A method and system for optimizing advertising delivery strategies based on preference data collaboration

Abstract

This invention discloses an advertising delivery strategy optimization method and system based on preference data collaboration, relating to the field of advertising delivery technology. The method includes: collecting internal user data from a local advertising delivery platform and obtaining external user data from a third-party platform through a secure data interface; integrating and preprocessing the above data to extract explicit and implicit preference data; constructing a user-ad rating matrix based on the preference data and generating user preference vectors after decomposition; acquiring the copy text, images, and category information of the advertisement to be delivered, constructing an ad feature expression model, and generating ad characteristic vectors; calculating a matching score based on the user preference vector and ad characteristic vectors, and dynamically optimizing the advertising delivery strategy based on the score. This method, by comprehensively utilizing explicit and implicit user preferences, achieves more accurate user interest modeling and ad matching, effectively improving the advertising delivery effect and efficiency.

Description

Technical Field

[0001] This invention relates to the field of advertising delivery technology, specifically to an advertising delivery strategy optimization method and system based on preference data collaboration. Background Technology

[0002] Preference data refers to data generated by collecting information on users' preferences or rankings of different options. It is usually presented in a comparative form rather than as absolute ratings, thus more closely resembling the characteristics of subjective human decision-making. Sources of this type of data include manual annotation, user behavior, and crowdsourcing platforms. Its core purpose is to help models learn "what humans prefer," rather than simply predicting the correct answer.

[0003] With the rapid development of the digital advertising market, advertisers and platforms face increasingly fierce competition. Achieving precise targeting and improving click-through rates and conversion rates with limited ad space resources has become a key challenge. Traditional advertising strategies primarily rely on users' historical behavioral data for personalized recommendations, but these methods often overlook the differences in user preferences for different ad content, leading to poor performance. Furthermore, the limited user data on a single platform makes it difficult to comprehensively characterize user interests, while the siloed nature of cross-platform data further restricts the accuracy of ad recommendations. Summary of the Invention

[0004] To address the aforementioned technical challenges, this paper provides a method and system for optimizing advertising delivery strategies based on preference data collaboration. This solution addresses the key challenge mentioned in the background section: with the rapid development of the digital advertising market, advertisers and platforms face increasingly fierce competition, making it crucial to achieve precise targeting and improve click-through rates and conversion rates with limited ad space resources. Traditional advertising delivery strategies primarily rely on users' historical behavioral data for personalized recommendations, but these methods often overlook the differences in user preferences for different ad content, leading to poor delivery results. Furthermore, the limited user data on a single platform makes it difficult to comprehensively characterize user interests, while the isolation of cross-platform data further restricts the accuracy of ad recommendations.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] An ad delivery strategy optimization method based on preference data collaboration includes:

[0007] Acquire internal user data from local advertising platforms and obtain external user data from third-party platforms through secure data interfaces;

[0008] Internal and external user data are integrated to obtain comprehensive user data. The comprehensive user data is then preprocessed to obtain explicit preference data and implicit preference data.

[0009] Based on explicit and implicit preference data, a user-advertisement rating matrix is ​​constructed, and the rating matrix is ​​decomposed to obtain a user preference vector.

[0010] We obtain advertising data from advertisers, then extract the copy text, images, and structural category information from the advertising data, analyze them, establish an advertising feature expression model, and obtain advertising characteristic vectors based on the advertising feature expression model.

[0011] The matching degree is calculated between the user preference vector and the advertisement feature vector to obtain the matching degree score;

[0012] The advertising strategy is dynamically adjusted based on the matching score.

[0013] In an optional embodiment, the process of integrating internal and external user data to obtain comprehensive user data, and preprocessing the comprehensive user data to obtain explicit preference data and implicit preference data, specifically includes:

[0014] Obtain internal user data from local advertising platforms;

[0015] External user data on third-party platforms is collected by establishing secure data interface connections based on encrypted communication protocols with third-party platforms.

[0016] The acquired external user data from third-party platforms is anonymized and standardized.

[0017] The processed external user data from third-party platforms is matched with internal user data within the advertising platform, and the matching is based on user account, device identifier, or behavioral characteristics.

[0018] Obtain comprehensive user data with a unified user identifier;

[0019] By preprocessing the comprehensive user data, the preprocessing includes data cleaning, format unification, normalization, and behavior weight calculation.

[0020] The preprocessed comprehensive user data is classified based on the initiative and passivity of user behavior. The initiative of users in rating, liking, collecting, and complaining about advertisements is identified as explicit preference data, while the passive behavior of users in browsing advertisements, such as click records, dwell time, conversion path, and page browsing depth, is identified as implicit preference data.

