Semantic-preserving serial attribute disentanglement personalization compatibility modeling method and system

Abstract

This invention belongs to the field of fashion item recommendation technology, and provides a semantically preserved serial attribute deconstruction personalized compatibility modeling method and system. For each fashion item, textual and visual features are extracted separately, and modality-consistent representations are learned from the textual and visual features. The Euclidean distance between the modality-consistent representations is minimized. For modality-specific representations, orthogonal constraints are introduced to suppress cross-modal redundancy. The attributes of the item are recursively deconstructed from the residuals of the modality representations. Simultaneously, while preserving semantic information, complete attributes are deconstructed, and under the premise of applying dual attribute independence constraints, a preference-aware hypergraph and a compatibility-aware hypergraph are constructed. The final predicted score is determined by combining the user's preference score for the item from the preference-aware hypergraph and the compatibility score of the item in the set from the compatibility-aware hypergraph. This solves the problem of attribute entanglement while avoiding attribute redundancy, and achieves the goal of personalized recommendation.

Description

Technical Field

[0001] This invention belongs to the field of fashion item recommendation technology, and in particular relates to a semantically preserved serial attribute deconstruction personalized compatibility modeling method and system. Background Technology

[0002] Compatibility modeling methods for fashion items such as tops, bottoms, shoes, and accessories are crucial to the development of the fashion industry. The purpose of compatibility modeling is to evaluate the compatibility score of a set of complementary fashion items, thereby providing people with reasonable dressing suggestions.

[0003] Current compatibility modeling methods mainly focus on decomposing the multimodal information of fashion items into a set of independent and interpretable attributes. The attributes decomposed from different modalities often overlap, leading to cross-modal attribute redundancy. Even when modality-consistent representations obtained by granular decomposition to alleviate attribute overlap between different modalities are obtained, attribute entanglement may still exist within modality-specific representations. Furthermore, most current compatibility modeling methods focus on characterizing the compatibility relationships within a set, failing to incorporate individual aesthetic perceptions, resulting in a lack of personalization in recommendation results. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a semantically preserved serial attribute deconstruction personalized compatibility modeling method and system. This invention employs coarse-grained and fine-grained deconstruction to deconstruct complete attributes while preserving semantic information. Under the constraint of dual attribute independence, it constructs a preference-aware hypergraph and a compatibility-aware hypergraph. The final predicted score is determined by combining the user's preference score for an individual item from the preference-aware hypergraph with the compatibility score of the individual item within a set from the compatibility-aware hypergraph. This approach avoids attribute redundancy, solves the problem of attribute entanglement, and achieves the goal of personalized recommendation.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: In a first aspect, the present invention provides a semantically preserved serial attribute deconstruction personalized compatibility modeling method, comprising: Obtain visual images and textual information about fashion items; Based on visual images and text information, textual and visual features of each fashion item are extracted through coarse-grained decomposition, and orthogonal constraints are introduced to suppress cross-modal redundancy. On the basis of coarse-grained decomposition, the attributes of the item are recursively decomposed from the residuals of the modal representation through fine-grained decomposition. Based on the deconstruction results, complete attributes are deconstructed while preserving semantic information, and dual attribute independence constraints are applied. A preference-aware hypergraph is constructed with the user as the hyperedge, and a compatibility-aware hypergraph is constructed with the set as the hyperedge. Based on the preference-aware hypergraph and the compatibility-aware hypergraph, the user's preference score for an individual item and the compatibility score of the individual item in the set are determined respectively; the final predicted score is determined based on the preference score and the compatibility score.

[0006] Furthermore, the loss function of the coarse-grained decomposition for: ; The loss function of fine-grained dissociation for: ; in, and To learn modality-consistent representations from textual and visual features; and Modality-specific representations for textual and visual modalities; For the project No. Representation of a dissociative property; For the project Residual properties in modally consistent or modally specific representations; The number of attributes that are dissociated from modally consistent or modally specific representations; The number of attributes dissociated from the modal consistent representation; The number of attributes to be dissected from a specific representation of the text; The number of attributes dissociated from a visually specific representation.

