An intelligent recommendation method for personalized marketing based on a knowledge graph

Abstract

The application discloses a kind of personalized marketing intelligent recommendation methods based on knowledge graph, comprising the following steps: S1, marketing knowledge graph is encoded, generates user vector, commodity vector and relationship vector, and constructs uniform coding index table;S2, according to the user information to be recommended, target user vector is filtered, and associated commodity vector is extracted, and target commodity set is constructed;S3, path reasoning and coding are carried out, and path coding set is generated;S4, feature aggregation is carried out to path coding set and target commodity set, and path matching feature is generated;S5, recommendation score set is generated by using the calculation of recommendation score of Transform network and factor decomposition machine;S6, according to recommendation score set, intelligent recommendation result is obtained, and model and atlas are updated according to user feedback data.The present application fuses TransE model and has the advantages of high recommendation accuracy, strong interest recognition ability and strong dynamic adaptation ability.

Description

Technical Field

[0001] This invention relates to the field of intelligent marketing technology, and in particular to a personalized marketing intelligent recommendation method based on knowledge graphs. Background Technology

[0002] As the number of users and products on internet platforms continues to increase, marketing methods are gradually shifting from manual, experience-based selection to intelligent recommendation methods based on data analysis. Existing technologies typically construct user behavior datasets by collecting user browsing, click, and purchase records. These datasets are then modeled using collaborative filtering, matrix factorization, or deep learning methods. Based on the modeling results, the model predicts the user's level of interest in different products, thereby generating product recommendations. This type of method can improve recommendation efficiency and reduce the cost of manual selection.

[0003] With the development of knowledge graph technology, some technical solutions construct marketing knowledge graphs containing user nodes and product nodes, and use graph embedding algorithms to map user nodes and product nodes into vector form. Then, they calculate recommendation results based on the similarity between vectors, thereby leveraging the relational information in the knowledge graph to improve the semantic relevance of the recommendation results. However, existing knowledge graph-based recommendation methods still have shortcomings in practical applications.

[0004] Existing technologies in knowledge graph embedding typically only map node vectors, lacking a unified encoding mechanism. This results in a lack of consistent constraints on the representation spaces of different node vectors, leading to inconsistencies in vector representations during path reasoning and affecting the accuracy of the reasoning results. In user filtering, most methods rely solely on vector similarity for matching, failing to incorporate knowledge graph structure information, resulting in low target user location accuracy and impacting the quality of recommendation results.

[0005] Meanwhile, in the path reasoning and feature modeling process, existing technologies typically use simple path concatenation to construct path features, failing to fully utilize path structure information for deep feature modeling, resulting in insufficient accuracy of recommendation scoring results. Furthermore, some methods do not update the knowledge graph and recommendation model based on user feedback data after recommendation completion, causing the recommendation results to fail to reflect changes in user interests in a timely manner.

[0006] Therefore, how to provide a personalized marketing intelligent recommendation method based on knowledge graphs is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a personalized marketing intelligent recommendation method based on knowledge graphs. This invention fully utilizes the TransE model, Locality Sensitive Hash algorithm, DeepWalk algorithm, MINERVA algorithm, capsule network, Transformer network, and factorization machine to realize the generation and feedback update of recommendation results. It improves the consistency of entity vector representation in the knowledge graph, the accuracy of user interest identification, and the accuracy of recommendation scoring. At the same time, it realizes the dynamic updating of knowledge graph and recommendation model through user feedback data, and has the advantages of high recommendation accuracy, strong interest identification ability, and strong dynamic adaptability.

[0008] A personalized marketing intelligent recommendation method based on knowledge graph according to an embodiment of the present invention includes the following steps: S1. The pre-built marketing knowledge graph is uniformly encoded using the TransE model to generate user vectors, product vectors, and relationship vectors, and a unified encoding index table is constructed. S2. Based on the collected user information to be recommended, use the Locality Sensitive Hash algorithm to match multiple candidate user vectors from the unified coding index table, and combine the DeepWalk algorithm to filter the target user vectors. At the same time, extract the product vectors associated with the target user vectors to construct a target product set. S3. Starting with the node corresponding to the target user vector, perform path traversal in the marketing knowledge graph, and use the MINERVA algorithm to perform path reasoning and encoding on the traversal results to generate a path encoding set. S4. Associate and match the path encoding set with the target product set, and perform feature aggregation operation on the matching results through capsule network to generate path matching features for each target product vector; S5. Based on path matching features, a Transformer network is used to perform feature modeling operations, and a recommendation score is calculated using a factorization machine to generate a set of recommendation scores. S6. Sort the target product set according to the recommended rating set to obtain intelligent recommendation results, collect user feedback data, and update the Transformer network, factor decomposition machine and marketing knowledge graph according to the user feedback data.

[0009] Optionally, the user information to be recommended represents a data set used to describe the attribute characteristics of the user to be recommended, the node represents an entity unit in the marketing knowledge graph used to describe user entities or product entities, and the target product vector represents any product vector in the target product set.

[0010] Optionally, S1 specifically includes: S11. Read all node sets and edge sets in the pre-built marketing knowledge graph, extract the node identifiers of user nodes and product nodes in the node set, and perform relationship identifier extraction operation on the relationship edges in the edge set. S12. Use the node identifiers corresponding to each user node and product node as the head entity identifier and tail entity identifier respectively, and use the relationship identifiers of the corresponding relationship edges as relationship entity identifiers. Summarize the three types of entity identifiers to form a set of triples. S13. Input the set of triples into the TransE model. In the TransE model, perform vector space mapping operation on the nodes and edges corresponding to the three types of entity identifiers to generate the corresponding user vector, relation vector and product vector. S14. Based on the marketing knowledge graph, perform vector numbering operations on user vectors, product vectors and relationship vectors, associate and store the corresponding node identifiers, relationship identifiers and vector numbers, and generate an initial index table. S15. Perform vector integrity check on the initial index table, and arrange the user vectors, product vectors and relationship vectors that pass the check in the order of node identifiers to generate a unified coded index table.

[0011] Optionally, the process of constructing the marketing knowledge graph specifically includes: Acquire user behavior data and product data from multiple sources, perform time sorting and behavior type labeling operations on user behavior data, and perform product identifier extraction and field consistency operations on product data to obtain user behavior sets and product record sets; Based on the user behavior set and the product record set, user identifiers and product identifiers are extracted respectively, and encoded into user nodes and product nodes to form a node set; Based on the correspondence between each user identifier and product identifier, construct the interaction relationship edges between user nodes and product nodes, and assign a corresponding relationship identifier to each interaction relationship edge to form an edge set; Perform node numbering operations on user nodes and product nodes in the node set, perform relationship numbering operations on interaction edges in the edge set, and establish the association between node numbers and relationship numbers; Write the set of numbered nodes and the set of edges into the preset initial graph structure, and bind the relationships to generate a marketing knowledge graph.

