A Precision Marketing Association Analysis Method Based on a Three-Dimensional Knowledge Graph of Advertising-User-Product

Abstract

This invention provides a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising, users, and products. The method includes collecting multi-dimensional heterogeneous data from the marketing ecosystem, encompassing user behavior, product information, and advertising creative domains. Semantic feature extraction and entity alignment are performed on the collected multi-dimensional heterogeneous data. An initial three-dimensional association graph is constructed based on the semantic relationships and interaction behaviors between entities. This invention constructs a three-dimensional knowledge graph integrating multi-dimensional heterogeneous data from users, products, and advertisements. It utilizes natural language processing and computer vision technologies to achieve precise semantic alignment between advertising creatives and product attributes. Simultaneously, it leverages the feature propagation mechanism of graph neural networks to uncover implicit associations between deep user preferences and advertising creatives. Finally, by combining path structure reasoning and real-time context awareness, it generates a personalized recommendation list with high interpretability and accuracy, significantly improving the matching efficiency and conversion effect of advertising placement.

Description

Technical Field

[0001] This invention relates to the field of advertising and marketing technology, specifically to a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product. Background Technology

[0002] With the rapid development of mobile internet and e-commerce, digital marketing has shifted from the traditional "broad-based" model to a precision marketing model based on big data. In the actual marketing ecosystem, users, advertisements, and products constitute the three core elements. Existing recommendation systems and advertising technologies typically employ collaborative filtering or content-based filtering algorithms. While these have achieved some success, they still face significant technical bottlenecks when dealing with complex marketing scenarios: Data silos and entity fragmentation: Existing data processing solutions often model user behavior, product information, and advertising creative domains separately. For example, traditional recommendation systems primarily rely on historical user-product interactions, while advertising delivery systems focus on predicting click-through rates between users and ads. This fragmentation leads to advertising delivery often neglecting the deep connection between the attributes of the promoted products and user preferences, making it difficult to form a complete marketing loop. The semantic gap between advertising creative and product attributes: In existing technologies, there is a lack of effective semantic alignment mechanisms between advertising creatives (such as images and text) and product entities (such as categories and attributes). Advertisers' creative content may not match the actual selling points of the product, or algorithms may fail to accurately identify the relationship between the visual content of the advertisement and the product description. This leads to the system's inability to determine whether an advertisement is truly suitable for the product it promotes, resulting in problems such as "creative misrepresentation" or "low matching degree," which reduces the user's click-through experience and conversion rate. Insufficient ability to uncover implicit interests: A user's true purchase intention is often implicit in their choice of specific product attributes, while the attractiveness of an advertisement depends on its creative content. Traditional collaborative filtering algorithms rely solely on explicit click behavior and cannot bridge the semantic gap between "product attributes" and "ad creative content." For example, a user may like "retro style" furniture but has never clicked on an ad promoting that style. Traditional algorithms would assume the user is not interested in the ad, thus failing to uncover the user's potential, unexpressed, deep-seated needs. Lack of explainable recommendation logic: While deep learning models improve prediction accuracy, they are often considered "black boxes," making it difficult to explain to operations personnel or advertisers why a particular ad is recommended to a specific user. This lack of explainability based on path and semantic relationships leaves marketing strategy adjustments and optimizations without a theoretical basis.

[0003] Therefore, there is an urgent need for a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising, users, and products. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising, users, and products, in order to solve the problems mentioned in the background.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-users-products, comprising the following steps: S1: Collect multi-dimensional heterogeneous data from the marketing ecosystem, which covers user behavior domain, product information domain and advertising creative domain; S2: Extract semantic features and align entities from the collected multi-dimensional heterogeneous data, and construct an initial three-dimensional association map based on the semantic associations and interaction behaviors between entities; The initial three-dimensional relational graph includes user nodes, product nodes, advertising nodes, and the initial connecting edges between them; S3: Based on the graph neural network algorithm, feature propagation and neighbor aggregation are performed on the initial three-dimensional association graph to mine the implicit associations between nodes. Based on the mined implicit associations, the initial three-dimensional association graph is completed and optimized to obtain the final three-dimensional marketing knowledge graph. S4: Based on the path structure and aggregation features between nodes in the final 3D marketing knowledge graph, calculate the precision marketing match degree of the target user with the candidate ads, generate a personalized recommendation list based on the precision marketing match degree, and execute precision marketing delivery based on the recommendation list.

