Intelligent marketing system based on expanding consumption frequency and consumption scale

Abstract

The present invention belongs to the technical field of data analysis and processing, and specifically relates to an intelligent marketing system based on expanding consumption frequency and consumption scale, the system comprising: a data acquisition unit, a graph construction unit, a feature extraction unit, a multi-relation graph convolution model analysis unit and a promotion strategy generation unit; the data acquisition unit is used to collect user data and product data; the graph construction unit is used to create a user node set and a product node set according to the user data to obtain a multi-dimensional product and relationship graph; the feature extraction unit is used to extract the features of the multi-dimensional product and relationship graph to obtain a feature extraction result; the multi-relation graph convolution model analysis unit is used to obtain an embedded representation of the user and an embedded representation of the product; the promotion strategy generation unit is used to recommend products with a matching degree exceeding a set threshold for each user. The present invention helps to achieve the goal of expanding consumption frequency and consumption scale.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing and analysis, and specifically relates to an intelligent marketing system based on expanding consumption frequency and consumption scale. Background Art

[0002] With the rapid development of digitalization and information technology, intelligent marketing systems are increasingly being used in e-commerce and retail. Traditional marketing methods are gradually being replaced by data-driven intelligent recommendation systems. These systems analyze user behavior data, interest preferences, and consumption habits to push the right products to the right users at the right time, greatly improving the accuracy and efficiency of marketing. Existing technologies mainly rely on recommendation algorithms, such as collaborative filtering algorithms, content-based recommendations, and hybrid recommendation methods. Although these technologies have achieved personalized recommendations to a certain extent, they still have many limitations and problems in practical applications and fail to fully meet the ever-changing market needs and user expectations.

[0003] Collaborative filtering is the most commonly used recommendation method in the prior art. It makes recommendations based on the similarity between users or between users and items. Collaborative filtering can be divided into user-based collaborative filtering and item-based collaborative filtering. User-based collaborative filtering recommends items that are similar to the target user by calculating the similarity between users; while item-based collaborative filtering recommends other items that are similar to the target user by calculating the similarity between items. However, these methods rely on historical data and are susceptible to data sparsity and cold start problems. For new users or new products, the lack of sufficient historical interaction data greatly reduces the accuracy of recommendation results. In addition, the collaborative filtering method also has the problem of simplification of recommendation results and cannot fully tap the potential interests and diverse needs of users. Content-based recommendation is another common recommendation method that recommends similar items to users by analyzing the content information of items (such as item descriptions, tags, features, etc.). This method can alleviate the cold start problem of collaborative filtering to a certain extent because it does not rely on the user's historical behavior data. However, content-based recommendation also faces some challenges. First, it requires a detailed description and structuring of the characteristics of the items, which is often difficult to achieve in practical applications, especially for items without clear characteristics. Second, content-based recommendation methods tend to lead to too narrow recommendation results. Users can only receive recommendations that are very similar to the items they already like, lacking diversity and novelty. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide an intelligent marketing system based on expanding consumption frequency and consumption scale, which significantly improves the intelligent marketing system's modeling and optimization analysis capabilities of various interactive relationships between users and products, thereby helping to achieve the goal of expanding consumption frequency and consumption scale.

[0005] The technical solution adopted by the present invention is as follows:

[0006] An intelligent marketing system based on expanding consumption frequency and consumption scale, the system includes: a data collection unit, a graph construction unit, a feature extraction unit, a multi-relationship graph convolution model analysis unit and a promotion strategy generation unit; the data collection unit is used to collect user data and product data; the graph construction unit is used to create a user node set and a product node set according to the user data, each node in the product node set represents a product, each node in the user node set represents a user, and the user node set and the product node set are merged to form a total node set; according to the user data, all possible user-product interaction types are determined, and for each relationship type, all user-product interactions are traversed. If there is an interaction, an edge is created between the user node and the product node, and the weight of the edge is calculated to obtain a multi-dimensional product and relationship graph; the feature extraction unit is used to extract the features of the multi-dimensional product and relationship graph to obtain a feature extraction result; the multi-relationship graph convolution model analysis unit is used to create a multi-relationship graph convolution network based on the constructed multi-dimensional product and relationship graph and the feature extraction result, and analyze the multi-relationship convolution network to obtain the user's embedded representation and the product's embedded representation; the promotion strategy generation unit is used to calculate the matching degree of each user and product based on the user's embedded representation and the product's embedded representation, and recommend products whose matching degree exceeds a set threshold to each user.

[0007] Furthermore, the user data includes: user unique identifier, user behavior and user product evaluation index; the product data includes: product unique identifier and product price; the user behavior includes: purchase, browsing, collection, comment and add to shopping cart; each user behavior corresponds to a type of interaction between the user and the product.

[0008] Furthermore, the execution process of the data collection unit is as follows: collecting the user's unique identifier ,in , The number of users; collecting user behavior data ,in , is the product quantity, Indicates 5 types of behavior; records user product evaluation index ; Collect unique product identifiers ,in ; Record product prices ; Get the data collection process for:

[0009] ;

[0010] in, Represents a collection of users, Represents a collection of products.

[0011] Furthermore, the execution process of the graph construction unit is as follows: Create a user node set ; Create a product node set: ; Get the merged node set ; For each user behavior , create edge sets :

[0012] ;

[0013] in, Is a user To the product In user behavior type The weight of the edge under is calculated by the following formula:

[0014] ;

[0015] in, User behavior type The frequency of interaction under User behavior type The amount of interaction under For users To the product The last user behavior type The time since the interaction; For users To the product Product evaluation index; is the preset time decay factor; For users The set of neighbor nodes; is the subscript index, which takes a positive integer value; merge all edge sets to get the edge set ; Build multi-dimensional product and relationship diagrams ; .

[0016] Furthermore, the feature extraction unit first uses The convolutional neural network of the layer extracts the features of multi-dimensional products and relationship graphs. The formula is as follows:

[0017] ;

[0018] Among them, GRU is a gated recurrent unit; It is The node feature matrix of the layer; It is The node feature matrix of the layer; is a nonlinear activation function; Self-connection is added The adjacency matrix of is the adjacency matrix; yes The degree matrix of is the weight matrix, where Row, No. The elements of the column are ; BatchNorm is a batch normalization operation; and are preset hyperparameters that balance the contributions of different components; the final feature extraction result is the output of the last layer ; Set product weight to , for the The node feature matrix of the layer ; , They are The user node feature matrix and product node feature matrix of the layer.

[0019] Furthermore, the execution process of the multi-relation graph convolution model analysis unit specifically includes: initializing the user node feature matrix according to the output result of the feature extraction unit and product node feature matrix ,in is the number of users, is the number of products, is the feature dimension; for each relationship type , construct a graph convolution layer, and the update rules of each layer are as follows:

[0020] ;

[0021] ;

[0022] in: It is The adjacency matrix of the interaction types plus self-connection; yes The degree matrix of ; The final output of the graph convolution layer is and ; Introduce a relation-level attention mechanism to calculate the importance of each interaction type ; According to the importance of each relationship , use the multi-head attention mechanism to aggregate the representations of different interaction types and obtain the multi-head attention results; perform normalization on each multi-head attention result to obtain the final embedding representation of the user , and the final embedded representation of the product ; Represents the field of real numbers.

