Precision marketing recommendation method fused with privacy-enhancing technologies

Abstract

The application discloses a precision marketing recommendation method fusing privacy enhancement technology, and relates to the technical field of marketing recommendation.In the process of active browsing of a user, the minimum necessary data such as category, price, sliding speed and stay duration are collected silently, the false touch record is traced back and eliminated, the subsequent modeling is ensured to be clean and credible, the cleaned data is abstracted as double-dimension features, and the double-dimension features are uploaded to the cloud after encryption;the cloud adjusts the threshold value adaptively according to the coefficient of variation, continuously calibrates the false touch filtering strength, realizes the local sensitive data, and only the encrypted features are used for similarity calculation;highly similar groups are selected through the price preference distance and interest distribution difference weighting, and the recommended goods are dynamically distributed according to the similarity ratio;after the mixed arrangement, the recommended goods are sent to an independent recommendation page to reduce the homogenization exposure;the features and candidate set are updated incrementally as soon as the user finishes each screen sliding, a closed loop is formed, the original log is not landed throughout the whole process, there is no cross-page tracking, and there is no three-party data sharing, so that the dual requirements of compliance and experience are met.

Description

Technical Field

[0001] This invention belongs to the field of marketing recommendation technology, specifically, it relates to a precise marketing recommendation method that integrates privacy enhancement technology. Background Technology

[0002] Finding a balance between unlocking the value of data and protecting user privacy is a pressing issue that the marketing industry needs to address. Application CN120852015A discloses a method and system for accurate product recommendation based on data analysis. The method includes: acquiring users' historical browsing data, dividing operation thresholds, identifying accidental touch behaviors in the browsing data, and outputting accidental touch behavior data; filtering out accidental touch behavior data, acquiring target behavior data, constructing user features, and making product recommendations; detecting users' current browsing data, dynamically updating user features, acquiring updated product recommendation data, and outputting it. The system includes: an accidental touch behavior identification module, a user feature construction module, and a user feature update module; by identifying whether there are accidental touch behaviors in the browsing data, and then filtering out target behavior data, product recommendations are made for users based on the target behavior data, thereby improving the accuracy of product recommendations.

[0003] In practice, real-time monitoring of user actions can easily lead to privacy leaks during transmission, storage, and analysis. Secondly, the algorithm for judging accidental touches is prone to bias and discrimination, which can easily harm users' true interests. It judges accidental touches based solely on a single dimension such as dwell time and swipe speed combined with static thresholds, ignoring users' real behaviors such as fast browsing habits. This results in users' true interests being filtered out incorrectly. When dealing with elderly users, new users, and various special groups, the algorithm for judging accidental touches can cause recommendation bias for special groups, thus creating an information cocoon effect.

[0004] To address the aforementioned issues, this invention proposes a precise marketing recommendation method that integrates privacy enhancement technology. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a precise marketing recommendation method that integrates privacy enhancement technologies, solving the problems of excessive data noise, high privacy risks, and poor real-time recommendation performance in existing technologies.

[0006] The objective of this invention can be achieved through the following technical solutions: A precision marketing recommendation method incorporating privacy-enhancing technologies, the method comprising: Step 1: When the target user's main browsing interface is on the product page, start the data collection process to extract the target user's product browsing data and user behavior data, and build a combination of related data. Step 2: Based on the combination of related data of target users within the retrospective period, conduct comprehensive analysis to identify the accidental product browsing data of target users within the retrospective period, and perform separation operation. Based on the separated product browsing data and user behavior data, construct the target user characteristics. Step 3: Extract user features from several users and perform similarity verification with target user features to determine preferred user features, extract preferred products corresponding to preferred user features, combine target user features to generate a recommended list of preferred products associated with the target user within the backtracking period, and execute the recommendation operation; Step 4: Continuously extract the target user's characteristics and update the recommended product list when the target user finishes browsing the list of recommended products, and continue to perform the recommendation operation.

[0007] As a further embodiment of the present invention, the data acquisition process in step one is a sub-process of the product APP, which is controlled by the product APP; When the target user is using the product app and the target user is browsing the product page, the product app will initiate the data collection process and perform data collection operations.

[0008] As a further aspect of the present invention, the specific method for extracting the target user's product browsing data and user behavior data and constructing the associated data combination in step one is as follows: The current moment of the data acquisition process is obtained as the end time of a backtracking cycle. The duration of a backtracking cycle is traced back to the past data acquisition operations to determine the start time of the backtracking cycle, denoted as t1. Based on the duration within a backtracking cycle, the end time tj of the backtracking cycle is determined, where j is an integer preset by the operator, representing the total number of moments within the backtracking cycle. Moments not in the data acquisition process are not included in the backtracking cycle. Extract product browsing data and user behavior data of target users during the retrospective period. Product browsing data includes product category and product price, while user behavior data includes page scrolling speed and page dwell time. Extract all product categories of the target user within the backtracking period, arrange them randomly, and denote them as the product category sequence L=L1,L2,...,Lm, where m is the total number of product categories; Get the average price of all products under any product category Ln viewed by the target user, and denot it as the average price of the category associated with product category Ln, where n is the counting index, 1≤n≤m; Simultaneously determine the average price of the remaining categories to obtain the category average price sequence G=G1,G2,...,Gm; Within the backtracking period, obtain the page scrolling speed of the target user associated with the time period, and take the average value, which is recorded as the average page scrolling speed V associated with the target user within the backtracking period; Extract the browsing time of the target user for all products under any product category Ln, take the average value, and record it as the average time of the category associated with product category Ln. Simultaneously determine the average time of the other categories to obtain the category average time sequence T=T1,T2,...,Tm; The product category sequence L, the average price sequence G, the average page scrolling speed V, and the average time sequence T are summarized and denoted as the associated data combination ZH={L,G,V,T}.

