Intelligent clothing recommendation system and method based on multi-modal knowledge graph
By constructing a knowledge graph of users and clothing and performing cross-graph fusion, and using graph neural networks for recommendation, the problem of insufficient user preference modeling in traditional recommendation systems is solved, and high-precision and interpretable personalized clothing recommendations are achieved.
Patent Information
- Application Number
- CN202511089853.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2026-01-02
AI Technical Summary
Traditional recommendation systems cannot effectively uncover the deep preference features behind user behavior, ignore the extension of users' interests on other platforms, and have difficulty modeling the complex relationships between clothing items, lacking a unified information fusion framework.
We construct user knowledge graphs and clothing knowledge graphs, and through cross-graph semantic alignment and heterogeneous graph fusion, we use graph neural networks to perform end-to-end recommendations and generate interpretable recommendation paths.
It achieves high-precision and interpretable personalized recommendations, supports multi-dimensional information linkage, improves the accuracy and reliability of recommendations, and is suitable for various recommendation scenarios.
Smart Images
Figure CN121258618A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of commodity recommendation, more particularly to a clothing intelligent recommendation system and method based on a multi-modal knowledge graph. BACKGROUND
[0002] With the rise of Internet e-commerce platforms, personalized recommendation has become an important means to improve user experience and business conversion rate. Traditional recommendation systems are mainly based on collaborative filtering or content analysis, usually only considering user behavior records (clicks, purchases, collections, etc.) and basic attributes of items. However, this kind of method has certain limitations in the current use context.
[0003] Due to the lack of semantic understanding and reasoning ability, traditional recommendation methods cannot mine the deep preference characteristics behind user behavior, making it difficult to structure model potential interests. Because the user portrait information is not comprehensive, the user-item interaction centered model ignores the user's interest extension in other platforms and scenarios, such as occupation, scene preference, style change trend, etc. Secondly, in the clothing recommendation task, there are complex style matching, scene adaptation, fabric functionality and other multi-level correlations between different clothes, and traditional models are difficult to model them. Finally, user preferences, clothing semantics, image vision, social networks and other information are often distributed in different systems, and there is still a lack of a unified information fusion framework.
[0004] In recent years, knowledge graphs have been widely used in the field of recommendation systems. By structuring users, items and their relationships, it helps to supplement semantic information in recommendation models. However, most current work only constructs user knowledge graphs or item knowledge graphs, with few studies on fusion modeling between the two.
[0005] Therefore, there is an urgent need for a recommendation system architecture that integrates user knowledge graphs and clothing knowledge graphs to achieve deep interest modeling, multi-dimensional information linkage and more interpretable recommendations. SUMMARY
[0006] The purpose of the present application is to provide a clothing intelligent recommendation system and method based on a multi-modal knowledge graph, which can obtain high-precision, interpretable personalized recommendation services.
[0007] The present application provides a clothing intelligent recommendation method based on a multi-modal knowledge graph, comprising the following steps: S1: Extracting entities, attributes and relationships from multi-source user data to construct a user knowledge representation graph; S2: Constructing a clothing knowledge graph according to clothing inventory units; S3: Performing cross-graph semantic alignment and heterogeneous graph fusion according to the user knowledge representation graph and the clothing knowledge graph to obtain a fusion graph; S4: Use a graph neural network model to learn the fused graph to obtain user and clothing nodes, and obtain recommendation candidates based on the user and clothing nodes; S5: Based on the recommended candidates, backtrack the recommended path to obtain path information, and convert the path information into natural language interpretation.
[0008] This invention also provides an intelligent clothing recommendation system based on a multimodal knowledge graph, the system comprising the following modules: The user knowledge graph construction module is configured to extract entities, attributes, and relationships from multi-source user data and construct a user knowledge representation graph. The apparel knowledge graph construction module is configured to: construct an apparel knowledge graph based on the unit of apparel inventory quantity; The cross-graph fusion module is configured to: perform cross-graph semantic alignment and heterogeneous graph fusion based on the user knowledge representation graph and the clothing knowledge graph to obtain a fused graph; The graph neural network recommendation module is configured to: learn the fused graph using a graph neural network model to obtain user and clothing nodes, and obtain recommendation candidates based on the user and clothing nodes; The explanation generation module is configured to: perform recommended path backtracking based on the recommended candidates to obtain path information, and convert the path information into natural language explanation.