[0021] In an optional embodiment, constructing the user-advertisement rating matrix based on explicit and implicit preference data specifically includes:

[0022] By normalizing the explicit preference data, the original score weights are set according to different behavior types;

[0023] The original rating weights are linearly normalized to map all explicit preference data to a uniform [0,1] value range. The normalization formula is as follows:

[0024] ;

[0025] In the formula, Indicates user For advertising Explicit preference ratings, This represents the user's original explicit behavior score. and These are the minimum and maximum values ​​among all the original explicit behavior scores, respectively;

[0026] By modeling behavioral intensity using implicit preference data;

[0027] Based on implicit preference data, we obtain data on ad click count, page dwell time, browsing depth, and conversion events;

[0028] By using a linear weighted fusion method, different weight coefficients are assigned to ad clicks, page dwell time, browsing depth, and conversion event data to obtain the user's implicit preference score for the ad. The calculation formula is as follows:

[0029] ;

[0030] In the formula, This represents the user's implicit preference rating for the advertisement. Indicates user For advertising Click data rating, This indicates a score based on the duration of stay. This indicates the depth of the browsing data score. This indicates a score for conversion-related data. The weights corresponding to each data item satisfy the following conditions: ;

[0031] A user-advertisement rating matrix is ​​constructed by weighting and fusing explicit preference scores and implicit standard scores. The formula for each element in the user-ad rating matrix R is:

[0032] ;

[0033] In the formula, element Indicates user For advertising Preference rating, Indicates user For advertising Explicit preference ratings, Indicates user For advertising Implicit preference scores, is the weighting coefficient, and its value ranges from [0,1].

[0034] In an optional embodiment, the step of performing matrix factorization on the rating matrix to obtain the user preference vector specifically includes:

[0035] Decompose the rating matrix R into a user preference matrix. and advertising feature matrix It satisfies the following approximation relationship:

[0036] ;

[0037] In the formula, The user preference matrix is ​​of dimension m×k. Let m be the number of users, n be the number of ads, and k be the dimension of the latent factors.

[0038] By minimizing the loss function on the user preference matrix and advertising feature matrix Conduct training:

[0039] ;

[0040] In the formula, For user preference vectors, For advertising potential vectors, For user preference matrix, For advertising feature matrix, Let A represent the rating matrix, where A represents the sample set of ratings, and λ is the regularization coefficient.

[0041] In an optional embodiment, the step of obtaining advertising data to be delivered from the advertiser, then extracting the copy text, images, and structural category information from the advertising data, analyzing them, establishing an advertising feature expression model, and obtaining an advertising feature vector specifically includes:

[0042] Based on the data of the ads to be delivered, obtain the text copy data, image data, and category information of the ads to be delivered;

[0043] Natural language processing is performed on the text copy data corresponding to the ads to be delivered, extracting keywords and entity information related to product functions, uses, selling points, and sentiment. This information is then converted into text feature vectors using a word vector embedding model. ;

[0044] The image data corresponding to the advertisement to be displayed is subjected to image preprocessing, object detection, and depth feature extraction operations. Multi-level image features are extracted through a convolutional neural network and represented as image feature vectors. ;

[0045] Encode the category information corresponding to the ads to be placed, and generate category vectors. ;

[0046] By fusing text feature vectors, image feature vectors, and category vectors, a feature representation model for advertisements is constructed.

[0047] Based on the feature representation model of the advertisement, the advertisement feature vector is obtained:

[0048] ;

[0049] In the formula, For advertising ... These are the text, image, and category feature vectors of the advertisement, respectively; For the corresponding weighting coefficients, satisfying .

[0050] In an optional embodiment, the step of calculating the matching degree between the user preference vector and the advertising feature vector to obtain a matching degree score specifically includes:

[0051] User preference vector With advertising feature vector Based on this, a matching score is obtained through a dot product scoring model;

[0052] The formula for the dot product scoring model is as follows:

[0053] ;

[0054] In the formula, For users With advertising Match score, For users The preference vector, Advertising feature vector Transpose of;

[0055] Based on the matching score, when the matching score is higher than the set upper threshold, the user is identified as a high-potential user. The system automatically increases the display frequency of the advertisement among the target user group, adjusts the exposure position to the first screen of the page, the top of the page, the middle embedded position, and the front position of the application homepage recommendation stream, increases the real-time bidding bid, and prioritizes the allocation of high exposure resources. When the matching score is in the middle range, the user is identified as a potential responding customer. The system will control the display time period and duration of the advertisement, prioritize scheduling it in time windows with higher conversion rates, and limit its daily display frequency and exposure amount. When the matching score is lower than the lower threshold, the user is identified as a low-relevance user. The system will stop showing the corresponding advertisement to the user, reclaim the allocated resources, and exclude the user from the subsequent retargeting or expanded audience of the advertisement.

[0056] Furthermore, an advertising delivery strategy optimization system based on preference data collaboration is proposed to implement the advertising delivery strategy optimization method described above, specifically including:

[0057] The data acquisition and fusion module is used to acquire internal user data from the local advertising platform and external user data from third-party platforms through a secure data interface; to perform anonymization, standardization, and correlation matching on the internal and external user data to obtain comprehensive user data; and to preprocess the comprehensive user data to obtain explicit preference data and implicit preference data.

[0058] The feature modeling and vector generation module constructs a user-ad rating matrix based on explicit and implicit preference data, performs matrix decomposition on the rating matrix, and generates user preference vectors. It is used to receive text, image, and category information of the advertisement to be delivered, extract features from the advertisement content, construct an advertisement feature expression model, and generate advertisement feature vectors.

