A digital marketing recommendation method and system based on user behavior data analysis

Abstract

The application provides a digital marketing recommendation method and system based on user behavior data analysis, constructs a personal behavior knowledge graph and a marketing public knowledge graph to generate a fusion knowledge graph of users and goods; introduces a user interest transmission mechanism based on time decay on the fusion knowledge graph, extracts a user dynamic behavior code representing the current behavior preference tendency of the user; generates a user preference feature vector based on the user dynamic behavior code and historical interaction preference information of the user and the goods; generates a product inherent feature vector of the marketing attribute of each to-be-recommended product according to the fusion knowledge graph, and generates a personalized recommendation list of each to-be-recommended product from the user preference feature vector and the product inherent feature vector of each to-be-recommended product to perform marketing recommendation on each to-be-recommended product. The scheme of the application can effectively fuse the individual behavior dynamics and the structured constraint of the general correlation knowledge between goods to perform digital marketing recommendation on the goods.

Description

Technical Field

[0001] This application relates to the field of digital marketing recommendation technology, and more specifically, to a digital marketing recommendation method and system based on user behavior data analysis. Background Technology

[0002] Digital marketing recommendation is an intelligent decision-making technology system driven by massive user behavior data and built on machine learning and data mining algorithms. It aims to filter and sort a set of personalized items that best match the current intent of the target user from a large set of product or content candidates in real time.

[0003] With the continuous evolution of digital marketing technologies, personalized recommendation systems based on user behavior data analysis have become a core means to improve user stickiness and conversion efficiency. Existing methods often rely on single data sources such as historical click sequences or collaborative filtering for modeling, making it difficult to effectively integrate general domain knowledge such as product category hierarchy, functional complementarity, and market co-occurrence. This results in insufficient generalization ability in scenarios involving user cold starts and long-tail products. Furthermore, user consumption decisions are influenced by both short-term interest fluctuations and long-term steady-state preferences. Existing models do not adequately characterize the timeliness of behavior, lack a time-decay-based interest transfer mechanism, and fail to adaptively and collaboratively express short- and long-term preferences. In addition, recommendation ranking often uses click probability prediction as the sole optimization objective, failing to organically integrate dynamic signals from business operation dimensions such as product inventory, promotional intensity, and advertising bids. This makes it difficult to flexibly balance user experience optimization and maximizing platform marketing revenue. Therefore, how to effectively integrate the structured constraints of individual behavioral dynamics and general correlation knowledge between products for digital marketing recommendations has become a challenge for the industry. Summary of the Invention

[0004] This application provides a digital marketing recommendation method and system based on user behavior data analysis, which can effectively integrate the structured constraints of individual behavioral dynamics and general correlation knowledge between products to make digital marketing recommendations for products.

[0005] Firstly, this application provides a digital marketing recommendation method based on user behavior data analysis, comprising the following steps:

[0006] Construct a personal behavior knowledge graph that includes user behavior trajectories and preference associations, as well as a marketing public knowledge graph that includes general product association rules;

[0007] The personal behavior knowledge graph and the marketing public knowledge graph are mapped to the association between products and user entities to obtain a fused knowledge graph of users and products.

[0008] A time-decay-based user interest transfer mechanism is introduced into the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and to extract user dynamic behavior codes that represent the user's current behavioral preference tendencies.

[0009] Based on the user dynamic behavior encoding and the user's historical interaction preference information with the product, a user preference feature vector of the user's short-term interest in the product is generated.

[0010] Based on the fused knowledge graph, product-specific feature vectors of marketing attributes for each product to be recommended are generated. Then, a personalized recommendation list for each product to be recommended is generated from the user preference feature vector and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

[0011] In some embodiments, constructing a personal behavior knowledge graph containing user behavior trajectories and preference associations, and a marketing public knowledge graph containing general product association rules specifically includes:

[0012] Obtain the user's historical interaction sequence across multiple devices;

[0013] Based on the historical interaction behavior sequence, a behavior subgraph rooted at the user node is constructed, and directed edges between behavior nodes and product nodes are introduced in the behavior subgraph to construct a personal behavior knowledge graph.

[0014] Extract the category hierarchy, functional complementarity, and market co-occurrence relationships between products from external marketing knowledge bases and product attribute databases;

[0015] Based on the aforementioned category hierarchy, functional complementarity, and market co-occurrence relationships, a marketing public knowledge graph containing the strength of general association rules between products is constructed, with product nodes as entities.

[0016] In some embodiments, mapping the personal behavior knowledge graph and the marketing public knowledge graph to associate products with user entities to obtain a fused knowledge graph of users and products specifically includes:

[0017] The identifiers of product node entities in the personal behavior knowledge graph and product node entities in the marketing public knowledge graph are identified to be consistent. The attribute information of the same product node in the two graphs is merged through entity alignment operation.

[0018] Identify the products that user nodes are indirectly associated with through behavioral edges in the personal behavior knowledge graph, and establish virtual interest edges between user nodes and extended product nodes in the marketing public knowledge graph by extending the product set through general association rules.

[0019] Based on the merged product nodes, user nodes, and a set of mixed edges including behavioral edges and virtual interest edges, a fused knowledge graph of users and products is constructed.

[0020] In some embodiments, a time-decay-based user interest transfer mechanism is introduced into the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and to extract user dynamic behavior codes representing the user's current behavioral preference tendencies. Specifically, this includes:

[0021] Starting with the user node as the central node, perform a multi-level neighborhood random walk along the behavior edges and virtual interest edges on the fused knowledge graph;

[0022] During the walk, for each behavior edge traversed, a time decay factor is calculated based on the time difference between the occurrence time of the behavior event and the current recommended time.

[0023] For each product node reached by the terminal of a traversal path, the weights of each edge on the aggregated path and the time decay factor determine the contribution score of that product node to the user's current preference.

[0024] Based on the contribution score, each candidate product node is sorted and a vector representation of a preset dimension is extracted to form a dynamic user behavior code that represents the user's current behavioral preference.

[0025] In some embodiments, generating a user preference feature vector of short-term interest in a product based on the user dynamic behavior encoding and the user's historical interaction preference information with the product specifically includes:

[0026] Obtain information on users' historical interaction preferences with products;

[0027] Extract the user's long-term steady-state preference vector from the historical interaction preference information;

[0028] The user dynamic behavior encoding is fused with the user's long-term steady-state preference vector element by element to generate a user preference feature vector of the user's short-term interest in the product.