[0009] As a further aspect of the present invention, in step two, the specific method for identifying the target user's accidental product browsing data within the traceback period and performing the separation operation is as follows: Based on the associated data combination built by the target user in the historical period, extract the time series of all historical categories and the historical average page scrolling speed data in the historical period. The historical period includes x backtracking periods, where x is a preset integer. Extract the page dwell time of the target user when browsing all products under product category Ln within the historical period, and calculate the mean and standard deviation, which are denoted as historical average browsing time μ_t and historical browsing time standard deviation σ_t, respectively. The false touch detection time threshold δ1 is calculated using δ1=μ_t-α*σ_t, where α is the first adjustment coefficient; Calculate the mean and standard deviation of the average page scrolling speed of all target users, denoted as historical average scrolling speed μ_v and historical scrolling speed standard deviation σ_v, respectively; The false touch detection sliding speed threshold δ2 is calculated using δ2=μ_v+β*σ_v, where β is the second adjustment coefficient; Get the page dwell time of any product under product category Ln of the target user within the traceback period, and compare it with the accidental touch judgment time threshold δ1. If the page dwell time of the corresponding product is <δ1, then mark the product as a pending accidental touch product; otherwise, do not process it. Compare the average page scrolling speed V within the backtracking period with the accidental touch judgment scrolling speed threshold δ2. If V is greater than δ2, generate a global rapid marker F. If a pending accidental touch product and a global urgent flag F exist in the same backtracking period, the product is determined to be an accidental touch product. The time the target user spends on the page corresponding to the product within the backtracking period is extracted and marked as accidental touch product viewing data. Obtain the product category of the product, and calculate the total page dwell time of all products under the product category viewed by the target user within the retrospective period. Calculate the ratio of the page dwell time of the product to the total page dwell time. If the ratio is greater than θ%, mark the product category of the product as the data of the product viewed by accident, where θ% is a preset percentage. Repeat the above steps to identify all accidental product viewing data of the target user during the backtracking period, and perform a separation operation, which specifically involves deleting the accidental product viewing data from the product viewing data and user behavior data.

[0010] As a further aspect of the present invention, the specific method for determining the first adjustment coefficient α and the second adjustment coefficient β in step two is as follows: Calculate the coefficient of variation of page dwell time: CV_t = σ_t / μ_t; Calculate the coefficient of variation of the average page scrolling speed: CV_v = σ_v / μ_v; When CV_t is less than the preset first stability threshold, the first adjustment coefficient α is set to the first value; When CV_t is greater than or equal to the first stability threshold, the first adjustment coefficient α is set to a second value that is greater than the first value; When CV_v is less than the preset second stability threshold, the second adjustment coefficient β is set to the third value; When CV_v is greater than or equal to the second stability threshold, the second adjustment coefficient β is set to a fourth value that is greater than the third value.

[0011] As a further aspect of the present invention, the specific method for constructing target user characteristics based on the separated product browsing data and user behavior data in step two is as follows: The product category sequence L', category average price sequence G', and category average time sequence T' after separation are combined with the average page scrolling speed V to form a related data combination ZH'={L',G',V,T'}; Let k be the total number of product categories in the product category sequence L'. Then L' = L1', L2', ..., Lk', G' = G1', G2', ..., Gk', and T' = T1', T2', ..., Tk'. Using PG=(∑ k o=1 The mean preference price PG is calculated as (Go') / k, where o is the counting index, and 1 ≤ o ≤ k; CI=[∑ k o=1 (To'-(∑ k o=1 To' / k))] / k calculates the concentration of browsing interest CI of the target user for any product category Lo', and forms a combination CI_Lo' of product category browsing interest concentration with product category Lo'; Similarly, determine the combination of browsing interest concentration for k product categories to form the product category browsing interest concentration set RCI={CI_L1',CI_L2',...,CI_Lk'}; The target user characteristics FM={PG,RCI} are constructed based on the concentration of browsing interest in product categories (RCI) and the mean preferred price (PG), and are stored in encrypted cloud storage.

[0012] As a further aspect of the present invention, the specific method for determining the preferred user characteristics in step three is as follows: Based on the cloud, extract the user features of any non-target user, denoted as FM_u={PG_u,RCI_u}, where u is the user identifier; The price preference similarity score S_pri is calculated using S_pri=1 / (1+|PG_u-PG|). Extract the product categories that are common to RCI_u and RCI, form a common category set CL, count the total number of common categories, denoted as cl, and denote the common category proportion factor as |cl| / k; For any product category in the common category set CL, take its browsing interest concentration in RCI_u and RCI, and denote them as CI and CI_u respectively; The difference in interest concentration (DCI) is calculated using DCI = |CI_u-CI|. Similarly, determine the difference in interest concentration across all common categories, take the average, and denot it as the average interest difference degree AD; The interest distribution similarity score S_int is calculated using S_int=(|cl| / k)×(1 / (1+AD)). The comprehensive similarity Sim is calculated using Sim=w1×S_pri+w2×S_int, where w1 and w2 are preset weight coefficients, and w1+w2=1; Based on the above method, the comprehensive similarity between b users and the target user is obtained, sorted in descending order, and the user features of the top h users are taken as the preferred user features, which are combined into the preferred user feature set {FM_1,FM_2,...,FM_h}, where b and h are preset integers, and h≤b.