[0009] The intelligent clothing recommendation system and method based on multimodal knowledge graph provided by this invention has the following beneficial effects: This invention separates the recommendation process into a graph construction process and a recommendation generation process; two graphs are modeled separately, fused and aligned across graphs, and then end-to-end recommendations are performed using graph neural networks, achieving pluggable construction methods and selectable recommendation strategies; the recommendation results are traced back to improve the interpretability of the recommendations, and finally a high-precision, interpretable personalized recommendation service is obtained.
[0010] This invention supports the modeling of higher-order latent preferences by structuring user graph behavior and semantic features; it also achieves high accuracy in semantic fusion recommendations: after fusing the two graphs, the semantic paths between users and products are richer, and the recommendations are more aligned with interests; and it supports highly interpretable recommendations: through the path structure in the graph, it can output reasons for recommendations, thereby increasing user trust.
[0011] This invention supports the integration of multiple modal information, such as image features, text tags, and social networks, into a knowledge graph. By constructing a fused user knowledge graph and a clothing knowledge graph, this invention achieves multi-dimensional and highly semantic personalized recommendations. It is applicable to various recommendation scenarios, such as outfit recommendations, scene recommendations, and seasonal recommendations, all of which can be flexibly expanded through graph paths. Attached Figure Description
[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the intelligent clothing recommendation method based on multimodal knowledge graph provided by the present invention; Figure 2 This is a schematic diagram illustrating the execution process of the intelligent clothing recommendation method based on multimodal knowledge graph provided by the present invention. Detailed Implementation
[0013] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0014] Figure 1 This diagram illustrates the intelligent clothing recommendation method based on a multimodal knowledge graph, as shown in this embodiment. In this embodiment, the intelligent clothing recommendation method based on a multimodal knowledge graph includes the following steps: S1: User Knowledge Graph Construction Module: Extracts entities, attributes, and relationships from multi-source user data to construct a user knowledge representation graph; In one exemplary embodiment, the multi-source user data includes user behavior records, social network relationships, and user tags; In one exemplary embodiment, step S1 specifically includes: S11: Perform user entity identification based on the multi-source user data to obtain user entities; S12: Based on the user entity, extract user attributes and assign attribute information to each user entity; In one exemplary embodiment, the attribute information includes gender, age group, style preference, common clothing occasions, historical preference categories, and physical characteristics; S13: Based on the user entity and its attribute information, multimodal feature embedding is performed using multimodal information to obtain a user multimodal knowledge graph; In one exemplary embodiment, the multimodal information includes visual information and textual information; As an exemplary embodiment, in step S13, in a scenario with visual acquisition support, image modalities or text information such as user body shape and skin color can be further integrated to construct a user multimodal knowledge graph; S14: Based on the user multimodal knowledge graph, user relationships are constructed by building edges according to the dimensions of interest similarity, geographical proximity, and social connection, and a user knowledge representation graph is established. S2: Clothing Knowledge Graph Construction Module: Constructs a clothing knowledge graph based on clothing inventory units; In one exemplary embodiment, step S2 specifically includes: constructing a clothing knowledge graph based on clothing inventory units, as shown in the formula: , , , , in, Represents clothing entity The final embedding is used to construct a clothing knowledge graph; For original embedding; The modality coding module is used to fuse the raw vector representations of image, text, and structural modalities. Image feature representation extracted from VGG Semantic feature representation extracted for BERT This represents the learnable structure embedding vector generated during TransE initialization; This indicates semantic edge message propagation; , , , These represent the current clothing node, the relationship type of the current edge, the neighboring entities connected to the current clothing node, and the current multimodal clothing knowledge graph, respectively. Indicates edge weight; Represents a linear mapping specific to a relation; This is the initial representation of adjacent clothing nodes; This indicates a weak label prior; Brand labels representing physical garments; This represents a brand-style mapping function; As an exemplary embodiment, in step S2, a clothing knowledge graph is constructed, building a structured graph covering a wide range of clothing entities, attributes, and