[0029] In some embodiments, generating product-specific feature vectors for each marketing attribute of the product to be recommended based on the fused knowledge graph specifically includes:

[0030] For each product to be recommended, extract other product nodes and their corresponding association types within the adjacent one-hop range of the product to be recommended from the fused knowledge graph;

[0031] Based on the other product nodes within the adjacent one-hop range of each product to be recommended and the corresponding association type, calculate the context-enhanced embedding vector of the corresponding product to be recommended;

[0032] The original attribute encoding vector of each product to be recommended is concatenated with the corresponding context-enhanced embedding vector to determine the product-specific feature vector of the marketing attributes of each product to be recommended.

[0033] In some embodiments, generating a personalized recommendation list for each product to be recommended from the user preference feature vector and the product-specific feature vector of each product to be recommended specifically includes:

[0034] Based on the user preference feature vector and the inherent feature vector of each product to be recommended, the estimated click probability rating of the user for each product to be recommended is determined.

[0035] Obtain the marketing strategy weighting factors for each product to be recommended;

[0036] The overall recommendation priority score for each product to be recommended is determined based on the estimated click probability score and marketing strategy weighting factors.

[0037] A personalized recommendation list for each product is generated based on all the overall recommendation priority scores.

[0038] Secondly, this application provides a digital marketing recommendation system based on user behavior data analysis, comprising:

[0039] The building module is used to construct a personal behavior knowledge graph containing user behavior trajectories and preference associations, as well as a marketing public knowledge graph containing general product association rules;

[0040] The processing module is used to perform association mapping between products and user entities on the personal behavior knowledge graph and the marketing public knowledge graph to obtain a fused knowledge graph of users and products.

[0041] The processing module is also used to introduce a time-decay-based user interest transmission mechanism on the fused knowledge graph, perform preference propagation analysis on the interaction relationship between users and product nodes, and extract user dynamic behavior codes that represent the user's current behavioral preference tendencies.

[0042] The processing module is also used to generate a user preference feature vector of short-term interest in products based on the user dynamic behavior encoding and the user's historical interaction preference information with products;

[0043] The execution module is used to generate product-specific feature vectors of marketing attributes for each product to be recommended based on the fused knowledge graph, and to generate a personalized recommendation list for each product to be recommended based on the user preference feature vectors and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

[0044] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the aforementioned digital marketing recommendation method based on user behavior data analysis.

[0045] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned digital marketing recommendation method based on user behavior data analysis.

[0046] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0047] The digital marketing recommendation method and system based on user behavior data analysis provided in this application constructs a personal behavior knowledge graph and a marketing public knowledge graph. The former accurately captures the individual dynamic trajectory and preference intensity evolution of user-product interactions, while the latter extracts structured prior knowledge such as category hierarchy, functional complementarity, and market co-occurrence from a global perspective. Both graphs respectively serve the dual functions of behavioral timeliness modeling and domain knowledge constraint. Subsequently, through entity alignment and the introduction of virtual interest edges, a dual-graph association mapping is achieved, seamlessly weaving isolated behavioral subgraphs and the public semantic network into a unified fusion knowledge graph. This allows the propagation path of individual dynamic interests to be structurally extended along the objective association rules between products, fundamentally breaking down the barriers to interest exploration in sparse behavior scenarios. On top of the fusion graph, a time-decay-based user interest transfer mechanism is introduced, using multi-level weighted walks to transfer user interests... The immediate intensity of current behavior and the natural decay effect of historical behavior diffuse along semantic relationships such as complementary and frequently purchased to the potential product space. Therefore, the extracted user dynamic behavior encoding possesses both time sensitivity and knowledge guidance, achieving accurate projection of individual behavioral dynamics under structured knowledge constraints. Subsequently, this dynamic encoding is adaptively weighted and fused with long-term historical preference vectors to generate user preference feature vectors. This effectively suppresses recommendation drift caused by single, accidental behaviors by sensitively capturing short-term interest fluctuations and using long-term steady-state preferences as correction anchors, ensuring the robustness and consistency of preference representation. On the product side, a graph attention network is used to perform differentiated aggregation encoding of product neighborhood nodes and relationship types in the fused knowledge graph. This ensures that the inherent feature vector of a product includes both its own attributes and the semantics of the neighborhood structure, thereby strengthening the product representation's ability to perceive general association rules. Ultimately, a personalized recommendation list is generated collaboratively from user preference vectors and product feature vectors. A dual-objective composite ranking system, incorporating click probability and marketing strategy, is introduced. This achieves an organic unity between individual behavioral dynamics, product association knowledge, and business operational needs, significantly improving the accuracy and business adaptability of digital marketing recommendations under complex conditions such as cold starts, sparse interactions, and scene transitions. This approach effectively integrates the structured constraints of individual behavioral dynamics and general association knowledge between products for digital marketing recommendations. Attached Figure Description

[0048] Figure 1 This is an exemplary flowchart of a digital marketing recommendation method based on user behavior data analysis, according to some embodiments of this application;

[0049] Figure 2 This is an exemplary flowchart illustrating the determination of a fused knowledge graph according to some embodiments of this application;

[0050] Figure 3This is a flowchart illustrating the working process of a digital marketing recommendation system based on user behavior data analysis, according to some embodiments of this application.

[0051] Figure 4 This is a schematic diagram of the structure of a digital marketing recommendation system based on user behavior data analysis, according to some embodiments of this application;

[0052] Figure 5 This is a schematic diagram of the structure of a computer device that implements a digital marketing recommendation method based on user behavior data analysis, according to some embodiments of this application. Detailed Implementation

[0053] To better understand the technical solution of this application, the technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] refer to Figure 1 The figure is an exemplary flowchart of a digital marketing recommendation method based on user behavior data analysis according to some embodiments of this application. The digital marketing recommendation method based on user behavior data analysis mainly includes the following steps:

[0055] In step 101, a personal behavior knowledge graph containing user behavior trajectories and preference associations is constructed, as well as a marketing public knowledge graph containing general association rules for products.

[0056] In some embodiments, constructing a personal behavior knowledge graph containing user behavior trajectories and preference associations, as well as a marketing public knowledge graph containing general product association rules, can be achieved through the following steps:

[0057] Obtain the user's historical interaction sequence across multiple devices;

[0058] Based on the historical interaction behavior sequence, a behavior subgraph rooted at the user node is constructed, and directed edges between behavior nodes and product nodes are introduced in the behavior subgraph to construct a personal behavior knowledge graph.

[0059] Extract the category hierarchy, functional complementarity, and market co-occurrence relationships between products from external marketing knowledge bases and product attribute databases;

[0060] Based on the aforementioned category hierarchy, functional complementarity, and market co-occurrence relationships, a marketing public knowledge graph containing the strength of general association rules between products is constructed, with product nodes as entities.