[0012] Optionally, S2 specifically includes: S21. Obtain the information of users to be recommended, perform field parsing, discrete field encoding and continuous field normalization operations on the information of users to be recommended, and concatenate the normalization results according to the preset field order to generate user feature vectors. S22. Perform dimension alignment and vector space mapping operations on the user feature vector to generate a query user vector with the same dimension as the user vector in the unified coding index table. S23. Perform locality-sensitive hashing calculation based on the query user vector, generate a hash code for the query user vector using a preset hash function, and retrieve multiple candidate user vectors from the unified coding index table based on the hash code, specifically including: Perform a vector dimension partitioning operation on the query user vector, dividing the query user vector into multiple sub-vector segments according to a preset dimension length; Perform product quantization encoding operation on each sub-vector segment, perform nearest neighbor matching operation on each sub-vector segment using a preset sub-quantization codebook to determine the corresponding codeword number, and arrange all codeword numbers in the order of the sub-vector segments to generate the product quantization vector of the query user vector; A locality-sensitive hash mapping operation is performed on the product quantized vector. A hash calculation is performed on the product quantized vector using a preset hash function to generate the corresponding hash code. Read all user vectors from the unified coding index table, and perform product quantization coding and hash calculation operations on all user vectors to generate a user code set; Perform Hamming distance calculation on the hash code and user code set, and select user vectors whose Hamming distance values ​​meet the preset conditions as candidate user vectors; Read the node identifier corresponding to the candidate user vector and bind it to the candidate user vector; S24. Using the user nodes corresponding to each candidate user vector as the starting nodes, the DeepWalk algorithm is used to perform random walk operations in the marketing knowledge graph to generate node access sequences. Based on the node access sequences, the graph neighborhood score of the corresponding candidate user vector is calculated, and the target user vector is selected based on the graph neighborhood score. S25. Locate the target user node corresponding to the target user vector, read the set of associated edges of the target user node and extract the set of product nodes pointed to by the set of associated edges, extract the product vector corresponding to the set of product nodes from the unified coding index table, and construct the target product set.

[0013] Optionally, the process of filtering the target user vector specifically includes: Read the node identifiers corresponding to each candidate user vector, locate the corresponding user nodes in the marketing knowledge graph, and use each user node as the starting node for the random walk to generate a set of starting nodes. Based on the set of starting nodes, DeepWalk random walk operation is performed in the marketing knowledge graph. According to the preset walk step size, each starting node visits the adjacent nodes along the associated edges in turn, and records the node identifiers visited in each step to generate the corresponding set of access paths. A sliding window truncation operation is performed on the access path set to extract continuous node segments from each access path. The node co-occurrence frequency is calculated based on the number of times nodes co-occur in each node segment, and a node co-occurrence set is generated. The number of co-occurrences represents the cumulative number of times any candidate user node and a node in the access path set co-occur in the same node segment within a preset sliding window range. The neighborhood association strength between candidate user nodes and nodes in the access path set other than the candidate user node is calculated based on the node co-occurrence set. The neighborhood association strength is used as the graph neighborhood score of the corresponding candidate user vector to generate a graph neighborhood score set. The graph neighborhood score set is sorted, and the candidate user vector with the highest graph neighborhood score is selected as the target user vector. At the same time, the node identifiers corresponding to the target user vectors are read, and the correspondence between the target user vectors and the node identifiers is established.

[0014] Optionally, S3 specifically includes: S31. Read the target user node corresponding to the target user vector in the marketing knowledge graph, extract the node identifier of the target user node, and locate the set of adjacent nodes in the marketing knowledge graph that have a relationship edge with the target user node. S32. Starting from the target user node, perform a path traversal operation on the set of adjacent nodes, expand the node access range layer by layer along the relation edges according to the preset traversal depth, record the relation identifiers and node identifiers passed through each expansion, and generate a candidate path set. Each path in the candidate path set contains a sequence of node identifiers and a sequence of relation identifiers arranged in order. S33. Perform a unified encoding index table query operation on each candidate path, extract the corresponding user vector, product vector and relationship vector respectively, and perform vector concatenation operation in the order of the candidate path set to generate a path sequence set; S34. The MINERVA algorithm is used to perform path reasoning on each path sequence in the path sequence set. Action selection calculations are performed on each path sequence to generate a relation jump sequence. The reasoning path set is then determined based on the relation jump sequence. Specifically, this includes: Read each path sequence in the path sequence set, extract the user vector in the path sequence as the initial path vector, and read the corresponding relation vector and product vector in the order of arrangement in the path sequence. Arrange the user vector, relation vector and product vector in the path order to establish the path access sequence. Based on the path access sequence, the MINERVA algorithm is used. The user vector is used as the starting vector, and the relation vector is read step by step in the order of the path access sequence. The user vector of the current step is combined with the corresponding relation vector in sequence to generate the path intermediate vector of the current step, and the path intermediate vector is used as the starting vector of the next step. After the intermediate vector of each path is generated, the intermediate vector of the path is matched one by one with all the relation vectors in the unified coding index table. Based on the matching results, the relation vector corresponding to the current step is determined and the corresponding relation identifier is recorded. The relation jump sequence is formed by arranging them in the path order. Based on the relationship jump sequence, locate the corresponding relationship edge in the marketing knowledge graph, and visit the corresponding node identifiers in sequence according to the relationship identifiers in the relationship jump sequence to generate a node access sequence; The user vector and product vector corresponding to the node access sequence and the relation vector corresponding to the relation jump sequence are rearranged according to the access order, and a corresponding relationship is established with the path sequence to generate a set of inference paths. S35. Perform path encoding operation on each inference path in the inference path set, and stack the corresponding user vector, product vector and relation vector in order of node access in the inference path to generate a path encoding set.

