Personalized visual design recommendation system based on big data analysis

Abstract

The invention relates to the technical field of design recommendation, in particular to a personalized visual design recommendation system based on big data analysis, and the system comprises a bipartite graph modeling unit which is used for constructing a user-design bipartite graph according to historical interaction data between a user and a visual design product, and obtaining a visual design product image; and the embedded propagation unit is used for carrying out first-order to high-order embedded propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network so as to obtain a first embedded feature vector of the user and a second embedded feature vector of the visual design product. According to the method, the personalized user-design bipartite graph can be accurately constructed for each user by combining the bipartite graph modeling unit with historical interaction data between the user and the visual design product, so that user preferences are effectively captured.

Description

Technical Field

[0001] The present invention relates to the technical field of design recommendation, and particularly relates to a personalized visual design recommendation system based on big data analysis. Background Art

[0002] Traditional systems mainly rely on users' historical behaviors (such as clicks, purchases, etc.) for recommendation, ignoring users' aesthetic and personalized needs for design works. Therefore, they may only recommend similar items based on users' historical behaviors; and traditional systems often only rely on users' historical behaviors, ignoring the aesthetic value of design products themselves, resulting in sometimes lack of visual attractiveness in the recommendation results. Users may receive some designs related to their interests, but they may not meet users' aesthetic needs; and traditional systems usually rely on relatively simple feature extraction methods and cannot deeply explore users' multi-dimensional needs and the complex features of design products. For example, some traditional systems may only rely on basic tags or categories of designs for recommendation and cannot capture users' interests in multiple aspects such as design styles and image aesthetics; and traditional systems often use static user feature models and cannot flexibly adapt to changes in users' preferences. Over time, users' interests may change, but traditional methods may not be able to reflect this change in a timely manner. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a personalized visual design recommendation system based on big data analysis.

[0004] The technical solution adopted to solve the above technical problem is: A personalized visual design recommendation system based on big data analysis, comprising:

[0005] A bipartite graph modeling unit, which is used to construct a user-design bipartite graph according to the historical interaction data between users and visual design products, and obtain visual design product images;

[0006] An embedding propagation unit, which is used to perform first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network to obtain a first embedding feature vector of the user and a second embedding feature vector of the visual design product;

[0007] A style classification unit, which is used to perform style classification on the visual design product images according to a style classifier to obtain style labels of the visual design products, and extract the feature vector of the fully connected layer of the style classifier as the style feature vector of the visual design products;

[0008] An aesthetic evaluation unit, which is used to perform aesthetic evaluation on the visual design product image according to an aesthetic evaluation model to obtain an aesthetic evaluation label of the visual design product, and extract a feature vector of an intermediate layer of the aesthetic evaluation model as an aesthetic feature vector of the visual design product.

[0009] Preferably, the system further includes:

[0010] A vector splicing unit, which is used to perform text encoding on the style label and the aesthetic evaluation label to obtain a first label vector of the style label and a second label vector of the aesthetic evaluation label, and splice the second embedded feature vector, the style feature vector, the aesthetic feature vector, the first label vector and the second label vector to obtain a fusion vector of the visual design product;

[0011] A design recommendation unit, which is used to calculate a preference score of the user for the visual design product according to the first embedded feature vector of the user and the fusion vector of the visual design product, and perform personalized visual design recommendation for the user according to the preference score.

[0012] Preferably, the user-design bipartite graph includes a user node set, a design node set and an edge set, where the user node corresponds to the user, the design node corresponds to the visual design product, and the edge corresponds to the historical interaction data.