[0023] Furthermore, the importance of each interaction type is calculated using the following formula: :

[0024] ;

[0025] in, Represents a splicing operation; An operation to find the rank of a matrix.

[0026] Furthermore, the following formula is used to use the multi-head attention mechanism to aggregate the representations of different interaction types and obtain the multi-head attention result:

[0027] ;

[0028] ;

[0029] in, The multi-head attention result for the user; The multi-head attention results for the product; The function is defined as:

[0030] ;

[0031] ;

[0032] ;

[0033] in, is the number of attention heads; is the linear transformation matrix of the multi-head attention output; Calculate the attention function, calculate the similarity between the query and the key, and use it in the weighted value matrix; is the key matrix; is the value matrix; is the query matrix; For the The output of an attention head; For the The linear transformation matrix of the attention heads; for function; is the matrix transpose operation.

[0034] Furthermore, the promotion strategy generation unit calculates the matching degree of each user and product according to the embedded representation of the user and the embedded representation of the product by the following formula: :

[0035] ;

[0036] in, is the current time; For users To the product The last user behavior type The time since the interaction; This is an operation to find the modulus of a vector.

[0037] By adopting the above technical scheme, the present invention produces the following beneficial effects: the present invention comprehensively captures the multi-dimensional interaction information between users and products by constructing a multi-relationship graph. In the existing recommendation system, usually only one or several simple interaction relationships between users and products are considered, such as purchase and browsing, while ignoring other valuable interaction types that may exist between users and products (such as comments, collections, sharing, etc.). The present invention utilizes the modeling ability of multi-relationship graphs to simultaneously process multiple types of interaction information, so as to more comprehensively understand the interests and preferences of users. For example, by analyzing the browsing history and comment behavior of users, the system can more accurately judge the depth of user interest in a product, rather than just making judgments based on purchase behavior. This multi-dimensional relationship modeling not only improves the relevance and coverage of the recommendation results, but also significantly improves the user experience and satisfaction. Secondly, by using a multi-relationship graph convolutional network, the present invention realizes refined modeling and feature extraction of different interaction types. Traditional graph convolutional networks have good performance when processing a single relationship graph, but when facing a multi-relationship graph, it is often difficult to effectively model the interactions between different relationship types. The MR-GCN of the present invention can perform convolution operations on each relationship type separately to capture the characteristics and patterns of each interaction type by introducing the adjacency matrix and degree matrix of multiple relationship types in the graph convolution layer. In this way, the system can more accurately learn the influence of different interaction types on user consumption decisions in the high-dimensional feature space, making the recommendation strategy more accurate. For example, for those users who are active in browsing and collecting behaviors but have less purchasing behaviors, the system can identify their potential interests and stimulate their willingness to buy through personalized recommendation strategies, thereby increasing the frequency of consumption. Another major beneficial effect of the present invention is reflected in the full consideration of time factors and the improvement of dynamic response capabilities. User interests and behavioral habits are dynamically changing, especially in a highly competitive market environment, user preferences may change rapidly over time. Traditional recommendation systems are usually difficult to quickly adapt to such changes because they lack the ability to dynamically model time factors. The present invention can adjust the recommendation strategy according to the time dimension of user behavior by introducing a time decay factor in the matching degree calculation formula, so that the system gives a higher weight to recent user behaviors and gradually reduces the impact on earlier behaviors. This time-sensitive recommendation mechanism enables the system to capture changes in user interests in real time and provide more timely and relevant recommendations, thereby increasing user stickiness and consumption frequency. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of the system structure of an intelligent marketing system based on expanding consumption frequency and consumption scale provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0040] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0041] Example 1: Reference Figure 1 , an intelligent marketing system based on expanding consumption frequency and consumption scale, the system includes: a data collection unit, a graph construction unit, a feature extraction unit, a multi-relation graph convolution model analysis unit and a promotion strategy generation unit; the data collection unit is used to collect user data and product data; the graph construction unit is used to create a user node set and a product node set according to the user data, each node in the product node set represents a product, each node in the user node set represents a user, and the user node set and the product node set are merged to form a total node set; according to the user data, all possible interaction types between users and products are determined, and for each relationship type, all user and product interactions are traversed. Yes, if there is an interaction, an edge is created between the user node and the product node, and the weight of the edge is calculated to obtain a multi-dimensional product and relationship graph; the feature extraction unit is used to extract the features of the multi-dimensional product and relationship graph to obtain the feature extraction result; the multi-relationship graph convolution model analysis unit is used to create a multi-relationship graph convolution network based on the constructed multi-dimensional product and relationship graph and the feature extraction result, and analyze the multi-relationship convolution network to obtain the user's embedded representation and the product's embedded representation; the promotion strategy generation unit is used to calculate the matching degree of each user and product based on the user's embedded representation and the product's embedded representation, and recommend products whose matching degree exceeds a set threshold to each user.

[0042] Specifically, the data collection unit plays a vital role in the intelligent marketing system based on expanding consumption frequency and consumption scale. Its core principle is to provide a rich and accurate data basis for the subsequent analysis and decision-making of the system through multi-channel and multi-dimensional data collection. In this unit, the system first collects user behavior data and product-related data in real time through multiple channels online and offline. User behavior data mainly includes user purchase records, browsing records, click-through rates, favorites, and evaluations, which reflect the user's interest and preference for different products. In addition, the system also collects basic demographic data of users, such as age, gender, geographic location, occupation, and income level, which can help the system better understand the user's background and consumption ability, so as to design a more accurate marketing strategy. On the other hand, product data includes information such as product category, price, sales volume, and evaluation. These data are not only used to describe the market performance of the product, but also to identify the competitiveness and user acceptance of the product. By integrating these multi-dimensional data, the data collection unit can build a comprehensive user portrait and product portrait. In practical applications, the data collection unit not only relies on traditional data acquisition methods, such as transaction records in shopping mall POS systems and CRM systems, but also combines modern Internet data sources, such as user behavior data from e-commerce platforms, social media, and mobile applications. This multi-channel data collection method can ensure the diversity and comprehensiveness of data, thereby enhancing the intelligent analysis capabilities of the system. More importantly, the data collection unit introduces a real-time data update mechanism, which means that the system can immediately obtain and update the corresponding data while user behavior occurs. This real-time data collection capability significantly improves the system's sensitivity and response speed to changes in user needs, enabling the system to dynamically adjust marketing strategies and seize market opportunities in a timely manner. In addition, in order to improve the accuracy and effectiveness of data, the data collection unit also uses a variety of data cleaning and preprocessing technologies. For example, the system automatically filters out invalid data and noise data, and uses data standardization and normalization techniques to handle outliers and data deviations to ensure that the quality of input data meets the requirements of subsequent analysis. These data preprocessing technologies not only optimize data quality, but also help improve the overall operating efficiency of the system. Through the all-round data collection and processing of the data collection unit, the system can obtain accurate and comprehensive user and product data, thereby providing strong data support for subsequent graph construction, feature extraction, multi-relation graph convolution analysis and other modules. This multi-channel, multi-dimensional, real-time dynamic update data collection mechanism enables the intelligent marketing system to more accurately identify and predict user consumption behavior, formulate more targeted marketing strategies, improve user stickiness and willingness to buy, and ultimately maximize marketing effectiveness in order to expand consumption frequency and scale.