[0061] In practice, the system first acquires users' historical interaction behavior sequences through a multi-terminal tracking module. These multi-terminals include mobile applications, web browsers, and mini-program clients. The collected behavioral data includes at least a user unique identifier field, a behavior type field, a behavior trigger timestamp field, a dwell time field, and a unique identifier field for the interacted product. Behavior types cover key marketing actions such as clicking to browse, adding to cart, adding to favorites / following, submitting orders, and sharing / forwarding. The collected raw behavior sequences are grouped and aggregated according to user dimensions, and the behavior events of the same user are sorted in ascending order based on the behavior trigger timestamp, forming a historical interaction behavior sequence that evolves over time.

[0062] In addition, in specific implementation, a behavior subgraph rooted at the user node is constructed based on the historical interaction behavior sequence. Directed edges between behavior nodes and product nodes are introduced into this behavior subgraph to construct the personal behavior knowledge graph. Specifically, with the user node as the root node, each interaction behavior in the historical interaction behavior sequence is instantiated as a behavior node, and the interacting product is instantiated as a product node. Directed edges are drawn from the user node to the behavior node, and the attribute field of this edge records the occurrence time information of the behavior event. Directed edges are drawn from the behavior node to the product node, and the weight field of this edge is assigned a value by the behavior type according to a preset preference intensity mapping table. For example, clicking is mapped to a lower intensity value, adding to cart is mapped to a medium intensity value, and placing an order is mapped to a higher intensity value. For multiple similar interaction behaviors of the same user on the same product within a preset time window, an exponentially weighted moving average method is used to smoothly aggregate and update the edge weights to avoid excessive influence from a single accidental behavior. This forms a heterogeneous subgraph with the user as the root node, containing behavior nodes and product nodes and their directed connections; this subgraph is the personal behavior knowledge graph.

[0063] It should be noted that the external marketing knowledge base in this application refers to a structured knowledge set maintained by the e-commerce platform operator or a third-party data service provider, containing content such as product category hierarchy, brand relationship network, and product pairing recommendation rules; the product attribute database refers to a database system that stores basic attribute information of products sold on the platform, including fields such as product identifier, product name, category affiliation, specifications, functional description, and official pairing recommendations; the category hierarchy relationship refers to the hierarchical association relationship formed by the category affiliation of products in the standard product category tree, specifically manifested as the inclusion and being included relationship between a specific product and its category node, as well as between category nodes at different levels. This relationship reflects the semantic hierarchy structure of products in a taxonomic sense; the functional complementarity relationship refers to the interdependence and combination of two different products in their use functions. The relationships that fully leverage product utility are determined based on the semantic matching degree of product function descriptions and industry-standard pairing rules. Functional complementarity relationships have strong scenario-specific relevance and serve as important knowledge bases for cross-category recommendations. Market co-occurrence relationships refer to the statistical association between two different products that are frequently purchased by the same user in the same shopping session in historical transaction orders. This relationship is mined from large-scale user actual consumption behavior, reflecting market-level product combination purchase patterns and serving as a core knowledge source for collaborative filtering recommendations. General association rule strength refers to a quantitative indicator in the marketing public knowledge graph that characterizes the tightness of a certain type of relationship between two product nodes. This strength value is calculated from the confidence level during the relationship extraction stage; a higher value indicates that the association between the two products in that relationship dimension is more reliable and significant.

[0064] In addition, during the specific implementation, when constructing the marketing public knowledge graph, structured relationship information between products is extracted in batches from the external marketing knowledge base and the product attribute database. The external marketing knowledge base includes product category hierarchy trees, brand association relationship databases, etc., and the product attribute database includes product specification attribute tables and product description text information. The specific relationship types extracted include: constructing "belongs to" type triples based on category hierarchy relationships extracted from the category hierarchy tree; constructing "complementary to" type triples based on functional complementarity relationships extracted from product function matching data; and constructing "frequently purchased" type triples based on market co-occurrence relationships extracted from statistical analysis results of co-occurrence of products in historical orders, where the co-occurrence strength is obtained by normalizing the co-occurrence frequency of the same order. Using the extracted product nodes as entities, relationship types as predicates, and association strength as confidence, a set of triples covering all products is constructed to form the marketing public knowledge graph.

[0065] It should be noted that the personal behavior knowledge graph in this application refers to a heterogeneous directed graph structure built around a single user's historical interaction behavior sequence with a single user as the root node. The behavior nodes in the graph record the type and time of the interaction action, the product nodes represent the interaction objects, and the weights of the directed edges reflect the strength of the user's behavioral preferences for different products. The marketing public knowledge graph refers to a structured semantic network that describes the general association rules between products, extracted and constructed from global marketing data sources and product attribute information databases. The graph uses product nodes as basic entities and predicates such as category affiliation, functional complementarity, and market co-occurrence. The confidence of the relationship edges is determined by statistical frequency or business rules. This graph provides global product association prior knowledge independent of individual users. The behavior subgraph refers to the local graph structure in the personal behavior knowledge graph corresponding to a specific user, which is composed of user nodes, behavior nodes, product nodes, and directed connection edges between the three.

[0066] In addition, it should be noted that in the process of constructing the personal behavior knowledge graph containing user behavior trajectories and preference associations, the collection, storage and processing of user data involved in this application strictly comply with the requirements of the Personal Information Protection Law of the People's Republic of China and other relevant laws and regulations, and are carried out on the premise of obtaining the explicit authorization and consent of users. Furthermore, sensitive personal information is desensitized and anonymized to fully protect users' privacy rights and does not constitute an infringement of users' privacy.

[0067] In step 102, the personal behavior knowledge graph and the marketing public knowledge graph are mapped to the association between products and user entities to obtain a fused knowledge graph of users and products.

[0068] In some embodiments, reference Figure 2 The diagram is an exemplary flowchart of determining the fused knowledge graph in some embodiments of this application. In this embodiment, the association mapping between the personal behavior knowledge graph and the marketing public knowledge graph to obtain the fused knowledge graph of users and products can be achieved by the following steps:

[0069] In step 1021, the identifiers of product node entities in the personal behavior knowledge graph and product node entities in the marketing public knowledge graph are identified to be consistent, and the attribute information of the same product node in the two graphs is merged through entity alignment operation;

[0070] In step 1022, the products that user nodes are indirectly associated with through behavioral edges in the personal behavior knowledge graph are identified, and the product set in the marketing public knowledge graph is expanded by general association rules, and a virtual interest edge is established between the user node and the expanded product node.

[0071] In step 1023, the user and product fusion knowledge graph is constructed based on the merged product nodes, user nodes, and a set of mixed edges including behavior edges and virtual interest edges.