relationships; specifically, this includes: clothing entity modeling, using clothing inventory units (Stock Keeping) Using units (SKUs) or products as node units, basic information such as product name, product category, and style tags is extracted; clothing attribute extraction is performed by extracting clothing attributes from product text descriptions, image analysis, expert tags, etc., including style, season, applicable scenarios, material, pattern, color, price, brand, etc.; multimodal feature embedding is performed by integrating product text descriptions and images to construct multimodal clothing nodes to realize a multimodal clothing knowledge graph; clothing relationship construction is performed by establishing semantic edges between clothing entities, such as "similar style", "matchable", "functional substitution", "fabric complementarity", etc.; implicit semantic mining is performed by extracting implicit semantic sources, such as the style preferences implied by brands. Some clothing brands often exhibit stable style orientations, such as "minimalist", "sweet", or "outdoor". Therefore, when clothing lacks explicit style labels, other implicit semantic information can be combined as a supplementary basis for style attributes. This type of weak semantic association helps to improve the integrity and generalization ability of style modeling of the graph. It should be noted that the traditional construction of clothing knowledge graphs is consistent with that of user knowledge graphs. It also involves inputting images, descriptions, or related SKU information of clothing to generate an embedding. This embodiment also takes into account some implicit information of clothing, such as features with inherent style attributes like brands, as well as the relationships between clothing for subsequent matching recommendations. S3: Cross-graph fusion module: Based on the user knowledge representation graph and the clothing knowledge graph, perform cross-graph semantic alignment and heterogeneous graph fusion to obtain a fused graph; In one exemplary embodiment, step S3 specifically includes: performing cross-graph semantic alignment and heterogeneous graph fusion based on the user knowledge representation graph and the clothing knowledge graph to obtain a fused graph, as shown in the formula: , , in, To integrate the maps; A collection of user entity nodes; A collection of clothing entity nodes; For users, clothing, and the set of edges across the graph; To create a unified diagram structure by integrating user diagrams and clothing diagrams; A semantically mappable group of nodes for user knowledge representation graphs and clothing knowledge graphs; Indicates cosine similarity; The representation vector for style class nodes in the user knowledge representation graph; For style-related nodes in the clothing knowledge graph The representation vector; To construct a similarity threshold for edges in heterogeneous graphs; , , These represent user style nodes, semantically aligned cross-graph edges, and clothing style nodes, respectively. As an exemplary embodiment, in step S3, cross-graph semantic alignment is first performed to identify semantically mappable node groups in the two graphs, specifically including: entity semantic matching, attribute semantic mapping, and relation semantic alignment; Then, heterogeneous graph fusion is performed to establish the connection between the user graph and the clothing graph, specifically including: cross-graph connection edge construction, fusion hypergraph generation, relation weight calculation, and semantic consistency verification. It should be noted that step S3 primarily addresses the information gap between the user graph and the clothing graph, achieving preference-product semantic alignment; its functions include: Entity semantic alignment identifies semantically mappable groups of nodes in two graphs, such as “user preference style” and “clothing style”. like:
[0015] This is the representation vector of style class nodes in the user graph; Style nodes in the clothing atlas The representation vector; Indicates cosine similarity; To construct a similarity threshold for edges in heterogeneous graphs; Cross-relationship construction involves establishing heterogeneous graph connection edges such as "user-preference style-clothing style" and "user-applicable scenario-clothing application scenario" through methods such as user historical behavior co-occurrence and similar path mining. The graph fusion structure is generated to form a reasonable fusion hypergraph structure between user entities and clothing entities, and to construct multi-hop semantic paths; To ensure semantic consistency, co-occurrence statistics or graph embedding alignment are used to filter and weight relationships during the fusion process to avoid pseudo-associations; This embodiment enables user profiling to move beyond sparse label matching and instead leverages structural path modeling to achieve high-dimensional semantic connections between preferences, attributes, and products; the final graph format is as follows:
[0016] in, : A unified graph structure that integrates user graphs and clothing graphs; The set of user entity nodes, that is, all nodes in the user graph; The set of clothing entity nodes, that is, all nodes in the clothing atlas; Users, clothing, and sets of edges across the graph; S4: Graph Neural Network Recommendation Module: The fused graph is learned using a graph neural network model to obtain user and clothing nodes, and recommendation candidates are obtained based on the user and clothing nodes; In one exemplary embodiment, step S4 specifically includes: learning the fused graph using a graph neural network model to obtain user and clothing nodes; and obtaining recommendation candidates based on the user and clothing nodes, as shown in the formula:
[0017]
[0018]
[0019] in, Indicates the initial input of the node; express Node at the Representation in layered graph neural networks; Represents a non-linear activation function; Representing relation type Attention weights; Represents a linear mapping matrix specific to a relation; Indicates the number of propagation layers; This indicates that the neighbor node u is at the 1st rank. Representation of layer -1; Represents user node With clothing nodes The recommendation score is calculated from the cosine similarity of the final embedding and used to generate the recommendation candidate set; This indicates the calculation of cosine similarity; and User nodes Clothing nodes go through Representation after layer propagation; As an exemplary embodiment, in step S4, graph neural network embedding learning is first performed to learn node representations on the fused graph, specifically including: graph structure modeling, multi-hop information propagation, node embedding update, and embedding quality evaluation; Then, recommendation candidates are generated, and recommendation scoring and filtering are performed based on graph structure paths. Specifically, this includes: path score calculation, multi-path aggregation, candidate ranking and filtering, and path diversity optimization. It should be noted that in step S4, regarding the fusion graph structure, this embodiment uses a graph neural network model to learn the embeddings of users and clothing. The main process is as follows: Graph structure modeling: The fusion graph is input into the graph neural network, and the initial features of the nodes are jointly composed of structural attributes and modal features; Information propagation mechanism: The neighbor aggregation mechanism is used to learn the contextual representation of nodes, supporting high-order path modeling (e.g., user → preferred style → spring → cotton and linen → a certain clothing); Node embedding generation: The representation of each user and clothing node after multi-hop graph propagation is obtained; Recommendation generation: The similarity between the representations of user and clothing nodes is calculated, and the highest-scoring clothing nodes are selected as recommendation candidates.
[0020] 1. Initial input for the node: Use the final embedding from the graph construction phase as the initial input; 2. Multi-hop neighbor aggregation:
[0021] in, Attention weights (or edge weights) for relation type r; : A linear mapping matrix specific to a relationship; : Number of propagation layers, supporting multi-hop path information modeling; 3. The final embedding is used for recommendations:
[0022] in, User nodes, clothing nodes; : The representation after propagation through layer L; S5: Explanation generation module: Based on the recommended candidates, perform recommended path backtracking to obtain path information, and convert the path information into natural language explanation; In one exemplary embodiment, step S5 specifically includes: performing a recommendation path backtracking based on the recommended candidates to obtain path information, and converting the path information into a natural language explanation, such as the formula:
[0023]
[0024] in, Indicates connected users With clothing A set of high-weighted paths; This represents an intermediate node in the path; This indicates that the path has been activated for recommendation; This represents the set of keywords extracted from the semantics of path nodes, for example... , , These represent the semantic relationships involved in the path, such as clothing style matching, seasonal adaptation, and material alignment. As an exemplary embodiment, in step S5, the recommended path backtracking refers to extracting the decision path that generates the recommendation, specifically including: key path identification, path importance assessment, decision element extraction, and path semantic annotation; converting the path information into natural language interpretation, specifically including: template matching, keyword filling, language optimization, and interpretation verification. It should be noted that step S5 is responsible for providing a structural explanation for the recommendation results, enhancing user trust. Specifically, this includes a path backtracking mechanism, which extracts the user-clothing path that generated the recommendation from the fusion graph; semantic tag extraction, which extracts key decision-making elements from path node attributes, such as style matching, scene adaptation, and material compatibility; and natural language generation, which converts the path and tags into standardized recommendation statements, such as: "This clothing is recommended because you prefer 'artistic style' and it is suitable for 'spring commuting'." This ensures that each recommendation can be traced back to the graph path, enhancing system interpretability and user experience. 1. Path backtracking:
[0025] in, The set of high-weighted paths connecting user u and clothing i; 2. Extract key semantics from path nodes:
[0026] 3. Generate natural language recommendations using templates.