[0013] Preferably, performing first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network to obtain a first embedded feature vector of the user and a second embedded feature vector of the visual design product includes:

[0014] Embedding the ID identifications of the user nodes and the design nodes according to the user-design bipartite graph to obtain a first initial embedding vector of the user and a second initial embedding vector of the visual design product;

[0015] Performing message encoding from the design nodes to the user nodes in the user-design bipartite graph to obtain a first first-order message, where the expression of the first first-order message is as follows:

[0016]

[0017] where m u←i represents the first first-order message from design node i to user node u, N u represents the set of design nodes interacted by user node u, N iDenote the set of user nodes interacting with design node i, W1 represents the first trainable weight matrix, W2 represents the second trainable weight matrix, e i Denote the second initial embedding vector of the design node, e u Denote the first initial embedding vector of the user node;

[0018] Calculate the self-connection message of the user node in the user-design bipartite graph, perform a weighted sum of the first-order messages of the set of design nodes interacting with the user node to obtain a first weighted sum, and perform an embedding representation on the self-connection message and the first weighted sum according to a non-linear activation function to obtain the first-order embedding feature vector of the user node, where the expression of the first-order embedding feature vector of the user node is as follows:

[0019]

[0020] Where, Denote the first-order embedding feature vector of the user node, m u←u Denote the self-connection message of the user node, LeakReLu represents the non-linear activation function;

[0021] Determine the first high-order message from the design node to the user node according to the embedding feature vector of the previous order of the user node, where the expression of the first high-order message of the user node is as follows:

[0022]

[0023] Where, Denote the l-th order message from design node i to user node u, Denote the first trainable weight matrix of the l-th order, Denote the second trainable weight matrix of the l-th order, Denote the (l - 1)-th order embedding feature vector of user node u, Denote the (l - 1)-th order embedding feature vector of design node i;

[0024] Determine the high-order embedding feature vector of the user node according to the first high-order message of the user node, where the expression of the high-order embedding feature vector of the user node is as follows:

[0025]

[0026] Where, Denote the l-th order embedding feature vector of the user node.

[0027] Preferably, performing first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network to obtain a first embedding feature vector of the user and a second embedding feature vector of the visual design product, further comprising:

[0028] Encoding messages from the user nodes to the design nodes in the user-design bipartite graph to obtain second first-order messages, where the expression of the second first-order messages is as follows:

[0029]

[0030] where, m i←u represents the second first-order message from user node u to design node i, W3 represents the third trainable weight matrix, and W4 represents the fourth trainable weight matrix;

[0031] Calculating the self-connection message of the design nodes in the user-design bipartite graph, performing a weighted sum on the second first-order messages of the set of user nodes interacting with the design nodes to obtain a second weighted sum, and performing an embedding representation on the second self-connection message and the second weighted sum according to a non-linear activation function to obtain the first-order embedding feature vector of the design nodes, where the expression of the first-order embedding feature vector of the design nodes is as follows:

[0032]

[0033] where, represents the first-order embedding feature vector of the design nodes, m i←i represents the self-connection message of the design nodes;

[0034] Determining the second high-order message from the user nodes to the design nodes according to the previous-order embedding feature vector of the design nodes, where the expression of the second high-order message is as follows:

[0035]

[0036] where, represents the l-th order message from user node u to design node i, represents the l-th order third trainable weight matrix, represents the l-th order fourth trainable weight matrix;

[0037] Determining the high-order embedding feature vector of the design nodes according to the first high-order message of the design nodes, where the expression of the high-order embedding feature vector of the design nodes is as follows:

[0038]

[0039] where, Represents the l-th order embedding feature vector of the design node.

[0040] Preferably, according to the graph neural network, perform first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product, further including:

[0041] Concatenate the first-order embedding feature vector to the high-order embedding feature vector of the user node to obtain the first embedding feature vector of the user;

[0042] Concatenate the first-order embedding feature vector to the high-order embedding feature vector of the design node to obtain the second embedding feature vector of the visual design product.

[0043] Preferably, the style classifier uses a pre-trained ResNet-50 model, the aesthetic evaluation model uses a NIMA model, and the text encoding uses BERT encoding.

[0044] Preferably, the calculation formula of the preference score is as follows:

[0045]

[0046] where y ui Represents the preference score of the user for the visual design product, Represents the transpose of the first embedding feature vector, Represents the fusion vector of the visual design product.