[0006] As a preferred embodiment, the multi-dimensional heterogeneous data includes user basic data, product static data, advertising dynamic data, and interaction behavior data; The user basic data includes user demographic attributes, historical preference tags, real-time geographical location, and device environment information; The static product data includes product category data, product attribute data, product sales data, and product review data; The advertising dynamic data includes advertising material data, delivery strategy data, bidding data, and creative copy data; The interactive behavior data includes exposure data, click data, add-to-cart data, and conversion data.

[0007] As a preferred embodiment, the specific process of step S1 includes: Set up a data cleaning and integration module, and link the marketing data warehouse through the data cleaning and integration module to obtain basic user data, static product data, dynamic advertising data, and interaction behavior data; The collected data is denoised and normalized, and a unique identifier index for users, products, and advertisements is established. The denoising process includes removing duplicate logs and filtering crawler traffic, and the normalization process includes performing Min-Max normalization or Z-Score standardization on numerical features.

[0008] As a preferred embodiment, the specific process of extracting semantic features and aligning entities from the collected multi-dimensional heterogeneous data, and constructing an initial three-dimensional association map based on the semantic relationships and interaction behaviors between entities includes: Natural language processing technology is used to extract text features from advertising creative copy data and product review data to obtain text semantic vectors; Using computer vision technology, visual features are extracted from advertising material data and product images to obtain visual feature vectors; The comprehensive semantic similarity between the advertisement and the product is calculated based on the text semantic vector and visual feature vector. If the overall semantic similarity exceeds a preset threshold, a promotion-attribution association edge is established between the advertising node and the product node, and the weight of the association edge is determined by the semantic similarity value. Based on user identifiers, product identifiers, and advertising identifiers in the interaction behavior data, a browsing-purchase behavior edge is established between user nodes and product nodes, and a click-feedback behavior edge is established between user nodes and advertising nodes.

[0009] In a preferred embodiment, the step of calculating the comprehensive semantic similarity between the advertisement and the product based on the text semantic vector and the visual feature vector includes: The specific process for calculating the semantic similarity between an advertisement and a product based on text semantic vectors and visual feature vectors includes: calculating the text similarity between the advertisement copy and the product description, as well as the visual similarity between the advertisement image and the main product image, and obtaining the comprehensive semantic similarity through weighted fusion; The calculation formula is as follows: ; in, Indicates the weighting coefficient. Represents cosine similarity. and These represent the text semantic vectors for advertisements and products, respectively. and These represent the visual feature vectors of the advertisement and the product, respectively.

[0010] As a preferred embodiment, the specific process of step S3 includes: A heterogeneous graph neural network model is constructed, which includes multiple relation aggregation layers, each of which is used to aggregate the features of neighbor nodes of a specific relation type. For user nodes, the attribute features of their neighboring product nodes and the creative features of their neighboring advertising nodes are aggregated through the relationship aggregation layer to update the feature vector of the user node. For advertising nodes, aggregate the selling features of their associated product nodes and the user profile features of historical user nodes that have interacted with them, and update the feature vector of the advertising node. Based on the updated node feature vectors, the decoder is used to calculate the link probability between user nodes and advertising nodes that are not directly connected in the graph. If the link probability is greater than the preset implicit association threshold, then potential interest association edges are added between user nodes and advertising nodes, and the graph is completed and dynamically updated.