[0059] The matching calculation and strategy optimization module calculates the matching degree based on the user preference vector and the ad characteristic vector, and dynamically adjusts the ad delivery strategy based on the matching degree score.

[0060] In an optional embodiment, the data acquisition and fusion module specifically includes:

[0061] An internal data acquisition unit is used to collect internal user data from the advertising platform to be placed.

[0062] An external data acquisition unit is used to obtain external user data from a third-party platform through a secure data interface.

[0063] The data processing unit merges internal user data and external user data to obtain comprehensive user data, and preprocesses the comprehensive user data to obtain explicit preference data and implicit preference data.

[0064] In an optional embodiment, the feature modeling and vector generation module specifically includes:

[0065] The rating matrix construction unit is used to normalize explicit preference data and implicit preference data and model behavioral intensity to obtain explicit preference scores and implicit preference scores, and to construct a user advertising rating matrix based on the explicit preference scores and implicit preference scores.

[0066] A matrix decomposition unit is used to decompose the rating matrix to obtain a user potential interest matrix and an advertising feature matrix, and to train a user preference vector by minimizing a loss function.

[0067] The advertising feature extraction and vector generation unit is used to analyze the text features, image features and category information in the advertising data, construct an advertising feature expression model, and obtain advertising feature vectors.

[0068] In an optional embodiment, the matching calculation and strategy optimization module specifically includes:

[0069] The matching degree calculation unit calculates the matching degree score based on the user preference vector and the advertisement characteristic vector using the dot product model method;

[0070] The threshold judgment unit is used to divide the matching score into three intervals: upper, middle, and lower, to determine the degree of association between the user and the advertisement.

[0071] The delivery adjustment unit, the delivery strategy production unit is used to formulate delivery strategies based on the matching degree interval results, including increasing the display frequency, adjusting the display position, controlling the frequency or stopping the display.

[0072] Compared with the prior art, the beneficial effects of the present invention are:

[0073] This solution proposes an advertising strategy optimization method and system based on preference data collaboration. By integrating internal user data from local advertising platforms with external user data from third-party platforms, a unified comprehensive user data is constructed. Preprocessing is performed after anonymization and standardization to identify explicit and implicit preference data. A user-ad rating matrix is ​​built based on these explicit and implicit preference data, and user preference vectors are obtained through matrix decomposition. This avoids the recommendation failure problem caused by limited data and insufficient preference understanding in traditional recommendation strategies, fundamentally improving the relevance and user acceptance of advertising.

[0074] This solution proposes an advertising strategy optimization method and system based on preference data collaboration. It constructs an advertising feature expression model by combining advertising text, images, and structural category information, generating advertising characteristic vectors. By calculating the matching degree between user preference vectors and advertising characteristic vectors, it automatically adjusts display frequency, exposure position, and delivery time based on the matching degree score, rationally allocating advertising resources and avoiding ineffective exposure and budget waste. This method can achieve adaptive optimization of advertising strategies without relying on manual rules, effectively improving click-through rate, conversion rate, and overall delivery efficiency, providing advertising platforms with more real-time, intelligent, and revenue-oriented delivery strategy support. Attached Figure Description

[0075] Figure 1 This is a flowchart of an advertising delivery strategy optimization method based on preference data collaboration proposed in this invention;

[0076] Figure 2 This is a flowchart illustrating the acquisition of explicit and implicit preference data in this invention.

[0077] Figure 3 This is a flowchart illustrating the process of obtaining the matching score in this invention.

[0078] Figure 4 This is a system framework diagram of an advertising delivery strategy optimization system based on preference data collaboration proposed in this invention. Detailed Implementation

[0079] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0080] Reference Figures 1-4 As shown, an advertising delivery strategy optimization method based on preference data collaboration includes:

[0081] Acquire internal user data from local advertising platforms and obtain external user data from third-party platforms through secure data interfaces;

[0082] Internal and external user data are integrated to obtain comprehensive user data. The comprehensive user data is then preprocessed to obtain explicit preference data and implicit preference data.

[0083] Based on explicit and implicit preference data, a user-advertisement rating matrix is ​​constructed, and the rating matrix is ​​decomposed to obtain a user preference vector.

[0084] We obtain advertising data from advertisers, then extract the copy text, images, and structural category information from the advertising data, analyze them, establish an advertising feature expression model, and obtain advertising characteristic vectors based on the advertising feature expression model.

[0085] The matching degree is calculated between the user preference vector and the advertisement feature vector to obtain the matching degree score;

[0086] The advertising strategy is dynamically adjusted based on the matching score.

[0087] Furthermore, internal and external user data are integrated to obtain comprehensive user data. This comprehensive user data is then preprocessed to obtain explicit and implicit preference data, specifically including:

[0088] Obtain internal user data from local advertising platforms;

[0089] External user data on third-party platforms is collected by establishing secure data interface connections based on encrypted communication protocols with third-party platforms.

[0090] The acquired external user data from third-party platforms is anonymized and standardized.