[0047] The beneficial effects of the present invention are as follows: (1) By combining the bipartite graph modeling unit with the historical interaction data between users and visual design products, the present invention can accurately construct a personalized user-design bipartite graph for each user, thereby effectively capturing user preferences. Through first-order to high-order embedding propagation, the graph neural network can extract richer features from complex user-design relationships, and then accurately reflect the personalized needs and preferences of users, improving the accuracy of personalized recommendations. Moreover, different features of visual design products, including style features, image aesthetics features, and label vectors, are extracted by multiple units. The extraction of these multi-dimensional features provides a more comprehensive and profound understanding of visual design products. After fusing these features, it is possible to recommend design works that not only meet the personal preferences of each user but also meet their aesthetic needs; (2) The present invention evaluates the visual design product images through the aesthetics evaluation unit to provide an aesthetics score, which enables recommendations to not only be based on user historical behavior but also consider the aesthetic value of the design, avoiding recommendation biases that may occur by relying solely on user click history. This function can ensure that the recommended designs not only match user preferences but also have a certain visual appeal. Moreover, the style classification unit classifies the visual design product images to extract style feature vectors, which provides design recommendations with specific styles for users. For example, users like minimalist style, modern style, or retro style, etc. Combining the features of the embedding propagation unit, the recommendation system can recommend design products that are more in line with the user's style preferences based on the user's past preferences; (3) Through the vector splicing unit, the present invention converts the text encoding of style labels and aesthetics evaluation labels into vector form and splices them with the feature vectors of visual design products. This method can combine the semantic information of the labels with design features, providing more context information for the model and enhancing the model's personalized understanding of different types of users. Moreover, according to the historical interaction data and real-time feedback between users and design products, the embedding feature vectors of users are dynamically updated, and combined with the real-time data of style and aesthetics evaluation, personalized recommendations are provided for users. This means that the system can continuously adapt to changes in user interests and optimize recommendations based on the latest user behavior. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 FIG. is a schematic diagram of the system architecture of the overall system in an embodiment proposed by the present invention.

[0049] Reference numerals: 1, bipartite graph modeling unit; 2, embedding propagation unit; 3, style classification unit; 4, aesthetics evaluation unit; 5, vector splicing unit; 6, design recommendation unit. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] Embodiment 1, as Figure 1 shown, a personalized visual design recommendation system based on big data analysis proposed by the present invention includes:

[0051] The bipartite graph modeling unit 1 is used to construct a user-design bipartite graph based on the historical interaction data between users and visual design products, and obtain visual design product images;

[0052] The embedding propagation unit 2 is used to perform first-order to high-order embedding propagation on user nodes and design nodes in the user-design bipartite graph according to the graph neural network, so as to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product;

[0053] The style classification unit 3 is used to classify the style of the visual design product image according to the style classifier, so as to obtain the style label of the visual design product, and extract the feature vector of the fully connected layer of the style classifier as the style feature vector of the visual design product;

[0054] The aesthetic evaluation unit 4 is used to evaluate the aesthetics of the visual design product image according to the aesthetic evaluation model, so as to obtain the aesthetic evaluation label of the visual design product, and extract the feature vector of the middle layer of the aesthetic evaluation model as the aesthetic feature vector of the visual design product.