[0013] As a further aspect of the present invention, in step three, a list of preferred products associated with the target user within the retrospective period is generated, and the specific method for performing the recommendation operation is as follows: Extract the preferred user feature set {FM_1,FM_2,...,FM_h}, and obtain the comprehensive similarity corresponding to each preferred user feature, denoted as Sim_1,Sim_2,...,Sim_h; Calculate the ratio between h comprehensive similarities and perform normalization so that the sum of the ratios is one, to obtain the ratio sequence bz_1, bz_2, ..., bz_h; Generate a list of recommended preferred products to be populated (List), and determine the total number of preferred products, denoted as sum; The preferred product quantity is allocated to each user feature according to the ratio sequence bz_1, bz_2, ..., bz_h, resulting in sum_1, sum_2, ..., sum_h. If there are decimals, they are rounded up, and the total number of preferred products in the preferred product recommendation list List is expanded accordingly: sum_1 + sum_2 + ... + sum_h = sum. Obtain the number of preferred products sum_1 corresponding to the preferred user feature FM_1, extract the product category browsing interest concentration set RCI_1 from the preferred user feature FM_1, allocate the number of preferred products sum_1 to each product category based on the ratio between the browsing interest concentrations of each product category, and extract the corresponding number of products that the user has browsed according to the preferred user feature FM_1 from each product category, which are denoted as preferred products. If the ratio allocation has a decimal, it is rounded down. Similarly, determine the preferred products associated with all users corresponding to the preferred user feature set {FM_1,FM_2,...,FM_h}, add them to the preferred product recommendation list List in random order, display them on the target user's recommended product page, and perform the recommendation operation. The recommended product page is different from the product page.

[0014] As a further aspect of the present invention, in step four, when the target user finishes browsing the list of recommended products, the target user's product browsing data on the product page is continuously extracted, the target user's characteristics are updated, the list of recommended products is updated, and the recommendation operation is continuously performed on the target user.

[0015] The beneficial effects of this invention are: This invention, by integrating privacy enhancement technologies, achieves a dual improvement in the security and effectiveness of user data during the precision marketing recommendation process. Its advantages lie in its simultaneous emphasis on privacy protection when collecting user behavior data. By identifying and separating accidental touches, it effectively filters noisy data, thereby constructing more authentic and high-quality user characteristics. Based on this, it generates dynamically updated recommendation lists through similarity matching, which not only improves the accuracy of personalized recommendations but also enhances the system's adaptability through continuous feedback and optimization. While protecting user privacy, it improves the relevance of recommendations and user experience, achieving a balance between privacy protection and marketing effectiveness. This invention achieves accurate and efficient user behavior capture by setting data collection as a controllable subprocess of the product app. Its advantages include triggering data collection only when users browse product pages, avoiding unnecessary resource consumption and illegal user behavior monitoring. Furthermore, it integrates time-series behavioral data through a backtracking cycle, ensuring the continuity and completeness of information. By constructing associated data combinations, it not only links static attributes such as product category and average price but also integrates dynamic behavioral indicators such as page scrolling speed and dwell time, making the analysis of user interests and consumption intentions more comprehensive and accurate. This invention purifies user behavior data by accurately identifying and separating accidental product browsing data, removing noise introduced by erroneous operations, and improving the accuracy of subsequent analysis. Based on historical data, it dynamically calculates the accidental touch judgment threshold and uses the coefficient of variation to adaptively adjust the adjustment coefficient, making the accidental touch identification process more flexible and reliable, and able to adapt to the stable behavior patterns of different users. On this basis, it constructs target user characteristics to deeply reflect users' real interests and consumption tendencies, providing a data foundation for personalized recommendations. This invention achieves more accurate and personalized product recommendations by comprehensively considering users' price preferences and the similarity of their interest distribution. Its advantages lie in not only basing recommendations on shared product categories but also deeply analyzing the differences in user interest concentration across these categories, thus capturing interest preferences more meticulously. It also incorporates price sensitivity to improve similarity calculation; selects superior users through weighted comprehensive similarity screening; and dynamically allocates the number of recommended products based on similarity ratios and interest concentration, effectively integrating group behavioral characteristics to generate a diverse and relevant recommendation list for target users. Furthermore, by establishing a separate recommendation page, it avoids interference with regular browsing pages, thus optimizing the user experience. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart illustrating the method described in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this application provides a precise marketing recommendation method that integrates privacy enhancement technologies; As an embodiment 1 of this application, it specifically includes: Step 1: When the target user's main browsing interface is on the product page, start the data collection process to extract the target user's product browsing data and user behavior data, and build a combination of related data. Step 2: Based on the combination of related data of target users within the retrospective period, conduct comprehensive analysis to identify the accidental product browsing data of target users within the retrospective period, and perform separation operation. Based on the separated product browsing data and user behavior data, construct the target user characteristics. Step 3: Extract user features from several users and perform similarity verification with target user features to determine preferred user features, extract preferred products corresponding to preferred user features, combine target user features to generate a recommended list of preferred products associated with the target user within the backtracking period, and execute the recommendation operation; Step 4: Continuously extract the target user's characteristics and update the recommended product list when the target user finishes browsing the list of recommended products, and continue to perform the recommendation operation.