[0011] As a preferred embodiment, the specific process for calculating the precise marketing match degree of the target user for candidate advertisements based on the path structure and aggregation features between nodes in the final three-dimensional marketing knowledge graph includes: In the final three-dimensional marketing knowledge graph, a multi-hop semantic path is defined between the target user node and the candidate ad node; Extract the final low-dimensional embedding vectors of the target user node and candidate ad node after aggregation by graph neural network, and calculate the vector cosine similarity between the two as the basic semantic matching score; Calculate the path weights for all connected paths between the target user node and the candidate ad node, and use the sum of the path weights of all connected paths as the structural association enhancement score. The accuracy of marketing matching is calculated by combining the basic semantic matching score with the structural association enhancement score.

[0012] As a preferred embodiment, the formula for calculating the precision marketing matching degree is: ; in, Indicates the accuracy of targeted marketing. and This indicates an adjustable weighting coefficient. This represents the basic semantic matching score. Represents the set of all connected paths. This represents the weight value of path p, where p represents a specific connected path between the target user node U and the candidate advertisement node A. Represents the set of all connected paths; The formula for calculating the basic semantic score is: ; in This represents the final low-dimensional embedding vector obtained after the target user node U undergoes multi-layer aggregation in a graph neural network. This represents the final low-dimensional embedding vector obtained after candidate ad node A undergoes multi-layer aggregation in a graph neural network. Represents the user node vector The Euclidean norm, Represents the vector of advertising nodes The Euclidean norm.

[0013] As a preferred embodiment, the specific process of generating a personalized recommendation list based on the precision marketing matching degree and executing precision marketing delivery based on the recommendation list includes: Candidate ads are sorted from high to low based on their precision marketing matching accuracy, and the top N ads are selected to generate a personalized recommendation list. Obtain the current user's real-time context information, reorder it in conjunction with the personalized recommendation list, remove ads that do not fit the current context, and obtain the final recommendation list; Based on the final recommendation list, the corresponding advertising creative materials are loaded in real time on the user's terminal interface to complete precise marketing delivery, and the user's subsequent interaction behavior is recorded to feed back into the knowledge graph for iterative updates.

[0014] This invention provides a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising, users, and products, which has the following beneficial effects: This invention breaks down data silos and semantic fragmentation between entities in traditional marketing by constructing a three-dimensional knowledge graph that integrates multi-dimensional heterogeneous data from users, products, and advertisements. It utilizes natural language processing and computer vision technologies to achieve precise semantic alignment between advertising creatives and product attributes, effectively solving the problem of "mismatched content" in ad placement. At the same time, it leverages the feature propagation mechanism of graph neural networks to uncover the implicit relationship between deep user preferences and advertising creatives, bridging the "intent-creativity gap." Finally, by combining path structure reasoning and real-time context awareness, it generates a personalized recommendation list with high interpretability and accuracy, significantly improving the matching efficiency and conversion effect of ad placement. Attached Figure Description

[0015] Figure 1 This is a flowchart of the precision marketing association analysis method based on a three-dimensional knowledge graph of advertising, users, and products, as presented in this invention. Detailed Implementation