[0043] The graph construction unit is one of the most critical modules in the intelligent marketing system based on expanding consumption frequency and consumption scale. Its main principle is to convert complex user and product data into a graph structure for further analysis and optimization of marketing strategies. The design of this unit is based on the core concept that by constructing a multi-dimensional interactive relationship graph between users and products, the complex relationship between user behavior and product characteristics can be more comprehensively and finely reflected, thereby providing stronger support for intelligent recommendation and precision marketing. In the intelligent marketing system, the graph construction unit first needs to establish a user node set and a product node set based on the user and product data provided by the data collection unit. Each user node represents an individual user, including its basic information, behavior data and related feature attributes, while each product node represents a specific product, including product classification, price, evaluation and other information. The user node set and the product node set together constitute the basic node set of the system. Through these node sets, the system can clearly depict the basic information of users and products and their respective characteristics.

[0044] The core of graph construction is to create edges, i.e., relationships, between user nodes and product nodes. These edges represent the interaction between users and products, such as purchases, browsing, comments, etc. In order to build these edges, the system will first analyze all possible types of user-product interactions based on user behavior data. These types may include purchase behavior, adding to shopping carts, product collections, browsing history, product reviews, etc. For each type of interaction, the system will traverse all user-product pairs. If there is a certain type of interaction between a user and a product, the system will create an edge between the corresponding user node and product node. The establishment of these edges is not only a simple mapping of the direct relationship between users and products, but also a deep portrayal of the complex preferences and potential needs behind user behavior. Each edge will be assigned a certain weight, which is usually determined by the intensity, frequency, or importance of the interaction between users and products. The setting of weights is an important step in graph construction because it directly affects the subsequent graph analysis and marketing strategy generation. In the process of weight setting, the system will consider a variety of factors, such as the number of times a user has purchased a product, the time of the last purchase, and the user's rating of the product. Through these multi-dimensional data, the system can calculate the reasonable weight of each edge to ensure that the graph structure can truly and accurately reflect the interactive relationship and preference between users and products. After completing the edge creation of user nodes and product nodes, the graph construction unit will further optimize the entire graph structure. This optimization not only includes the simple addition or reduction of nodes and edges, but also involves adjusting the global structure of the graph, such as eliminating redundant connections, merging similar nodes, and balancing the density of the graph. Through these optimization operations, the system can ensure the simplicity and effectiveness of the graph structure and improve the accuracy and computational efficiency of subsequent analysis. An important innovation of the graph construction unit is that it can support the construction of multidimensional relationship graphs. Unlike traditional single relationship graphs, multidimensional relationship graphs can simultaneously describe multiple different types of relationships in one graph structure, which is particularly important for complex user behavior analysis. For example, the user's purchase behavior and browsing behavior for the same product can be represented by two different types of edges respectively. Such a design can capture the diversity and complexity of user behavior, thereby more accurately predicting the user's future behavior and preferences. In a multidimensional relationship graph, the graph construction unit creates independent subgraphs for each relationship type. These subgraphs share the same set of nodes but different sets of edges, and each subgraph represents a specific type of user-product interaction. In this way, the system can analyze the impact of multiple relationships at the same time under the same framework, thereby considering more dimensional factors when generating promotional strategies. The graph construction unit also has the ability to update dynamically, which enables it to respond to changes in user behavior and market environment in real time. As user behavior data is continuously updated, the system can dynamically adjust the graph structure, add new nodes and edges, or update the weights of existing nodes and edges.This dynamic adjustment mechanism enables the system to maintain high efficiency and accuracy in a rapidly changing market environment, respond to changes in user needs and market trends in a timely manner, and optimize marketing strategies. In specific applications, these characteristics of the graph construction unit enable it to fully characterize the complex relationship between users and products, providing strong data support for subsequent multi-relationship graph convolution analysis and intelligent recommendations. In this way, the system can better understand user behavior patterns and consumption habits, predict users' future consumption tendencies, and formulate personalized promotion strategies, ultimately achieving the goal of expanding consumption frequency and scale.

[0045] The feature extraction unit first performs an in-depth analysis of the multi-dimensional product and relationship graph, which involves the combined application of multiple graph analysis technologies and machine learning methods. In the initial stage of feature extraction, the system analyzes the node features and edge features in the graph. Node features include attribute information of user nodes and product nodes, such as basic demographic characteristics of users, purchase behavior characteristics, and product categories, prices, and market performance. Edge features reflect the intensity and frequency of interaction between users and products, such as the number of times users purchase a product, browsing time, and ratings. By extracting and analyzing these node and edge features, the system can preliminarily construct feature vectors of users and products, providing a basis for further analysis. However, simple node and edge features cannot fully reveal the complex relationship between users and products, so the feature extraction unit further adopts advanced graph convolution operations to better capture high-order relationships and feature interactions. Graph convolutional network (GCN) is a neural network model suitable for graph structured data. By aggregating information about nodes and their neighboring nodes in the graph, GCN can capture the importance of the node's position in the graph and its relationship with other nodes. In the intelligent marketing system of the present invention, a multi-relation graph convolutional network (MR-GCN) is applied to the feature extraction process. Unlike traditional GCN, MR-GCN can perform feature aggregation and learning on multiple relationship dimensions, thereby extracting comprehensive feature representations of users and products under different types of interactions. This method can effectively identify multiple behavioral patterns of users. For example, in some cases, users may be more inclined to buy a specific type of product, while in other cases, they are more inclined to browse and compare multiple products. By learning these complex behavioral patterns and feature interactions, the system can more accurately predict the user's consumption intention and adjust the marketing strategy accordingly. In addition, the feature extraction unit also uses a self-attention mechanism to enhance the effect of feature learning. The self-attention mechanism can automatically adjust the weight of the feature according to the importance of the node, so that the system pays more attention to the features closely related to the marketing goal during the feature extraction process, thereby improving the accuracy of the marketing strategy. For example, if a user frequently purchases a certain type of product in the near future, the self-attention mechanism will automatically increase the weight of the user node and its related edges, so that these features have a higher influence in subsequent analysis. Another important function of the feature extraction unit is to process noise and redundant features in the graph. In practical applications, due to the complexity of user behavior and the diversity of data collection, some irrelevant or redundant features are inevitable in the graph structure. In order to ensure the analysis accuracy and efficiency of the system, the feature extraction unit introduces feature selection and dimensionality reduction techniques, such as principal component analysis (PCA) and graph regularization methods. These techniques can effectively remove noise features and redundant information, retaining only the most representative and core features that best reflect user preferences, thereby optimizing the overall performance of the system.