[0091] Specifically, after acquiring external user data from third-party platforms, the system first performs anonymization and standardization on this data. The core objective of anonymization is to ensure user privacy and security. In this step, the system encrypts data containing personally identifiable information, such as using encryption algorithms to encrypt and store sensitive information like users' real names and phone numbers. Furthermore, for data items that cannot be directly used, the system employs pseudo-anonymization, using random identifiers to replace real identity information, thereby ensuring that the user's personal identity cannot be directly deduced in the event of a data breach.

[0092] Next, the system standardizes and transforms external user data to ensure consistency across different sources. During this process, the system uniformly converts date, time, and numerical formats. For example, time fields are uniformly converted to the standard ISO 8601 format, and numerical data (such as purchase amount and browsing duration) are normalized according to preset intervals, ensuring that data from different platforms have the same dimensions and scale. Furthermore, text data is processed using a unified encoding method to ensure it can be correctly parsed and applied in subsequent analysis.

[0093] The processed external user data from third-party platforms is matched with internal user data within the advertising platform, based on user account, device identifier, or behavioral characteristics.

[0094] Understandably, once the anonymization and standardization processes are complete, the system begins data association and matching. First, the system matches user accounts, comparing user account information on advertising platforms and third-party platforms to determine if the two data sources belong to the same user. For example, if a user logs in using the same account on both advertising platforms and third-party platforms, the system merges this data to generate a unified user profile. If user accounts cannot be directly matched, the system matches based on device identifiers. This means the system compares device IDs or other device identification information from different platforms to try and identify if they are used by the same device. For example, when a user uses the same device on multiple platforms, the system can use device identifiers to associate this data. If an accurate match still cannot be achieved, the system switches to a matching method based on user behavior characteristics. Specifically, the system analyzes the user's behavioral characteristics on each platform, calculating the similarity of behaviors to determine if they belong to the same user.

[0095] Obtain comprehensive user data with a unified user identifier;

[0096] By preprocessing comprehensive user data, including data cleaning, format unification, normalization, and behavior weight calculation;

[0097] The preprocessed comprehensive user data is classified based on the initiative and passivity of user behavior. The initiative of users in rating, liking, collecting, and complaining about advertisements is identified as explicit preference data, while the passive behavior of users in browsing advertisements, such as click records, dwell time, conversion path, and page browsing depth, is identified as implicit preference data.

[0098] Furthermore, based on explicit and implicit preference data, a user-advertisement rating matrix is ​​constructed, specifically including:

[0099] By normalizing the explicit preference data, the original score weights are set according to different behavior types;

[0100] The original rating weights are linearly normalized to map all explicit preference data to a uniform [0,1] value range. The normalization formula is as follows:

[0101] ;

[0102] In the formula, Indicates user For advertising Explicit preference ratings, This represents the user's original explicit behavior score. and These are the minimum and maximum values ​​among all the original explicit behavior scores, respectively;

[0103] Specifically, for different types of user behavior in explicit preference data, the difference in the intensity of behavioral response can be used as the basis for scoring weights. When analyzing different types of user interaction with ads, the system evaluates them based on the subjectivity of the feedback and the depth of behavioral reach, and sets corresponding original scoring weight values. For example, behaviors with stronger subjective intent (such as ad rating operations) are given higher original weights, while operations with lower operational thresholds and shallower behavioral reach (such as collection or liking) are given relatively lower original weights. These scoring weight values ​​serve as the initial quantitative basis for user preference tendencies, and are used for subsequent normalization and scoring matrix construction.

[0104] By modeling behavioral intensity using implicit preference data;

[0105] Based on implicit preference data, we obtain data on ad click count, page dwell time, browsing depth, and conversion events;

[0106] By using a linear weighted fusion method, different weight coefficients are assigned to ad clicks, page dwell time, browsing depth, and conversion event data to obtain the user's implicit preference score for the ad. The calculation formula is as follows:

[0107] ;

[0108] In the formula, This represents the user's implicit preference rating for the advertisement. Indicates user For advertising Click data rating, This indicates a score based on the duration of stay. This indicates the depth of the browsing data score. This indicates a score for conversion-related data. The weights corresponding to each data item satisfy the following conditions: ;

[0109] Specifically, when calculating users' implicit preference scores for advertisements, different types of behavioral data are integrated using a linear weighting method. To improve the discriminativeness and practical guidance of the scores, the system assigns differentiated weight coefficients to different data based on the strength of each implicit behavior's representation of the user's true interest. Ad clicks, as a behavioral signal indicating initial user interest, have a low threshold and a relatively small weight coefficient. Page dwell time reflects the user's level of attention to the ad content and can represent the content's attractiveness to a certain extent, thus it is given a medium weight. Browsing depth reflects the user's extended exploration behavior within the ad content, representing the user's intention to actively obtain information, and has a slightly higher weight. Conversion events are behaviors where users complete specific target actions such as registration, purchase, or redirection, possessing the strongest behavioral intent and commercial value, and therefore are assigned the highest weight. For example, based on empirical rules, the weight of click behavior is preset to 0.1, dwell time to 0.2, browsing depth to 0.3, and conversion behavior to 0.4. Adaptive adjustments can also be made through data analysis and machine learning to optimize the campaign's effectiveness.