[0072] In specific implementation, the consistency of identifiers of product node entities in the personal behavior knowledge graph and the marketing public knowledge graph is identified. Attribute information of the same product node in both graphs is merged through entity alignment. Specifically, this involves performing entity alignment on product node entities in both the personal behavior knowledge graph and the marketing public knowledge graph. Using a globally unified product identifier as the alignment anchor, all product nodes in both graphs are traversed, and product nodes with consistent identifier values ​​are identified as pointing to the same objective product entity. For product nodes identified as the same entity, their connection relationships with user behavior edges in the personal behavior knowledge graph and their general association relationships with other product nodes in the marketing public knowledge graph are merged. Attribute information merging adopts a complement fusion strategy; if conflicting values ​​occur in the same attribute field, the newer value with the data update timestamp is used for overwriting. Through entity alignment, semantic unity is achieved at the product node level in both graphs. Other methods can be used in other embodiments, which are not limited here.

[0073] In addition, in specific implementation, identifying the products indirectly associated with user nodes through behavioral edges in the personal behavior knowledge graph, and establishing a virtual interest edge between user nodes and extended product nodes in the marketing public knowledge graph through general association rules, specifically involves: in the personal behavior knowledge graph, starting from the user node, the product nodes indirectly reachable along behavioral edges constitute the user's direct interest product set. For each source product in the direct interest product set, a one-hop neighborhood expansion is performed along the general association edge of the source product in the marketing public knowledge graph, incorporating the associated products reached by the expansion into the candidate extended product set. For product nodes in the candidate extended product set that have not been actually interacted with by the user, a directed edge is created in the fusion graph to be constructed, directly pointing from the user node to the product node; this directed edge is defined as the virtual interest edge. The initial weight of the virtual interest edge is determined by two factors: the user's behavioral edge weight for the source product and the confidence level of the general association rule between the source and target products. The initial weights are the product of the two values, and then normalized to the maximum and minimum values ​​to map them to a uniform numerical range. Other methods can be used in other embodiments, which are not limited here.

[0074] Furthermore, in specific implementation, the user-product fusion knowledge graph is constructed based on the merged product nodes, user nodes, and a hybrid edge set containing behavioral edges and virtual interest edges. Specifically, this involves integrating the entity-aligned and merged product node set, the complete user node set, and the hybrid edge set containing original behavioral edges and newly added virtual interest edges, and storing them uniformly in a graph database engine to form the user-product fusion knowledge graph. This graph possesses both individual behavioral connections and global semantic associations. Other implementation methods can also be used in other embodiments, which are not limited here.

[0075] It should be noted that the fused knowledge graph in this application refers to a unified heterogeneous information network obtained by combining a personal behavior knowledge graph and a marketing public knowledge graph through entity alignment and virtual relationship completion operations. This graph simultaneously contains actual behavioral connections between users and products, as well as potential interest connections obtained based on knowledge reasoning, providing complete structural support for cross-domain interest propagation calculations within a unified graph space. Virtual interest edges refer to directed edges in the fused knowledge graph that directly connect user nodes to product nodes where users have never actually interacted. The establishment of these edges is not based on real behavior logs, but rather on the inference extension of general association rules between products in the marketing public knowledge graph. Virtual interest edges provide a reasonable path for the outward diffusion of user interests in the graph structure, and their weights are jointly determined by the source product behavior strength and the confidence of the association rules. Entity alignment refers to the process of identifying and merging nodes belonging to different knowledge graphs but pointing to the same objective entity. Typically, a globally unique identifier for the entity is used as the matching basis, and the attribute descriptions and relationship connections of the entity in multiple graphs are merged to eliminate information redundancy and inconsistency caused by the dispersion of data sources.

[0076] In step 103, a time-decay-based user interest transmission mechanism is introduced into the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and to extract user dynamic behavior codes that represent the user's current behavioral preference tendencies.

[0077] In some embodiments, introducing a time-decay-based user interest transfer mechanism on the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and extracting user dynamic behavior codes representing the user's current behavioral preference tendencies, can be achieved through the following steps:

[0078] Starting with the user node as the central node, perform a multi-level neighborhood random walk along the behavior edges and virtual interest edges on the fused knowledge graph;

[0079] During the walk, for each behavior edge traversed, a time decay factor is calculated based on the time difference between the occurrence time of the behavior event and the current recommended time.

[0080] For each product node reached by the terminal of a traversal path, the weights of each edge on the aggregated path and the time decay factor determine the contribution score of that product node to the user's current preference.

[0081] Based on the contribution score, each candidate product node is sorted and a vector representation of a preset dimension is extracted to form a dynamic user behavior code that represents the user's current behavioral preference.

[0082] In specific implementation, taking the user node as the starting center node, the multi-level neighborhood random walk along the behavioral edges and virtual interest edges on the fused knowledge graph is performed as follows: On the fused knowledge graph, the target user node to be analyzed is designated as the starting center node, and the depth of the walk is set to a preset value, such as three layers. Simultaneously, the upper limit of the number of neighboring nodes sampled in each layer is set to a preset value, such as ten. In each layer of the walk, starting from the current node, all neighboring nodes directly connected to it through directed edges are obtained. Using the weight values ​​of the connecting edges as the probability distribution basis, a weighted random sampling operation is performed to determine the specific target node for the next hop.

[0083] Furthermore, in practical implementation, during the walkthrough, for each behavior edge traversed, a corresponding time decay factor is calculated based on the time difference between the occurrence time of the behavior event associated with that edge and the current recommendation request time. The calculation of the time decay factor follows an exponential decay law: the larger the time difference, the smaller the decay factor value, indicating that the influence of the behavior on the current moment naturally weakens over time. The decay rate is adjusted by a preset decay parameter, and the specific calculation formula for the time decay factor is as follows: ,in, Indicates the time decay factor. This is a preset decay rate constant greater than zero, used to control the rate of decay. This is the absolute value of the time difference between the occurrence time of the behavioral event and the current recommended time, with the unit set to hours or days depending on the actual scenario; from this formula, it can be seen that when the time difference... When the value is 0, the attenuation factor is 1 (i.e., no attenuation); as... As the decay factor increases, it decreases exponentially and smoothly, approaching 0, thus achieving a timeliness modeling effect where older behaviors contribute less to current preferences. If the edges traversed by the walking path are virtual interest edges, since they are not triggered by real behaviors and have no corresponding behavior time records, they are assigned a fixed time decay constant. Other methods can be used in other embodiments, which are not limited here.