[0046] The multi-relationship graph convolution model analysis unit first obtains a multi-dimensional product and relationship graph from the graph construction unit. This graph not only contains nodes of users and products, but also contains edges representing various interactive relationships between users and products, such as purchase, browsing, evaluation, collection, etc. Each interactive relationship is represented as a different type of edge in the graph, and these edges reflect the connection between users and products in different dimensions. When processing such graphs, traditional graph convolutional networks (GCNs) usually only consider a single type of relationship, which is obviously not enough for complex user behaviors in the real world. Therefore, the present invention introduces a multi-relationship graph convolutional network (MR-GCN), which can simultaneously perform feature learning on multiple relationship dimensions to capture the complex and diverse interaction patterns between users and products.

[0047] In the multi-relation graph convolution model analysis unit, MR-GCN gradually aggregates the feature information of nodes through layer-by-layer convolution operations. This aggregation is not limited to simple feature integration between directly adjacent nodes, but can jointly learn features from different relationship types. Specifically, MR-GCN constructs convolution kernels for each relationship type, which are used to capture node feature interactions under specific relationships. In each layer of convolution operation, the node updates its own feature representation based on the information of its neighbor nodes and the weight of the edge. As the number of convolution layers increases, the feature representation of the node not only contains the information of its direct neighbors, but also gradually integrates the features of more distant neighbors, thus forming a multi-level and multi-angle feature expression. This feature expression can reveal the user's behavior patterns under different relationship types. For example, high-frequency interactions under purchasing behavior may reflect the user's strong purchasing intention, while long-term stay under browsing behavior may point to the user's potential interest. By capturing these complex behavior patterns, the system can more accurately predict the user's future behavior and formulate corresponding marketing strategies. In addition, the multi-relation graph convolution model analysis unit further improves the accuracy and effect of feature learning through the self-attention mechanism. The self-attention mechanism can automatically adjust the importance weights of each relationship type during the feature aggregation process, so that the system can pay more attention to the relationship types that contribute most to the prediction task during the learning process. For example, when the user's purchase behavior is highly correlated with the system's marketing goals, the self-attention mechanism will automatically increase the weight of the purchase behavior edge, thereby increasing the influence of this relationship type in the feature learning process. This dynamic adjustment mechanism enables the multi-relation graph convolution network to better adapt to complex user behavior patterns and achieve more efficient and accurate feature learning in multiple relationship dimensions. Based on feature learning, the multi-relation graph convolution model analysis unit further generates embedded representations of users and products. Embedded representation is a low-dimensional vector representation that can express multiple features of users and products in a compact vector space. This low-dimensional representation not only reduces the storage and computing overhead of data, but also significantly improves the computing efficiency of subsequent models. After generating the embedded representation, the system will further calculate the matching degree between each user and product. The calculation of matching degree is usually based on the similarity of the embedding vectors of users and products, for example, measured by cosine similarity, Euclidean distance and other metrics. A high matching degree usually indicates that the user has a high interest or purchase intention in a certain product. The system can recommend the most suitable product to the user based on this information, thereby achieving the goal of precision marketing and sales conversion. It is worth mentioning that the multi-relational graph convolution model analysis unit also has the ability to dynamically adjust and self-optimize. As time goes by and user behavior data is continuously updated, the system can dynamically update the graph structure and node feature representation through an online learning mechanism, thereby continuously optimizing its prediction accuracy of user behavior.This dynamic adjustment capability enables the system to maintain efficient analysis and prediction capabilities in the face of rapidly changing market environments and changing user behaviors, respond to market changes in a timely manner, optimize marketing strategies, and achieve the goal of expanding consumption frequency and scale.

[0048] After determining the matching degree, the system will filter out those products that are highly matched with the user according to the set threshold as candidate promotion objects. Unlike traditional promotion strategies, the system of the present invention can achieve accurate user positioning and product recommendation. By giving priority to recommending products with high matching degree to corresponding users, the system not only improves the effectiveness of promotion, but also significantly improves the user's willingness to buy and consumption frequency. Furthermore, the promotion strategy generation unit can also dynamically adjust the promotion strategy according to the characteristics of different user groups and the product life cycle. For example, for users with high loyalty, the system may recommend new products or high-profit products to stimulate the user's desire to consume; while for users with low consumption frequency, the system may provide discounts or coupons and other attractive measures to increase the user's activity and participation. In addition, the unit can also self-optimize and adjust the strategy according to real-time market data and user feedback. This means that when the market environment changes or new trends appear in user behavior patterns, the system can adjust the promotion strategy in time to ensure the flexibility and efficiency of marketing activities. Through such a dynamic optimization mechanism, the promotion strategy generation unit can always maintain a high degree of consistency with user needs and market trends, and maximize sales conversion rate and customer satisfaction. Overall, the promotion strategy generation unit achieved the core goal of the intelligent marketing system in expanding consumption frequency and scale through precise matching calculation and dynamic adjustment capabilities, demonstrating its significant advantages in precision marketing and user experience optimization.

[0049] Embodiment 2: The user data includes: user unique identifier, user behavior and user product evaluation index; the product data includes: product unique identifier and product price; the user behavior includes: purchase, browsing, collection, comment and add to shopping cart; each user behavior corresponds to a type of interaction between the user and the product.

[0050] Specifically, user data includes user unique identifier, user behavior and user product evaluation index, which together constitute a complete user portrait. User unique identifier is a unique ID used to mark each user, which ensures that the system can accurately track and record all behaviors and interactions of each user. User behavior data includes purchase, browsing, collection, comment and add to shopping cart, etc. These behavior data represent various ways of interaction between users and products, and each behavior corresponds to different stages and psychological states of users in the shopping process. For example, purchase behavior directly reflects the user's consumption decision and preference, while browsing and collection behavior may indicate the user's potential interest in a certain product. In addition, comment behavior can reveal the user's experience and satisfaction with the product, and adding to shopping cart is usually a preparatory behavior before making a purchase decision. By analyzing these behavior data, the system can more accurately predict the user's consumption intention and formulate corresponding marketing strategies. At the same time, as an important part of user data, the user product evaluation index provides users with subjective ratings of different products, which can be in the form of star ratings, text comments or scoring systems. This subjective evaluation data provides an important reference dimension for the system, so that the system can not only make recommendations based on the user's behavior trajectory, but also optimize the accuracy of recommendations in combination with the user's perceptual evaluation. Combined with user behavior and evaluation index, the system can have a more comprehensive understanding of user preferences, needs and pain points, so as to formulate more personalized and targeted promotion strategies. On the other hand, product data mainly includes product unique identifiers and product prices. Product unique identifiers are used to distinguish different products, which enables the system to accurately track the sales and user feedback of each product. Product price is one of the important factors affecting user purchase decisions. In intelligent marketing systems, price information can be used to analyze user preferences for products in different price ranges, and can also be used to formulate differentiated pricing strategies and promotional activities. For example, for price-sensitive user groups, the system may tend to recommend lower-priced products or provide price discount information to improve the purchase conversion rate of these users. Each user behavior corresponds to a type of interaction between users and products, forming a rich and diverse user-product relationship data. These relationship data are modeled in the form of multi-relationship graphs, providing a solid data foundation for subsequent multi-relationship graph convolution model analysis. By analyzing the various types of interactions formed by different user behaviors, the system can identify the consumption patterns and preference characteristics of users in different scenarios, and then generate product recommendations and promotion strategies for each user that best suit their consumption habits and needs. In general, Example 2, through refined user data and product data design, enables the intelligent marketing system to capture user behavior and product characteristics more comprehensively and accurately, thereby achieving more accurate marketing decisions and strategy optimization, and truly achieving the goal of expanding consumption frequency and scale.