[0015] Optionally, S4 specifically includes: S41. Read the path code set, perform node identifier parsing operation on each path code in the path code set, locate the corresponding product vector in the target product set according to the node identifier, and establish an association mapping relationship according to the correspondence between the path code and the product vector to generate a mapping set. S42. Perform vector expansion operation on each path code in the mapping set, read the user vector, relation vector and product vector arranged in the node access order in the path code in sequence, and arrange them in the path access order to generate the corresponding path vector sequence. S43. Input the path vector sequence into the capsule network, perform vector transformation operations sequentially according to the order of the path vector sequence, perform dynamic route calculation operations on the vector transformation results, iteratively calculate and update the coupling weights between the path vector sequences, and perform weighted aggregation operations on the path vector sequences based on the coupling weights to generate path aggregation vectors. S44. Perform vector matching operation on the path aggregation vector and the product vector in the target product set. Classify the path aggregation vector according to the corresponding product vector. Based on the association mapping relationship between the path code and the product vector, arrange all path aggregation vectors belonging to the same product vector in a centralized manner to generate a product path set. S45. Concatenate the path aggregation vectors in the product path set according to the order of the path codes in the path code set to generate the path matching features of the corresponding product vectors, and establish a number correspondence between the path matching features and the corresponding product vectors.

[0016] Optionally, S5 specifically includes: S51. Read the path matching features and the corresponding target product vectors, arrange them according to the vector number order in the unified coding index table, perform vector transformation operation on the path matching features in the Transformer network, and generate the corresponding transformation result sequence according to the arrangement order of the path matching features. S52. According to the order of the transformation result sequence, perform attention calculation operation on the transformation result corresponding to each path matching feature, iterate through any two transformation results to perform association calculation, and perform weighted processing on the transformation result sequence based on the association calculation result to generate a feature modeling sequence. The association calculation specifically includes: Read each transformation result sequentially according to the order of the transformation result sequence, and use the read transformation result as the current calculation vector; Perform a multiplication operation on the current computation vector and the values ​​at the same dimension positions of the non-current computation vectors in the transformation result sequence, and accumulate all multiplication results in dimensional order to generate associated values; Arrange the associated values ​​according to the order of the transformation result sequence to generate an associated value sequence; Perform Softmax normalization on each associated value in the associated value sequence to generate an associated weight sequence; S53. Align the feature modeling sequence with the corresponding target commodity vector according to the vector numbering order, and input the alignment result into the factorization machine to perform feature interaction calculation; S54. Using the interactive calculation results as the basis for scoring, calculate the recommended score value and generate a set of recommended scores.

[0017] Optionally, S6 specifically includes: S61. Bind each recommended rating value in the recommended rating set to the corresponding target product vector and product node identifier to generate a rating-corresponding set; S62. Sort all recommended rating values ​​in the rating set in descending order according to their numerical values, and adjust the order of the target product vector and product node identifiers accordingly. S63. Extract the corresponding product node identifiers according to the sorting results and generate intelligent recommendation results; S64. Collect feedback data from users to be recommended on the intelligent recommendation results, and perform parameter update operations on the Transformer network and factorization machine based on the feedback data. At the same time, update the relationship edges between the corresponding user nodes and product nodes in the marketing knowledge graph.

[0018] The beneficial effects of this invention are: First, this invention performs a unified encoding operation on the marketing knowledge graph using the TransE model, mapping user nodes, product nodes, and relationship edges to a unified vector space and constructing a unified encoding index table. This enables different types of entities in the knowledge graph to form a consistent vector representation under the same encoding rules, solving the problem of reduced accuracy of path reasoning results due to inconsistent node vector expression spaces in the prior art. This improves the accuracy and consistency of entity relationship expression in the knowledge graph.

[0019] Secondly, this invention performs filtering operations on user vectors by combining the Locality Sensitive Hash algorithm and the DeepWalk algorithm, and performs path reasoning on target user nodes using the MINERVA algorithm. At the same time, it uses capsule networks to aggregate features of the path encoding set, so that the structural information in the path can be effectively fused through the dynamic routing mechanism. This enhances the ability of path features to express user interests and improves the accuracy of target user interest identification and the effectiveness of path reasoning.

[0020] Finally, this invention performs deep feature modeling on path matching features through a Transformer network and completes the recommendation score calculation by combining it with a factorization machine. At the same time, it performs update operations on the Transformer network, factorization machine, and marketing knowledge graph based on user feedback data, so that the recommendation model parameters and knowledge graph structure can be adjusted synchronously according to user feedback. This solves the problem that the recommendation results cannot reflect changes in user interests in a timely manner in the prior art, thereby improving the accuracy and dynamic adaptability of the recommendation results and realizing continuous optimization of the personalized marketing recommendation process. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a personalized marketing intelligent recommendation method based on knowledge graphs proposed in this invention; Figure 2 This is a flowchart illustrating the overall closed-loop process of a knowledge graph-based intelligent recommendation method for personalized marketing proposed in this invention. Figure 3 This is a flowchart illustrating the recommendation scoring and model update process of a knowledge graph-based personalized marketing intelligent recommendation method proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 A personalized marketing intelligent recommendation method based on knowledge graphs includes the following steps: S1. The pre-built marketing knowledge graph is uniformly encoded using the TransE model to generate user vectors, product vectors, and relationship vectors, and a unified encoding index table is constructed. S2. Based on the collected user information to be recommended, use the Locality Sensitive Hash algorithm to match multiple candidate user vectors from the unified coding index table, and combine the DeepWalk algorithm to filter the target user vectors. At the same time, extract the product vectors associated with the target user vectors to construct a target product set. S3. Starting with the node corresponding to the target user vector, perform path traversal in the marketing knowledge graph, and use the MINERVA algorithm to perform path reasoning and encoding on the traversal results to generate a path encoding set. S4. Associate and match the path encoding set with the target product set, and perform feature aggregation operation on the matching results through capsule network to generate path matching features for each target product vector; S5. Based on path matching features, a Transformer network is used to perform feature modeling operations, and a recommendation score is calculated using a factorization machine to generate a set of recommendation scores. S6. Sort the target product set according to the recommended rating set to obtain intelligent recommendation results, collect user feedback data, and update the Transformer network, factor decomposition machine and marketing knowledge graph according to the user feedback data.

[0024] In this embodiment, the user information to be recommended represents a data set used to describe the attribute characteristics of the user to be recommended, the node represents an entity unit in the marketing knowledge graph used to describe user entities or product entities, and the target product vector represents any product vector in the target product set.