[0084] In addition, in specific implementation, for each product node reached by the terminal of a traversal path, the contribution score of that product node to the user's current preference is determined by aggregating the weights of each edge on the path and the time decay factor. Specifically, for each complete traversal path starting from the user node and finally reaching a certain product node, the original weights of each edge traversed in sequence on the path are multiplied by the corresponding time decay factor item by item, and the product result is used as the contribution score of the terminal product node on a single path. If multiple different traversal paths eventually reach the same product node, the contribution scores of that product node on each path are summed to obtain the contribution score of that product node to the user's current preference. Other methods can also be used in other embodiments, which are not limited here.

[0085] In addition, in specific implementation, the process of sorting each candidate product node based on the contribution score and extracting a vector representation of a preset dimension to form a user dynamic behavior code representing the user's current behavioral preferences is as follows: All candidate product nodes are sorted in descending order according to their contribution scores, and a preset number of product nodes with the highest scores are extracted. The preset number is usually determined based on the actual capacity of the recommended display positions or engineering practice experience, such as the top 64 or 128 product nodes. The node embedding vectors generated by the graph embedding pre-trained model for these product nodes are extracted. The extracted vector set is then concatenated sequentially according to the ranking order or compressed into a unified dimension through pooling operations to form a fixed-length numerical vector representation. This vector is defined as the user dynamic behavior code. Other methods can be used in other embodiments, which are not limited here.

[0086] It should be noted that the user interest transmission mechanism in this application refers to the calculation process in which, starting from the target user node, a weighted multi-level neighborhood random walk is used to gradually propagate the user's preference intensity for interacted products along the behavior edges and virtual interest edges in the graph structure to nodes of products that have not been directly interacted with. This mechanism enables the global product association rules in the marketing public knowledge graph to effectively serve the interest exploration and expansion of individual users. The time decay factor represents the natural decreasing law of the influence of the user's historical behavior on the current interest representation. The longer the time interval, the smaller the value of the decay factor. The contribution score reflects the relative strength of the user's current behavioral preference after it has spread to products that have not been directly contacted through the direct interaction relationship and general association rule in the graph structure. The higher the score, the closer the association between the product node and the user's recent interest state and the more in line with the user's potential attention tendency. The user dynamic behavior encoding refers to a fixed-dimensional numerical vector that represents the user's instantaneous interest distribution and potential preference tendency at the current moment. This encoding comprehensively integrates the intensity of the user's recent behavior, the timeliness information of the behavior occurrence time, and the structural information of the product association path in the fused knowledge graph.

[0087] In step 104, a user preference feature vector of the user’s short-term interest in the product is generated based on the user’s dynamic behavior encoding and the user’s historical interaction preference information with the product.

[0088] In some embodiments, generating a user preference feature vector of a user's short-term interest in a product based on the user's dynamic behavior encoding and the user's historical interaction preference information with the product can be achieved through the following steps:

[0089] Obtain information on users' historical interaction preferences with products;

[0090] Extract the user's long-term steady-state preference vector from the historical interaction preference information;

[0091] The user dynamic behavior encoding is fused with the user's long-term steady-state preference vector element by element to generate a user preference feature vector of the user's short-term interest in the product.

[0092] In practice, the system acquires historical interaction preferences between users and products. This information comes from users' long-term accumulated behavioral log data, specifically extracting a list of all products for which the user has engaged in positive interactions within a preset long-term period, such as the past ninety days. Positive interactions include clicking to browse, adding to cart, adding to favorites / following, and submitting orders. Each record contains at least the product identifier, the type of interaction, and the specific timestamp of the interaction.

[0093] Furthermore, in specific implementation, extracting the user's long-term steady-state preference vector from the historical interaction preference information involves: for each product in the product list within the historical interaction preference information, obtaining its pre-trained inherent attribute vector, which includes embedded representations of key marketing attributes such as product category code, brand identifier, and price range. Using the basic weight corresponding to the interaction behavior type of each product as a benchmark, multiplying it by a reverse time decay coefficient calculated based on the distance between the interaction time and the current moment (i.e., the longer the interaction time, the smaller the coefficient), a weighted average is calculated for the attribute vectors of all products in the list. The resulting vector after weighted averaging is the user's long-term steady-state preference vector, reflecting the user's stable consumption preference pattern unaffected by recent incidental behavior.

[0094] Furthermore, in specific implementation, the user dynamic behavior encoding and the user's long-term steady-state preference vector are fused element-by-element to generate a user preference feature vector for the user's short-term interest in products. Specifically, this involves calculating the user's activity evaluation index within the current interaction session, such as the ratio of the number of interactions in the current session to the user's historical average number of interactions during the same period. This ratio is then normalized and used as the fusion weight coefficient for the dynamic preference component. A higher fusion weight value indicates a greater proportion of the current short-term dynamic behavior in the final preference expression. Based on this fusion weight coefficient, the user dynamic behavior encoding and the user's long-term steady-state preference vector are weighted and summed to obtain the user preference feature vector for the user's short-term interest in products.

[0095] It should be noted that the historical interaction preference information in this application refers to the collection of positive interaction records between users and various products within a relatively long historical time window. This information includes fields such as the identifier of the interacting product, the type of interaction behavior, and the interaction timestamp. The user's long-term steady-state preference vector represents the user's inherent consumption preference tendency that remains relatively stable over a long period of time and is not easily affected by short-term behavioral fluctuations or accidental clicks. The user preference feature vector represents the basic directional guidance of the user's long-term preferences and the interest migration signals reflected by recent behavioral fluctuations, which can more accurately match the user's true purchase intention at the time of the recommendation request.

[0096] In step 105, product-specific feature vectors of marketing attributes for each product to be recommended are generated based on the fused knowledge graph. A personalized recommendation list for each product to be recommended is generated from the user preference feature vector and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

[0097] In some embodiments, generating product-specific feature vectors for each marketing attribute of a product to be recommended based on the fused knowledge graph can be achieved using the following steps:

[0098] For each product to be recommended, extract other product nodes and their corresponding association types within the adjacent one-hop range of the product to be recommended from the fused knowledge graph;

[0099] Based on the other product nodes within the adjacent one-hop range of each product to be recommended and the corresponding association type, calculate the context-enhanced embedding vector of the corresponding product to be recommended;

[0100] The original attribute encoding vector of each product to be recommended is concatenated with the corresponding context-enhanced embedding vector to determine the product-specific feature vector of the marketing attributes of each product to be recommended.

[0101] In specific implementation, for each product to be recommended, extracting other product nodes within a one-hop range adjacent to the product to be recommended and their corresponding association types from the fused knowledge graph involves: for each product to be recommended in the candidate product pool, locating the product node corresponding to the product in the fused knowledge graph, extracting all neighboring product nodes directly connected to the product node via directed edges and located within a one-hop neighborhood, and simultaneously recording the association type labeled by the directed edges connecting each neighboring node, such as complementary relationship or frequently purchased relationship. For each node in the set of neighboring product nodes, obtaining its pre-trained original attribute embedding vector; other methods can be used in other embodiments, which are not limited here.