[0027] In some embodiments, the above-described intelligent clothing recommendation system based on multimodal knowledge graphs can also be implemented in the following ways.
[0028] In this embodiment, the system includes a user knowledge graph construction module, a clothing knowledge graph construction module, a cross-graph fusion module, a graph neural network recommendation module, and a recommendation explanation generation module. These modules collaborate to achieve high-precision, interpretable clothing recommendation services through multimodal knowledge modeling and graph structure computation. The User Knowledge Graph Construction Module is used to build a structured, multimodal fusion user knowledge representation graph. Its main functions include: entity extraction, extracting user entities from multi-source data such as user behavior records (e.g., browsing, clicks, purchases), social network relationships, and user tags; attribute modeling, assigning attribute information to each user entity, including but not limited to: gender, age group, style preference, common clothing scenarios, historical preference categories, and physical features; multimodal information extraction, which, in scenarios with visual acquisition support, can further integrate image modalities such as user body shape and skin color, or text information, to construct a user multimodal knowledge graph; and user relationship edge building, establishing various relationship connections between users based on dimensions such as interest similarity, geographical proximity, and social connections, forming a user graph with scalability. The design purpose of this module is to provide a knowledge carrier that can structurally express users' personalized characteristics, providing a structural foundation for subsequent semantic alignment and graph reasoning. It should be noted that this embodiment considers more dynamic relationships between users, rather than just a single node representing a user's personal style. Traditional user knowledge graphs are built using simple user profiles, user descriptions, and behavioral data inputs. Here, the relationships between users are specifically considered. By connecting users with those who have similar interests, geographical proximity, or close social relationships, the user profile is strengthened. The Clothing Knowledge Graph Construction Module aims to build a structured graph covering a wide range of clothing entities, attributes, and relationships. It features: Clothing entity modeling, using stock keeping units (SKUs) or products as node units to extract basic information such as product name, product category, and style tags; attribute extraction, extracting clothing attributes from product text descriptions, image analysis, and expert tags, including style, season, applicable scenarios, material, pattern, color, price, and brand; and multimodal feature embedding, integrating product text descriptions and images to construct multimodal clothing nodes to achieve a multimodal clothing knowledge graph.
[0029] Among them, relationship construction involves establishing semantic edges between clothing entities, such as "similar style", "matchable", "functional substitution", "complementary fabric", etc.; and the extraction of implicit semantic sources, such as the style preferences implied by the brand. It should be noted that some clothing brands tend to exhibit stable style orientations, such as "minimalist," "sweet," or "outdoor." Therefore, in the absence of explicit style labels for clothing, other implicit semantic information can be used as supplementary basis for style attributes. This type of weak semantic association helps to improve the completeness and generalization ability of style modeling in the graph. It should be noted that the traditional construction of clothing knowledge graphs is consistent with that of user knowledge graphs, which also involves inputting images, descriptions, or related SKU information of clothing to generate an embedding; this embodiment also takes into account some implicit information of clothing, such as features with inherent style attributes like brands, as well as the relationships between clothing for subsequent matching recommendations; Cross-graph fusion module; primarily addresses the information gap between user graphs and clothing graphs, achieving semantic alignment of preferences and products; its functions include: Entity semantic alignment identifies semantically mappable groups of nodes in two graphs, such as “user preference style” and “clothing style”. Cross-relationship construction involves establishing heterogeneous graph connection edges such as "user-preference style-clothing style" and "user-applicable scenario-clothing application scenario" through methods such as user historical behavior co-occurrence and similar path mining. The graph fusion structure is generated to form a reasonable fusion hypergraph structure between user entities and clothing entities, and to construct multi-hop semantic paths; To ensure semantic