[0055] In the present invention, the bipartite graph modeling unit constructs a user-design bipartite graph and obtains the relationship between users and visual design products based on historical interaction data (such as user viewing, liking, purchasing, etc.). A bipartite graph refers to a graph structure in which the nodes are divided into two categories: user nodes and design nodes, and the connection of edges represents the interaction between a certain user and a certain design product (for example, a user is interested in a certain design product). Constructing the bipartite graph is to capture the interaction information between users and visual design products and provide input for the subsequent embedding propagation of the graph neural network (GNN); the bipartite graph model can effectively represent the relationship network between users and design products. Through the graph neural network, the preferences of users for different design products can be learned from the historical interactions between users and design products, and the designs that users may like can be predicted. The embedding propagation unit, based on the graph neural network (GNN), performs embedding propagation on the user nodes and design nodes in the user-design bipartite graph, from first-order to high-order propagation; the GNN iteratively propagates information, transfers the interest preference information of users to the design product nodes, and transfers the feature information of the design products to the user nodes, and finally obtains the embedding feature vectors of users and the embedding feature vectors of design products; through the embedding propagation of the GNN, the system can accurately align the features of users and design products according to the historical behaviors of users and the features of design products, so as to achieve personalized recommendation; the style classification unit classifies the images of visual design products by style, and classifies the design products into different styles according to the content of the images (such as colors, shapes, element layouts, etc.); the task of the style classifier is to identify the style of the design product (such as modern, retro, simple, luxurious, etc.) and map it to a style label; the aesthetics evaluation unit uses the aesthetics evaluation model to evaluate the aesthetics of the images of visual design products and evaluate the visual beauty of the design products; the aesthetics evaluation model can identify aesthetic features such as the visual quality of the design, color matching, and element coordination, and give an aesthetics evaluation label (such as beautiful, ordinary, and poor, etc.) according to these features; the aesthetic preferences of users are often personalized and subjective. Some users prefer visually appealing designs, and some users prefer simple designs. Through the aesthetics evaluation model, the system can identify and push design products that meet the aesthetic preferences of users.

[0056] In an alternative embodiment, the system further includes:

[0057] A vector splicing unit 5, which is used to perform text encoding on the style label and the aesthetics evaluation label to obtain a first label vector of the style label and a second label vector of the aesthetics evaluation label, and splice the second embedding feature vector, the style feature vector, the aesthetics feature vector, the first label vector and the second label vector to obtain a fusion vector of the visual design product;

[0058] The design recommendation unit 6 is used to calculate the preference score of the user for the visual design product according to the first embedded feature vector of the user and the fusion vector of the visual design product, and perform personalized visual design recommendation for the user according to the preference score.

[0059] It should be noted that the splicing operation refers to splicing all the above vectors (the second embedded feature vector, the style feature vector, the aesthetic feature vector, the first label vector, and the second label vector) (that is, arranging them in order into a long vector); the vector splicing unit converts the style label and the aesthetic evaluation label into vectors and combines them with other feature vectors to construct a comprehensive representation of the design product. This fusion vector not only contains the style and aesthetic information of the design product but also integrates the user's interest characteristics, providing more comprehensive input features for the recommendation system; the task of the design recommendation unit is to calculate the preference score of the user for each design product according to the user's interest vector and the fusion vector of the design product, so as to perform personalized recommendation; the first embedded feature vector of the user and the fusion vector of the design product are combined (usually through dot product, weighted sum or other calculation methods) to obtain a preference score; the preference score reflects the degree of interest of the user in a certain design product. The higher the score, the stronger the user's preference for the design product; according to the preference scores of the user for each visual design product, the design recommendation unit sorts all the design products; the design products with higher scores will be preferentially recommended to the user, thus realizing personalized recommendation.

[0060] Embodiment 2. A personalized visual design recommendation system proposed by the present invention. Compared with Embodiment 1, this embodiment further includes: The user-design bipartite graph includes a user node set, a design node set, and an edge set. Among them, the user nodes correspond to users, the design nodes correspond to visual design products, and the edges correspond to historical interaction data.

[0061] In an optional embodiment, performing first-order to high-order embedding propagation on the user nodes and design nodes in the user-design bipartite graph according to the graph neural network to obtain the first embedded feature vector of the user and the second embedded feature vector of the visual design product includes:

[0062] Embedding the ID identifications of the user nodes and design nodes according to the user-design bipartite graph to obtain the first initial embedded vector of the user and the second initial embedded vector of the visual design product;

[0063] Performing message encoding from the design nodes to the user nodes in the user-design bipartite graph to obtain the first first-order message, where the expression of the first first-order message is as follows:

[0064]

[0065] where, mu←i Denote the first-order message from design node \(i\) to user node \(u\), \(N\) u Denote the set of design nodes interacting with user node \(u\), \(N\) i Denote the set of user nodes interacting with design node \(i\), \(W1\) represents the first trainable weight matrix, \(W2\) represents the second trainable weight matrix, \(e\) i Denote the second initial embedding vector of the design node, \(e\) u Denote the first initial embedding vector of the user node;