[0016] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0017] like Figure 1As shown, this embodiment of the invention provides a precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-users-products, including the following steps: S1: Collect multi-dimensional heterogeneous data from the marketing ecosystem, which covers user behavior domain, product information domain and advertising creative domain; Specifically, multi-dimensional heterogeneous data includes basic user data, static product data, dynamic advertising data, and interaction behavior data; The user basic data includes user demographic attributes, historical preference tags, real-time geographical location, and device environment information. The user demographic attributes include age, gender, occupation, etc. The historical preference tags include category preference, brand preference, price sensitivity, etc. The real-time geographical location includes province, city, business district, etc. The device environment information includes operating system, network environment, etc. The static product data includes product category data, product attribute data, product sales data, and product review data. The product category data includes first-level categories, second-level categories, third-level categories, etc. The product attribute data includes material, style, applicable scenarios, etc. The product sales data includes monthly sales volume, inventory status, etc. The product review data includes positive review rate, keyword tags, etc. The advertising dynamic data includes advertising material data, delivery strategy data, bidding data, and creative copy data. The advertising material data includes images, videos, logos, etc. The delivery strategy data includes targeted audiences, delivery time periods, etc. The bidding data includes bid levels, budget consumption, etc. The creative copy data includes titles, promotional text, etc. The interactive behavior data includes exposure data, click data, add-to-cart data, and conversion data. The exposure data includes exposure time and exposure location. The click data includes click duration and click frequency. The add-to-cart data includes the time and quantity added to the cart. The conversion data includes the order time and payment amount. In this embodiment, the process of collecting multi-dimensional heterogeneous data in the marketing ecosystem includes: A data cleaning and integration module is set up, which is used to acquire basic user data, static product data, dynamic advertising data, and interactive behavior data. By linking the marketing data warehouse through the data cleaning and integration module, the collected user basic data, product static data, advertising dynamic data and interaction behavior data are denoised and normalized, and a unique identifier index for user-product-advertisement is established. Specifically, raw log data is captured in real time through the event tracking SDK, business database interface, and third-party monitoring API. Secondly, the raw data is preprocessed using data cleaning scripts, including removing duplicate logs, filtering crawler traffic, filling missing values ​​(such as using the mean to fill missing population attribute values), and performing Min-Max normalization or Z-Score standardization on numerical features. Finally, a global ID mapping table is constructed to map user IDs (User_ID), product IDs (Item_ID), and ad IDs (Ad_ID) scattered in different data tables to unified index identifiers required for graph construction, thereby achieving entity alignment of multi-source data.

[0018] S2: Extract semantic features and align entities from the collected multi-dimensional heterogeneous data, and construct an initial three-dimensional association graph based on the semantic associations and interaction behaviors between entities. The initial three-dimensional association graph includes user nodes, product nodes, advertising nodes and the initial connection edges between them. In this embodiment, the process of extracting semantic features and aligning entities from the collected multi-dimensional heterogeneous data, and constructing an initial three-dimensional association map based on the semantic relationships and interaction behaviors between entities includes: Natural language processing technology is used to extract text features from advertising creative copy data and product review data to obtain text semantic vectors; Using computer vision technology, visual features are extracted from advertising material data and product images to obtain visual feature vectors; Specifically, visual feature vectors are obtained by using pre-trained language models (such as BERT or RoBERTa) to segment and encode the text of advertising titles, promotional texts, and product reviews, extracting 768-dimensional or 1024-dimensional text feature vectors; and by using convolutional neural network models (such as ResNet-50 or VGG-16) to extract the fully connected features of the advertising material images and product main images in the penultimate layer, as 2048-dimensional visual feature vectors. The semantic similarity between advertisements and products is calculated based on text semantic vectors and visual feature vectors; The specific process for calculating the semantic similarity between an advertisement and a product based on text semantic vectors and visual feature vectors includes: calculating the text similarity between the advertisement copy and the product description, as well as the visual similarity between the advertisement image and the main product image, and obtaining the comprehensive semantic similarity through weighted fusion; The calculation formula is as follows: ; in, Indicates the weighting coefficient. Represents cosine similarity. and These represent the text semantic vectors for advertisements and products, respectively. and These represent the visual feature vectors of the advertisement and the product, respectively. If the overall semantic similarity exceeds a preset threshold, a "promotion-attribution" association edge is established between the ad node and the product node. The weight of this association edge is determined by the semantic similarity value. Based on user identifiers, product identifiers, and advertising identifiers in the interaction behavior data, a "browse-purchase" behavior edge is established between user nodes and product nodes, and a "click-feedback" behavior edge is established between user nodes and advertising nodes. The weight of the behavior edge is assigned according to the conversion rate or click frequency of the interaction behavior. The specific assignment logic is as follows: for the "Browse-Purchase" edge, the purchase behavior is given a higher weight than the browsing behavior (e.g., purchase = 5.0, browsing = 1.0); for the "Click-Feedback" edge, the weight is set according to the click duration (the weight is 2.0 if the duration is >3s, otherwise it is 1.0). The initial three-dimensional association graph consists of an entity layer and a relationship layer. The entity layer includes user entity nodes, product entity nodes, and advertisement entity nodes. The relationship layer includes explicit interaction relationships and semantic association relationships. The processed multi-dimensional heterogeneous data is mapped to the corresponding entity nodes, and the entity nodes are connected according to the established association edges to form the initial three-dimensional network topology and obtain the initial three-dimensional association graph. It should be noted that traditional methods, when processing multi-source heterogeneous data, often model users, products, and advertisements separately, ignoring the inherent semantic consistency between advertising creative content and product attributes. This often results in advertisements being "off-topic." This invention, by constructing an initial three-dimensional association graph, unifies advertisements, users, and products in the same semantic space. Through entity alignment and the establishment of association edges, it explicitly expresses the complete marketing chain of "users recognizing products through advertisements," solving the problem of the separation between advertising creative content and product entity attributes, and laying a data foundation for subsequent deep association analysis.