consistency, co-occurrence statistics or graph embedding alignment are used to filter and weight relationships during the fusion process to avoid false associations. The design of this module enables user profiling to no longer rely solely on sparse label matching, but to achieve high-dimensional semantic connection between preferences, attributes, and products through structural path modeling. Graph Neural Network Recommendation Module; In terms of the fused graph structure, this module uses a graph neural network model to perform embedding learning on users and clothing, the main process of which is as follows: Graph structure modeling involves inputting a fused graph spectrum into a graph neural network, where the initial features of nodes are jointly composed of structural attributes and modal features. The information propagation mechanism uses a neighbor aggregation mechanism to learn node context representations, supporting high-order path modeling (such as user → preferred style → spring → cotton and linen → a certain clothing item). Node embedding generation: Obtain the representation of each user and clothing node after propagation in a multi-hop graph; Recommendation generation involves calculating the similarity between the user and the clothing node representation, and selecting the clothing nodes with the highest scores as recommendation candidates. The explanation generation module provides a structured explanation for the recommendation results, enhancing user trust. Specifically, it includes a path backtracking mechanism, which extracts the user-clothing path that generated the recommendation from the fusion graph; semantic tag extraction, which extracts key decision-making elements from path node attributes, such as style matching, scene adaptation, and material compatibility; and natural language generation, which converts the path and tags into standardized recommendation statements, such as: "This clothing is recommended because you prefer 'artistic style' and it's suitable for 'spring commuting'." This module ensures that every recommendation can be traced back to the graph path, enhancing system explainability and user experience.
[0030] In some embodiments, the above-described intelligent clothing recommendation method based on multimodal knowledge graphs can also be implemented in the following ways.
[0031] In this embodiment, as Figure 2 As shown, the intelligent clothing recommendation method based on multimodal knowledge graphs includes the following steps: Step 1. Constructing the user knowledge graph; Step 2. Construction of a clothing knowledge graph; Step 3. Cross-graph semantic alignment; Step 4. Heterogeneous graph fusion; Step 5. Graph neural network embedding learning; Step 6. Recommendation candidate generation; Step 7. Recommended path backtracking; Step 8. Explain the text generation; In step 1, user knowledge graph construction refers to extracting entities, attributes, and relationships from multi-source user data to build a user profile graph. Specifically, this includes: user entity recognition, user attribute extraction, multimodal feature embedding, and user relationship edge construction.
[0032] In step 2, the construction of the clothing knowledge graph refers to building a structured graph covering clothing entities, attributes, and relationships. Specifically, this includes: clothing entity modeling, clothing attribute extraction, clothing relationship construction, implicit semantic mining, and multimodal feature embedding.
[0033] In step 3, cross-graph semantic alignment refers to identifying semantically mappable node groups in two graphs. Specifically, this includes entity semantic matching, attribute semantic mapping, and relation semantic alignment.
[0034] In step 4, heterogeneous graph fusion refers to establishing connections between the user graph and the clothing graph. Specifically, this includes: constructing cross-graph connection edges, generating a fusion hypergraph, calculating relation weights, and verifying semantic consistency.
[0035] In step 5, graph neural network embedding learning refers to learning node representations on the fused graph. Specifically, this includes: graph structure modeling, multi-hop information propagation, node embedding updates, and embedding quality evaluation.
[0036] In step 6, candidate generation refers to scoring and filtering recommendations based on graph-structured paths. Specifically, this includes: path score calculation, multi-path aggregation, candidate ranking and filtering, and path diversity optimization.
[0037] In step 7, the recommendation path backtracking refers to extracting the decision path that generated the recommendation. Specifically, this includes: critical path identification, path importance assessment, decision element extraction, and path semantic annotation.
[0038] In step 8, explanatory text generation refers to converting the path information into natural language explanations. This specifically includes: template matching, keyword filling, language optimization, and explanation verification.