[0007] Furthermore, the construction of the preference-aware hypergraph includes: [details of the preference-aware hypergraph construction]. Define the correlation matrix , where each element Indicates user With single item Interaction and the first k The degree of correlation of each attribute: ; in, and They represent the first time. Layered single items With users Embedded representation; in preference-aware hypergraph The formalized node-hyperedge-node information aggregation process is used in the first... Layer update user embedding With single item embedding : ; in, It is a non-linear activation function. This represents the weights used when aggregating information from the hyperedge to the node; average pooling is performed on the embeddings obtained from different layers to generate the first... k The final representation of users and individual products under each attribute: ; in, This indicates the number of convolutional layers in the hypergraph.

[0008] Furthermore, the dissociation includes: introducing a method to maximize the reduction in coding rate, by measuring the difference in coding rate between the overall single-item embedding and the corresponding embedding of each attribute, to promote a clearer separation structure of different attributes in the attribute space; Given the embedding matrix of all individual items, determine the overall coding rate; The overall compactness of an attribute is defined as the sum of the encoding rates of all attribute embedding matrices; By maximizing the difference between the overall coding rate and the overall compactness, different attributes exhibit stronger separability in the attribute space; the attribute constraint loss between users and sets is determined separately, and the loss of individual items is integrated to obtain the final attribute loss.

[0009] Furthermore, given the embedding matrix of all individual items... Overall coding rate for: ; in, To tolerate hyperparameters; Indicates the number of dissociative properties; For identity matrix; to make the overall attributes more compact. Defined as an all-attribute embedding matrix Sum of coding rates: ; By maximizing and The difference between them makes different attributes exhibit stronger separability in the attribute space: .

[0010] Furthermore, modality-consistent representations are learned from textual and visual features to minimize the Euclidean distance between textual and visual modal representations; statistical correlation is reduced by minimizing mutual information between different items under the same attribute, making the representations within each attribute more dispersed and information-rich; the attribute independence loss between users and sets is defined separately and combined with the loss of individual items to obtain the final attribute loss.

[0011] Furthermore, the loss of the individual item is: ; in, Representing variables The joint distribution; These represent their marginal distributions.

[0012] Furthermore, the determination of the final predicted score includes: by integrating user... With single item Personalized preference scores between them, and sets With single item The compatibility score between them is used to obtain the final predicted score. : ; in, For hyperparameters; For the set In the Representation on each attribute; Indicates user For the first Preference for each attribute; A Bayesian personalized ranking loss is employed to ensure the observed triples. The predicted score was higher than that of the unobserved triplet. The predicted score, the loss function is defined as: ; The total loss for independence within and between attributes is defined as: ; The recommendation loss, dissociation loss, and pairwise constraint loss are integrated to obtain the overall optimization objective.

[0013] Secondly, the present invention also provides a semantically preserved serial attribute deconstruction personalized compatibility modeling system, comprising: The data acquisition module is configured to acquire visual images and text information of fashion items; The deconstruction module is configured to: extract the text features and visual features of each fashion item through coarse-grained deconstruction based on visual images and text information, and introduce orthogonal constraints to suppress cross-modal redundancy; and recursively deconstruct the attributes of the item from the residuals of the modal representation through fine-grained deconstruction based on the coarse-grained deconstruction. The hypergraph building module is configured to: based on the dissociation results, dissociate complete attributes while preserving semantic information, apply dual attribute independence constraints, and construct a preference-aware hypergraph with users as hyperedges; and construct a compatibility-aware hypergraph with sets as hyperedges. The compatibility score determination module is configured to: determine the user's preference score for an individual item and the compatibility score of the individual item in the set based on the preference-aware hypermap and the compatibility-aware hypermap, respectively; and determine the final predicted score based on the preference score and the compatibility score.