[0051] Embodiment 3: The execution process of the data collection unit is as follows: Collecting user unique identifier ,in , The number of users; collecting user behavior data ,in , is the product quantity, Indicates 5 types of behavior; records user product evaluation index ; Collect unique product identifiers ,in ; Record product prices ; Get the data collection process for:

[0052] ;

[0053] in, Represents a collection of users, Represents a collection of products.

[0054] Specifically, in this system, the design of the data collection unit is based on a profound understanding that the user's consumption behavior is diverse and complex, and it is difficult to accurately depict the user's real needs and interests by relying on a single data dimension. Therefore, Example 3 establishes a multi-dimensional data model by collecting multiple data types such as user unique identifiers, user behavior data, and user product evaluation indexes. This data model not only records the user's basic information and historical behavior, but also covers the user's subjective evaluation and emotional feedback on different products. Through this multi-dimensional data collection method, the system can better understand the user's consumption path and decision-making process. For example, the types of behavior such as purchase, browsing, collection, comment, and adding to the shopping cart in the user behavior data reflect the different stages and decision-making intentions of the user in the shopping process, and the relationship between these behaviors can reveal the user's purchase motivation and potential needs. At the same time, the system also pays special attention to the user's product evaluation index, which is the subjective data generated by the user in the consumption process, and can provide richer emotional and preference information than simple behavior data. By combining behavior data and evaluation data, the system can not only analyze the user's interest in a specific product, but also gain an in-depth understanding of the user's attitude towards factors such as product quality, price, and service, thereby more accurately predicting the user's future consumption behavior. This combination of data enables the system to capture user preference patterns in multiple dimensions, providing a more reliable basis for the formulation of marketing strategies. In terms of product data, the system further enriches the depth of the data model by collecting information such as the product's unique identifier and price. Product identifiers enable the system to track the market performance and user feedback of each product, while price data provides an important reference for understanding users' price sensitivity and market positioning. Price information plays a key role in the user decision-making process, and users' sensitivity to price directly affects their purchasing decisions and consumption frequency. By analyzing this data, the system can design differentiated pricing strategies and promotional activities for different user groups to maximize sales conversion rates and user engagement. The core of the data collection process lies in its structured and standardized design, which ensures the integrity and consistency of the data, enabling the system to efficiently process and utilize this data in subsequent analysis. Through the function , the system can organically combine user and product data to form a structured data set. This set not only provides the relationship between users and products, but also reveals the potential patterns and trends behind these relationships, providing the necessary data support for the multi-relationship graph convolution model analysis unit. This method effectively improves the efficiency of data utilization, enabling the intelligent marketing system to quickly and accurately identify user needs and market opportunities in a large and complex data environment.

[0055] Example 4: The execution process of the graph construction unit is as follows: Create a user node set ; Create a product node set: ; Get the merged node set ; For each user behavior , create edge sets :

[0056] ;

[0057] in, Is a user To the product In user behavior type The weight of the edge under is calculated by the following formula:

[0058] ;

[0059] in, User behavior type The frequency of interaction under User behavior type The amount of interaction under For users To the product The last user behavior type The time since the interaction; For users To the product Product evaluation index; is the preset time decay factor; For users The set of neighbor nodes of is the subscript index, which takes a positive integer value; merge all edge sets to get the edge set ; Build multi-dimensional product and relationship diagrams ; .

[0060] Specifically, in Example 4, the graph construction unit maps the various interactive behaviors between users and products into edges and their weights in the graph structure through a carefully designed formula. This design provides multi-dimensional relational data for the intelligent marketing system, allowing the system to more accurately capture the user's behavioral characteristics and consumption preferences. The core principle of the formula is to calculate the weight of the edge by comprehensively considering factors of multiple dimensions such as interaction frequency, interaction amount, time factor, and user evaluation, so as to achieve accurate modeling of the complex relationship between users and products. First, the formula takes into account the frequency of interaction between users and products. , which is a key factor reflecting the strength of user behavior. The higher the frequency, the greater the user's attention and interest in the product. Therefore, in the weight calculation, the interaction frequency acts as a direct multiplier and significantly affects the weight of the edge. However, the frequency itself cannot fully reflect the strength of the relationship between the user and the product, especially when it comes to high-value products or high-quality user interactions. Therefore, the formula introduces the interaction amount The logarithmic correction term , this part aims to nonlinearly amplify the influence of high-value transactions. By taking the logarithm of the amount, the formula can mitigate the impact of the amount growth to a certain extent, but will not completely ignore the special significance of high-value interactions. This design allows the weight calculation to take into account both the breadth (frequency) of user behavior and the depth and value (amount) of user behavior, thereby more accurately reflecting the user's true interest and preference for the product.

[0061] In addition, the time factor The introduction of reflects the system's attention to the dynamics of user behavior. In real scenarios, the influence of user behavior will gradually decay over time. For this reason, a time decay factor is added to the formula and an exponential function , which is used to simulate the time decay effect of the behavior. Controls the influence of time on weight. When the value is large, the weight will decay quickly, and vice versa. This design takes into account the timeliness of user behavior, that is, the more recent the behavior, the greater the impact on the user's current preference, while the earlier behavior will have a lower reference value for the current decision. In this way, the system can dynamically adjust the weight calculation to ensure that the recommendation results are more in line with the user's current interests and needs. In addition, the user product evaluation index The addition of further enhances the accuracy of weight calculation. This index reflects the user's subjective satisfaction or preference for the product, which is an important feedback information left by the user during the consumption process. By multiplying the evaluation index with the weight calculation, the system can reflect the user's emotions and attitudes to a certain extent. Positive evaluations will increase the weight, thereby increasing the possibility of the product being recommended, while negative evaluations will reduce the weight and reduce the chance of recommendation. In this way, the system not only relies on objective behavioral data, but also combines the user's subjective feedback to make the recommendation results more personalized and humane. The denominator in the formula plays a normalization role. It standardizes the weights of all possible user-product interactions to ensure that the calculated weights are comparable between different users and products. This normalization process avoids weight distortion caused by extreme behaviors of individual users or products (such as high frequency or high amount of extreme transactions), making the entire graph model more robust and fair. By integrating these multi-dimensional factors, the weight formula realizes a multi-angle characterization of the user-product relationship and can fully reflect the complex interaction patterns of users under different behavior types. Ultimately, these weighted edge sets The combination of forms a complete multi-dimensional product and relationship diagram , where each edge represents a certain relationship between the user and the product and its strength. This multi-relationship graph is not just a simple summary of the user-product interaction data. It structurally reflects the diversity and complexity of user behavior and provides rich semantic information for subsequent intelligent analysis and strategy formulation.