[0110] A user-advertisement rating matrix is ​​constructed by weighting and fusing explicit preference scores and implicit standard scores. The formula for each element in the user-ad rating matrix R is:

[0111] ;

[0112] In the formula, element Indicates user For advertising Preference rating, Indicates user For advertising Explicit preference ratings, Indicates user For advertising Implicit preference scores, is the weighting coefficient, and its value ranges from [0,1].

[0113] Furthermore, the rating matrix is ​​decomposed to obtain the user preference vector, which specifically includes:

[0114] Decompose the rating matrix R into a user preference matrix. and advertising feature matrix It satisfies the following approximation relationship:

[0115] ;

[0116] In the formula, The user preference matrix is ​​of dimension m×k. Let m be the number of users, n be the number of ads, and k be the dimension of the latent factors.

[0117] Specifically, when decomposing the user-ad rating matrix R into a user preference matrix and an ad feature matrix, a latent factor dimension parameter k is introduced to limit the representation capabilities of users and ads in the low-dimensional feature space. The parameter k represents the dimension of the latent semantic factor, and its value determines the dimensionality of the user preference vector and the ad latent vector. By appropriately setting the value of k, a balance can be achieved between model complexity and prediction accuracy, ensuring that the user preference representation and ad feature representation after matrix decomposition can effectively extract key features and have good adaptability to new data.

[0118] By minimizing the loss function on the user preference matrix and advertising feature matrix Conduct training:

[0119] ;

[0120] In the formula, For user preference vectors, For advertising potential vectors, For user preference matrix, For advertising feature matrix, Let A represent the rating matrix, where A represents the sample set of ratings, and λ is the regularization coefficient to prevent overfitting.

[0121] Preferably, user preference vectors, by compressing complex user behavior data and preference information into low-dimensional vector representations, can more comprehensively and accurately reflect users' interest characteristics, significantly outperforming traditional single-behavior modeling methods. Compared to relying solely on single-dimensional or single-type behavior data, user preference vectors integrate multiple explicit and implicit preferences, enhancing the relevance of personalization and recommendations. At the same time, vectorized representation improves computational efficiency, facilitating large-scale real-time matching and dynamic optimization, thereby significantly enhancing the accuracy and intelligence of the advertising delivery system.

[0122] Furthermore, advertising data to be delivered is obtained from advertisers. Then, the copywriting text, images, and structural category information are extracted from the advertising data, analyzed, and an advertising feature representation model is established to obtain an advertising feature vector, specifically including:

[0123] Based on the data of the ads to be delivered, obtain the text copy data, image data, and category information of the ads to be delivered;

[0124] Natural language processing is performed on the text copy data corresponding to the ads to be delivered, extracting keywords and entity information related to product functions, uses, selling points, and sentiment. This information is then converted into text feature vectors using a word vector embedding model. ;

[0125] Specifically, the system first performs Chinese word segmentation, part-of-speech tagging, and syntactic dependency analysis on the text copy of the advertisement to be displayed, in order to identify the core content of the sentences. Based on the constructed product knowledge graph and sentiment dictionary, it identifies and extracts keywords, product attribute words, usage scenario words, and positive and negative sentiment words in the copy, thereby extracting entities and semantic fragments that reflect the product's functions, uses, selling points, and user emotional preferences. After completing the extraction of keyword and entity information, the system calls a pre-trained word vector embedding model to vectorize the above keywords and contextual information. By performing average pooling, weighted combination, or Transformer encoding on the word vectors, a unified-dimensional text feature vector is generated to represent the semantic information and communication focus of the advertisement copy, providing input support for the subsequent fusion calculation of advertisement feature vectors.

[0126] The image data corresponding to the advertisement to be displayed is subjected to image preprocessing, object detection, and depth feature extraction operations. Multi-level image features are extracted through a convolutional neural network and represented as image feature vectors. ;

[0127] Understandably, the image data corresponding to the advertisement is first processed through image preprocessing. Image preprocessing includes operations such as unifying the size, normalizing grayscale, and normalizing pixels of the original image to improve the consistency and clarity of the image data. After preprocessing, the image is used by an object detection model to identify the product, person, or brand regions in the image, and key regions are extracted for subsequent analysis. Subsequently, a convolutional neural network (CNN) model is constructed. The CNN model uses a network architecture composed of multiple convolutional layers, pooling layers, and nonlinear activation functions to extract multi-level semantic features of the image layer by layer, including low-level texture edges, mid-level graphic contours, and high-level semantic information. Finally, the output feature vector serves as the image representation and is used to obtain the advertisement feature vector in subsequent steps.