[0019] S3: Based on the graph neural network algorithm, feature propagation and neighbor aggregation are performed on the initial three-dimensional association graph to mine the implicit associations between nodes. Based on the mined implicit associations, the initial three-dimensional association graph is completed and optimized to obtain the final three-dimensional marketing knowledge graph. In this embodiment, the process of performing feature propagation and neighbor aggregation on the initial 3D association graph based on graph neural network algorithms to mine implicit relationships between nodes, and then completing and optimizing the initial 3D association graph based on the mined implicit relationships to obtain the final 3D marketing knowledge graph includes: A heterogeneous graph neural network model is constructed, which includes multiple relation aggregation layers, each of which is used to aggregate the features of neighbor nodes of a specific relation type. The specific implementation process is as follows: using Relational Graph Convolutional Network (RGCN) or Graph Attention Network (GAT) as the basic architecture, and defining specific aggregation functions for different meta-paths such as "user-product", "user-advertisement", and "advertisement-product"; For user nodes in the graph, the attribute features (such as category, price, style tags) of their neighboring product nodes and the creative features (such as visual style, copywriting emotion) of their neighboring advertising nodes are aggregated through the relationship aggregation layer to update the feature vector of the user node, so as to represent the user's comprehensive preferences in the product dimension and advertising dimension. The formula for calculating the updated feature vector of a user node is: ; in, This represents the feature vector updated by the target user node u after computation through a k+1 layer graph neural network. Let R represent a non-linear activation function, and let R represent the set of all relation types in the graph. This represents the set of neighboring nodes of the target user node u under relation type r. Represents the neighbor set The number of nodes in the middle, This represents the trainable weight matrix for relation type r in the k-th layer of the network. This represents the feature vector of neighbor node v at the k-th layer, where k represents the layer index of the graph neural network; For advertising nodes in the graph, aggregate the selling features of their associated product nodes and the user profile features of historical user nodes that have interacted with them, and update the feature vector of the advertising node to represent the appeal of the advertisement to different groups of people and the core competitiveness of the promoted products. Based on the updated node feature vectors, the decoder is used to calculate the link probability between user nodes and advertising nodes that are not directly connected in the graph. The link probability represents the user’s potential interest in the advertisement. The specific process for calculating the link probability between user nodes and ad nodes that are not directly connected in the graph using a decoder is as follows: An inner product decoder or a multilayer perceptron (MLP) decoder is used to calculate the user feature vector. With candidate ad feature vectors The dot product or the score after MLP transformation is used to map the link probability to the (0,1) interval using the Sigmoid function. ; If the link probability is greater than the preset implicit association threshold, it is determined that there is an implicit association between the two, that is, a "potential interest" association edge is added between the user node and the advertising node, and the weight of the edge is set to the value of the link probability. The initial three-dimensional association graph is completed based on the added "potential interest" related edges. At the same time, the edge weights in the graph are dynamically updated according to the interaction frequency and timeliness between nodes. Connections with weights below the noise threshold are removed to obtain the final three-dimensional marketing knowledge graph. Specifically, the update strategy is to introduce a time decay factor. The weights of the old interaction edges are decayed. τ is the time difference from the current time, and τ is the time constant; edges with weights lower than a preset noise threshold (e.g., 0.01) are pruned and deleted; It should be noted that in actual marketing scenarios, there is a serious "intent-creativity gap" problem. That is, users' interests are often implicitly expressed in their choice of specific product attributes, while the attractiveness of an advertisement depends on its creative material. Traditional collaborative filtering algorithms cannot bridge this gap. This invention utilizes the feature propagation mechanism of graph neural networks to transmit the attribute features of products to users, while aligning the creative features of advertisements with product attributes. This enables the model to uncover implicit associations between users who have not yet clicked on advertisements but whose creative style is highly consistent with their product preferences. Specifically, this solves the problem in existing technologies that cannot infer the attractiveness of advertisement creatives based on users' deep preferences.