[0102] Furthermore, in the specific implementation, the context-enhanced embedding vector for each recommended product is calculated based on other product nodes within a one-hop radius of the product to be recommended and their corresponding association types. Specifically, this involves locating the product node corresponding to the recommended product in the fused knowledge graph, extracting all neighboring product nodes directly connected to it via directed edges and located within a one-hop radius, and recording the association type labeled on each connection edge, such as complementary or frequently purchased. Subsequently, a graph attention network is used to aggregate and encode the neighborhood information: the embedding representation of the recommended product node itself is used as the query vector, and the embedding representations of each neighboring product node are used as the key and value vectors, respectively. An initial attention score is obtained by calculating the semantic matching degree between the query vector and each key vector. Based on this, the encoded vector of the association type is incorporated as a modifier into the attention score calculation process, allowing different types of associations to have a differentiated impact on the final weight. For example, the functional matching semantics represented by the complementary relationship and the market co-occurrence semantics represented by the frequently purchased relationship differ in attention response strength. After normalization, the attention weight coefficients corresponding to each neighboring product node are obtained. Then, the value vectors of each neighbor are weighted and summed using these coefficients to generate the context-enhanced embedding vector of the product to be recommended in the local graph structure. Other methods can be used in other embodiments, which are not limited here.

[0103] In addition, in specific implementation, the original attribute encoding vectors of each product to be recommended are concatenated with the corresponding context-enhanced embedding vectors to determine the inherent feature vectors of the marketing attributes of each product to be recommended. Specifically, static description fields of products are collected from the product attribute database, including structured data such as product category identifiers, brand identifiers, price ranges, specifications, and functional tags. These categorical fields are converted into low-dimensional dense vectors through one-hot encoding or embedding layer mapping. At the same time, semantic vector representations of the product's textual description information are extracted using a pre-trained language model. Then, the above multi-source attribute vectors are concatenated or weighted and fused to form the original attribute encoding vectors that can characterize the inherent characteristics of the product itself. After obtaining the original attribute encoding vectors and the context-enhanced embedding vectors, they are concatenated end-to-end in the vector dimension direction to obtain a higher-dimensional combined vector. The combined vector is then fed into a fully connected neural network. The weight parameters in the network are optimized and adjusted through the training process of historical recommendation data. During the calculation, the fully connected layer performs linear combination and non-linear activation transformation on the elements of each dimension of the input vector, thereby achieving dimensional compression and remapping of the feature space. Finally, it outputs a product inherent feature vector that is completely consistent with the dimension of the user preference feature vector. Other methods can be used in other embodiments, which are not limited here.

[0104] It should be noted that the product-inherent feature vector in this application refers to the vectorized representation generated by combining the product's inherent attribute description information and its local neighborhood context association information in the graph structure under the structured semantic environment of the fused knowledge graph. This vector contains both the product's basic marketing attribute label information and the product's structural position features and neighborhood product semantic information in the product association network. The context-enhanced embedding vector refers to the supplementary feature vector obtained by weighted aggregation encoding of the product's neighbor node set and its connection relationship in the fused knowledge graph through a graph neural network. This vector effectively captures the local graph structure semantic environment in which the product is located, enabling the numerical representation of the product to transcend the limitations of single attribute description and possess networked association cognitive ability. The association relationship type refers to the semantic label marked on the directed edge connecting two product nodes, used to identify the specific category of the association property between the products represented by the edge.

[0105] In some embodiments, generating a personalized recommendation list of each product to be recommended from the user preference feature vector and the product-specific feature vector of each product to be recommended can be achieved by the following steps:

[0106] Based on the user preference feature vector and the inherent feature vector of each product to be recommended, the estimated click probability rating of the user for each product to be recommended is determined.

[0107] Obtain the marketing strategy weighting factors for each product to be recommended;

[0108] The overall recommendation priority score for each product to be recommended is determined based on the estimated click probability score and marketing strategy weighting factors.

[0109] A personalized recommendation list for each product is generated based on all the overall recommendation priority scores.

[0110] In specific implementation, determining the user's estimated click probability rating for each recommended product based on the user preference feature vector and the inherent feature vector of each product to be recommended involves performing an inner product operation on the user preference feature vector and the inherent feature vector of each recommended product, i.e., multiplying corresponding elements of the two vectors and then summing them. The result of the inner product operation is then processed by a non-linear activation function, such as using the logistic function to map any real number to a probability value range between zero and one. This probability value is the user's estimated click probability rating for the recommended product. Other methods can be used in other embodiments, which are not limited here.

[0111] In addition, in specific implementation, the marketing strategy weight factors for each recommended product within the current marketing cycle are obtained. These weight factors are calculated from the product inventory depth coefficient, promotional intensity level coefficient, and advertising bid coefficient. The product inventory depth coefficient reflects the ratio of the product's current available inventory to a safety stock baseline; the more abundant the inventory, the higher the coefficient. The promotional intensity level coefficient reflects the discount level of the product's current promotional activities; the greater the discount, the higher the coefficient. The advertising bid coefficient reflects the bid level of the advertiser for the product in the real-time bidding system; the higher the bid amount, the higher the coefficient. These three coefficients are normalized, and the normalized coefficients are then weighted and summed according to a preset weight ratio. For example, the weight of the inventory dimension is set to 0.2, the promotion dimension to 0.3, and the bid dimension to 0.5. The weighted sum, after normalization, yields the final marketing strategy weight factors. The preset weight ratios are determined through a multi-dimensional trade-off analysis of the platform's marketing objectives and offline optimization experiments. First, the business operations team initially determines the relative importance of each dimension based on the core strategic objectives within the current marketing cycle. For example, in a scenario where clearing inventory is the primary objective, the weight of the inventory dimension can be increased; in a scenario where expanding revenue is the primary objective, the weight of the advertising bid dimension can be increased. Then, in an offline experimental environment, using historical recommendation log data, the effects of different weighting combinations are simulated and evaluated. The weighted composite score of key business indicators such as click-through rate, conversion rate, and advertising revenue of the recommendation list is used as the evaluation criterion. A weighting combination that achieves a relatively optimal overall business benefit is selected through grid search or manual parameter tuning. Other methods may be used in other embodiments, and are not limited here.