[0025] In this embodiment, S1 specifically includes: S11. Read all node sets and edge sets in the pre-built marketing knowledge graph, extract the node identifiers of user nodes and product nodes in the node set, and perform relationship identifier extraction operation on the relationship edges in the edge set. S12. Use the node identifiers corresponding to each user node and product node as the head entity identifier and tail entity identifier respectively, and use the relationship identifiers of the corresponding relationship edges as relationship entity identifiers. Summarize the three types of entity identifiers to form a set of triples. S13. Input the set of triples into the TransE model. In the TransE model, perform vector space mapping operation on the nodes and edges corresponding to the three types of entity identifiers to generate the corresponding user vector, relation vector and product vector. S14. Based on the marketing knowledge graph, perform vector numbering operations on user vectors, product vectors and relationship vectors, associate and store the corresponding node identifiers, relationship identifiers and vector numbers, and generate an initial index table. S15. Perform vector integrity check on the initial index table, and arrange the user vectors, product vectors and relationship vectors that pass the check in the order of node identifiers to generate a unified coded index table.

[0026] In this embodiment, the process of constructing the marketing knowledge graph specifically includes: Acquire user behavior data and product data from multiple sources, perform time sorting and behavior type labeling operations on user behavior data, and perform product identifier extraction and field consistency operations on product data to obtain user behavior sets and product record sets; Based on the user behavior set and the product record set, user identifiers and product identifiers are extracted respectively, and encoded into user nodes and product nodes to form a node set; Based on the correspondence between each user identifier and product identifier, construct the interaction relationship edges between user nodes and product nodes, and assign a corresponding relationship identifier to each interaction relationship edge to form an edge set; Perform node numbering operations on user nodes and product nodes in the node set, perform relationship numbering operations on interaction edges in the edge set, and establish the association between node numbers and relationship numbers; Write the set of numbered nodes and the set of edges into the preset initial graph structure, and bind the relationships to generate a marketing knowledge graph.

[0027] In this embodiment, S13 specifically includes: S131. Read each triplet in the triplet set in a fixed order, and establish a position mapping sequence based on the position index of the head entity identifier and the tail entity identifier in the node set, and the position index of the relation entity identifier in the edge set. S132. Based on the location mapping sequence, map the user nodes corresponding to each head entity identifier to the head storage location in the entity vector space of the TransE model, and arrange the head storage locations according to the node identifier order to form a user node sequence. S133. Based on the position mapping sequence, map the relation edges corresponding to each relation entity identifier to the relation storage location in the relation vector space of the TransE model, and arrange each relation storage location according to the relation identifier order to form a relation edge sequence. S134. Based on the location mapping sequence, map the product nodes corresponding to each tail entity identifier to the tail storage location in the entity vector space of the TransE model, and arrange each tail storage location according to the node identifier order to form a product node sequence. S135. In the TransE model, according to the order of the triplet set, perform vector superposition operation on the entity vector in each head storage location and the relation vector in the corresponding relation storage location. Establish a positional correspondence between the superposition result and the entity vector in the corresponding tail storage location, so that the head storage location, relation storage location and tail storage location form a one-to-one vector mapping relationship, and generate user vector, relation vector and product vector.

[0028] In this embodiment, S2 specifically includes: S21. Obtain the information of users to be recommended, perform field parsing, discrete field encoding and continuous field normalization operations on the information of users to be recommended, and concatenate the normalization results according to the preset field order to generate user feature vectors. S22. Perform dimension alignment and vector space mapping operations on the user feature vector to generate a query user vector with the same dimension as the user vector in the unified coding index table. S23. Perform locality-sensitive hashing calculation based on the query user vector, generate a hash code for the query user vector using a preset hash function, and retrieve multiple candidate user vectors from the unified coding index table based on the hash code, specifically including: Perform a vector dimension partitioning operation on the query user vector, dividing the query user vector into multiple sub-vector segments according to a preset dimension length; Perform product quantization encoding operation on each sub-vector segment, perform nearest neighbor matching operation on each sub-vector segment using a preset sub-quantization codebook to determine the corresponding codeword number, and arrange all codeword numbers in the order of the sub-vector segments to generate the product quantization vector of the query user vector; A locality-sensitive hash mapping operation is performed on the product quantized vector. A hash calculation is performed on the product quantized vector using a preset hash function to generate the corresponding hash code. Read all user vectors from the unified coding index table, and perform product quantization coding and hash calculation operations on all user vectors to generate a user code set; Perform Hamming distance calculation on the hash code and user code set, and select user vectors whose Hamming distance values ​​meet the preset conditions as candidate user vectors; Read the node identifier corresponding to the candidate user vector and bind it to the candidate user vector; S24. Using the user nodes corresponding to each candidate user vector as the starting nodes, the DeepWalk algorithm is used to perform random walk operations in the marketing knowledge graph to generate node access sequences. Based on the node access sequences, the graph neighborhood score of the corresponding candidate user vector is calculated, and the target user vector is selected based on the graph neighborhood score. S25. Locate the target user node corresponding to the target user vector, read the set of associated edges of the target user node and extract the set of product nodes pointed to by the set of associated edges, extract the product vector corresponding to the set of product nodes from the unified coding index table, and construct the target product set.

[0029] In this embodiment, the process of filtering the target user vector specifically includes: Read the node identifiers corresponding to each candidate user vector, locate the corresponding user nodes in the marketing knowledge graph, and use each user node as the starting node for the random walk to generate a set of starting nodes. Based on the set of starting nodes, DeepWalk random walk operation is performed in the marketing knowledge graph. According to the preset walk step size, each starting node visits the adjacent nodes along the associated edges in turn, and records the node identifiers visited in each step to generate the corresponding set of access paths. A sliding window truncation operation is performed on the access path set to extract continuous node segments from each access path. The node co-occurrence frequency is calculated based on the number of times nodes co-occur in each node segment, and a node co-occurrence set is generated. The number of co-occurrences represents the cumulative number of times any candidate user node and a node in the access path set co-occur in the same node segment within a preset sliding window range. The neighborhood association strength between candidate user nodes and nodes in the access path set other than the candidate user node is calculated based on the node co-occurrence set. The neighborhood association strength is used as the graph neighborhood score of the corresponding candidate user vector to generate a graph neighborhood score set. The graph neighborhood score set is sorted, and the candidate user vector with the highest graph neighborhood score is selected as the target user vector. At the same time, the node identifiers corresponding to the target user vectors are read, and the correspondence between the target user vectors and the node identifiers is established.