[0020] S4: Based on the path structure and aggregation features between nodes in the final 3D marketing knowledge graph, calculate the precision marketing match degree of the target user with the candidate ads, generate a personalized recommendation list based on the precision marketing match degree, and execute precision marketing delivery based on the recommendation list.

[0021] Specifically, the process of calculating the precise marketing match degree of the target user for candidate advertisements based on the path structure and aggregation features between nodes in the final three-dimensional marketing knowledge graph, generating a personalized recommendation list based on the precise marketing match degree, and executing precise marketing delivery based on the recommendation list includes: In the final three-dimensional marketing knowledge graph, a multi-hop semantic path is defined between the target user node and the candidate ad node. The semantic path includes: user -> interest product -> similar product -> related ad, or user -> similar user -> click ad -> similar ad. The specific process for defining the multi-hop semantic path between the target user node and the candidate ad node is as follows: using the breadth-first search (BFS) algorithm, starting from the target user node and ending at the candidate ad node, search for all connected paths with no more than 4 hops, and filter out the set of valid paths containing specific meta-paths (such as UIA or UUA). Extract the final low-dimensional embedding vectors of the target user node and candidate ad node after aggregation by the graph neural network, and calculate the vector cosine similarity between the two as the basic semantic matching score; The calculation formula is: ;in This represents the basic semantic matching score. This represents the final low-dimensional embedding vector obtained after the target user node U undergoes multi-layer aggregation in a graph neural network. This represents the final low-dimensional embedding vector obtained after candidate ad node A undergoes multi-layer aggregation in a graph neural network. Represents the user node vector The Euclidean norm, Represents the vector of advertising nodes The Euclidean norm; Path weights are calculated for all connected paths between the target user node and the candidate ad node. The path weight is the product of the weights of each edge on the path. The sum of the path weights of all connected paths is used as the structural association enhancement score. The calculation formula is: ,in This represents the weight of edge e on path p. Let p represent the set of all connected paths, where p represents a specific connected path between the target user node U and the candidate advertisement node A, and e represents a specific edge on path p. By combining the basic semantic matching score with the structural association enhancement score, the precision marketing matching degree is calculated. The calculation formula is: ,in, Indicates the accuracy of targeted marketing. and This indicates an adjustable weighting coefficient. Represents the set of all connected paths. This represents the weight value of path p, where p represents a specific connected path between the target user node U and the candidate advertisement node A. Represents the set of all connected paths; It should be noted that, and Determining the optimal ratio through offline A / B testing is typically... Higher values ​​are set to enhance the reasoning effect of structural associations; Candidate ads are sorted from high to low based on their precision marketing matching accuracy, and the top N ads are selected to generate a personalized recommendation list. Obtain the current user's real-time context information (such as the current time period and geographical location), reorder the recommendations based on the personalized recommendation list, remove ads that do not fit the current context (such as not pushing food and beverage ads late at night), and obtain the final recommendation list; In this embodiment, the reordering logic is as follows: Define a set of context filtering rules. If the current time is between 22:00 and 06:00 and the advertising category is "catering / food", then multiply its matching score by a penalty coefficient of 0.1 or filter it directly. Based on the final recommendation list, the corresponding advertising creative materials are loaded in real time in the terminal interface accessed by the user (such as the APP homepage and information feed) to complete the precise marketing delivery, and the user's subsequent interaction behavior is recorded to feed back into the knowledge graph for iterative updates; The specific feedback mechanism is as follows: the user's exposure, click and conversion data after this campaign are written into the data warehouse in real time, and the feature update in step S2 and the incremental training of the knowledge graph in step S3 are triggered to realize the dynamic evolution of the knowledge graph. It's important to note that precise marketing matching not only considers the direct semantic similarity between users and ads, but also incorporates higher-order connectivity in the graph structure. This means that if a user's favorite product has a close structural relationship with a product promoted by an ad in the graph (such as complementary products or products from the same series), the system will give it a high matching score even if the user has never seen the ad before. This path-structure-based reasoning method gives marketing recommendations strong interpretability, clearly telling operators "why this ad is recommended," while also significantly improving the accuracy and conversion rate of recommendations.