[0020] Example 2 This embodiment, based on embodiment 1, further discloses a method for extracting product browsing data and user behavior data of target users and constructing a combination of related data, specifically including the following: This solution is mainly applied to product apps. When the target user is using the product app, and only when the main browsing interface of the target user's device terminal (such as mobile phone, tablet, etc.) is on the product page, the product app will activate the data collection process and perform data collection operations. The data acquisition process runs as an independent sub-process of the product APP, interacting with the main process of the product APP through a secure inter-process communication mechanism. The start, stop, data acquisition frequency, and content range of this sub-process are controlled by the main logic of the product APP according to the operator's preset strategy and the user's authorization status, ensuring that the acquisition behavior is within the controllable range of the main application.

[0021] It should also be noted that the data collection process is only granted the minimum necessary permissions, and can only access the data necessary to complete its function, such as the current product page information, and cannot directly access sensitive personal data such as user identity information, address book, and SMS messages.

[0022] The determination method for determining whether a target user primarily uses the product app requires that the product app is in the foreground of the terminal, and the product app displays a product page, which can be further divided into: product details page, product list page, and search results page, rather than login page, personal center, customer service chat page, etc. This logic can be implemented through the internal verification process of the product app.

[0023] When the conditions for using the main product app are met simultaneously, the main process of the product app sends a wake-up command containing a session ID to the data collection subprocess as a temporary unique identifier for this data collection, and it is not directly bound to the user's permanent identity.

[0024] Next, when the data acquisition process begins, the first moment of the data acquisition process execution, i.e. the current moment, is extracted as the end time of a backtracking cycle. Then, the duration of a backtracking cycle is traced back to the past data acquisition operations (only backtracking data acquisition operations) to determine the start time of the backtracking cycle, denoted as t1. Next, based on the start time t1, mark the end time tj, which is the end time. Here, j is an integer preset by the operator, representing the total number of times within the backtracking period. Times not in the data acquisition process are not included in the backtracking period. For example, if the user has not been in a data collection process within the past 3 minutes, the process continues to backtrack until it is confirmed to be a data collection process, excluding the time when the target user switches applications or browses non-product pages.

[0025] After determining the backtracking period, extract the target user's product browsing data and user behavior data within the backtracking period. The product browsing data includes product category and product price, while the user behavior data includes page scrolling speed and page dwell time. It should be noted that the product category is a standardized classification inherent to the product app and is displayed hierarchically, such as: Electronics -> Mobile Phones -> Smartphones. The product price is the listed price of the product. The page scrolling speed is calculated by listening to touch events or scroll view events and calculating the pixel displacement per unit time. The page dwell time is the time interval between when the product page enters the terminal's display area and when it leaves the display area. All of the above data can be implemented based on existing technologies. Therefore, this solution will not elaborate on them and will consider them as known data.

[0026] Next, extract all product categories of the target user within the backtracking period, and randomly arrange all product categories to obtain a sequence, denoted as product category sequence L=L1,L2,...,Lm, where m is the total number of product categories browsed by the target user within the backtracking period.

[0027] Then, obtain the prices of all products under any product category Ln in the product category sequence L=L1,L2,...,Lm that the target user browses within the backtracking period, and perform average processing to obtain an average value, which is denoted as the average price of the category associated with product category Ln, where n is the counting index, and the value of n ranges from 1 to m.

[0028] Following the above method, the average price of all product categories in the product category sequence L=L1,L2,...,Lm is determined, and the average price sequence is denoted as G according to the order of the product category sequence L, which is: G=G1,G2,...,Gm; Next, the page scrolling speed of the target user associated with time is obtained within the backtracking period, and then averaged to obtain the average page scrolling speed V associated with the target user within the backtracking period. Finally, extract the browsing time of the target user for all products under any product category Ln, and average it to form the average browsing time of the category associated with product category Ln. Then, in the same way, determine the average browsing time of all product categories associated with each other to obtain the average browsing time sequence T=T1,T2,...,Tm. The results obtained above are summarized, including the product category sequence L, the category average price sequence G, the average page scrolling speed V, and the category average time sequence T. The combination is denoted as the associated data combination ZH={L,G,V,T}.

[0029] Example 3 This embodiment, based on embodiment 2, further discloses a method for constructing target user characteristics based on product browsing data and user behavior data, specifically including the following: Extract the relevant data groups of the target user within the backtracking period, and perform accidental product viewing data identification, extract the accidental product viewing data, and update the relevant data groups; First, based on the associated data combinations that the target user has built within the historical period, extract all historical category time series and historical average page scrolling speed data within the historical period. Here, a historical period includes x backtracking periods, where x is a preset integer. It should be noted that at this time, the associated data combinations built within the historical period include the associated data combinations within the current backtracking period. In other words, the time span of the historical period includes the entire backtracking period.

[0030] Next, the page dwell time of the target user in browsing all products under product category Ln within the historical period is extracted, and the average and standard deviation of the page dwell time are calculated and denoted as historical average browsing time μ_t and historical browsing time standard deviation σ_t, respectively. The historical average browsing time μ_t represents the normal average time of the target user browsing a certain product category within the historical period, and the historical browsing time standard deviation σ_t reflects the fluctuation of the target user's browsing time in that product category. Based on the historical average browsing time μ_t and the historical browsing time standard deviation σ_t, the threshold δ1 for accidental touch judgment associated with the target user is calculated by using: δ1=μ_t-α*σ_t, where α is the first adjustment coefficient used to adjust the historical browsing time standard deviation σ_t. The larger the value of α, the stricter the subsequent judgment conditions, and vice versa.

