Research on personalized clothing recommendation system based on AIGC

In modern society, personalized fashion has become a life attitude and consumption trend. People pay more and more attention to show their uniqueness and personalized style, and the application of personalized clothing recommendation system in the apparel industry precisely meets this demand [1-4]. Personalized clothing recommendation system is a kind of technical means applied to the clothing industry, through the analysis of the user’s personal preferences and needs, to provide users with personalized clothing recommendations [5-7]. Through this system, users can get the most suitable clothing and matching suggestions according to their own preferences and styles, and improve their image and fashion taste. Among many recommendation systems, personalized clothing recommendation system based on AIGC is gradually applied with its unique advantages [8-11].

AIGC, i.e. Artificial Intelligence in Generalized Commodities, is a comprehensive method of using artificial intelligence technology [12]. In intelligent marketing personalized recommendation, AIGC analyzes a large amount of user data and uses machine learning and deep learning algorithms to achieve accurate prediction of consumer interest and provide personalized product or service recommendations [13-16]. AIGC has many advantages in intelligent marketing personalized recommendation: first, high accuracy: AIGC can analyze the user’s historical browsing First, high accuracy: AIGC can analyze the user’s history, browsing, purchasing behavior, social media data and other information to accurately predict the user’s interests and needs [17-20]. Second, high efficiency: AIGC can automate the analysis and processing of massive amounts of data, which improves recommendation efficiency and reduces labor costs [21-22]. Third, immediacy: AIGC can monitor user behavior and feedback in real time, and recommend relevant products or services in time when users are browsing or purchasing. AIGC can play an important value in e-commerce, tourism, social media and other fields [23-26].

Literature [27] describes a mathematical model-based recommendation system for apparel design that predicts and controls the style and structural parameters of garments based on consumers’ individualized requirements. It was emphasized that the system facilitates designers to generate optimal design solutions quickly, accurately, intelligently and automatically. Literature [28] created a user knowledge graph, a garment knowledge graph, and a contextual knowledge graph, and utilized the Apriori algorithm to capture the intrinsic associations between garment attributes and contextual attributes. It revealed that this approach possesses higher recommendation quality. Literature [29] proposed an intelligent fashion recommendation system to deliver personalized apparel products and the performance of this system was validated in custom design and mass market selection. Literature [30] proposed an automated personalized recommendation method for apparel based on user sentiment analysis, which was shown to perform better compared to other algorithms by proposing personalized recommendations based on multiple interest values. Literature [31] emphasized that the fashion style consistency of the user’s social circle and clothing items is an important factor influencing the user’s decision in clothing recommendation systems. The effectiveness of this personalized clothing recommendation approach was also demonstrated through a series of analyses. Literature [32] introduces the cluster analysis subsystem of genetic algorithm in the traditional collaborative filtering recommendation system, which illustrates that the collaborative filtering recommendation system based on genetic clustering can effectively solve the scalability problem of collaborative filtering recommendation system. Literature [33] pointed out the deficiencies of traditional commercial clothing recommendation systems and proposed a fit-oriented clothing pattern intelligent recommendation system to support the design of personalized clothing products, which was successfully applied. Literature [34] proposed FRS, which improves user satisfaction by integrating user preferences and historical data and generating personalized fashion recommendations based on the user’s style and preferences.

The above studies introduced clothing recommendation systems based on sentiment analysis, numerical modeling, etc., and pointed out the advantages of these systems over the traditional approaches, indicating that clothing recommendation systems are not only widely used in real life, but also receive the same attention from the academic community. Unfortunately, AIGC, as an emerging technology, has not been applied more in clothing recommendation, and even there are very few academic researches related to it, so the research on personalized clothing recommendation system based on AIGC is of historical significance.

This paper first explains Graph Neural Networks (GCN) from three perspectives: essence, core idea and main steps respectively, and describes in detail the techniques for extracting image and text features. The visual information of fashion items is extracted by convolutional neural network (CNN) model. A graph neural network between fashion items, clothing, and users is established to fuse rich node representations and filter the noisy interactions in the two-part graph using an improved GRCN. The potential shallow relationships are learned using Light GCN, and the relationships between user-user graph and clothing-garment graph are obtained by cosine similarity. A personalized clothing recommendation model fusing latent representations is proposed. Finally, the application effect of the personalized clothing recommendation system is evaluated by ablation test and simulation test.