[0022] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-users-products, characterized in that, Includes the following steps: S1: Collect multi-dimensional heterogeneous data from the marketing ecosystem, which covers user behavior domain, product information domain and advertising creative domain; S2: Extract semantic features and align entities from the collected multi-dimensional heterogeneous data, and construct an initial three-dimensional association map based on the semantic associations and interaction behaviors between entities; The initial three-dimensional relational graph includes user nodes, product nodes, advertising nodes, and the initial connecting edges between them; S3: Based on the graph neural network algorithm, feature propagation and neighbor aggregation are performed on the initial three-dimensional association graph to mine the implicit associations between nodes. Based on the mined implicit associations, the initial three-dimensional association graph is completed and optimized to obtain the final three-dimensional marketing knowledge graph. S4: Based on the path structure and aggregation features between nodes in the final 3D marketing knowledge graph, calculate the precision marketing match degree of the target user with the candidate ads, generate a personalized recommendation list based on the precision marketing match degree, and execute precision marketing delivery based on the recommendation list.

2. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The multi-dimensional heterogeneous data includes basic user data, static product data, dynamic advertising data, and interactive behavior data; The user basic data includes user demographic attributes, historical preference tags, real-time geographical location, and device environment information; The static product data includes product category data, product attribute data, product sales data, and product review data; The advertising dynamic data includes advertising material data, delivery strategy data, bidding data, and creative copy data; The interactive behavior data includes exposure data, click data, add-to-cart data, and conversion data.

3. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The specific process of step S1 includes: Set up a data cleaning and integration module, and link the marketing data warehouse through the data cleaning and integration module to obtain basic user data, static product data, dynamic advertising data, and interaction behavior data; The collected data is denoised and normalized, and a unique identifier index for users, products, and advertisements is established. The denoising process includes removing duplicate logs and filtering crawler traffic, and the normalization process includes performing Min-Max normalization or Z-Score standardization on numerical features.

4. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The specific process of extracting semantic features and aligning entities from the collected multi-dimensional heterogeneous data, and constructing an initial three-dimensional association map based on the semantic relationships and interaction behaviors between entities includes: Natural language processing technology is used to extract text features from advertising creative copy data and product review data to obtain text semantic vectors; Using computer vision technology, visual features are extracted from advertising material data and product images to obtain visual feature vectors; The comprehensive semantic similarity between the advertisement and the product is calculated based on the text semantic vector and visual feature vector. If the overall semantic similarity exceeds a preset threshold, a promotion-attribution association edge is established between the advertising node and the product node, and the weight of the association edge is determined by the semantic similarity value. Based on user identifiers, product identifiers, and advertising identifiers in the interaction behavior data, a browsing-purchase behavior edge is established between user nodes and product nodes, and a click-feedback behavior edge is established between user nodes and advertising nodes.

5. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 4, characterized in that, The calculation of the comprehensive semantic similarity between the advertisement and the product based on the text semantic vector and visual feature vector includes: The specific process for calculating the semantic similarity between an advertisement and a product based on text semantic vectors and visual feature vectors includes: calculating the text similarity between the advertisement copy and the product description, as well as the visual similarity between the advertisement image and the main product image, and obtaining the comprehensive semantic similarity through weighted fusion; The calculation formula is as follows: ; in, Indicates the weighting coefficient. Represents cosine similarity. and These represent the text semantic vectors for advertisements and products, respectively. and These represent the visual feature vectors of the advertisement and the product, respectively.

6. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The specific process of step S3 includes: A heterogeneous graph neural network model is constructed, which includes multiple relation aggregation layers, each of which is used to aggregate the features of neighbor nodes of a specific relation type. For user nodes, the attribute features of their neighboring product nodes and the creative features of their neighboring advertising nodes are aggregated through the relationship aggregation layer to update the feature vector of the user node. For advertising nodes, aggregate the selling features of their associated product nodes and the user profile features of historical user nodes that have interacted with them, and update the feature vector of the advertising node. Based on the updated node feature vectors, the decoder is used to calculate the link probability between user nodes and advertising nodes that are not directly connected in the graph. If the link probability is greater than the preset implicit association threshold, then potential interest association edges are added between user nodes and advertising nodes, and the graph is completed and dynamically updated.

7. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The specific process for calculating the precise marketing match degree of the target user for candidate advertisements based on the path structure and aggregation features between nodes in the final three-dimensional marketing knowledge graph includes: In the final three-dimensional marketing knowledge graph, a multi-hop semantic path is defined between the target user node and the candidate ad node; Extract the final low-dimensional embedding vectors of the target user node and candidate ad node after aggregation by graph neural network, and calculate the vector cosine similarity between the two as the basic semantic matching score; Calculate the path weights for all connected paths between the target user node and the candidate ad node, and use the sum of the path weights of all connected paths as the structural association enhancement score. The accuracy of marketing matching is calculated by combining the basic semantic matching score with the structural association enhancement score.

8. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 7, characterized in that, The formula for calculating the precision marketing matching degree is: ; in, Indicates the accuracy of targeted marketing. and This indicates an adjustable weighting coefficient. This represents the basic semantic matching score. Represents the set of all connected paths. This represents the weight value of path p, where p represents a specific connected path between the target user node U and the candidate advertisement node A. Represents the set of all connected paths; The formula for calculating the basic semantic score is: ; in This represents the final low-dimensional embedding vector obtained after the target user node U undergoes multi-layer aggregation in a graph neural network. This represents the final low-dimensional embedding vector obtained after candidate ad node A undergoes multi-layer aggregation in a graph neural network. Represents the user node vector The Euclidean norm, Represents the vector of advertising nodes The Euclidean norm.

9. The precise marketing association analysis method based on a three-dimensional knowledge graph of advertising-user-product as described in claim 1, characterized in that, The specific process of generating a personalized recommendation list based on the precision marketing match and executing precision marketing campaigns based on the recommendation list includes: Candidate ads are sorted from high to low based on their precision marketing matching accuracy, and the top N ads are selected to generate a personalized recommendation list. Obtain the current user's real-time context information, reorder it in conjunction with the personalized recommendation list, remove ads that do not fit the current context, and obtain the final recommendation list; Based on the final recommendation list, the corresponding advertising creative materials are loaded in real time on the user's terminal interface to complete precise marketing delivery, and the user's subsequent interaction behavior is recorded to feed back into the knowledge graph for iterative updates.

Images