[0030] In this embodiment, S3 specifically includes: S31. Read the target user node corresponding to the target user vector in the marketing knowledge graph, extract the node identifier of the target user node, and locate the set of adjacent nodes in the marketing knowledge graph that have a relationship edge with the target user node. S32. Starting from the target user node, perform a path traversal operation on the set of adjacent nodes. Expand the node access range layer by layer along the relation edges according to the preset traversal depth. Record the relation identifiers and node identifiers passed through each expansion and generate a candidate path set. Each path in the candidate path set contains a sequence of node identifiers and a sequence of relation identifiers arranged in order. S33. Perform a unified encoding index table query operation on each candidate path, extract the corresponding user vector, product vector and relationship vector respectively, and perform vector concatenation operation in the order of the candidate path set to generate a path sequence set; S34. The MINERVA algorithm is used to perform path reasoning on each path sequence in the path sequence set. Action selection calculations are performed on each path sequence to generate a relation jump sequence. The reasoning path set is then determined based on the relation jump sequence. Specifically, this includes: Read each path sequence in the path sequence set, extract the user vector in the path sequence as the initial path vector, and read the corresponding relation vector and product vector in the order of arrangement in the path sequence. Arrange the user vector, relation vector and product vector in the path order to establish the path access sequence. Based on the path access sequence, the MINERVA algorithm is used. The user vector is used as the starting vector, and the relation vector is read step by step in the order of the path access sequence. The user vector of the current step is combined with the corresponding relation vector in sequence to generate the path intermediate vector of the current step, and the path intermediate vector is used as the starting vector of the next step. After the intermediate vector of each path is generated, the intermediate vector of the path is matched one by one with all the relation vectors in the unified coding index table. Based on the matching results, the relation vector corresponding to the current step is determined and the corresponding relation identifier is recorded. The relation jump sequence is formed by arranging them in the path order. Based on the relationship jump sequence, locate the corresponding relationship edge in the marketing knowledge graph, and visit the corresponding node identifiers in sequence according to the relationship identifiers in the relationship jump sequence to generate a node access sequence; The user vector and product vector corresponding to the node access sequence and the relation vector corresponding to the relation jump sequence are rearranged according to the access order, and a corresponding relationship is established with the path sequence to generate a set of inference paths. S35. Perform path encoding operation on each inference path in the inference path set, and stack the corresponding user vector, product vector and relation vector in order of node access in the inference path to generate a path encoding set.

[0031] In this embodiment, S4 specifically includes: S41. Read the path code set, perform node identifier parsing operation on each path code in the path code set, locate the corresponding product vector in the target product set according to the node identifier, and establish an association mapping relationship according to the correspondence between the path code and the product vector to generate a mapping set. S42. Perform vector expansion operation on each path code in the mapping set, read the user vector, relation vector and product vector arranged in the node access order in the path code in sequence, and arrange them in the path access order to generate the corresponding path vector sequence. S43. Input the path vector sequence into the capsule network, perform vector transformation operations sequentially according to the order of the path vector sequence, perform dynamic route calculation operations on the vector transformation results, iteratively calculate and update the coupling weights between the path vector sequences, and perform weighted aggregation operations on the path vector sequences based on the coupling weights to generate path aggregation vectors. S44. Perform vector matching operation on the path aggregation vector and the product vector in the target product set. Classify the path aggregation vector according to the corresponding product vector. Based on the association mapping relationship between the path code and the product vector, arrange all path aggregation vectors belonging to the same product vector in a centralized manner to generate a product path set. S45. Concatenate the path aggregation vectors in the product path set according to the order of the path codes in the path code set to generate the path matching features of the corresponding product vectors, and establish a number correspondence between the path matching features and the corresponding product vectors.

[0032] In this embodiment, S43 specifically includes: S431. Arrange the user vector, relationship vector and product vector in the path vector sequence according to the node access order, establish the vector correspondence, and generate the path arrangement vector. S432. Perform vector transformation operation on each path arrangement vector. In the capsule network, perform vector space mapping processing on each path arrangement vector according to the preset vector transformation rules, convert the path arrangement vector into the corresponding transformation vector, and generate a set of transformation vectors. S433. Perform a coupling relationship establishment operation on the transformation vector set, establish coupling relationships between each transformation vector according to the path arrangement order, and assign a corresponding coupling weight value to each coupling relationship to generate a coupling weight set. S434. Perform dynamic routing iterative calculation based on the coupling weight set. In each iteration, perform weighted calculation on each transformation vector and the corresponding coupling weight value to generate an intermediate aggregate vector. Update the coupling weight value according to the degree of vector correlation between the intermediate aggregate vector and each transformation vector to generate an updated coupling weight set. S435. After reaching the preset number of dynamic routing iterations, perform a weighted aggregation operation on the updated set of coupling weights and the set of transformation vectors to generate a path aggregation vector for the corresponding path vector sequence.

[0033] In this embodiment, S5 specifically includes: S51. Read the path matching features and the corresponding target product vectors, arrange them according to the vector number order in the unified coding index table, perform vector transformation operation on the path matching features in the Transformer network, and generate the corresponding transformation result sequence according to the arrangement order of the path matching features. S52. According to the order of the transformation result sequence, perform attention calculation operation on the transformation result corresponding to each path matching feature, iterate through any two transformation results to perform association calculation, and perform weighted processing on the transformation result sequence based on the association calculation result to generate a feature modeling sequence. The association calculation specifically includes: Read each transformation result sequentially according to the order of the transformation result sequence, and use the read transformation result as the current calculation vector; Perform a multiplication operation on the current computation vector and the values ​​at the same dimension positions of the non-current computation vectors in the transformation result sequence, and accumulate all multiplication results in dimensional order to generate associated values; Arrange the associated values ​​according to the order of the transformation result sequence to generate an associated value sequence; Perform Softmax normalization on each associated value in the associated value sequence to generate an associated weight sequence; S53. Align the feature modeling sequence with the corresponding target commodity vector according to the vector numbering order, and input the alignment result into the factorization machine to perform feature interaction calculation; S54. Using the interactive calculation results as the basis for scoring, calculate the recommended score value and generate a set of recommended scores.

[0034] In this embodiment, the vector transformation operation specifically includes: In the Transformer network, a corresponding vector position is assigned to each path matching feature; Perform a weight matrix multiplication operation on each path matching feature, multiplying the values ​​of each dimension of the path matching feature with the corresponding weight values ​​in the preset transformation weight matrix, and arranging all multiplication results in the dimensional order of the path matching feature to generate a transformation vector. The bias superposition operation is performed on the transformation vector. The values ​​of each dimension in the transformation vector are added to the corresponding bias values ​​in the preset bias vector. The addition results are arranged in the order of the dimensions of the path matching features to generate the bias transformation vector. The bias transformation vector is normalized by adjusting the values ​​of all dimensions in the bias transformation vector proportionally to generate a normalized vector. The normalized vector is written into the vector position of the corresponding path matching feature to generate a sequence of transformation results.