[0066] Calculate the self-connection message of the user node in the user-design bipartite graph, perform a weighted sum of the first-order messages of the set of design nodes interacting with the user node to obtain the first weighted sum, and perform an embedding representation of the self-connection message and the first weighted sum according to the non-linear activation function to obtain the first-order embedding feature vector of the user node. Among them, the expression of the first-order embedding feature vector of the user node is as follows:

[0067]

[0068] Among them, Denote the first-order embedding feature vector of the user node, \(m\) u←u Denote the self-connection message of the user node, LeakReLu represents the non-linear activation function;

[0069] Determine the first high-order message from the design node to the user node according to the embedding feature vector of the previous order of the user node. Among them, the expression of the first high-order message of the user node is as follows:

[0070]

[0071] Among them, Denote the \(l\)-th order message from design node \(i\) to user node \(u\), Denote the first trainable weight matrix of the \(l\)-th order, Denote the second trainable weight matrix of the \(l\)-th order, Denote the \((l - 1)\)-th order embedding feature vector of user node \(u\), Denote the \((l - 1)\)-th order embedding feature vector of design node \(i\);

[0072] Determine the high-order embedding feature vector of the user node according to the first high-order message of the user node. Among them, the expression of the high-order embedding feature vector of the user node is as follows:

[0073]

[0074] Among them, Denote the \(l\)-th order embedding feature vector of the user node.

[0075] It should be noted that the message propagation from the design node to the user node means that each design node interacts its embedding vector with the user node, and the message propagation method is calculated through the set of design nodes that interact with the user node (i.e., the set of design products associated with the user); the trainable weight matrix means that during the message propagation process, the weight matrix is used to adjust the information transmission method, and the training of the weight matrix will optimize the information transmission effect of the model at different levels; the self-connection message means that each user node will transmit messages to the set of design nodes directly connected to itself, and these messages represent the user's feedback or preference for the design product.

[0076] In an optional embodiment, according to the graph neural network, perform first-order to higher-order embedding propagation on the user nodes and design nodes in the user-design bipartite graph to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product, further including:

[0077] Perform message encoding on the user nodes to design nodes in the user-design bipartite graph to obtain the second first-order message, where the expression of the second first-order message is as follows:

[0078]

[0079] where m i←u represents the second first-order message from the user node u to the design node i, W3 represents the third trainable weight matrix, and W4 represents the fourth trainable weight matrix;

[0080] Calculate the self-connection message of the design nodes in the user-design bipartite graph, perform weighted summation on the second first-order messages of the set of user nodes that interact with the design nodes to obtain the second weighted sum, and perform embedding representation on the second self-connection message and the second weighted sum according to the non-linear activation function to obtain the first-order embedding feature vector of the design nodes, where the expression of the first-order embedding feature vector of the design nodes is as follows:

[0081]

[0082] where, represents the first-order embedding feature vector of the design node, and m i←i represents the self-connection message of the design node;

[0083] Determine the second higher-order message from the user node to the design node according to the previous-order first-order embedding feature vector of the design node, where the expression of the second higher-order message is as follows:

[0084]

[0085] where, represents the l-th order message from the user node u to the design node i, Denote the third trainable weight matrix of the l-th order, Denote the fourth trainable weight matrix of the l-th order;

[0086] Determine the high-order embedding feature vector of the design node according to the first high-order message of the design node, where the expression of the high-order embedding feature vector of the design node is as follows:

[0087]

[0088] where, Denote the l-th order embedding feature vector of the design node.

[0089] In an optional embodiment, perform first-order to high-order embedding propagation on the user nodes and design nodes in the user-design bipartite graph according to the graph neural network to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product, further including:

[0090] Concatenate the first-order embedding feature vector to the high-order embedding feature vector of the user node to obtain the first embedding feature vector of the user;

[0091] Concatenate the first-order embedding feature vector to the high-order embedding feature vector of the design node to obtain the second embedding feature vector of the visual design product.