[0039] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A method for intelligent clothing recommendation based on multimodal knowledge graphs, characterized in that, Includes the following steps: S1: Extract entities, attributes, and relationships from multi-source user data to construct a user knowledge representation graph; S2: Construct a clothing knowledge graph based on clothing inventory units; S3: Based on the user knowledge representation graph and the clothing knowledge graph, perform cross-graph semantic alignment and heterogeneous graph fusion to obtain a fused graph; S4: Use a graph neural network model to learn the fused graph to obtain user and clothing nodes, and obtain recommendation candidates based on the user and clothing nodes; S5: Based on the recommended candidates, backtrack the recommended path to obtain path information, and convert the path information into natural language interpretation.
2. The intelligent clothing recommendation method based on multimodal knowledge graph according to claim 1, characterized in that, Step S2 specifically includes: constructing a clothing knowledge graph based on clothing inventory units, as shown in the formula: ， ， ， ， in, Represents clothing entity The final embedding is used to construct a clothing knowledge graph; For original embedding; The modality coding module is used to fuse the raw vector representations of image, text, and structural modalities. Image feature representation extracted from VGG Semantic feature representation extracted for BERT This represents the learnable structure embedding vector generated during TransE initialization; Indicates semantic edge message propagation; , , , These represent the current clothing node, the relationship type of the current edge, the neighboring entities connected to the current clothing node, and the current multimodal clothing knowledge graph, respectively. Indicates edge weight; Represents a linear mapping specific to a relation; This is the initial representation of adjacent clothing nodes; This indicates a weak label prior; Brand labels representing physical garments; This represents the brand-style mapping function.
3. The intelligent clothing recommendation method based on multimodal knowledge graph according to claim 1, characterized in that, Step S3 specifically includes: performing cross-graph semantic alignment and heterogeneous graph fusion based on the user knowledge representation graph and the clothing knowledge graph to obtain a fused graph, as shown in the formula: ， ， in, To integrate the maps; A collection of user entity nodes; A collection of clothing entity nodes; For users, clothing, and the set of edges across the graph; To create a unified diagram structure by integrating user diagrams and clothing diagrams; A semantically mappable group of nodes for user knowledge representation graphs and clothing knowledge graphs; Indicates cosine similarity; The representation vector for style class nodes in the user knowledge representation graph; For style-related nodes in the clothing knowledge graph The representation vector; To construct a similarity threshold for edges in heterogeneous graphs; , , These represent user style nodes, semantically aligned cross-graph edges, and clothing style nodes, respectively.
4. The intelligent clothing recommendation method based on multimodal knowledge graph according to claim 1, characterized in that, Step S4 specifically includes: learning the fused graph using a graph neural network model to obtain user and clothing nodes; and obtaining recommendation candidates based on the user and clothing nodes, as shown in the formula: in, Indicates the initial input of the node; express Node at the Representation in layered graph neural networks; Represents a nonlinear activation function; Representing relation type Attention weights; Represents a linear mapping matrix specific to a relation; Indicates the number of propagation layers; This indicates that the neighbor node u is at the 1st rank. Representation of layer -1; Represents user node With clothing nodes The recommendation score is calculated from the cosine similarity of the final embedding and used to generate the recommendation candidate set; This indicates the calculation of cosine similarity; and User nodes Clothing nodes go through The representation after layer propagation.
5. The intelligent clothing recommendation method based on multimodal knowledge graph according to claim 1, characterized in that, Step S5 specifically includes: performing a backtracking of the recommended path based on the recommended candidates to obtain path information, and converting the path information into a natural language explanation, as shown in the formula: in, Indicates connected users With clothing A set of high-weighted paths; This represents an intermediate node in the path; This indicates that the path has been activated for recommendation; This represents the set of keywords extracted from the semantics of path nodes, for example... , , These represent semantic relationships such as clothing style matching, seasonal adaptation, and material alignment involved in the path.