[0035] In this embodiment, feature interaction calculation specifically includes: Read each feature modeling vector in the feature modeling sequence and the corresponding target product vector in the unified coding index table in sequence. Perform a multiplication operation on the values ​​of each feature modeling vector and the corresponding target product vector at the same dimension position, and accumulate all multiplication results in dimensional order to generate the corresponding feature interaction value. Perform a vector concatenation operation between each feature modeling vector and the corresponding target product vector to generate a combined vector; Perform multiplication on the values ​​at any two dimensions in the combined vector, and accumulate all multiplication results according to the dimensional order of the combined vector to generate a combined interactive value; The feature interaction values ​​and combined interaction values ​​are arranged in vector number order to obtain the interaction calculation results.

[0036] In this embodiment, the calculation process of the recommended score specifically includes: Read the feature interaction values ​​and combined interaction values ​​corresponding to each vector number in the interaction calculation results in sequence, and establish a correspondence according to the vector number order; For each vector number, perform an addition operation on the feature interaction value and the combined interaction value to generate the corresponding score value; The rating values ​​are compressed and mapped to a preset rating range to generate corresponding recommended rating values.

[0037] In this embodiment, S6 specifically includes: S61. Bind each recommended rating value in the recommended rating set to the corresponding target product vector and product node identifier to generate a rating-corresponding set; S62. Sort all recommended rating values ​​in the rating set in descending order according to their numerical values, and adjust the order of the target product vector and product node identifiers accordingly. S63. Extract the corresponding product node identifiers according to the sorting results and generate intelligent recommendation results; S64. Collect feedback data from users to be recommended on the intelligent recommendation results, and perform parameter update operations on the Transformer network and factorization machine based on the feedback data. At the same time, update the relationship edges between the corresponding user nodes and product nodes in the marketing knowledge graph.

[0038] Example 1: To verify the feasibility of this invention in practice, it was applied to an online product marketing recommendation scenario. This scenario has long faced challenges such as rapid growth in user numbers, a continuous increase in product categories, and frequent changes in user interests. In this scenario, traditional recommendation methods mainly rely on similarity calculations based on users' historical click data. Due to the lack of a unified coding structure and path reasoning capabilities, the recommendation results deviate from users' true interests. Specifically, this manifests as low click-through rates for recommended products, short user dwell times, and unstable product conversion rates, making it difficult to meet personalized marketing needs. Therefore, in this scenario, the knowledge graph-based personalized marketing intelligent recommendation method proposed in this invention is adopted to improve recommendation accuracy and the degree of matching with user interests.

[0039] In this scenario, user behavior data and product data are first collected. User behavior data includes browsing behavior records, click behavior records, and purchase behavior records, while product data includes product category information, attribute information, and association information. A marketing knowledge graph is then constructed based on this user behavior data and product data. After the marketing knowledge graph is constructed, the TransE model is used to perform a unified encoding operation on the user nodes, product nodes, and relationship edges in the marketing knowledge graph, generating user vector sets, product vector sets, and relationship vector sets. A unified encoding index table is then constructed, ensuring that all entity vectors reside in a unified vector space.

[0040] Subsequently, during the recommendation process, user information is collected, and candidate user vectors are filtered in a unified coding index table using the Locality Sensitive Hash (LSH) algorithm. Then, target user vectors are filtered using the DeepWalk algorithm combined with structural information from the marketing knowledge graph, ensuring that the target user vectors accurately reflect the current user's interest characteristics. After the target user vector filtering is completed, path traversal is performed in the marketing knowledge graph based on the target user nodes, and path reasoning is performed using the MINERVA algorithm to generate a path coding set that reflects the propagation relationships of user interests.

[0041] After the path encoding set is generated, a capsule network is used to perform dynamic routing calculations on the path encoding set and generate path matching features, effectively aggregating the structural information in the path. Subsequently, the path matching features are input into a Transformer network to perform deep feature modeling, and a factorization machine is used to calculate recommendation scores, generating a recommendation score set. Intelligent recommendation results are then generated based on this set. After the recommendation results are generated, user click behavior data and purchase behavior data are collected, and the Transformer network, factorization machine, and marketing knowledge graph are updated based on user feedback data, enabling the recommendation model to continuously optimize according to changes in user interests.

[0042] To verify the practical effect of this invention, a user set and a product set of the same size were selected for testing in this scenario. The number of users was 12,000, the number of products was 8,500, the marketing knowledge graph contained 20,500 nodes and 96,800 relationship edges, and the recommendation effect was compared between traditional recommendation methods and the method of this invention. The following results were obtained: Table 1. Comparison of Personalized Marketing Recommendation Effectiveness

[0043] As shown in Table 1, the method of this invention significantly outperforms traditional recommendation methods in all key indicators. Regarding recommendation accuracy, the method of this invention achieves 89.7%, a 23.9% improvement compared to the 72.4% of traditional recommendation methods, indicating that the unified coding and path reasoning mechanism of this invention can more accurately identify user interests. In terms of click-through rate, the method of this invention achieves 16.8%, a 75.0% improvement compared to the 9.6% of traditional recommendation methods, demonstrating that the recommendation results generated by this invention are more attractive to users.

[0044] Regarding product conversion rate, the method of this invention achieved 7.2%, an 89.5% increase compared to the 3.8% of traditional recommendation methods, indicating that this invention can effectively increase the actual purchase probability of recommended products. As for average user dwell time, the method of this invention reached 67 seconds, a significant increase compared to the 38 seconds of traditional recommendation methods, demonstrating that the recommendation results of this invention can improve user engagement.

[0045] Furthermore, regarding the duplicate recommendation rate, the method of this invention reduces it to 6.4%, a significant decrease compared to the 18.2% of traditional recommendation methods, indicating that the path reasoning mechanism of this invention can reduce duplicate recommendations. In terms of the proportion of new product recommendations, the method of this invention reaches 28.9%, a significant increase compared to the 12.5% ​​of traditional recommendation methods, demonstrating that this invention can more effectively uncover potential product interests.

[0046] The above data demonstrates that this invention, through a unified coding mechanism, path reasoning mechanism, and deep feature modeling mechanism, can significantly improve recommendation accuracy, recommendation click-through rate, and product conversion rate, while reducing duplicate recommendation rate and enhancing the recommendation system's adaptability to changes in user interests. This verifies the effectiveness and practicality of this invention in personalized marketing recommendations.