[0062] Embodiment 5: The feature extraction unit first uses The convolutional neural network of the layer extracts the features of multi-dimensional products and relationship graphs. The formula is as follows:

[0063] ;

[0064] Among them, GRU is a gated recurrent unit; It is The node feature matrix of the layer; It is The node feature matrix of the layer; is a nonlinear activation function; Self-connection is added The adjacency matrix of is the adjacency matrix; yes The degree matrix of is the weight matrix, where Row, No. The elements of the column are ; BatchNorm is a batch normalization operation; and are preset hyperparameters that balance the contributions of different components; the final feature extraction result is the output of the last layer ; Set product weight to , for the The node feature matrix of the layer ; , They are The user node feature matrix and product node feature matrix of the layer.

[0065] Specifically, the key to feature extraction lies in the design of a multi-layer graph convolutional network, which can aggregate the information of a node and its neighboring nodes in each layer of convolution operations, and mine the local and global structural features of the graph layer by layer. First, in the core formula of the convolutional neural network, the node feature matrix The update process is based on the normalized adjacency matrix Sum degree matrix This design introduces the structural information of the graph, where It not only contains the connection relationship of the original graph, but also adds the self-connection matrix , ensuring that each node not only integrates the information of neighboring nodes in each convolution layer, but also retains its own feature information. The addition of self-connection is to prevent the information from being over-smoothed in multiple layers of convolution, so that the node features can still maintain a certain degree of individuality and distinction, which is particularly important for capturing the unique behavior patterns of users and products. In the feature update process, the weight matrix Linear transformation used to learn features. These weights are continuously optimized through the training process to minimize a specific objective function (such as prediction error). This process enables the convolutional neural network to learn the most recognizable feature combination in the node feature space, thereby improving the system's ability to identify user preferences and product relevance. The optimization and learning of these features not only takes into account the direct interaction behavior of users, but also captures the potential influence between users and the mutual correlation between products through the graph structure, allowing the system to understand market behavior at a higher dimension. The formula also introduces a combination of batch normalization (BatchNorm) and nonlinear activation functions (such as ReLU), which are key technologies for improving model performance and stability in deep learning networks. Batch normalization standardizes the output of each layer to eliminate the distribution differences between different input batches, thereby accelerating model training and improving generalization capabilities. Activation function The introduction of gives the model nonlinear fitting capabilities, enabling it to handle more complex patterns and data distributions. In the feature extraction process, this nonlinear transformation can effectively enhance the expressiveness of the model and capture more subtle interaction features between users and products, which is crucial for accurate user behavior prediction and product recommendation. Further enhancing the complexity and expressiveness of the model is the integration of the gated recurrent unit (GRU). As a simplified recursive neural network structure, GRU can efficiently process long-term and short-term dependencies in sequence data. In the feature extraction unit of the present invention, the introduction of GRU enables the system to capture the temporal dynamics of user behavior and product interaction. By combining the node features of the current layer and the previous layer, GRU can memorize the historical patterns of user behavior and predict future behavior based on these patterns. The gate mechanism of GRU allows the network to automatically regulate the flow of information, which is particularly useful for processing complex user behavior sequence data. Through this mechanism, the system can not only capture users' short-term behavior preferences, but also identify users' long-term consumption habits and interest changes, so that the recommendation results are more in line with users' actual needs and market trends. In addition, the differentiated processing of user nodes and product nodes in the feature extraction unit further improves the flexibility and adaptability of the system. By introducing product weights , the system can adjust the relative importance of user features and product features during feature extraction. This weight adjustment is set based on different application scenarios and marketing needs. For example, in some situations, user behavior features may be more critical, so the weight of product features can be reduced. , and vice versa. This weight adjustment mechanism ensures that the system can be flexibly applied in different market environments and marketing strategies, maximizes the use of information in the data, and improves the effectiveness of recommendations and marketing. The final output of the formula is the node feature matrix after multi-layer convolution and nonlinear transformation. , this matrix integrates the feature information of all L layers, forming a highly condensed feature representation that covers the multi-dimensional interaction information between users and products. Through this high-dimensional feature representation, the intelligent marketing system can more accurately locate user needs in a complex market environment, optimize product recommendation strategies, and ultimately increase the frequency and scale of user consumption. This feature extraction capability of the system not only improves the accuracy and precision of the recommendation algorithm, but also significantly improves the user experience and enhances user stickiness and engagement.

[0066] Embodiment 6: The execution process of the multi-relation graph convolution model analysis unit specifically includes: initializing the user node feature matrix according to the output result of the feature extraction unit and product node feature matrix ,in is the number of users, is the number of products, is the feature dimension; for each relationship type , construct a graph convolution layer, and the update rules of each layer are as follows:

[0067] ;

[0068] ;

[0069] in: It is The adjacency matrix of the interaction types plus self-connection; yes The degree matrix of ; The final output of the graph convolution layer is and ; Introduce a relation-level attention mechanism to calculate the importance of each interaction type ; According to the importance of each relationship , use the multi-head attention mechanism to aggregate the representations of different interaction types and obtain the multi-head attention results; perform normalization on each multi-head attention result to obtain the final embedding representation of the user , and the final embedded representation of the product ; Represents the field of real numbers.

[0070] Specifically, the unit obtains the initial user node feature matrix from the feature extraction unit and product node feature matrix , each feature matrix represents the representation of users and products in high-dimensional space. These initial features provide basic input for subsequent graph convolution operations, allowing the system to mine and refine the deeper relationship between users and products layer by layer through further calculations. In the graph convolution operation, the user features and product features of each layer are connected through the adjacency matrix Sum degree matrix The purpose of these matrices is to capture the structured information between user and product nodes, especially in the context of interactions of multiple different types of relationships. For example, a user may have different types of interactions with multiple products in the system, such as purchase, browsing, and commenting, and each interaction can provide different feature information for the node. In each layer of convolution, by normalizing the adjacency matrix, the system can ensure that the information propagation between nodes is balanced and avoid excessive concentration of features due to the high connectivity of some nodes. This normalization operation helps to maintain the diversity of features, so that the feature update of each node takes into account both its own characteristics and the influence of its neighboring nodes. The introduction of the gated recurrent unit (GRU) is another key part of the formula. GRU is essentially a recursive neural network that controls the flow and memory of information by introducing a gate mechanism, which is particularly effective for processing time series data. In the context of the present invention, the user's behavior is dynamic and may change over time. Therefore, GRU can capture the temporal dynamics of user behavior and product interaction by combining the features of the current layer and the previous layer. This consideration of the time dimension enables the model to more accurately reflect the user's actual consumption pattern, thereby improving the accuracy of the recommendation. For example, if a user has recently shown a high interest in a certain category of products, the GRU can increase the weight of this information in the feature update, making the system more inclined to recommend related products.