[0014] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in the first aspect.

[0015] Fourthly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in the first aspect.

[0016] Fifthly, the present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs coarse-grained deconstruction, extracting textual and visual features for each fashion item, and learning a modal-consistent representation from these features. It minimizes the Euclidean distance between the modal-consistent representation and the target. For modality-specific representations, orthogonal constraints are introduced to suppress cross-modal redundancy. Building upon this coarse-grained deconstruction, fine-grained deconstruction recursively deconstructs item attributes from the residuals of the modal representations. Simultaneously, it deconstructs complete attributes while preserving semantic information. Under the constraint of dual attribute independence, it constructs a preference-aware hypergraph and a compatibility-aware hypergraph. The final predicted score is determined by combining the user's preference score for the item from the preference-aware hypergraph and the item's compatibility score within a set from the compatibility-aware hypergraph. This approach avoids attribute redundancy, solves the problem of attribute entanglement, and achieves personalized recommendation. Attached Figure Description

[0018] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0019] Figure 1 This is a schematic diagram of the model framework of Embodiment 1 of the present invention; Figure 2 This is an example of single-item multimodal information in Embodiment 1 of the present invention; Figure 3 The attribute embeddings learned by the complete model of Embodiment 1 of the present invention; Figure 4 The attribute embedding learned for the ablation attribute independence variant of Embodiment 1 of the present invention; Figure 5 The learned attributes of the complete model in Embodiment 1 of the present invention are embedded and projected into a two-dimensional space; Figure 6 The attribute embedding learned for the ablation attribute independence of Embodiment 1 of the present invention is projected into a two-dimensional space. Detailed Implementation

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0022] Example 1: This embodiment provides a semantically preserved serial attribute deconstruction personalized compatibility modeling method, such as... Figure 1 As shown, the semantically preserved serial attribute deconstruction personalized compatibility modeling framework includes three key modules: a serial multimodal deconstruction module, a preference-aware and compatibility-aware hypergraph neural network module, and a dual attribute independence constraint module. The language model is BERT (Bidirectional Encoder Representations from Transformers), and the visual model is ViT (Vision Transformer). The modeling methods include: S1. Problem Definition: Assumption Indicates including individual users The set. For any user Its interaction history is defined as including Each fashion item; the interaction history and user interaction sequence information are collected with user consent and in compliance with relevant information security regulations. Based on the interaction sequences of all users. Constructing a preference-aware hypergraph View fashion items as key points and regard users as hyper-edge This is to characterize user preference relationships. Furthermore, it allows... Indicates including A collection of sets, where each set... Depend on It consists of several complementary individual items. Accordingly, a compatibility-aware hypergraph is constructed. Treating the set as a super-edge To learn the higher-level compatibility relationships within the set. For any fashion item. Its text description and visual image are defined as follows: and Based on this, a personalized compatibility modeling framework based on dissociation is designed. This method deconstructs complete attributes while preserving semantic information and imposes dual attribute independence constraints. The final predicted score... By combining from the preference-aware hypergraph Users' preferences for individual items Preference score With compatibility-aware hypergraph single items In the set Compatibility score It is obtained in the following form: ; S2, Serial multimodal dissociation: Existing dissociation-based compatibility modeling methods primarily focus on deconstructing the multimodal information (such as visual and textual) of fashion items into a set of independent and interpretable attributes. However, the attributes deconstructed from different modalities often overlap to some extent. For example... Figure 2 As shown, both the visual image and the text description "floral pattern" convey the pattern attribute information "floral," leading to cross-modal attribute redundancy. To alleviate this problem, existing research has taken a coarse-grained approach, decomposing multimodal features into two types of representations: modality-consistent and modality-specific. Meanwhile, these methods typically employ parallel multilayer perceptrons or data partitioning strategies to decompose individual item attributes, but this process may introduce semantic information loss. To address these limitations, this embodiment proposes a serial multimodal decomposition framework. First, it suppresses cross-modal attribute overlap through coarse-grained decomposition, and then achieves fine-grained attribute decomposition without sacrificing the original semantic information.