[0047] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized marketing intelligent recommendation method based on knowledge graphs, characterized in that, Includes the following steps: S1. The pre-built marketing knowledge graph is uniformly encoded using the TransE model to generate user vectors, product vectors, and relationship vectors, and a unified encoding index table is constructed. S2. Based on the collected user information to be recommended, use the Locality Sensitive Hash algorithm to match multiple candidate user vectors from the unified coding index table, and combine the DeepWalk algorithm to filter the target user vectors. At the same time, extract the product vectors associated with the target user vectors to construct a target product set. S3. Starting with the node corresponding to the target user vector, perform path traversal in the marketing knowledge graph, and use the MINERVA algorithm to perform path reasoning and encoding on the traversal results to generate a path encoding set. S4. Associate and match the path encoding set with the target product set, and perform feature aggregation operation on the matching results through capsule network to generate path matching features for each target product vector; S5. Based on path matching features, a Transformer network is used to perform feature modeling operations, and a recommendation score is calculated using a factorization machine to generate a set of recommendation scores. S6. Sort the target product set according to the recommended rating set to obtain intelligent recommendation results, collect user feedback data, and update the Transformer network, factor decomposition machine and marketing knowledge graph according to the user feedback data.

2. The personalized marketing intelligent recommendation method based on knowledge graphs according to claim 1, characterized in that, The user information to be recommended represents a data set used to describe the attribute characteristics of the user to be recommended, the node represents an entity unit in the marketing knowledge graph used to describe user entities or product entities, and the target product vector represents any product vector in the target product set.

3. The personalized marketing intelligent recommendation method based on knowledge graphs according to claim 1, characterized in that, S1 specifically includes: S11. Read all node sets and edge sets in the pre-built marketing knowledge graph, extract the node identifiers of user nodes and product nodes in the node set, and perform relationship identifier extraction operation on the relationship edges in the edge set. S12. Use the node identifiers corresponding to each user node and product node as the head entity identifier and tail entity identifier respectively, and use the relationship identifiers of the corresponding relationship edges as relationship entity identifiers. Summarize the three types of entity identifiers to form a set of triples. S13. Input the set of triples into the TransE model. In the TransE model, perform vector space mapping operation on the nodes and edges corresponding to the three types of entity identifiers to generate the corresponding user vector, relation vector and product vector. S14. Based on the marketing knowledge graph, perform vector numbering operations on user vectors, product vectors and relationship vectors, associate and store the corresponding node identifiers, relationship identifiers and vector numbers, and generate an initial index table. S15. Perform vector integrity check on the initial index table, and arrange the user vectors, product vectors and relationship vectors that pass the check in the order of node identifiers to generate a unified coded index table.

4. The personalized marketing intelligent recommendation method based on knowledge graphs according to claim 3, characterized in that, The process of constructing the marketing knowledge graph specifically includes: Acquire user behavior data and product data from multiple sources, perform time sorting and behavior type labeling operations on user behavior data, and perform product identifier extraction and field consistency operations on product data to obtain user behavior sets and product record sets; Based on the user behavior set and the product record set, user identifiers and product identifiers are extracted respectively, and encoded into user nodes and product nodes to form a node set; Based on the correspondence between each user identifier and product identifier, construct the interaction relationship edges between user nodes and product nodes, and assign a corresponding relationship identifier to each interaction relationship edge to form an edge set; Perform node numbering operations on user nodes and product nodes in the node set, perform relationship numbering operations on interaction edges in the edge set, and establish the association between node numbers and relationship numbers; Write the set of numbered nodes and the set of edges into the preset initial graph structure, and bind the relationships to generate a marketing knowledge graph.

5. The personalized marketing intelligent recommendation method based on knowledge graph as described in claim 1, characterized in that, S2 specifically includes: S21. Obtain the information of users to be recommended, perform field parsing, discrete field encoding and continuous field normalization operations on the information of users to be recommended, and concatenate the normalization results according to the preset field order to generate user feature vectors. S22. Perform dimension alignment and vector space mapping operations on the user feature vector to generate a query user vector with the same dimension as the user vector in the unified coding index table. S23. Perform locality-sensitive hashing calculation based on the query user vector, generate a hash code for the query user vector using a preset hash function, and retrieve multiple candidate user vectors from the unified coding index table based on the hash code, specifically including: Perform a vector dimension partitioning operation on the query user vector, dividing the query user vector into multiple sub-vector segments according to a preset dimension length; Perform product quantization encoding operation on each sub-vector segment, perform nearest neighbor matching operation on each sub-vector segment using a preset sub-quantization codebook to determine the corresponding codeword number, and arrange all codeword numbers in the order of the sub-vector segments to generate the product quantization vector of the query user vector; A locality-sensitive hash mapping operation is performed on the product quantized vector. A hash calculation is performed on the product quantized vector using a preset hash function to generate the corresponding hash code. Read all user vectors from the unified coding index table, and perform product quantization coding and hash calculation operations on all user vectors to generate a user code set; Perform Hamming distance calculation on the hash code and user code set, and select user vectors whose Hamming distance values ​​meet the preset conditions as candidate user vectors; Read the node identifier corresponding to the candidate user vector and bind it to the candidate user vector; S24. Using the user nodes corresponding to each candidate user vector as the starting nodes, the DeepWalk algorithm is used to perform random walk operations in the marketing knowledge graph to generate node access sequences. Based on the node access sequences, the graph neighborhood score of the corresponding candidate user vector is calculated, and the target user vector is selected based on the graph neighborhood score. S25. Locate the target user node corresponding to the target user vector, read the set of associated edges of the target user node and extract the set of product nodes pointed to by the set of associated edges, extract the product vector corresponding to the set of product nodes from the unified coding index table, and construct the target product set.

6. The personalized marketing intelligent recommendation method based on knowledge graph according to claim 5, characterized in that, The process of filtering the target user vector specifically includes: Read the node identifiers corresponding to each candidate user vector, locate the corresponding user nodes in the marketing knowledge graph, and use each user node as the starting node for the random walk to generate a set of starting nodes. Based on the set of starting nodes, DeepWalk random walk operation is performed in the marketing knowledge graph. According to the preset walk step size, each starting node visits the adjacent nodes along the associated edges in turn, and records the node identifiers visited in each step to generate the corresponding set of access paths. A sliding window truncation operation is performed on the access path set to extract continuous node segments from each access path. The node co-occurrence frequency is calculated based on the number of times nodes co-occur in each node segment, and a node co-occurrence set is generated. The number of co-occurrences represents the cumulative number of times any candidate user node and a node in the access path set co-occur in the same node segment within a preset sliding window range. The neighborhood association strength between candidate user nodes and nodes in the access path set other than the candidate user node is calculated based on the node co-occurrence set. The neighborhood association strength is used as the graph neighborhood score of the corresponding candidate user vector to generate a graph neighborhood score set. The graph neighborhood score set is sorted, and the candidate user vector with the highest graph neighborhood score is selected as the target user vector. At the same time, the node identifiers corresponding to the target user vectors are read, and the correspondence between the target user vectors and the node identifiers is established.