[0071] In the graph convolution process, user node features and product node features are gradually extracted through multi-level convolution. As the number of convolution layers increases, the system is able to extract information from a wider range of graph structures, making the final feature expression of the node more comprehensive and rich. The convolution operation of each layer not only integrates the features of the current node, but also combines the features of neighboring nodes, and considers the influence of historical information through the GRU mechanism. This combination of multi-layer convolution and recursive mechanisms enables the system to capture more complex user-product relationships and behavior patterns, providing stronger data support for subsequent decision-making. In addition, in order to further improve the expressiveness and accuracy of the model, the system introduces a relationship-level attention mechanism. The role of this mechanism is to evaluate and weight the importance of different relationship types. In practical applications, different user behaviors may have different effects on marketing strategies. For example, purchasing behavior usually reflects the user's consumption intention more directly than browsing behavior. By calculating the importance weight of each relationship type , the system can identify which relationships are most critical to the current marketing goals. Subsequently, the multi-head attention mechanism applies these weights to different relationship types, aggregates the feature representations of these relationship types, and generates the final node embedding representation. This multi-head attention mechanism not only allows the model to consider the heterogeneity of multiple relationship types when extracting features, but also ensures that the contribution of each relationship type in the final feature representation is balanced and reasonable through normalization operations. Finally, the embedding representation of users and products and As the output of the model, they represent the location and characteristics of users and products in high-dimensional space. Through these embedded representations, the system can further perform personalized recommendations, build user portraits, and optimize marketing strategies.

[0072] Example 7: Calculate the importance of each interaction type using the following formula: :

[0073] ;

[0074] in, Represents a splicing operation; An operation to find the rank of a matrix.

[0075] Specifically, in Example 7, the multi-relation graph convolution model analysis unit calculates the importance of each interaction type , which provides a mechanism for intelligent marketing systems to deeply understand the multi-dimensional interactions between users and products. The core of this formula is to use nonlinear functions and matrix operations to measure and compare the relative contributions of different interaction types to user consumption behavior. By calculating these weights, the system can more accurately focus on the interaction types that most affect user behavior, thereby optimizing recommendation strategies, improving marketing effectiveness, and achieving the goal of expanding consumption frequency and scale. The design of the formula uses a nonlinear activation function , which is used in this formula to map the values ​​in the feature space to a range between -1 and 1. The activation function is used to perform nonlinear transformation on the input features, making the distribution of features more uniform, which helps the model better capture complex relationship patterns during the learning process. In the formula, The input of the function is a feature extraction result and The concatenated matrix, this concatenation operation ( ) combines user features and product features to form a comprehensive feature representation. Next, the formula introduces an operation to find the matrix rank , this operation is used to measure the The rank of a matrix represents the number of linearly independent rows or columns, reflecting the dimension and amount of information in the matrix. In this context, the calculation of the rank helps the system understand the diversity and complexity of the interaction data. For example, if an interaction type appears frequently in the data and is more evenly distributed, then its rank value is higher, which means that the interaction type provides more information. By multiplying the rank by the feature matrix, the system can amplify or reduce the influence of different features, thereby making subtle adjustments to the importance of each interaction type. Next, the formula uses an exponential function to further enhance the relative differences of these weighted features. The use of the exponential function here has a dual meaning: first, it converts all feature values ​​into positive numbers, ensuring that the final calculated weights are non-negative, which is necessary for the calculation of probability distributions; second, the exponential function has the effect of amplifying larger input values, thereby highlighting those interaction types that have a greater impact on user behavior. Through this amplification effect, the system can more keenly identify key user behavior patterns and product features. Finally, the normalization operation of the formula ensures that the sum of the importance weights of all interaction types is 1 by summing the exponential results of each interaction type. This normalization process is necessary because it converts the importance of each interaction type into a probability distribution, making these weights intuitively explanatory: a larger weight Indicates that this interaction type is more important for user behavior prediction and product recommendation. In this way, the system can dynamically adjust the attention paid to different interaction types, thereby giving priority to those behavior patterns that have the greatest impact on user decision-making in personalized recommendations and marketing strategy formulation.

[0076] Example 8: Using the following formula, a multi-head attention mechanism is used to aggregate representations of different interaction types to obtain a multi-head attention result:

[0077] ;

[0078] ;

[0079] in, The multi-head attention result for the user; The multi-head attention results for the product; The function is defined as:

[0080] ;

[0081] ;

[0082] ;

[0083] in, is the number of attention heads; is the linear transformation matrix of the multi-head attention output; Calculate the attention function, calculate the similarity between the query and the key, and use it in the weighted value matrix; is the key matrix; is the value matrix; is the query matrix; For the The output of an attention head; For the The linear transformation matrix of the attention heads; for function; is the matrix transpose operation.

[0084] Specifically, the multi-head attention mechanism in the formula is implemented by the function Aggregate representations of different interaction types to obtain multi-head attention results for users and products and These results are for user node features and product node features The core idea of ​​the multi-head attention mechanism is to capture the diversity and complexity of data through multiple attention heads. Each attention head independently learns a specific pattern or relationship in the data, and then generates a more comprehensive and accurate feature representation by combining these patterns. In the multi-head attention mechanism, the function First, the output of multiple attention heads Concatenate (Concat), and then pass a linear transformation matrix The output of each attention head is is calculated by attention function The principle of the attention calculation function is based on the query matrix , key matrix Sum Matrix The similarity between them is used for weighted calculation. In the formula, and Normalize the product of ,in is the feature dimension) and apply Function, calculates the attention weights between the query and the key. These weights reflect the relative importance of different features and are used to weight the value matrix In this way, the multi-head attention mechanism can focus on multiple aspects of the data at the same time, thereby effectively capturing a variety of complex relationships between users and products. For example, one attention head may focus on the user's purchasing behavior, while another attention head may pay more attention to the user's browsing history or product reviews. By considering these different types of interactions at the same time, the system can more comprehensively understand the user's interests and preferences, and adjust recommendation strategies and marketing activities accordingly. Another major advantage of the multi-head attention mechanism is that it can improve the robustness and generalization ability of the model. Since each attention head works independently and focuses on different aspects of the data, their combination makes the final feature representation richer and more diverse, reducing the reliance on a single feature or pattern. This feature is particularly suitable for intelligent marketing systems because it can handle noise and anomalies in large amounts of user and product data, ensuring the accuracy and reliability of recommendation results. In practical applications, the multi-head attention results of users Multi-head attention results for products It provides efficient aggregation and optimization of deep learning models on user and product feature representation. By taking user features and product features as query matrices respectively , key matrix Sum Matrix , the system can effectively calculate and utilize the complex interactions between these features, thereby improving the relevance of recommendation results and user experience. Finally, the attention calculation results integrate the outputs of multiple attention heads through linear transformation and splicing operations to form a rich multi-dimensional feature representation. These representations include not only the direct relationship between users and products, but also the indirect connections and potential patterns between them. Through this deep feature learning and representation, the multi-head attention mechanism provides powerful analysis and prediction capabilities for intelligent marketing systems, enabling the system to more accurately perform personalized recommendations and optimize marketing strategies, ultimately achieving the goal of expanding consumption frequency and scale.

[0085] Embodiment 9: The promotion strategy generation unit calculates the matching degree between each user and product according to the embedded representation of the user and the embedded representation of the product by the following formula: :

[0086] ;

[0087] in, is the current time; For users To the product The last user behavior type The time since the interaction; This is an operation to find the modulus of a vector.