2

Clothing matching recommendation system based on graph neural network

2.1

Feature extraction for graph neural networks

2.1.1

Graph Neural Networks

The pioneering application of graph data to neural networks is the graph neural network (GNN) [35-36]. Subsequently, GNNs have been widely used in various fields such as node classification, image retrieval and personalized recommendation and have derived various branches of the network due to their remarkable performance in handling unstructured data. Among them, graph convolutional network (GCN) is an important network branch of graph neural network. The essence of graph convolutional networks is to extract spatial features in topological graphs; the core idea is to utilize the message passing mechanism to aggregate and update the information of the nodes in the graph, and then to generate new node representations, which is accomplished by utilizing the information of the edges in the graph; the main step is to utilize the weighted summation function of the filter to update the representations of the features, which is operated for each pixel point in the space. In particular, the values of the weighting coefficients are parameters of the convolution kernel.

The input to the model consists of a feature matrix X represented by the initial features of all nodes and a neighbor matrix represented by the neighborhood graph of the nodes. The update of the features of this node in each convolutional layer is obtained by a message passing mechanism and a nonlinear transformation. The process of message passing mechanism can be defined by the following equation: (1) ${\begin{array}{l} L^{(1)} = ρ (\tilde{A} X W^{(0)}) \\ L^{(j + 1)} = ρ (\tilde{A} L^{j} W^{j}) \end{array}$

where $X \in ℝ^{n \times m}$ is the set of all nodes and feature matrices in the graph G = (V, E), n is the total number of nodes, m is the dimension of the feature vector, $V = (| V | = n)$ is the nodes in the graph, and E is the edges between nodes. The feature vector of v is for each row $x_{v} \in ℝ^{m}$ ; A and D are the adjacency and degree matrices, respectively, $D_{i i} = \sum_{i} A_{i j}$ ; $\tilde{A} = D^{- \frac{1}{2}} A D^{- \frac{1}{2}}$ is the normalized adjacency matrix, $W^{(0)} \in ℝ^{m \times k}$ is the weight matrix, ρ(·) is the activation function, j is the number of layers, L⁽⁰⁾ = X. L⁽¹⁾ is the feature matrix of the k-dimensional node after single-layer GCN aggregation that can only integrate the feature information of the first-order neighboring nodes, and L^(j+1) is the feature matrix of the k-dimensional node after multilayer GCN aggregation that captures the multi-order neighbor features.

2.1.2

Feature extraction

1)

Image Feature Extraction

The high-level semantic features of apparel emphasize the style and style of apparel, and have a pivotal role in personalized apparel recommendation. The methods often used in deep learning to extract high-level semantic features of clothing can be divided into three categories, the first category is the SVM algorithm based on convolutional neural networks combined with GCN, RestNet and Faster R-CNN, etc. The prerequisite for feature CNN models based on convolutional neural networks to perform training is that the clothing images should contain accurate labels while the training data should be sufficient; the second category is the bi-directional recurrent neural network based on recurrent neural networks; the third category is the deep belief network based on deep beliefs. Bidirectional Recurrent Neural Networks based on Recurrent Neural Networks; and the third category is Deep Belief Networks based on Deep Belief Networks (DBN), which requires input of fixed garment image dimensions although the algorithm does not require a large amount of garment labeling information to exhibit good feature extraction capabilities.