[0031] Then calculate the average and standard deviation of the target user's average page scrolling speed over the historical period, and denot them as historical average scrolling speed μ_v and historical scrolling speed standard deviation σ_v, respectively. Next, the swiping speed threshold δ2 associated with the target user is calculated by using: δ2=μ_v+β*σ_v, where β is the second adjustment coefficient used to adjust the standard deviation of historical swiping speed σ_v. The larger the value of β, the stricter the subsequent judgment conditions, and vice versa.

[0032] Next, extract the page dwell time of the target user when browsing any product under any product category Ln within the current backtracking period, and compare it with the accidental touch judgment time threshold δ1. If the page dwell time of the product is less than the accidental touch judgment time threshold δ1, the product is marked as a pending accidental touch product; otherwise, no processing is performed. Next, the average page scrolling speed V of the target user in the current backtracking period is compared with the accidental touch judgment scrolling speed threshold δ2. If the average page scrolling speed V is greater than δ2, a global urgent flag F is generated; otherwise, no processing is performed. If a product to be identified as a mistakenly touched and a global urgent tag F exist in the same backtracking period, the product is determined to be a mistakenly touched product. At the same time, the page dwell time of the target user for the corresponding product in the backtracking period is extracted and marked as mistakenly touched product viewing data. That is, the logic of simultaneously satisfying the dual conditions of a short dwell time for a specific product and an overall browsing pace that is too fast is used to improve the accuracy of mistaken touch identification and reduce the possibility of misjudging the target user's real points of interest as noise.

[0033] At the same time, obtain the product category of the product; The system calculates the total time a target user spends browsing all products in the product category within this backtracking period, and then calculates the ratio of the time spent browsing the product's page to the total time spent browsing the page. If the calculated ratio is greater than θ%, the product category is marked as accidentally viewed product data; otherwise, no action is taken. Here, θ% is a preset percentage for the operator. For example, if a target user accidentally clicks on the mobile phone category and only browses one mobile phone, then the mobile phone product and the mobile phone category will be used together as the data for the accidentally clicked product. By repeating the above steps, all accidental product viewing data of the target user during the tracing period can be identified, and separation operations can be performed on this accidental product viewing data. It should be noted that the separation operation specifically involves deleting the data of accidentally viewed products from the product viewing data and user behavior data. This step is carried out based on the local data processing pipeline of the product APP. The data of accidentally viewed products will not enter the subsequent feature building process, nor will it be persistently stored or uploaded.

[0034] In the above method, the second adjustment coefficient β and the first adjustment coefficient α are determined as follows: First, the historical average browsing time μ_t and the historical browsing time standard deviation σ_t are extracted. The coefficient of variation CV_t associated with the page dwell time is calculated by using CV_t=σ_t / μ_t. Then, extract the historical average scroll speed μ_v and the historical scroll speed standard deviation σ_v, and calculate the coefficient of variation CV_v of the average page scroll speed by using CV_v=σ_v / μ_v.

[0035] Next, perform the following verification operation: If CV_t is less than the first stability threshold preset by the operator, the first adjustment coefficient α is set to the first value. If CV_t is greater than or equal to the first stability threshold, the first adjustment coefficient α is set to the second value, which is greater than the first value. Both the first and second values ​​are preset. It should be explained that if CV_t is less than the first stability threshold preset by the operator, it means that the user's browsing history is very stable. At this time, even a slight deviation may be an abnormal accidental touch. Therefore, a larger α value is used to tighten the threshold. If CV_t is large, it means that the user's browsing time itself fluctuates greatly. The judgment of accidental touch needs to be more cautious. Therefore, a smaller α value is used to relax the threshold.

[0036] If CV_v is less than the preset second stability threshold, the second adjustment coefficient β is set to the third value; if CV_v is greater than or equal to the second stability threshold, the second adjustment coefficient β is set to a fourth value greater than the third value. The third and fourth values ​​are both preset and are independent of the first and second values.

[0037] At this point, having identified all the accidentally viewed product data, we can obtain the product category sequence after the separation operation, denoted as L', the category average price sequence, denoted as G', and the category average time sequence, denoted as T'. Finally, the above parameters and the average page scrolling speed V are combined to form the associated data combination ZH'={L',G',V,T'}; Next, count the total number of product categories in the product category sequence L', denoted as k. Then L'=L1',L2',...,Lk', G'=G1',G2',...,Gk',T'=T1',T2',...,Tk'; Next, by using PG=(∑ k o=1 The mean price preference (PG) is calculated using the formula Go') / k, where o is the counting index, 1 ≤ o ≤ k. The mean price preference (PG) reflects the overall price range preference of the target user for the products viewed in the current period and is considered a trend indicator.

[0038] Then by using CI=[∑ k o=1 (To'-(∑ k o=1 To' / k))] / k calculates the concentration of browsing interest CI for any product category Lo' of the target user. It represents the average absolute deviation between the browsing time of each category and the average browsing time in the period. It forms a combination of browsing interest concentration CI_Lo' with the corresponding product category Lo'. The larger the value of the concentration of browsing interest CI, the more uneven the distribution of the target user's browsing time among different categories in the retrospective period, and the more concentrated the interest is on certain specific categories. The smaller the CI value, the more even the distribution of browsing time and the more diffuse the interest. The combination CI_Lo', representing the concentration of browsing interest in product categories, indicates the correlation between the corresponding product category and the concentration of interest of the target user.