[0023] S2.1, Coarse-grained dissociation: To deconstruct core semantic components and mitigate cross-modal attribute overlap in the initial stage, coarse-grained deconstruction is first introduced to learn modality-consistent (shared across all modalities) and modality-specific (unique to each modality) representations. For each fashion item... The textual and visual features are extracted separately and denoted as follows: and From the perspective of text features With visual features Learning modal consistency representation and Accordingly, the modality-specific representations of textual and visual modalities are as follows: and It can be obtained from the following formula: ; ; in, This represents the learnable parameters.

[0024] To align modal consistency representations, minimize modal consistency representations. and The Euclidean distance between them. For modality-specific representations. and Orthogonal constraints are introduced to suppress cross-modal redundancy. Therefore, the coarse-grained decoupling loss function... It can be defined as: ; S2.1 Fine-grained dissociation: Coarse-grained deconstruction can effectively alleviate attribute overlap between different modalities. However, attribute entanglement may still exist between the resulting modality-consistent representation and the modality-specific representation. To achieve more complete deconstruction and obtain finer and more independent attribute representations, this embodiment further proposes a fine-grained deconstruction method based on residual refinement. Unlike parallel deconstruction strategies that may lead to information loss, this embodiment recursively deconstructs the attributes of individual items from the residuals of the modality representation, thereby better preserving the original semantic information. Specifically, using modality-consistent representation... Let's take an example to illustrate the attribute decomposition process based on residual refinement. In the first... k In the secondary refinement step, a predefined multilayer perceptron is used to refine the residuals. Dissociation properties And calculate the residual for the next step. : ; ; in, To simplify the expression, we define... ,in, This indicates the number of dissociation attributes. After fine-grained dissociation, the final modal consistency representation is obtained. Similarly, for modality-specific representations and Deconstruction is performed to obtain the specific representations of the deconstructed text. and visual specific representation .

[0025] To further ensure the independence between fine-grained dissociation attributes, orthogonal constraints are applied to the dissociated representations within each mode. Correspondingly, the fine-grained dissociation loss function... Defined as: ; in, The number of attributes that are dissociated from modally consistent or modally specific representations; The number of attributes dissociated from the modal consistent representation; The number of attributes to be dissected from a specific representation of the text; The number of attributes dissociated from a visually specific representation.

[0026] S2.3, Preference-aware and Compatibility-aware Hypergraph Neural Networks: Based on the advantage of hypergraphs in modeling higher-order associations beyond binary relations, previous studies have introduced hypergraph networks into the task of suit compatibility modeling. However, most current methods focus on characterizing the compatibility relationships within suits, failing to incorporate individual aesthetic perception, thus resulting in a lack of personalization in recommendation results. To remedy this deficiency, this embodiment constructs a preference-aware hypergraph and a compatibility-aware hypergraph to jointly model attribute-level user preferences and suit compatibility while preserving complete semantic information. Specifically, the modality-consistent representation obtained from deconstruction is used... Taking [example] as an example, we will further illustrate the hypergraph learning process.

[0027] S2.3.1, Preference-aware and Compatibility-aware Hypergraphs: To characterize diverse user preferences, a set of preference-aware hypergraphs is constructed. Each hypergraph corresponds to a dissociated attribute. The interaction between a user and an item is essentially driven by the difference in the intensity of the user's preference for different attributes. Therefore, to accurately model user preferences, the key lies in dynamically quantifying the contribution of each attribute in the interaction. Based on this, this embodiment uses a preference-aware hypergraph. Define its correlation matrix , where each element Indicates user With single item Interaction and the first k The degree of correlation between attributes.