[0112] In practice, the comprehensive recommendation priority score for each recommended product is determined based on its estimated click-through rate (CTR) and marketing strategy weighting factor. Specifically, the estimated CTR and marketing strategy weighting factor are weighted and compounded. A preset balancing coefficient is used to adjust the relative importance of the two factors. The value of the balancing coefficient is determined based on business needs. The balancing coefficient determines the proportion of the estimated CTR in the weighted result, while the proportion of the marketing strategy weighting factor is automatically supplemented by the remainder after subtracting the balancing coefficient from the unit value. A balancing coefficient closer to one indicates that the recommendation ranking prioritizes meeting users' personal interests and preferences, with maximizing click-through rate as the primary objective. A balancing coefficient closer to zero indicates that the recommendation ranking prioritizes achieving business operational goals such as inventory turnover or advertising revenue. The specific value of this balancing coefficient is not a fixed constant but is manually set or automatically adjusted by business operations personnel in the system configuration backend based on the core needs of the current marketing cycle. For example, a higher value can be set during daily operations to ensure user experience, while a lower value can be set during peak sales periods to increase ad revenue contribution. During end-of-quarter clearance sales, the value can be further reduced to accelerate the clearance of slow-moving inventory. For instance, a value of 0.7 indicates that user interest matching plays a dominant role in the ranking decision, while also considering a 30% proportion of marketing revenue. The weighted composite result is the overall recommendation priority score for the corresponding product to be recommended. Other implementation methods can also be used in other embodiments, which are not limited here.

[0113] In addition, the personalized recommendation list for each product to be recommended is generated based on all the comprehensive recommendation priority scores as follows: all products to be recommended are sorted in descending order according to the comprehensive recommendation priority scores, and the sorting result is used as the personalized recommendation list. Other methods can also be used in other embodiments, which are not limited here.

[0114] It should be noted that the click probability prediction score in this application reflects the predicted score of the likelihood of a user clicking on a specific product; the marketing strategy weight factor reflects the priority recommendation value of a product to be recommended from the perspective of platform business operations within the current marketing cycle. That is, the more abundant the inventory, the greater the promotion, or the higher the advertising bid, the higher the value of the marketing strategy weight factor, indicating that the platform is more willing to push the product to the front end to accelerate inventory turnover or obtain advertising revenue; the comprehensive recommendation priority score reflects the competitiveness of the product to be recommended to the user. This score, while ensuring that the recommendation results are highly matched with the user's interests and preferences, takes into account the goal of maximizing the platform's marketing revenue, and achieves a dynamic balance between personalized recommendation quality and commercial value return; the personalized recommendation list reflects the optimal combination of products selected after a comprehensive evaluation of the user's interests and preferences and the platform's business demands at the current request time. As the final decision output of the recommendation system, this list directly drives the product display order on the front-end interface or the product selection of marketing outreach messages, completing the precise personalized marketing delivery to the target users.

[0115] In practice, the marketing recommendation decision for each product to be recommended is as follows: After obtaining a personalized recommendation list, the system first encapsulates the product identifiers, corresponding ranking positions, estimated click probability scores, and marketing strategy weight factors in the list into a unified format recommendation response data packet, which is then transmitted to the front-end business system or marketing outreach engine via the application programming interface. Upon receiving the recommendation response data packet, the front-end business system performs differentiated rendering based on the user's current page context. For example, in the "You May Like" module on the application homepage, product cards are displayed in list order; in the related product pairings section on the shopping cart page, the top five paired products are displayed; or in the recommendation slot on the search results page, the top-ranked product is inserted. Simultaneously, the marketing outreach engine can select several top-ranked products from the recommendation list as push content materials and proactively reach users through in-app messages, SMS notifications, or emails. The outreach content includes product images, titles, promotional tags, and deep redirect links. After the recommendation decision takes effect, the system also collects user feedback on the recommended products in real time through an asynchronous log feedback mechanism, including key events such as exposure clicks, add-to-cart conversions, and completed orders. This feedback data is fed back to the data warehouse for incremental training and performance evaluation of the subsequent model, thus forming a closed loop of marketing recommendation decision-making from decision output, multi-channel execution to performance attribution.

[0116] In some embodiments, reference Figure 3This diagram illustrates the workflow of a digital marketing recommendation system based on user behavior data analysis in some embodiments of this application. In user terminal scenario 110, users interact with the system via mobile terminals by browsing, clicking, adding to favorites, and making purchases. Their behavioral patterns and interaction data are collected by the terminal device to record the user's operation sequence and preferences. Next, the network transmission server 120 transmits the collected user behavior data, user profile data, and product interaction data bidirectionally, enabling reliable data interaction between the terminal and the backend system to meet the transmission requirements of multi-source data fusion. Subsequently... The data storage and computing server 130 receives the transmitted data and realizes the construction of personal behavior knowledge graph and marketing public knowledge graph, entity association mapping and generation of fused knowledge graph. By introducing a user interest transmission mechanism based on time decay, it performs preference propagation analysis, extracts user dynamic behavior codes and generates user preference feature vectors and product inherent feature vectors. Finally, the recommendation processing terminal 140 generates a personalized recommendation list for each product to be recommended based on the user preference feature vector and product inherent feature vector, and completes the accurate marketing recommendation of products. The entire workflow provides users with product recommendation services that match their dynamic interests, and improves the accuracy and conversion efficiency of digital marketing.

[0117] Furthermore, in another aspect of this application, in some embodiments, this application provides a digital marketing recommendation system based on user behavior data analysis, referencing... Figure 4 The figure is a schematic diagram of the structure of a digital marketing recommendation system based on user behavior data analysis according to some embodiments of this application. The digital marketing recommendation system 400 based on user behavior data analysis includes: a construction module 401, a processing module 402, and an execution module 403, which are described below:

[0118] Construction module 401, in this application, is mainly used to construct a personal behavior knowledge graph containing user behavior trajectories and preference associations, and a marketing public knowledge graph containing general product association rules;

[0119] Processing module 402, in this application, is used to perform association mapping between products and user entities on the personal behavior knowledge graph and the marketing public knowledge graph to obtain a fused knowledge graph of users and products;

[0120] It should be noted that the processing module 402 in this application is also used to introduce a time decay-based user interest transmission mechanism on the fused knowledge graph, perform preference propagation analysis on the interaction relationship between users and product nodes, and extract user dynamic behavior codes that represent the user's current behavioral preference tendencies.

[0121] Additionally, it should be noted that the processing module 402 in this application is also used to generate a user preference feature vector of the user's short-term interest in the product based on the user dynamic behavior encoding and the user's historical interaction preference information with the product.

[0122] The execution module 403 in this application is mainly used to generate product-specific feature vectors of marketing attributes of each product to be recommended based on the fused knowledge graph, and to generate a personalized recommendation list of each product to be recommended based on the user preference feature vector and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

[0123] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described digital marketing recommendation method based on user behavior data analysis.