[0039] Based on the method of determining the product category browsing interest concentration combination CI_Lo', k product category browsing interest concentration combinations are determined to form the product category browsing interest concentration set RCI={CI_L1',CI_L2',...,CI_Lk'}; Finally, based on the concentration of browsing interest in product categories (RCI) and the mean preferred price (PG), the target user characteristics (FM={PG,RCI}) associated with the target user during this retrospective period are constructed and stored in the cloud with encryption.

[0040] The encrypted target user feature FM is uploaded to the cloud server. The cloud uses an anonymous feature identifier generated and encrypted by the device and associated with the target user on a periodic rotating basis to associate the encrypted target user feature FM.

[0041] In this way, the server can only store and aggregate these encrypted target user characteristics by identifier for subsequent group analysis.

[0042] Example 4 This embodiment, based on embodiment 3, further discloses a method for constructing a preferred product recommendation list and performing a recommendation operation on a target user, specifically including the following: First, when it is necessary to perform a recommendation operation for any target user, the user characteristics of any non-target user are obtained from the cloud and denoted as FM_u={PG_u,RCI_u}, where u is the user identifier of the user for easy differentiation; Next, extract the target user features FM={PG,RCI}; The price preference similarity score S_pri between the target user and the user u is calculated by using: S_pri=1 / (1+|PG_u-PG|). Then extract the product categories that are common to RCI_u and RCI, and form a common category set, denoted as CL. Count the total number of common categories in the common category set CL, denoted as cl, and denote the common category proportion factor as |cl| / k. For any product category in the common category set CL, extract the browsing interest concentration of that product category in RCI_u, denoted as CI_u, and then extract the browsing interest concentration of that product category in RCI, denoted as CI.

[0043] The interest concentration difference (DCI) between the target user and user u is calculated by using the formula: DCI = |CI_u - CI|. By repeating the above steps, the interest concentration differences of all common categories are determined, and the average value is taken as the average interest difference (AD). Then, the interest distribution similarity score S_int is calculated using S_int=(|cl| / k)×(1 / (1+AD)), and combined with S_pri, the comprehensive similarity Sim is calculated based on Sim=w1×S_pri+w2×S_int. Here, w1 and w2 are weight coefficients preset by the operator, and w1+w2=1. The weight coefficients w1 and w2 are adjusted according to the business scenario. For example, luxury goods recommendations focus on price similarity w1, while content-based products focus on interest similarity w2.

[0044] Using the above method, the comprehensive similarity between b users and the target user is extracted and calculated, resulting in b comprehensive similarity scores for the b users. The user features of the b users are then sorted from largest to smallest according to the comprehensive similarity scores, and the user features of the top h users are taken as the preferred user features. The h preferred user features are combined and denoted as the preferred user feature set {FM_1,FM_2,...,FM_h}, where b and h are preset integers, and h≤b.

[0045] Next, the comprehensive similarity corresponding to all preferred user features in the preferred user feature set {FM_1,FM_2,...,FM_h} is extracted and sorted according to the arrangement order of the preferred user feature set {FM_1,FM_2,...,FM_h}, denoted as Sim_1,Sim_2,...,Sim_h.

[0046] Next, the ratios between the h comprehensive similarities are calculated and normalized so that the sum of the h ratios is one, thus obtaining the ratio sequence bz_1, bz_2, ..., bz_h; A list of recommended products to be populated is generated. The order in which the target user scrolls through the product page to view the products will be the order of the recommended products list. The total number of recommended products is determined and denoted as sum. The value of sum can be preset by the operator, but the final specific value still needs to be determined by subsequent calculations.

[0047] Using the ratios of the ratio sequence bz_1, bz_2, ..., bz_h as the allocation proportions, the number of preferred products is allocated to each user feature. The final number of preferred products corresponding to each ratio in the ratio sequence bz_1, bz_2, ..., bz_h is represented as sum_1, sum_2, ..., sum_h. It's important to note that if decimals occur during the calculation, they are eliminated by rounding up (e.g., 10.1 becomes 11 after rounding up). Simultaneously, the total number of preferred products, sum, in the preferred product recommendation list List is expanded accordingly to satisfy sum_1 + sum_2 + ... + sum_h = sum.

[0048] Next, the number of preferred products sum_1 corresponding to the first preferred user feature FM_1 is extracted as an example for processing, and the number of preferred products corresponding to the other preferred user features is processed in the same way.

[0049] Obtain the set of product category browsing interest concentration in the preferred user feature FM_1, denoted as RCI_1. Based on the ratio between the browsing interest concentration of each product category, allocate the number of preferred products sum_1 to each product category. Extract the corresponding number of products that the user corresponding to the preferred user feature FM_1 has browsed from each product category, and denot them as preferred products. For example: If the preferred number of products sum_1 = 10, and there are 3 product categories, and the ratio of browsing interest concentration among product categories is 0.1:0.4:0.5, then the first product category is allocated 1 product, the second product category is allocated 4 products, and the first product category is allocated 5 products. If the ratio allocation has a decimal, it is rounded down. For example, if the ratio of the first product category is 0.11, then the allocated product quantity is 1.1. After rounding down, the allocated product quantity is 1.

[0050] Repeat the above steps to determine all the preferred products associated with the preferred user feature set {FM_1,FM_2,...,FM_h}, and add them to the preferred product recommendation list List in a random order. Based on the preferred product recommendation list List, display it to the target user's recommended product page and perform the recommendation operation. The recommended product page is different from the product page to ensure that the user knows that this is generated recommendation content, which complies with the principle of transparency.