[0028] ; in, and They represent the first time. Layered single items With users Embedded representation; user In the k The initial preference embedding on each attribute is defined as follows: Furthermore, in The formalized node-hyperedge-node information aggregation process is used in the first... Layer update user embedding With single item embedding The details are as follows: ; in, This is a non-linear activation function. It is important to note that... This represents the weights used when aggregating information from a hyperedge to a node. Since different semantic preferences are encoded in different hyperedges, these weights are used to distinguish and convey the corresponding preference information.

[0029] Furthermore, this embodiment performs average pooling on the embeddings obtained from different layers to generate the first... k The final representation of users and individual products under each attribute is as follows: ; in, This indicates the number of convolutional layers in the hypergraph.

[0030] Finally, the final embedding matrices for all users and individual products across each attribute dimension are obtained, as follows: and .

[0031] To characterize the complex compatibility relationships within the kit, this embodiment further constructs a compatibility-aware hypergraph. To model the overall coordination of the set at the attribute level, thereby obtaining the set embedding matrix. and the corresponding single-item embedding matrix .

[0032] S2.3.2, Cross-view fusion: Since the preference-aware hypergraph and the compatibility-aware hypergraph capture item information from different perspectives, this embodiment introduces a cross-attention mechanism for item embedding. and The fusion will be carried out as follows: ; in, Let it be the projection parameter matrix.

[0033] ; Ultimately, the merged single-item embedding matrix is ​​obtained. .

[0034] Following modal consistent representation The hypergraph learning process is applied in this embodiment to modality-specific representations. and This allows for the separate embedding of text-specific user, package, and individual item information. , , And visually specific user groups, sets and individual items embedded. , , .

[0035] S2.4 Dual Attribute Independence Constraint: Existing dissociation-based compatibility modeling methods primarily focus on constraining the independence between different attributes of individual items, while neglecting the independence between different categories of items under the same attribute. While imposing independence constraints only between attributes can effectively prevent mutual interference between the semantics of different attributes, the lack of intra-attribute constraints leads to insufficient discriminability between different categories of items within the same attribute, thus weakening the effectiveness of fine-grained user preference modeling and package compatibility modeling. Therefore, this embodiment proposes a pairwise independence constraint that simultaneously constrains the independence between and within attributes from both spatial and statistical perspectives.

[0036] S2.4.1 Independence constraint between attributes: Existing methods for attribute independence typically employ orthogonal regularization and distance correlation. However, these methods usually only impose pairwise constraints on attribute representations, making it difficult to impose holistic restrictions on attribute representations from a global perspective. Given that different attributes should encode distinct latent semantics, it is desirable to maximize the separability of different attribute representations within the same item in the attribute space, thereby reducing semantic interference and enhancing attribute independence. To this end, this embodiment introduces a method to maximize encoding rate reduction. By measuring the difference in encoding rate between the overall item embedding and the corresponding embeddings of each attribute, it promotes a clearer separation structure for different attributes in the attribute space. Formally, given the embedding matrices of all items... to its overall coding rate The calculation is as follows: ; in, To tolerate hyperparameters; This indicates the number of dissociated properties. Then, the overall compactness of the properties is considered. Defined as an all-attribute embedding matrix The sum of coding rates is in the following form: ; By maximizing and The difference between them makes different attributes exhibit stronger separability in the attribute space, as follows: ; Similarly, calculate the constraint loss between attributes for users and sets separately. and And integrate the losses of individual items Thus, the final inter-attribute loss is obtained: ; S2.4.2, Intra-attribute Independence Constraint: While attribute constraints can guarantee independence between different attributes at the representation space level, they typically do not explicitly restrict the statistical dependencies between different categories of items within each attribute. Therefore, under the same attribute, the embeddings of different categories of items may tend to be overly similar and exhibit excessive clustering, thus limiting the discriminability between attributes. To address this issue, this embodiment introduces attribute independence constraints from a statistical perspective. By minimizing the mutual information between different categories of items under the same attribute, its statistical correlation is reduced. This constraint promotes more dispersed and information-rich representations within each attribute, thereby improving the model's expressive power and fine-grained modeling performance.

