, as well as the degree of concentration of traffic associated with the candidate keyword among various competitors. The comprehensive scoring module is used to comprehensively score candidate keywords based on their contribution to pure conversion, usage coverage ratio, and degree of concentrated usage. The trend adjustment module is used to track the trend of changes in the coverage ratio within a recent time window and adjust the overall score based on the trend. The results output module is used to identify and output candidate keywords whose adjusted comprehensive scores reach a preset threshold as high conversion potential keywords.

Images