[0051] When the target user finishes browsing the recommended product list, the system continuously extracts the target user's product browsing data on the product page, updates the target user's characteristics and the recommended product list, and continuously performs recommendation operations to the target user. At the same time, the operator can also automatically set the cycle for updating the target user's characteristics and the recommended product list, and perform the update operation of recommended products before the target user finishes browsing the recommended product list.

[0052] All data in the formulas described above have been calculated with dimensions removed. Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0053] The above description is merely an example and illustration of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0054] It should be stated that all user data collected in this application was collected with the user's consent and authorization. Furthermore, the uses of user data are legal and compliant, and the use and processing of user data comply with the relevant laws, regulations, and standards of the relevant regions.

Claims

1. A precision marketing recommendation method integrating privacy enhancement technology, characterized in that, The method includes: Step 1: When the target user's main browsing interface is on the product page, start the data collection process to extract the target user's product browsing data and user behavior data, and build a combination of related data. Step 2: Based on the combination of related data of target users within the retrospective period, conduct comprehensive analysis to identify the accidental product browsing data of target users within the retrospective period, and perform separation operation. Based on the separated product browsing data and user behavior data, construct the target user characteristics. Step 3: Extract user features from several users and perform similarity verification with target user features to determine preferred user features, extract preferred products corresponding to preferred user features, combine target user features to generate a recommended list of preferred products associated with the target user within the backtracking period, and execute the recommendation operation; Step 4: Continuously extract the target user's characteristics and update the recommended product list when the target user finishes browsing the list of recommended products, and continue to perform the recommendation operation.

2. The method according to claim 1, characterized in that, The data acquisition process in step one is a sub-process of the product APP and is controlled by the product APP. When the target user is using the product app and the target user is browsing the product page, the product app will initiate the data collection process and perform data collection operations.

3. The method according to claim 2, characterized in that, In step one, the specific method for extracting the target user's product browsing data and user behavior data to construct the associated data combination is as follows: The current moment of the data acquisition process is obtained as the end time of a backtracking cycle. The duration of a backtracking cycle is traced back to the past data acquisition operations to determine the start time of the backtracking cycle, denoted as t1. Based on the duration within a backtracking cycle, the end time tj of the backtracking cycle is determined, where j is an integer preset by the operator, representing the total number of moments within the backtracking cycle. Moments not in the data acquisition process are not included in the backtracking cycle. Extract product browsing data and user behavior data of target users during the retrospective period. Product browsing data includes product category and product price, while user behavior data includes page scrolling speed and page dwell time. Extract all product categories of the target user within the backtracking period, arrange them randomly, and denot them as the product category sequence L=L1,L2,...,Lm, where m is the total number of product categories; Get the average price of all products under any product category Ln viewed by the target user, and denot it as the average price of the category associated with product category Ln, where n is the counting index, 1≤n≤m; Simultaneously determine the average price of the remaining categories to obtain the category average price sequence G=G1,G2,...,Gm; Within the backtracking period, obtain the page scrolling speed of the target user associated with the time period, and take the average value, which is recorded as the average page scrolling speed V associated with the target user within the backtracking period; Extract the browsing time of the target user for all products under any product category Ln, take the average value, and record it as the average time of the category associated with product category Ln. Simultaneously determine the average time of the other categories to obtain the category average time sequence T=T1,T2,...,Tm; The product category sequence L, the average price sequence G, the average page scrolling speed V, and the average time sequence T are summarized and denoted as the associated data combination ZH={L,G,V,T}.

4. The method according to claim 3, characterized in that, In step two, the specific method for identifying and separating the target user's accidental product browsing data within the traceback period is as follows: Based on the associated data combination built by the target user in the historical period, extract the time series of all historical categories and the historical average page scrolling speed data in the historical period. The historical period includes x backtracking periods, where x is a preset integer. Extract the page dwell time of the target user when browsing all products under product category Ln within the historical period, and calculate the mean and standard deviation, which are denoted as historical average browsing time μ_t and historical browsing time standard deviation σ_t, respectively. The false touch detection time threshold δ1 is calculated using δ1=μ_t-α*σ_t, where α is the first adjustment coefficient; Calculate the mean and standard deviation of the average page scrolling speed of all target users, denoted as historical average scrolling speed μ_v and historical scrolling speed standard deviation σ_v, respectively; The false touch detection sliding speed threshold δ2 is calculated using δ2=μ_v+β*σ_v, where β is the second adjustment coefficient; Get the page dwell time of any product under product category Ln of the target user within the traceback period, and compare it with the accidental touch judgment time threshold δ1. If the page dwell time of the corresponding product is <δ1, then mark the product as a pending accidental touch product; otherwise, do not process it. Compare the average page scrolling speed V within the backtracking period with the accidental touch judgment scrolling speed threshold δ2. If V is greater than δ2, generate a global rapid marker F. If a pending accidental touch product and a global urgent flag F exist in the same backtracking period, the product is determined to be an accidental touch product. The time the target user spends on the page corresponding to the product within the backtracking period is extracted and marked as accidental touch product viewing data. Obtain the product category of the product, and calculate the total page dwell time of all products under the product category viewed by the target user within the retrospective period. Calculate the ratio of the page dwell time of the product to the total page dwell time. If the ratio is greater than θ%, mark the product category of the product as the data of the product viewed by accident, where θ% is a preset percentage. Repeat the above steps to identify all accidental product viewing data of the target user during the backtracking period, and perform a separation operation, which specifically involves deleting the accidental product viewing data from the product viewing data and user behavior data.