2)

Text feature extraction

The text description of the clothing image not only contains the clothing style, style and color, but also contains features such as product title, and the extraction of text features of the clothing can make the clothing features more specific and effectively improve the performance of the model. Among them, encoding the words in the title of the item can facilitate the extraction of the category information of the item, which makes the model effectively determine the category of the item. This paper proposes a text classification model Text CNN, which first embeds the word vector obtained by text segmentation, and then performs a layer of convolution and pooling operation on the word vector in turn, and finally normalizes the output with full connection to achieve n classification. Because the model utilizes multiple convolutional kernels of different sizes to extract the key information in the sentence, it can better capture the local relevance. From Text CNN, we can conclude that the execution process of Text CNN is divided into four steps: in the first step, the length n of the sentence is set in the input layer; in the second step, the convolutional layer uses filters with different widths to realize the feature mapping; in the third step, after the pooling operation, a one-dimensional vector is obtained, and then the output is passed through the activation function, a Dropout layer is added to prevent the overfitting, and an L2 regularization is added in the fully-connected layer parameters; in the fourth step all the output values of the fully connected layer are connected to the softmax layer.

2.2

Multimodal personalized clothing recommendation under graph neural networks

2.2.1

Introduction to the model

For clothing compatibility modeling, let the user set be $U = {u_{1}, u_{2}, u_{3} \dots u_{g}}$ , g be the number of users. Let the set of clothing training sets be $O = {o_{1}, o_{2}, o_{3} \dots o_{n}}$ and n be the number of garments. Let the set of fashion items contained in all garments be V and the set of categories of fashion items be V_c. Given a set of garments o consisting of $| o |$ fashion items s, then $o = {s_{1}, s_{2}, s_{3} \dots s_{| o |}}$ , c ∈ V, where each fashion item is described by a picture I and text P, the goal of the proposed model in this paper is to predict the compatibility score y_o of that garment o.

2.2.2

Embedding Coding Initialization and Feature Extraction

In this paper, ResNet [37], an 18-layer deep residual network pre-trained on Image Net, is used as the feature extractor for the model. The visual information of fashion item s is extracted using Convolutional Neural Network (CNN) model as shown below: (2) $f_{I} = C N N (I_{s} | Θ_{c n n})$

f_p ∈ R^d is the text feature vector of fashion item c, where d is the size of the text feature vector and Θ_bert is the parameter of the BERT model. The concept of modality is introduced and f_m is used to represent the input modality, when m = I is the image input and when m = P is the text input. Textual information of fashion items is extracted using BERT model: (3) $f_{P} = B E R T (P_{s} | Θ_{b e r t})$

2.2.3

Graph data construction

1)

Constructing a clothing-fashion item graph

According to the inclusion relationship between clothing o and fashion item s in the dataset, the adjacency matrix is so constructed: (4) $A_{o s} = {\begin{array}{l} 1, s \in o \\ 0, s \notin o \end{array}$

2)

Constructing the garment-garment graph

$o_{f_{m} i}$ and $o_{f_{m} j}$ denote the modal feature embedding vectors of the two garments o_i and o_j, respectively, for which similarity needs to be computed, and S_oo denotes the cosine similarity between $o_{f_{m} i}$ and $o_{f_{m} j}$ . Similarity matrix: (5) $S_{o o} = \frac{{(o_{f_{m} i})}^{T} o_{f_{m j}}}{‖ o_{f_{m} i} ‖ ‖ o_{f_{m} j} ‖}$

For the modal feature embedding vectors $o_{f_{m} i}$ and $o_{f_{m} j}$ for the whole set of garments, both are represented using the fashion item embedding aggregation contained in the whole set of garments, if there is $o = {s_{1}, s_{2}, s_{3} \dots s_{| o |}}$ , then: (6) $o_{f_{m}} = s_{f_{m} 1} + s_{f_{m} 2} + s_{f_{m} 3} + \dots + s_{f_{m} | o |}$

Finally, for each garment node, this paper retains K_outit edge with the highest similarity score, where K_outfit is the hyperparameter, and constructs the adjacency matrix: (7) $A_{o o} = {\begin{array}{l} 1, S_{o o} \in t o p - K_{o u t f i t} (S_{o o}) \\ 0, S_{o o} \notin t o p - K_{o u t f i t} (S_{o o}) \end{array}$

3)