[0037] ; in, Representing variables The joint distribution; These represent their marginal distributions.

[0038] Similarly, the independence loss between user and set attributes is defined as follows: This is compared with the loss of individual items. By combining these attributes, we obtain the final inter-attribute loss: ; S2.5 Model Optimization: By integrating users With single item Personalized preference scores between them, and sets With single item The compatibility score between them is used to obtain the final predicted score. Its form is as follows: ; in, It is a hyperparameter, and .

[0039] Optionally, a Bayesian personalized ranking loss can be used to ensure the observed triples. The predicted score was higher than that of the unobserved triplet. The predicted score. Its loss function is defined as: ; in, It is a logarithmic sigmoid activation function. It is a sigmoid activation function; For users and the given interactive set Under the condition of positive sample single item The predicted score; For users and the given interactive set Under the condition of negative sample single item The predicted score.

[0040] The total loss for independence within and between attributes is defined as: ; in, To balance the hyperparameters; Finally, the recommendation loss, dissociation loss, and pairwise constraint loss are integrated to obtain the overall optimization objective: ; in, Indicates model parameters; , and These are hyperparameters used to control various loss rates, and .

[0041] The experimental results and performance comparisons are shown in Table 1: Table 1 Performance Comparison

[0042] Among them, the Area Under the Curve (AUC) is used to evaluate the model's global ranking ability, the Mean Reciprocal Rank (MRR) focuses on the ranking position of the first relevant item, the Normalized Discounted Cumulative Gain (NDCG) measures the overall ranking quality by considering both relevance and position, and the Hit Rate (HR) is used to evaluate whether a relevant item appears in the top K recommendations. These metrics comprehensively evaluate recommendation performance from multiple perspectives, including global ranking, top-ranking accuracy, overall ranking quality, and hit rate. Polyvore-U is a user-partitioned Polyvore dataset; IQON3000 is clothing matching data based on the IQON platform; and Alibaba-iFashion is the iFashion fashion recommendation dataset.

[0043] To verify the effectiveness of cross-attribute and intra-attribute independence in mitigating distribution bias, such as Figure 3 and Figure 4 As shown, this embodiment visualizes the attribute embeddings learned by the complete model and its ablation attribute-independent variants, as well as the attribute embeddings learned by the complete model and its ablation attribute-independent variants. First, for the comparison between the complete model and the ablation attribute-independent variants, principal component analysis is used to reduce the high-dimensional embeddings to a three-dimensional space for visualization. The attribute distribution is displayed in the form of a spherical plot, where each point represents an attribute embedding, and different colors correspond to different attributes. By comparing the ablation attribute-independent variants and the complete model, it can be observed that the embeddings of different attributes in the complete model exhibit a more obvious clustering trend, and the attribute distribution is clearer compared to the ablation attribute-independent variants. This result indicates that cross-attribute independence helps to achieve a more explicit attribute distribution in the representation space, thereby ensuring the independence of cross-attribute representations.

[0044] Subsequently, as Figure 5 and Figure 6 As shown, to compare the complete model and the intra-attribute independence variants, t-distributed random neighborhood embeddings are used to project the high-dimensional embeddings onto a two-dimensional space for visualization. To simplify the analysis, one attribute is randomly selected to evaluate the effect of intra-attribute independence constraints. Comparing the intra-attribute independence variants and the complete model reveals that the embedding distribution of the intra-attribute independence variants exhibits a more disordered structure, while the embeddings of the complete model are more dispersed in space. This result indicates that intra-attribute independence constraints can effectively alleviate the excessive clustering of representations within the same attribute, thereby improving the distinguishability of representations within the same attribute.