[0088] Specifically, the first part of the formula It is a classic cosine similarity calculation method used to measure user embedding representation and product embedding representation The cosine similarity ranges from -1 to 1. The closer the value is to 1, the higher the match between the user and the product. In the intelligent marketing system, the embedded representation of users and products is extracted from a complex multi-relational graph through a multi-layer convolutional network and a multi-head attention mechanism, which contains in-depth information about user behavior and product features. Therefore, cosine similarity can effectively capture the potential relationship between users and products and use it for recommendation decisions. This part of the calculation ensures that the recommendation system gives priority to products that are highly matched with user interests and needs. The second part of the formula The time decay factor is introduced to measure the impact of the timeliness of user behavior on the matching degree. Here, is the current time, Is a user About Products The last behavior type The interaction time. Time decay factor Controls the decay rate of the influence of user behavior. This design reflects an important marketing strategy principle: the user's latest behavior reflects the current interests and needs better than past behavior. Therefore, the more recent user behavior contributes to the matching degree, and the influence of behavior will gradually weaken over time. In this way, the system can adjust the recommendation strategy more dynamically to make it closer to the user's current needs and market environment. The third part of the formula It is a correction item based on behavioral diversity. This part uses the behavioral weights of users in different interaction types (such as purchase, browsing, commenting, etc.) The matching degree is adjusted by the sum of the norms of . The purpose of this design is to consider the comprehensive performance of users under various behavior types. If a user has high-frequency and high-weight interactions in multiple behavior types, this indicates that the user has a multi-faceted interest in the product and high interactivity, so the matching degree will be appropriately improved. On the contrary, if the user's behavior is concentrated in a certain type and lacks diversity, this may mean that the user's interests are limited or the behavior data is insufficient, and the matching degree will be adjusted accordingly. This matching degree calculation method reflects the multi-dimensional analysis capability of the intelligent marketing system. It not only considers the feature similarity between users and products, but also combines time factors and behavior diversity to ensure that the recommended results are more accurate and personalized. By introducing time decay and behavior diversity correction, the system can effectively cope with changes in user interests and the complexity of behavior patterns, and maintain sensitivity and responsiveness to user needs.

[0089] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. An intelligent marketing system based on expanding consumption frequency and consumption scale, characterized by: The system includes: a data collection unit, a graph construction unit, a feature extraction unit, a multi-relationship graph convolution model analysis unit and a promotion strategy generation unit; the data collection unit is used to collect user data and product data; the graph construction unit is used to create a user node set and a product node set according to the user data, each node in the product node set represents a product, each node in the user node set represents a user, and the user node set and the product node set are merged to form a total node set; according to the user data, all possible types of interaction between users and products are determined, for each type of relationship, all user and product pairs are traversed, if there is an interaction, an edge is created between the user node and the product node, and the weight of the edge is calculated to obtain a multidimensional product and relationship graph; the feature extraction unit is used to extract the features of the multidimensional product and relationship graph to obtain The multi-relation graph convolution model analysis unit is used to create a multi-relation graph convolution network based on the constructed multi-dimensional product and relationship graph and the feature extraction results, analyze the multi-relation convolution network, and obtain the user's embedded representation and the product's embedded representation; the promotion strategy generation unit is used to calculate the matching degree of each user and product based on the user's embedded representation and the product's embedded representation, and recommend products with a matching degree exceeding a set threshold for each user; the user data includes: user unique identifier, user behavior and user product evaluation index; the product data includes: product unique identifier and product price; the user behavior includes: purchase, browsing, collection, comment and add to shopping cart; each user behavior corresponds to a type of interaction between the user and the product; the execution process of the data acquisition unit is as follows: collect user unique identifier ,in , The number of users; collecting user behavior data ,in , is the product quantity, Indicates 5 types of behavior; records user product evaluation index ; Collect unique product identifiers ,in ; Record product prices ; Get the data collection process for: ； in, Represents a collection of users, Represents a product set; the execution process of the graph construction unit is as follows: Create a user node set ; Create a product node set: ; Get the merged node set ; For each user behavior , create edge sets : ； in, Is a user To the product In user behavior type The weight of the edge under is calculated by the following formula: ； in, User behavior type The frequency of interaction under User behavior type The amount of interaction under For users To the product The last user behavior type The time since the interaction; For users To the product Product evaluation index; is the preset time decay factor; For users The set of neighbor nodes of is the subscript index, which takes a positive integer value; merge all edge sets to get the edge set ; Indicates edge sets; build multidimensional product and relationship graphs ; .

2. The intelligent marketing system based on expanding consumption frequency and consumption scale as claimed in claim 1, characterized in that: The feature extraction unit first uses The convolutional neural network of the layer extracts the features of multi-dimensional products and relationship graphs. The formula is as follows: ； Among them, GRU is a gated recurrent unit; It is The node feature matrix of the layer; It is The node feature matrix of the layer; is a nonlinear activation function; Self-connection is added The adjacency matrix of is the adjacency matrix; yes The degree matrix of is the weight matrix, where Row, No. The elements of the column are ; BatchNorm is a batch normalization operation; and are preset hyperparameters that balance the contributions of different components; the final feature extraction result is the output of the last layer ; Set product weight to , for the The node feature matrix of the layer ; , They are The user node feature matrix and product node feature matrix of the layer.

3. The intelligent marketing system based on expanding consumption frequency and consumption scale as claimed in claim 2, characterized in that: The execution process of the multi-relation graph convolution model analysis unit specifically includes: initializing the user node feature matrix according to the output result of the feature extraction unit and product node feature matrix ,in is the number of users, is the number of products, is the feature dimension; for each relationship type , construct a graph convolution layer, and the update rules of each layer are as follows: ； ； in: It is The adjacency matrix of the interaction types plus self-connection; yes The degree matrix of ; The final output of the graph convolution layer is and ; Introduce a relation-level attention mechanism to calculate the importance of each interaction type ; According to the importance of each relationship , use the multi-head attention mechanism to aggregate the representations of different interaction types and obtain the multi-head attention results; perform normalization on each multi-head attention result to obtain the final embedding representation of the user , and the final embedded representation of the product ; Represents the field of real numbers.

4. The intelligent marketing system based on expanding consumption frequency and consumption scale as claimed in claim 3, characterized in that: The importance of each interaction type is calculated using the following formula: : ； in, Represents a splicing operation; An operation to find the rank of a matrix.

5. The intelligent marketing system based on expanding consumption frequency and consumption scale as claimed in claim 4, characterized in that: The following formula is used to use the multi-head attention mechanism to aggregate the representations of different interaction types and obtain the multi-head attention result: ； ； in, The multi-head attention result for the user; The multi-head attention results for the product; The function is defined as: ； ； ； in, is the number of attention heads; is the linear transformation matrix of the multi-head attention output; Calculate the attention function, calculate the similarity between the query and the key, and use it in the weighted value matrix; is the key matrix; is the value matrix; is the query matrix; For the The output of an attention head; For the The linear transformation matrix of the attention heads; for function; is the matrix transpose operation.

6. The intelligent marketing system based on expanding consumption frequency and consumption scale as claimed in claim 5, characterized in that: The promotion strategy generation unit calculates the matching degree of each user and product according to the embedded representation of the user and the embedded representation of the product by the following formula: : ； in, is the current time; For users To the product The last user behavior type The time since the interaction; This is an operation to find the modulus of a vector.

Images