[0128] Encode the category information corresponding to the ads to be placed, and generate category vectors. ;

[0129] Specifically, for the structural category information corresponding to the advertisement to be placed, the category data is first parsed hierarchically, and the category tags to which the advertisement belongs are hierarchically split and standardized. Then, one-hot encoding is used for processing. First, a predefined category dictionary is established, and all possible advertisement category tags are uniformly numbered. Then, according to the category position of each advertisement, a value of 1 is assigned to the corresponding dimension, and a value of 0 is assigned to the other dimensions to form a sparse category vector. This category vector can accurately represent the specific category to which the advertisement belongs, which is convenient for obtaining the advertisement characteristic vector later.

[0130] By fusing text feature vectors, image feature vectors, and category vectors, a feature representation model for advertisements is constructed.

[0131] Based on the feature representation model of the advertisement, the advertisement feature vector is obtained:

[0132] ;

[0133] In the formula, For advertising ... These are the text, image, and category feature vectors of the advertisement, respectively; For the corresponding weighting coefficients, satisfying .

[0134] Furthermore, the matching degree between the user preference vector and the advertisement feature vector is calculated to obtain a matching degree score, which specifically includes:

[0135] User preference vector With advertising feature vector Based on this, a matching score is obtained through a dot product scoring model;

[0136] The formula for the dot product scoring model is:

[0137] ;

[0138] In the formula, For users With advertising Match score, For users The preference vector, Advertising feature vector Transpose of;

[0139] Based on the matching score, when the matching score is higher than the set upper threshold, the user is identified as a high-potential user. The system automatically increases the display frequency of the ad among the target user group, adjusts the exposure position to the first screen of the page, the top of the page, the middle embedded position, and the front position of the application homepage recommendation stream, increases the real-time bidding bid, and prioritizes the allocation of high exposure resources. When the matching score is in the middle range, the user is identified as a potential responding customer. The system will control the display time period and duration of the ad, prioritize scheduling it within the time window with higher conversion rates, and limit its daily display frequency and exposure amount to test the user's potential response and balance resource investment with expected return. When the matching score is lower than the lower threshold, the user is identified as a low-relevance user. The system will stop showing the corresponding ad to the user, reclaim the allocated resources, and exclude the user from the subsequent retargeting or expanded audience of the ad to prevent invalid exposure and budget waste.

[0140] Specifically, for example, if the set threshold is [0.5, 0.8], when the matching score is greater than 0.8, the system identifies the user as a high-potential user, meaning the user has a high interest in or conversion probability of the current ad content. In this case, the system will perform ad delivery enhancement operations. When the matching score is between [0.5, 0.8], the system considers the user as a potential responder, meaning the user has some interest in the ad but has not yet clearly shown a strong preference. In this case, the system adopts a robust, tentative delivery strategy. When the matching score is less than 0.5, the system will determine the user as a low-relevance user, meaning the user lacks interest in or conversion potential of the ad. In this case, the system will execute a negative screening strategy.

[0141] Furthermore, an advertising delivery strategy optimization system based on preference data collaboration is proposed to implement the advertising delivery strategy optimization method described above, specifically including:

[0142] The data acquisition and fusion module is used to acquire internal user data from the local advertising platform and external user data from third-party platforms through a secure data interface; it performs anonymization, standardization, and correlation matching on the internal and external user data to obtain comprehensive user data; and it preprocesses the comprehensive user data to obtain explicit preference data and implicit preference data.

[0143] The feature modeling and vector generation module constructs a user-ad rating matrix based on explicit and implicit preference data, performs matrix decomposition on the rating matrix, and generates user preference vectors. It is used to receive text, image, and category information of the ads to be delivered, extract features from the ad content, construct an ad feature expression model, and generate ad feature vectors.

[0144] The matching calculation and strategy optimization module calculates the matching degree based on the user preference vector and the ad characteristic vector, and dynamically adjusts the ad delivery strategy based on the matching degree score.

[0145] Furthermore, the data acquisition and fusion module specifically includes:

[0146] Internal data acquisition unit, used to collect internal user data from the advertising platform to be placed;

[0147] External data acquisition unit, used to obtain external user data from third-party platforms through a secure data interface;

[0148] The data processing unit integrates internal and external user data to obtain comprehensive user data. It then preprocesses the comprehensive user data to obtain explicit preference data and implicit preference data.

[0149] Furthermore, the feature modeling and vector generation module specifically includes:

[0150] The rating matrix construction unit is used to normalize explicit preference data and implicit preference data and model behavioral intensity to obtain explicit preference scores and implicit preference scores, and to construct a user advertising rating matrix based on explicit preference scores and implicit preference scores.

[0151] The matrix factorization unit is used to perform matrix factorization on the rating matrix to obtain the user latent interest matrix and the advertising feature matrix. The user preference vector is obtained by training through minimizing the loss function.

[0152] The advertising feature extraction and vector generation unit is used to analyze the text features, image features, and category information in advertising data, construct an advertising feature expression model, and obtain advertising feature vectors.

[0153] Furthermore, the matching calculation and strategy optimization module specifically includes:

[0154] The matching score calculation unit calculates the matching score based on the user preference vector and the advertisement feature vector using the dot product model method.

[0155] The threshold judgment unit is used to divide the matching score into three categories: upper, middle, and lower, to determine the degree of association between the user and the advertisement.