[0045] Example 2: This embodiment provides a semantically preserved serial attribute deconstruction personalized compatibility modeling system, including: The data acquisition module is configured to acquire visual images and text information of fashion items; The deconstruction module is configured to: extract the text features and visual features of each fashion item through coarse-grained deconstruction based on visual images and text information, and introduce orthogonal constraints to suppress cross-modal redundancy; and recursively deconstruct the attributes of the item from the residuals of the modal representation through fine-grained deconstruction based on the coarse-grained deconstruction. The hypergraph building module is configured to: based on the dissociation results, dissociate complete attributes while preserving semantic information, apply dual attribute independence constraints, and construct a preference-aware hypergraph with users as hyperedges; and construct a compatibility-aware hypergraph with sets as hyperedges. The compatibility score determination module is configured to: determine the user's preference score for an individual item and the compatibility score of the individual item in the set based on the preference-aware hypermap and the compatibility-aware hypermap, respectively; and determine the final predicted score based on the preference score and the compatibility score.

[0046] The working method of the system is the same as the semantically preserved serial attribute deconstruction personalized compatibility modeling method in Embodiment 1, and will not be repeated here.

[0047] Example 3: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in Embodiment 1.

[0048] Example 4: This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the program, it implements the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in Embodiment 1.

[0049] Example 5: This embodiment provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the steps of the semantically preserved serial attribute deconstruction personalized compatibility modeling method described in Embodiment 1.

[0050] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A semantically preserved serial attribute deconstruction personalized compatibility modeling method, characterized in that, include: Obtain visual images and textual information about fashion items; Based on visual images and text information, textual and visual features of each fashion item are extracted through coarse-grained decomposition, and orthogonal constraints are introduced to suppress cross-modal redundancy. On the basis of coarse-grained decomposition, the attributes of the item are recursively decomposed from the residuals of the modal representation through fine-grained decomposition. Based on the deconstruction results, complete attributes are deconstructed while preserving semantic information, and dual attribute independence constraints are applied to construct a preference-aware hypergraph with the user as the hyperedge. Construct a compatibility-aware hypergraph using sets as hyperedges; Based on the preference-aware hypergraph and the compatibility-aware hypergraph, determine the user's preference score for an individual item and the compatibility score of the individual item in the set, respectively. The final predicted score is determined based on the preference score and compatibility score; The construction of the preference-aware hypergraph includes: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Define the correlation matrix , where each element Indicates user With single item Interaction and the first k The degree of correlation of each attribute: ； in, and They represent the first time. Layered single items With users Embedded representation; in preference-aware hypergraph The formalized node-hyperedge-node information aggregation process is used in the first... Layer update user embedding With single item embedding : ； in, It is a non-linear activation function. This represents the weights used when aggregating information from the hyperedge to the node; average pooling is performed on the embeddings obtained from different layers to generate the first... k The final representation of users and individual products under each attribute: ； wherein, denotes the number of hypergraph convolution layers.

2. The method for semantic-preserving, serial-attribute dissociation, personalization, compatibility modeling of claim 1, wherein, The loss function for the coarse-grained disentanglement is: ； The loss function of the fine-grained disambiguation is: ； in, and To learn modality-consistent representations from textual and visual features; and Modality-specific representations for textual and visual modalities; For the project No. Representation of a dissociative property; For the project Residual properties in modally consistent or modally specific representations; The number of attributes that are dissociated from modally consistent or modally specific representations; The number of attributes dissociated from the modal consistent representation; The number of attributes to be dissected from a specific representation of the text; The number of attributes dissociated from a visually specific representation.

3. The semantically preserved serial attribute deconstruction personalized compatibility modeling method as described in claim 1, characterized in that, The dissociation includes: introducing a method to maximize the reduction of coding rate, by measuring the difference in coding rate between the overall single-item embedding and the corresponding embedding of each attribute, to promote a clearer separation structure of different attributes in the attribute space; Given the embedding matrix of all individual items, determine the overall coding rate; The overall compactness of an attribute is defined as the sum of the encoding rates of all attribute embedding matrices; By maximizing the difference between the overall coding rate and the overall compactness, different attributes exhibit stronger separability in the attribute space; the attribute constraint loss between users and sets is determined separately, and the loss of individual items is integrated to obtain the final attribute loss.