6. A clothing intelligent recommendation system based on multimodal knowledge graph, characterized in that, The system includes the following modules: The user knowledge graph construction module is configured to extract entities, attributes, and relationships from multi-source user data and construct a user knowledge representation graph. The apparel knowledge graph construction module is configured to: construct an apparel knowledge graph based on the unit of apparel inventory quantity; The cross-graph fusion module is configured to: perform cross-graph semantic alignment and heterogeneous graph fusion based on the user knowledge representation graph and the clothing knowledge graph to obtain a fused graph; The graph neural network recommendation module is configured to: learn the fused graph using a graph neural network model to obtain user and clothing nodes, and obtain recommendation candidates based on the user and clothing nodes; The explanation generation module is configured to: perform recommended path backtracking based on the recommended candidates to obtain path information, and convert the path information into natural language explanation.
7. The intelligent clothing recommendation system based on multimodal knowledge graph according to claim 6, characterized in that, The specific configuration of the clothing knowledge graph construction module is as follows: Based on the clothing inventory unit, construct a clothing knowledge graph, as shown in the formula: ， ， ， ， in, Represents clothing entity The final embedding is used to construct a clothing knowledge graph; For original embedding; The modality coding module is used to fuse the raw vector representations of image, text, and structural modalities. Image feature representation extracted from VGG Semantic feature representation extracted for BERT This represents the learnable structure embedding vector generated during TransE initialization; Indicates semantic edge message propagation; , , , These represent the current clothing node, the relationship type of the current edge, the neighboring entities connected to the current clothing node, and the current multimodal clothing knowledge graph, respectively. Indicates edge weight; Represents a linear mapping specific to a relation; This is the initial representation of adjacent clothing nodes; This indicates a weak label prior; Brand labels representing physical garments; This represents the brand-style mapping function.
8. The intelligent clothing recommendation system based on multimodal knowledge graph according to claim 6, characterized in that, The cross-graph fusion module is specifically configured as follows: based on the user knowledge representation graph and the clothing knowledge graph, cross-graph semantic alignment and heterogeneous graph fusion are performed to obtain a fused graph, as shown in the formula: ， ， in, To integrate the maps; A collection of user entity nodes; A collection of clothing entity nodes; For users, clothing, and the set of edges across the graph; To create a unified diagram structure by integrating user diagrams and clothing diagrams; A semantically mappable group of nodes for user knowledge representation graphs and clothing knowledge graphs; Indicates cosine similarity; The representation vector for style class nodes in the user knowledge representation graph; For style-related nodes in the clothing knowledge graph The representation vector; To construct a similarity threshold for edges in heterogeneous graphs; , , These represent user style nodes, semantically aligned cross-graph edges, and clothing style nodes, respectively.
9. The intelligent clothing recommendation system based on multimodal knowledge graph according to claim 6, characterized in that, The graph neural network recommendation module is specifically configured as follows: It uses a graph neural network model to learn the fused graph to obtain user and clothing nodes; based on the user and clothing nodes, it obtains recommendation candidates, as shown in the formula: in, Indicates the initial input of the node; express Node at the Representation in layered graph neural networks; Represents a nonlinear activation function; Representing relation type Attention weights; Represents a linear mapping matrix specific to a relation; Indicates the number of propagation layers; This indicates that the neighbor node u is at the 1st rank. Representation of layer -1; Represents user node With clothing nodes The recommendation score is calculated from the cosine similarity of the final embedding and used to generate the recommendation candidate set; This indicates the calculation of cosine similarity; and User nodes Clothing nodes go through The representation after layer propagation.
10. The intelligent clothing recommendation system based on multimodal knowledge graph according to claim 6, characterized in that, The specific configuration of the explanation generation module is as follows: Based on the recommended candidates, perform recommended path backtracking to obtain path information, and convert the path information into a natural language explanation, as shown in the formula: in, Indicates connected users With clothing A set of high-weighted paths; This represents an intermediate node in the path; This indicates that the path has been activated for recommendation; This represents the set of keywords extracted from the semantics of path nodes, for example... , , These represent semantic relationships such as clothing style matching, seasonal adaptation, and material alignment involved in the path.