[0156] The campaign adjustment unit and the campaign strategy production unit are used to formulate campaign strategies based on the matching degree range results, including increasing the display frequency, adjusting the display position, controlling the frequency, or stopping the display.

[0157] In summary, the advantages of this invention are as follows: By integrating user data from the local platform with user data from third-party external sources, a unified comprehensive user data is constructed. Explicit and implicit preferences are extracted and a user preference vector is generated, avoiding the recommendation failure problem caused by single data and insufficient understanding of preferences in traditional recommendation strategies. At the same time, by integrating advertising text, images, and category information to construct advertising feature vectors, and dynamically optimizing display frequency, exposure position, and delivery time based on matching scores, intelligent advertising strategy adjustments with higher relevance and delivery efficiency are achieved, avoiding invalid exposure and improving click-through rate and conversion effect.

[0158] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for optimizing advertising delivery strategies based on preference data collaboration, characterized in that, include: Acquire internal user data from local advertising platforms and obtain external user data from third-party platforms through secure data interfaces; Internal and external user data are integrated to obtain comprehensive user data. The comprehensive user data is preprocessed and classified based on the initiative and passivity of user behavior to obtain explicit preference data and implicit preference data. Among them, the initiative of users in rating, liking, collecting and complaining about advertisements is identified as explicit preference data, while the passive behavior of users in the process of browsing advertisements, such as click records, dwell time, conversion path and page browsing depth, is identified as implicit preference data. Based on explicit and implicit preference data, a user-advertisement rating matrix is ​​constructed, and the rating matrix is ​​decomposed to obtain a user preference vector. We obtain advertising data from advertisers, then extract the copy text, images, and structural category information from the advertising data, analyze them, establish an advertising feature expression model, and obtain advertising characteristic vectors based on the advertising feature expression model. The matching degree is calculated between the user preference vector and the advertisement feature vector to obtain the matching degree score; The advertising strategy is dynamically adjusted based on the matching score. The construction of a user-advertisement rating matrix based on explicit and implicit preference data specifically includes: By normalizing the explicit preference data, the original score weights are set according to different behavior types; The original rating weights are linearly normalized to map all explicit preference data to a uniform [0,1] value range. The normalization formula is as follows: ； In the formula, Indicates user For advertising Explicit preference ratings, This represents the user's original explicit behavior score. and These are the minimum and maximum values ​​among all the original explicit behavior scores, respectively; By modeling behavioral intensity using implicit preference data; Based on implicit preference data, we obtain data on ad click count, page dwell time, browsing depth, and conversion events; By using a linear weighted fusion method, different weight coefficients are assigned to ad clicks, page dwell time, browsing depth, and conversion event data to obtain the user's implicit preference score for the ad. The calculation formula is as follows: ； In the formula, This represents the user's implicit preference rating for the advertisement. Indicates user For advertising Click data rating, This indicates a score based on the duration of stay. This indicates the depth of the browsing data score. This indicates a score for conversion-related data. The weights corresponding to each data item satisfy the following conditions: ; A user-advertisement rating matrix is ​​constructed by weighting and fusing explicit preference scores and implicit standard scores. The formula for each element in the user-ad rating matrix R is: ； In the formula, element Indicates user For advertising Preference rating, Indicates user For advertising Explicit preference ratings, Indicates user For advertising Implicit preference scores, These are weighting coefficients, with values ​​ranging from [0,1]. The process of calculating the matching degree between the user preference vector and the advertising feature vector to obtain a matching degree score specifically includes: User preference vector With advertising feature vector Based on this, a matching score is obtained through a dot product scoring model; The formula for the dot product scoring model is as follows: ； In the formula, For users With advertising Match score, For users The preference vector, Advertising feature vector Transpose of; Based on the matching score, when the matching score is higher than the set upper threshold, it is determined to be a high-potential user. The system will automatically increase the display frequency of the advertisement in the target user group, adjust the exposure position to the first screen of the page, the top of the page, the middle embedded position, and the front position of the application homepage recommendation stream, increase the real-time bidding bid, and prioritize the allocation of high exposure resources. When the matching score is in the middle range, it is determined to be a potential customer. The system will control the display time and duration of the advertisement, prioritize it in the time window with higher conversion rate, and limit its daily frequency and exposure. When the matching score is below the lower threshold, the user is judged as having low relevance. The system will stop showing the corresponding advertisement to the user, reclaim and allocate resources, and exclude the user from subsequent retargeting or expanded audiences for the advertisement.

2. The advertising delivery strategy optimization method based on preference data collaboration according to claim 1, characterized in that, The process involves integrating internal and external user data to obtain comprehensive user data. This comprehensive user data is then preprocessed, and categorized based on the initiative and passivity of user behavior to obtain explicit and implicit preference data. Specifically, this includes: Obtain internal user data from local advertising platforms; External user data on third-party platforms is collected by establishing secure data interface connections based on encrypted communication protocols with third-party platforms. The acquired external user data from third-party platforms is anonymized and standardized. The processed external user data from third-party platforms is matched with internal user data within the advertising platform, and the matching is based on user account, device identifier, or behavioral characteristics. Obtain comprehensive user data with a unified user identifier; By preprocessing comprehensive user data and classifying it according to the initiative and passivity of user behavior, explicit preference data and implicit preference data are obtained. The preprocessing includes data cleaning, format unification, normalization, and behavior weight calculation.