4. The method for semantic-preserving, serial-attribute dissociation, personalization, compatibility modeling of claim 3, wherein, Given the embedding matrix for all single items , the overall code rate is: ； where, is the tolerance hyperparameter; denotes the number of dissociation attributes; is the identity matrix; the overall compactness of the attributes is defined as the matrix of all attribute embeddings sum of encoding rates: ； By maximizing the difference between with encouraging different attributes to exhibit stronger separability in the attribute space: 。 5. The method for semantic-preserving, serial-attribute dissociation, personalization, and compatibility modeling of claim 3, wherein, Modality-consistent representations are learned from textual and visual features, minimizing the Euclidean distance between textual and visual modal representations. Statistical correlation is reduced by minimizing mutual information between different items under the same attribute, making the representations within each attribute more dispersed and information-rich. The attribute independence loss between users and sets is defined separately and combined with the loss of individual items to obtain the final attribute loss.

6. The method, as recited in claim 1, wherein the semantic-preserving, serial-attribute dissociation personalization compatibility modeling method is characterized by, The loss of the individual item is: ； wherein denotes the joint distribution of the variables ; and denotes their marginal distributions, respectively.

7. The semantically preserved serial attribute deconstruction personalized compatibility modeling method as described in claim 1, characterized in that, The determination of the final prediction score comprises integrating a personalized preference score between the user and the individual product and a compatibility score between the set and the individual product to obtain the final prediction score : ； wherein, is a hyperparameter; is a suit in the first attribute; represents a user preference for the first attribute; A Bayesian personalized ranking loss is employed to ensure that the predicted scores of observed triplets are higher than the predicted scores of unobserved triplets The loss function is defined as: ； The total loss for independence within and between attributes is defined as: ； By integrating the Bayesian personalized ranking loss, dissociation loss, and dual-attribute independence constraint, the overall optimization objective is obtained: ； wherein, denote model parameters; , and are hyperparameters for controlling the strength of each loss.

8. A semantic-preserving, serial attribute dissociation personalization compatibility modeling system, characterized by, include: The data acquisition module is configured to acquire visual images and text information of fashion items; The deconstruction module is configured to: extract the text features and visual features of each fashion item through coarse-grained deconstruction based on visual images and text information, and introduce orthogonal constraints to suppress cross-modal redundancy; and recursively deconstruct the attributes of the item from the residuals of the modal representation through fine-grained deconstruction based on the coarse-grained deconstruction. The hypergraph building module is configured to: based on the dissociation results, dissociate complete attributes while preserving semantic information, apply dual attribute independence constraints, and build a preference-aware hypergraph with the user as the hyperedge; Construct a compatibility-aware hypergraph using sets as hyperedges; The compatibility score determination module is configured to determine the user's preference score for an individual item and the compatibility score of the individual item in the set, based on the preference-aware hypermap and the compatibility-aware hypermap, respectively. The final predicted score is determined based on the preference score and compatibility score; The construction of the preference-aware hypergraph includes: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Define the correlation matrix , where each element Indicates user With single item Interaction and the first k The degree of correlation of each attribute: ； in, and They represent the first time. Layered single items With users Embedded representation; in preference-aware hypergraph The formalized node-hyperedge-node information aggregation process is used in the first... Layer update user embedding With single item embedding : ； wherein, is a non-linear activation function, denotes the weights used when aggregating information from the hyper-edges to the nodes; average pooling is performed on the embeddings obtained for different layers to generate the final representation of the user and the single item under the k th attribute: ； wherein, denotes the number of hypergraph convolution layers.

9. A computer program product, characterised in that, The computer program product comprises a computer program which, when executed by a processor, implements the steps of the semantic-preserving serial attribute dissociation personalization compatibility modeling method according to any one of claims 1-7.

Images