[0092] In an optional embodiment, the style classifier adopts a pre-trained ResNet-50 model, the aesthetics evaluation model adopts a NIMA model, and the text encoding adopts BERT encoding.

[0093] It should be noted that in image style classification, ResNet-50 is usually used to extract high-level features of images, especially the subtle differences between different styles, which may include artistic styles, design styles, or visual effects, etc. By pre-training the ResNet-50 model on a large image dataset, the model can extract useful features (such as shape, color, texture, pattern, etc.) from the input image and use them for the style classification task; NIMA understands the aesthetic features in images by learning a large number of labeled aesthetic score datasets. For example, the color combination, composition, light and shadow effects of the image will all affect the aesthetic score, and the model will output a score regarding the aesthetics of the image according to these factors. Usually, the score value ranges from 0 to 10, indicating the aesthetic quality of the image; in BERT, words or terms are converted into word vectors or subword vectors (Subword representations), and these vectors can capture the deep semantics and context relationships of the text. When processing text, BERT generates context-sensitive representations for each token (usually a word or subword) by inputting the tokens of the text. BERT can understand that the meaning of a word depends not only on the word itself but also on its surrounding context.

[0094] In an alternative embodiment, the calculation formula for the preference score is as follows:

[0095]

[0096] where y ui represents the preference score of the user for the visual design product, represents the transpose of the first embedded feature vector, represents the fusion vector of the visual design product.

[0097] The embodiments of the present invention have been described in detail above in conjunction with the accompanying drawings, but the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art to which the present invention pertains.

Claims

1. A personalized visual design recommendation system based on big data analysis, characterized in that, Including: A bipartite graph modeling unit (1), which is configured to construct a user-design bipartite graph according to historical interaction data between a user and a visual design product, and obtain a visual design product image; An embedding propagation unit (2), which is configured to perform first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network, so as to obtain a first embedding feature vector of the user and a second embedding feature vector of the visual design product; A style classification unit (3), which is configured to perform style classification on the visual design product image according to a style classifier, so as to obtain a style label of the visual design product, and extract a feature vector of a fully connected layer of the style classifier as a style feature vector of the visual design product; An aesthetic evaluation unit (4), which is configured to perform aesthetic evaluation on the visual design product image according to an aesthetic evaluation model, so as to obtain an aesthetic evaluation label of the visual design product, and extract a feature vector of an intermediate layer of the aesthetic evaluation model as an aesthetic feature vector of the visual design product.

2. The personalized visual design recommendation system based on big data analysis according to claim 1, characterized in that, The system further includes: A vector splicing unit (5), which is configured to perform text encoding on the style label and the aesthetic evaluation label, so as to obtain a first label vector of the style label and a second label vector of the aesthetic evaluation label, and splice the second embedding feature vector, the style feature vector, the aesthetic feature vector, the first label vector and the second label vector, so as to obtain a fusion vector of the visual design product; A design recommendation unit (6), which is configured to calculate a preference score of the user for the visual design product according to the first embedding feature vector of the user and the fusion vector of the visual design product, and perform personalized visual design recommendation for the user according to the preference score.

3. The personalized visual design recommendation system based on big data analysis according to claim 2, characterized in that The user-design bipartite graph includes a user node set, a design node set and an edge set, wherein the user nodes correspond to the users, the design nodes correspond to the visual design products, and the edges correspond to the historical interaction data.