3. The advertising delivery strategy optimization method based on preference data collaboration according to claim 1, characterized in that, The step of performing matrix decomposition on the rating matrix to obtain the user preference vector specifically includes: Decompose the rating matrix R into a user preference matrix. and advertising feature matrix It satisfies the following approximation relationship: ； In the formula, The user preference matrix is ​​of dimension m×k. Let m be the number of users, n be the number of ads, and k be the dimension of the latent factors. By minimizing the loss function on the user preference matrix and advertising feature matrix Conduct training: ； In the formula, For user preference vectors, For advertising potential vectors, For user preference matrix, For advertising feature matrix, Let A represent the rating matrix, where A represents the sample set of ratings, and λ is the regularization coefficient.

4. The advertising delivery strategy optimization method based on preference data collaboration according to claim 1, characterized in that, The process involves obtaining advertising data from advertisers, extracting copy text, images, and structural category information from the advertising data, analyzing this data, establishing an advertising feature expression model, and obtaining an advertising feature vector. Specifically, this includes: Based on the data of the ads to be delivered, obtain the text copy data, image data, and category information of the ads to be delivered; Natural language processing is performed on the text copy data corresponding to the ads to be delivered, extracting keywords and entity information related to product functions, uses, selling points, and sentiment. This information is then converted into text feature vectors using a word vector embedding model. ; The image data corresponding to the advertisement to be displayed is subjected to image preprocessing, object detection, and depth feature extraction operations. Multi-level image features are extracted through a convolutional neural network and represented as image feature vectors. ; Encode the category information corresponding to the ads to be placed, and generate category vectors. ; By fusing text feature vectors, image feature vectors, and category vectors, a feature representation model for advertisements is constructed. Based on the feature representation model of the advertisement, the advertisement feature vector is obtained: ； In the formula, For advertising ... These are the text, image, and category feature vectors of the advertisement, respectively; For the corresponding weighting coefficients, satisfying .

5. An Optimization of Advertising Targeting Strategy Based on Preference Data Collaboration The system, according to claims 1-4, is a method for optimizing advertising delivery strategies based on preference data collaboration, characterized in that it specifically includes: The data acquisition and fusion module is used to acquire internal user data from the local advertising platform and external user data from third-party platforms through a secure data interface; to perform anonymization, standardization, and correlation matching on the internal and external user data to obtain comprehensive user data; and to preprocess the comprehensive user data to obtain explicit preference data and implicit preference data. The feature modeling and vector generation module constructs a user-ad rating matrix based on explicit and implicit preference data, performs matrix decomposition on the rating matrix, and generates user preference vectors. It is used to receive text, image, and category information of the advertisement to be delivered, extract features from the advertisement content, construct an advertisement feature expression model, and generate advertisement feature vectors. The matching calculation and strategy optimization module calculates the matching degree based on the user preference vector and the ad characteristic vector, and dynamically adjusts the ad delivery strategy based on the matching degree score.

6. The advertising delivery strategy optimization system based on preference data collaboration according to claim 5, characterized in that, The data acquisition and fusion module specifically includes: An internal data acquisition unit is used to collect internal user data from the advertising platform to be placed and external user data from third-party platforms obtained through a secure data interface. An external data acquisition unit is used to obtain external user data from a third-party platform through a secure data interface. The data processing unit merges internal user data and external user data to obtain comprehensive user data, and preprocesses the comprehensive user data to obtain explicit preference data and implicit preference data.

7. The advertising delivery strategy optimization system based on preference data collaboration according to claim 5, characterized in that, The feature modeling and vector generation module specifically includes: The rating matrix construction unit is used to normalize explicit preference data and implicit preference data and model behavioral intensity to obtain explicit preference scores and implicit preference scores, and to construct a user advertising rating matrix based on the explicit preference scores and implicit preference scores. A matrix decomposition unit is used to decompose the rating matrix to obtain a user potential interest matrix and an advertising feature matrix, and to train a user preference vector by minimizing a loss function. The advertising feature extraction and vector generation unit is used to analyze the text features, image features and category information in the advertising data, construct an advertising feature expression model, and obtain advertising feature vectors.

8. The advertising delivery strategy optimization system based on preference data collaboration according to claim 5, characterized in that, The matching calculation and strategy optimization module specifically includes: The matching degree calculation unit calculates the matching degree score based on the user preference vector and the advertisement characteristic vector using the dot product model method; The threshold judgment unit is used to divide the matching score into three intervals: upper, middle, and lower, to determine the degree of association between the user and the advertisement. The delivery adjustment unit, the delivery strategy production unit is used to formulate delivery strategies based on the matching degree interval results, including increasing the display frequency, adjusting the display position, controlling the frequency or stopping the display.

Images