[0124] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a digital marketing recommendation method based on user behavior data analysis, according to some embodiments of this application. The digital marketing recommendation method based on user behavior data analysis in the above embodiments can... Figure 5 The computer device shown is used to implement this, and the computer device 500 includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0125] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0126] The communication bus 502 can be used to transmit information between the aforementioned components.

[0127] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0128] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The method used in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0129] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0130] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0131] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0132] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described digital marketing recommendation method based on user behavior data analysis.

[0133] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0134] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A digital marketing recommendation method based on user behavior data analysis, characterized in that, Includes the following steps: Construct a personal behavior knowledge graph that includes user behavior trajectories and preference associations, as well as a marketing public knowledge graph that includes general product association rules; The personal behavior knowledge graph and the marketing public knowledge graph are mapped to the association between products and user entities to obtain a fused knowledge graph of users and products. A time-decay-based user interest transfer mechanism is introduced into the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and to extract user dynamic behavior codes that represent the user's current behavioral preference tendencies. Based on the user dynamic behavior encoding and the user's historical interaction preference information with the product, a user preference feature vector of the user's short-term interest in the product is generated. Based on the fused knowledge graph, product-specific feature vectors of marketing attributes for each product to be recommended are generated. Then, a personalized recommendation list for each product to be recommended is generated from the user preference feature vector and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

2. The method as described in claim 1, characterized in that, Constructing a personal behavior knowledge graph that includes user behavior trajectories and preference associations, as well as a marketing public knowledge graph that includes general product association rules, specifically includes: Obtain the user's historical interaction sequence across multiple devices; Based on the historical interaction behavior sequence, a behavior subgraph rooted at the user node is constructed, and directed edges between behavior nodes and product nodes are introduced in the behavior subgraph to construct a personal behavior knowledge graph. Extract the category hierarchy, functional complementarity, and market co-occurrence relationships between products from external marketing knowledge bases and product attribute databases; Based on the aforementioned category hierarchy, functional complementarity, and market co-occurrence relationships, a marketing public knowledge graph containing the strength of general association rules between products is constructed, with product nodes as entities.

3. The method as described in claim 1, characterized in that, The process of mapping product and user entities together in the personal behavior knowledge graph and the marketing public knowledge graph to obtain a fused user and product knowledge graph specifically includes: The identifiers of product node entities in the personal behavior knowledge graph and product node entities in the marketing public knowledge graph are identified to be consistent. The attribute information of the same product node in the two graphs is merged through entity alignment operation. Identify the products that user nodes are indirectly associated with through behavioral edges in the personal behavior knowledge graph, and establish virtual interest edges between user nodes and extended product nodes in the marketing public knowledge graph by extending the product set through general association rules. Based on the merged product nodes, user nodes, and a set of mixed edges including behavioral edges and virtual interest edges, a fused knowledge graph of users and products is constructed.

4. The method as described in claim 1, characterized in that, A time-decay-based user interest transfer mechanism is introduced into the fused knowledge graph to perform preference propagation analysis on the interaction relationship between users and product nodes, and to extract user dynamic behavior codes that represent the user's current behavioral preference tendencies. Specifically, this includes: Starting with the user node as the central node, perform a multi-level neighborhood random walk along the behavior edges and virtual interest edges on the fused knowledge graph; During the walk, for each behavior edge traversed, a time decay factor is calculated based on the time difference between the occurrence time of the behavior event and the current recommended time. For each product node reached by the terminal of a traversal path, the weights of each edge on the aggregated path and the time decay factor determine the contribution score of that product node to the user's current preference. Based on the contribution score, each candidate product node is sorted and a vector representation of a preset dimension is extracted to form a dynamic user behavior code that represents the user's current behavioral preference.

5. The method as described in claim 1, characterized in that, Generating a user preference feature vector based on the user's dynamic behavior encoding and the user's historical interaction preference information with products specifically includes: Obtain information on users' historical interaction preferences with products; Extract the user's long-term steady-state preference vector from the historical interaction preference information; The user dynamic behavior encoding is fused with the user's long-term steady-state preference vector element by element to generate a user preference feature vector of the user's short-term interest in the product.

6. The method as described in claim 1, characterized in that, The generation of product-specific feature vectors for each product's marketing attributes based on the fused knowledge graph specifically includes: For each product to be recommended, extract other product nodes and their corresponding association types within the adjacent one-hop range of the product to be recommended from the fused knowledge graph; Based on the other product nodes within the adjacent one-hop range of each product to be recommended and the corresponding association type, calculate the context-enhanced embedding vector of the corresponding product to be recommended; The original attribute encoding vector of each product to be recommended is concatenated with the corresponding context-enhanced embedding vector to determine the product-specific feature vector of the marketing attributes of each product to be recommended.

7. The method as described in claim 1, characterized in that, Generating a personalized recommendation list for each product to be recommended from the user preference feature vector and the product-specific feature vector of each product to be recommended specifically includes: Based on the user preference feature vector and the inherent feature vector of each product to be recommended, the estimated click probability rating of the user for each product to be recommended is determined. Obtain the marketing strategy weighting factors for each product to be recommended; The overall recommendation priority score for each product to be recommended is determined based on the estimated click probability score and marketing strategy weighting factors. A personalized recommendation list for each product is generated based on all the overall recommendation priority scores.

8. A digital marketing recommendation system based on user behavior data analysis, characterized in that, include: The building module is used to construct a personal behavior knowledge graph containing user behavior trajectories and preference associations, as well as a marketing public knowledge graph containing general product association rules; The processing module is used to perform association mapping between products and user entities on the personal behavior knowledge graph and the marketing public knowledge graph to obtain a fused knowledge graph of users and products. The processing module is also used to introduce a time-decay-based user interest transmission mechanism on the fused knowledge graph, perform preference propagation analysis on the interaction relationship between users and product nodes, and extract user dynamic behavior codes that represent the user's current behavioral preference tendencies. The processing module is also used to generate a user preference feature vector of short-term interest in products based on the user dynamic behavior encoding and the user's historical interaction preference information with products; The execution module is used to generate product-specific feature vectors of marketing attributes for each product to be recommended based on the fused knowledge graph, and to generate a personalized recommendation list for each product to be recommended based on the user preference feature vectors and the product-specific feature vectors of each product to be recommended, so as to make marketing recommendations for each product to be recommended.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the digital marketing recommendation method based on user behavior data analysis as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the digital marketing recommendation method based on user behavior data analysis as described in any one of claims 1 to 7.

Images