5. The method according to claim 4, characterized in that, In step two, the specific method for determining the first adjustment coefficient α and the second adjustment coefficient β is as follows: Calculate the coefficient of variation of page dwell time: CV_t = σ_t / μ_t; Calculate the coefficient of variation of the average page scrolling speed: CV_v = σ_v / μ_v; When CV_t is less than the preset first stability threshold, the first adjustment coefficient α is set to the first value; When CV_t is greater than or equal to the first stability threshold, the first adjustment coefficient α is set to a second value that is greater than the first value; When CV_v is less than the preset second stability threshold, the second adjustment coefficient β is set to the third value; When CV_v is greater than or equal to the second stability threshold, the second adjustment coefficient β is set to a fourth value that is greater than the third value.

6. The method according to claim 4, characterized in that, In step two, the specific method for constructing target user characteristics based on the separated product browsing data and user behavior data is as follows: The product category sequence L', category average price sequence G', and category average time sequence T' after separation are combined with the average page scrolling speed V to form a related data combination ZH'={L',G',V,T'}; Let k be the total number of product categories in the product category sequence L'. Then L' = L1', L2', ..., Lk', G' = G1', G2', ..., Gk', and T' = T1', T2', ..., Tk'. Using PG=(∑ k o=1 The preferred price mean PG is calculated as Go') / k, where o is the counting index, and 1≤o≤k; CI=[∑ k o=1 (To'-(∑ k o=1 To' / k))] / k calculates the concentration of browsing interest CI of the target user for any product category Lo', and forms a combination CI_Lo' of product category browsing interest concentration with product category Lo'; Similarly, determine the combination of browsing interest concentration for k product categories to form the product category browsing interest concentration set RCI={CI_L1',CI_L2',...,CI_Lk'}; The target user characteristics FM={PG,RCI} are constructed based on the concentration of browsing interest in product categories (RCI) and the mean preferred price (PG), and are stored in encrypted cloud storage.

7. The method according to claim 6, characterized in that, In step three, the specific method for determining the preferred user characteristics is as follows: Based on the cloud, extract the user features of any non-target user, denoted as FM_u={PG_u,RCI_u}, where u is the user identifier; The price preference similarity score S_pri is calculated using the formula S_pri = 1 / (1 + |PG_u - PG|). Extract the common product categories between RCI_u and RCI to form a common category set CL. Count the total number of common categories, denoted as cl, and denote the common category proportion factor as |cl| / k. For any product category in the common category set CL, take its browsing interest concentration in RCI_u and RCI, and denote them as CI and CI_u respectively; The difference in interest concentration (DCI) is calculated using DCI = |CI_u-CI|. Similarly, determine the difference in interest concentration across all common categories, take the average, and denot it as the average interest difference degree AD; The interest distribution similarity score S_int is calculated using S_int=(|cl| / k)×(1 / (1+AD)). The comprehensive similarity Sim is calculated using Sim=w1×S_pri+w2×S_int, where w1 and w2 are preset weight coefficients, and w1+w2=1; Based on the above method, the comprehensive similarity between b users and the target user is obtained, sorted in descending order, and the user features of the top h users are taken as the preferred user features, which are combined into the preferred user feature set {FM_1,FM_2,...,FM_h}, where b and h are preset integers, and h≤b.

8. The method according to claim 7, characterized in that, In step three, a list of preferred products associated with the target user within the backtracking period is generated. The specific method for performing the recommendation operation is as follows: Extract the preferred user feature set {FM_1,FM_2,...,FM_h}, and obtain the comprehensive similarity corresponding to each preferred user feature, denoted as Sim_1,Sim_2,...,Sim_h; Calculate the ratio between h comprehensive similarities and perform normalization so that the sum of the ratios is one, to obtain the ratio sequence bz_1, bz_2, ..., bz_h; Generate a list of recommended preferred products to be populated (List), and determine the total number of preferred products, denoted as sum; The preferred product quantity is allocated to each user feature according to the ratio sequence bz_1, bz_2, ..., bz_h, resulting in sum_1, sum_2, ..., sum_h. If there are decimals, they are rounded up, and the total number of preferred products in the preferred product recommendation list List is expanded accordingly: sum_1 + sum_2 + ... + sum_h = sum. Obtain the number of preferred products sum_1 corresponding to the preferred user feature FM_1, extract the product category browsing interest concentration set RCI_1 from the preferred user feature FM_1, allocate the number of preferred products sum_1 to each product category based on the ratio between the browsing interest concentrations of each product category, and extract the corresponding number of products that the user has browsed according to the preferred user feature FM_1 from each product category, which are denoted as preferred products. If the ratio allocation has a decimal, it is rounded down. Similarly, determine the preferred products associated with all users corresponding to the preferred user feature set {FM_1,FM_2,...,FM_h}, add them to the preferred product recommendation list List in random order, display them on the target user's recommended product page, and perform the recommendation operation. The recommended product page is different from the product page.

9. The method according to claim 8, characterized in that, In step four, when the target user finishes browsing the recommended list of preferred products, the system continuously extracts the target user's product browsing data on the product page, updates the target user's characteristics and the recommended list of preferred products, and continuously performs recommendation operations to the target user.

Images