Constructing a user-user graph

The construction is done using the dynamic graph approach, where the similarity between users is first calculated using the cosine similarity: (8) $S_{u u}^{n} = \frac{{(u_{f_{m} i})}^{T} u_{f_{m} j}}{‖ u_{f_{m} i} ‖ ‖ u_{f_{m} j} ‖}$

where u_fmi and u_fmj denote the modal feature embedding vectors of two users u_i and u_j, respectively, for which similarity needs to be computed. $S_{u u}^{n}$ denotes the cosine similarity computed for the nth time between $u_{f_{m} i}$ and $u_{f_{m} j}$ , n is the epoch of model training, given the initial n = 0 of model training. For each user node, the K_user edge with the highest similarity is retained, where K_user is the hyperparameter, and the adjacency matrix is constructed: (9) $A_{u u}^{n} = {\begin{array}{l} 1, S_{u u}^{n} \in t o p - K_{u s e r} (S_{u u}^{n}) \\ 0, S_{u u}^{n} \notin t o p - K_{u s e r} (S_{u u}^{n}) \end{array}$

4)

Construct a user-clothing graph

Such that when there is an interaction between the user and the garment, Interactions_u0 = 1. If there is no interaction, then Interactions_u0 = 0. Then. The adjacency matrix is specified as: (10) $A_{u o} = {\begin{array}{l} 1, I n t e r a c t i o n s_{u 0} = 1 \\ 0, I n t e r a c t i o n s_{u 0} = 0 \end{array}$

2.2.4

Graph Relationship Learning

Firstly, the adjacency matrix is normalized for user-user graph and clothing-clothing graph [38], where $D_{u u}^{n}$ is the diagonal matrix of the corresponding dynamic graph $A_{u u}^{n}$ and D_oo is the diagonal matrix of the static graph A_oo. The normalization results are ${\tilde{A}}_{u u}^{n}$ and ${\tilde{A}}_{o o}$ . Eq: (11) ${\begin{matrix} {\tilde{A}}_{u u}^{n} & = & {(D_{u u}^{n})}^{- \frac{1}{2}} A_{u u}^{n} {(D_{u u}^{n})}^{- \frac{1}{2}} \\ {\tilde{A}}_{o o} & = & {(D_{o o})}^{- \frac{1}{2}} A_{o o} {(D_{o o})}^{- \frac{1}{2}} \end{matrix}$

A graph convolution operation is performed on two homogeneous graphs using Light GCN with a potential relationship between the user and the user: (12) ${\begin{array}{l} {\hat{h}}_{i}^{u u} = \sum_{j \in N (i)} {\tilde{A}}_{u u}^{n} h_{j} \\ {\hat{h}}_{i}^{o o} = \sum_{j \in N (i)} {\tilde{A}}_{o o} h_{j} \end{array}$

Where, i is the table target node and h_j is the feature vector of the neighboring node.

Determine the weight matrix of FCSA-GNN: c_i denotes the category of the target node fashion item s_i, c_other denotes the other categories of the non-target node, and $γ_{c_{i}} o$ denotes the weight of the fashion item with respect to the edges of the garment graph in that garment graph. $ω (c_{i}, o)$ denotes the frequency of occurrence of the category of the target fashion item s_i on the dataset V_c. Then there are: (13) $γ_{c_{i} o} = \frac{ω (c_{s_{i}}, o)}{\sum_{c_{s_{j}} \in c_{s_{o t h e r}}} ω (c_{s_{j}}, o)}$

Using FIRP in FCSA-GNN for aggregation, given that the feature vector of a homogeneous node is h, then there: (14) $h^{(h + 1)} = F F N (\sum_{j \in N (i)} γ_{i j} A_{o s} h^{(t)})$

where h^(t) denotes the homogeneous node vector in layer t of the graph neural network, h^t ∈ R^d, d are the dimensions of the feature vectors, t ∈ [0, L], and L is the number of layers in the FCSA-GNN. $h^{0} \in {o_{i d}, o_{f_{l}}, o_{f_{p}}, s_{i d}, s_{f_{l}}, s_{f p}}$ .