4. A personalized visual design recommendation system based on big data analysis according to claim 3, characterized in that, Performing first-order to high-order embedding propagation on the user nodes and the design nodes in the user-design bipartite graph according to a graph neural network, so as to obtain a first embedding feature vector of the user and a second embedding feature vector of the visual design product, includes: Embedding the ID identifications of the user nodes and the design nodes according to the user-design bipartite graph, so as to obtain a first initial embedding vector of the user and a second initial embedding vector of the visual design product; Performing message encoding from the design nodes to the user nodes in the user-design bipartite graph, so as to obtain a first first-order message, wherein the expression of the first first-order message is as follows: Among them, m u←i represents the first-order message from design node i to user node u, N u represents the set of design nodes interacted by user node u, N i represents the set of user nodes interacted by design node i, W1 represents the first trainable weight matrix, W2 represents the second trainable weight matrix, e i represents the second initial embedding vector of the design node, e u represents the first initial embedding vector of the user node; Calculate the self-connection message of the user node in the user-design bipartite graph, perform a weighted sum of the first-order messages of the set of design nodes interacting with the user node to obtain a first weighted sum, and perform an embedding representation of the self-connection message and the first weighted sum according to a non-linear activation function to obtain the first-order embedding feature vector of the user node, where the expression of the first-order embedding feature vector of the user node is as follows: Among them, represents the first-order embedding feature vector of the user node, and m u←u represents the self-connection message of the user node, and LeakReLu represents the non-linear activation function; Determine the first high-order message from the design node to the user node according to the embedding feature vector of the previous order of the user node, where the expression of the first high-order message of the user node is as follows: Among them, represents the l-th order message from the design node i to the user node u, represents the first trainable weight matrix of the l-th order, represents the second trainable weight matrix of the l-th order, represents the (l - 1)-th order embedding feature vector of the user node u, represents the (l - 1)-th order embedding feature vector of the design node i; Determine the high-order embedding feature vector of the user node according to the first high-order message of the user node, where the expression of the high-order embedding feature vector of the user node is as follows: Among them, represents the l-th order embedding feature vector of the user node.

5. A personalized visual design recommendation system based on big data analysis according to claim 4, characterized in that, Perform first-order to high-order embedding propagation on the user node and the design node in the user-design bipartite graph according to the graph neural network to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product, and further include: Perform message encoding on the user node to the design node in the user-design bipartite graph to obtain a second first-order message, where the expression of the second first-order message is as follows: where m i←u represents the second first-order message from the user node u to the design node i, W3 represents the third trainable weight matrix, and W4 represents the fourth trainable weight matrix; Calculate the self-connection message of the design node in the user-design bipartite graph, perform a weighted sum of the second first-order messages of the set of user nodes interacting with the design node to obtain a second weighted sum, and perform an embedding representation of the second self-connection message and the second weighted sum according to a non-linear activation function to obtain the first-order embedding feature vector of the design node, where the expression of the first-order embedding feature vector of the design node is as follows: Among them, represents the first-order embedded feature vector of the design node, and m i←i represents the self-connection message of the design node; Determine the second high-order message from the user node to the design node according to the embedding feature vector of the previous order of the design node, where the expression of the second high-order message is as follows: Among them, represents the l-th order message from user node u to design node i, represents the third trainable weight matrix of the l-th order, represents the fourth trainable weight matrix of the l-th order; Determine the high-order embedding feature vector of the design node according to the first high-order message of the design node, where the expression of the high-order embedding feature vector of the design node is as follows: Among them, represents the l-th order embedding feature vector of the design node.

6. The personalized visual design recommendation system based on big data analysis according to claim 5, wherein Perform first-order to high-order embedding propagation on the user node and the design node in the user-design bipartite graph according to the graph neural network to obtain the first embedding feature vector of the user and the second embedding feature vector of the visual design product, and further include: Cascade and splice the first-order embedding feature vector to the high-order embedding feature vector of the user node to obtain the first embedding feature vector of the user; Cascade and splice the first-order embedding feature vector to the high-order embedding feature vector of the design node to obtain the second embedding feature vector of the visual design product.

7. An individualized visual design recommendation system based on big data analysis according to claim 6, characterized in that, The style classifier uses a pre-trained ResNet-50 model, the aesthetic evaluation model uses a NIMA model, and the text encoding uses BERT encoding.

8. The personalized visual design recommendation system based on big data analysis according to claim 7, characterized in that The calculation formula of the preference score is as follows: Among them, y ui represents the preference score of the user for the visual design product, represents the transpose of the first embedded feature vector, represents the fusion vector of the visual design product.

Images