7. The personalized marketing intelligent recommendation method based on knowledge graph as described in claim 1, characterized in that, S3 specifically includes: S31. Read the target user node corresponding to the target user vector in the marketing knowledge graph, extract the node identifier of the target user node, and locate the set of adjacent nodes in the marketing knowledge graph that have a relationship edge with the target user node. S32. Starting from the target user node, perform a path traversal operation on the set of adjacent nodes, expand the node access range layer by layer along the relation edges according to the preset traversal depth, record the relation identifiers and node identifiers passed through each expansion, and generate a candidate path set. Each path in the candidate path set contains a sequence of node identifiers and a sequence of relation identifiers arranged in order. S33. Perform a unified encoding index table query operation on each candidate path, extract the corresponding user vector, product vector and relationship vector respectively, and perform vector concatenation operation in the order of the candidate path set to generate a path sequence set; S34. The MINERVA algorithm is used to perform path reasoning on each path sequence in the path sequence set. Action selection calculations are performed on each path sequence to generate a relation jump sequence. The reasoning path set is then determined based on the relation jump sequence. Specifically, this includes: Read each path sequence in the path sequence set, extract the user vector in the path sequence as the initial path vector, and read the corresponding relation vector and product vector in the order of arrangement in the path sequence. Arrange the user vector, relation vector and product vector in the path order to establish the path access sequence. Based on the path access sequence, the MINERVA algorithm is used. The user vector is used as the starting vector, and the relation vector is read step by step in the order of the path access sequence. The user vector of the current step is combined with the corresponding relation vector in sequence to generate the path intermediate vector of the current step, and the path intermediate vector is used as the starting vector of the next step. After the intermediate vector of each path is generated, the intermediate vector of the path is matched one by one with all the relation vectors in the unified coding index table. Based on the matching results, the relation vector corresponding to the current step is determined and the corresponding relation identifier is recorded. The relation jump sequence is formed by arranging them in the path order. Based on the relationship jump sequence, locate the corresponding relationship edge in the marketing knowledge graph, and visit the corresponding node identifiers in sequence according to the relationship identifiers in the relationship jump sequence to generate a node access sequence; The user vector and product vector corresponding to the node access sequence and the relation vector corresponding to the relation jump sequence are rearranged according to the access order, and a corresponding relationship is established with the path sequence to generate a set of inference paths. S35. Perform path encoding operation on each inference path in the inference path set, and stack the corresponding user vector, product vector and relation vector in order of node access in the inference path to generate a path encoding set.

8. The personalized marketing intelligent recommendation method based on knowledge graph according to claim 1, characterized in that, S4 specifically includes: S41. Read the path code set, perform node identifier parsing operation on each path code in the path code set, locate the corresponding product vector in the target product set according to the node identifier, and establish an association mapping relationship according to the correspondence between the path code and the product vector to generate a mapping set. S42. Perform vector expansion operation on each path code in the mapping set, read the user vector, relation vector and product vector arranged in the node access order in the path code in sequence, and arrange them in the path access order to generate the corresponding path vector sequence. S43. Input the path vector sequence into the capsule network, perform vector transformation operations sequentially according to the order of the path vector sequence, perform dynamic route calculation operations on the vector transformation results, iteratively calculate and update the coupling weights between the path vector sequences, and perform weighted aggregation operations on the path vector sequences based on the coupling weights to generate path aggregation vectors. S44. Perform vector matching operation on the path aggregation vector and the product vector in the target product set. Classify the path aggregation vector according to the corresponding product vector. Based on the association mapping relationship between the path code and the product vector, arrange all path aggregation vectors belonging to the same product vector in a centralized manner to generate a product path set. S45. Concatenate the path aggregation vectors in the product path set according to the order of the path codes in the path code set to generate the path matching features of the corresponding product vectors, and establish a number correspondence between the path matching features and the corresponding product vectors.

9. The personalized marketing intelligent recommendation method based on knowledge graph according to claim 1, characterized in that, S5 specifically includes: S51. Read the path matching features and the corresponding target product vectors, arrange them according to the vector number order in the unified coding index table, perform vector transformation operation on the path matching features in the Transformer network, and generate the corresponding transformation result sequence according to the arrangement order of the path matching features. S52. According to the order of the transformation result sequence, perform attention calculation operation on the transformation result corresponding to each path matching feature, iterate through any two transformation results to perform association calculation, and perform weighted processing on the transformation result sequence based on the association calculation result to generate a feature modeling sequence. The association calculation specifically includes: Read each transformation result sequentially according to the order of the transformation result sequence, and use the read transformation result as the current calculation vector; Perform a multiplication operation on the current computation vector and the values ​​at the same dimension positions of the non-current computation vectors in the transformation result sequence, and accumulate all multiplication results in dimensional order to generate associated values; Arrange the associated values ​​according to the order of the transformation result sequence to generate an associated value sequence; Perform Softmax normalization on each associated value in the associated value sequence to generate an associated weight sequence; S53. Align the feature modeling sequence with the corresponding target commodity vector according to the vector numbering order, and input the alignment result into the factorization machine to perform feature interaction calculation; S54. Using the interactive calculation results as the basis for scoring, calculate the recommended score value and generate a set of recommended scores.

10. The personalized marketing intelligent recommendation method based on knowledge graph according to claim 1, characterized in that, S6 specifically includes: S61. Bind each recommended rating value in the recommended rating set to the corresponding target product vector and product node identifier to generate a rating-corresponding set; S62. Sort all recommended rating values ​​in the rating set in descending order according to their numerical values, and adjust the order of the target product vector and product node identifiers accordingly. S63. Extract the corresponding product node identifiers according to the sorting results and generate intelligent recommendation results; S64. Collect feedback data from users to be recommended on the intelligent recommendation results, and perform parameter update operations on the Transformer network and factorization machine based on the feedback data. At the same time, update the relationship edges between the corresponding user nodes and product nodes in the marketing knowledge graph.

Images