The dimension reduction is then performed as in the following equation, ∥ denoting the series operation, and the FCSA-GNN output is ${\hat{h}}_{j}$ . FC is the fully linked layer, and ${\hat{h}}_{j}$ is the final output. The final output is denoted by symbols as ${\hat{o}}_{f_{I}}^{o s}$ , ${\hat{o}}_{f_{P}}^{o s}$ , ${\hat{o}}_{i d}^{o s}$ . i.e: (15) ${\hat{h}}_{j} = F C (h^{0} ‖ h^{1} \dots ‖ h^{L})$

2.2.5

Interactive Relational Learning-Improved GRCN

1)

Clothing content node updating module: the feature vectors of the input clothing modality are $o_{f_{m}} = s_{f_{m} 1} + s_{f_{m} 2} + s_{f_{m} 3} + \dots + s_{f_{m} | 0 |}$ , $m \in {I, P}$ , $l e a k y_Re l u (\cdot)$ denote the activation function, W_f is the trainable weight, and b_f is the deviation vector. The final clothing modal eigenvector obtained by this module is denoted by ${\hat{o}}_{f_{m}}^{u o}$ . i.e: (16) ${\bar{o}}_{f_{m}} = l e a k y R e l u (W_{f} o_{f_{m}} + b_{f})$

2)

User content node update module: using the neighboring features in the user-clothing bipartite graph to adjust the user features obtained by random initialization, node update of user features by aggregating neighboring clothing features through iterative routing operations: (17) ${\begin{array}{l} k_{u o} = \frac{\exp ({\bar{o}}_{f_{m}}^{T} u_{f_{m}}^{(l - 1)})}{\sum_{j \in N (u)} \exp ({\bar{o}}_{f_{m} j}^{T} u_{f_{m}}^{(l - 1)})} \\ u_{f_{m}}^{(l + 1)} = u_{f_{m}}^{(l)} + \sum_{i \in N (u)} k_{u o} {\bar{o}}_{f_{m}} \end{array}$

Where l = 0 when, $u_{f_{m}}^{0} = u_{f_{m}}$ , m ∈ {P, I}, k_uo is the similarity between user features and clothing features. The end-user modal feature vector obtained through this module is denoted by ${\hat{u}}_{f_{m}}^{u^{u}}$ .

Noise cropping: based on the obtained user preferences and the refined item features, their relative distances in both directions are computed: (18) ${\begin{array}{l} {\bar{s}}_{u \leftarrow o}^{m} = \frac{\exp ({\hat{u}}_{f_{m}}^{T} {\hat{o}}_{f_{m}})}{\sum_{j \in N (u)} \exp ({\hat{u}}_{f_{m}}^{T} {\hat{o}}_{f_{m} j})} \\ {\bar{s}}_{o \leftarrow u}^{m} = \frac{\exp ({\hat{o}}_{f_{m}}^{T} {\hat{u}}_{f_{m}})}{\sum_{j \in N (o)} \exp ({\hat{o}}_{f_{m}}^{T} {\hat{u}}_{f_{m} j})} \end{array}$

where N(u) and N(o) denote the set of neighboring nodes of user node u and clothing node o in the user-clothing graph, respectively, and ${\bar{s}}_{u \leftarrow o}^{m}$ and ${\bar{s}}_{o + u}^{m}$ are a quantity reflecting the affinity between ${\hat{u}}_{f_{m}}$ and ${\hat{o}}_{f_{m}}$ in the m modality. Define a basic vector for each user or item as follows: (19) $ρ = [ρ^{I}, ρ^{P}]$

where ρ denotes the basis vector which is used to measure the relative preference of the node for different modalities. Combining the basis vectors, the weights of the edges are calculated by fusing the multimodal scores: (20) ${\begin{array}{l} s_{u \leftarrow o} = \max (ρ_{u}^{l} {\bar{s}}_{u \leftarrow o}^{l}, ρ_{u}^{P} {\bar{s}}_{u \leftarrow o}^{p}) \\ s_{o \leftarrow u} = \max (ρ_{o}^{l} {\bar{s}}_{o \leftarrow u}^{I}, ρ_{o}^{P} {\bar{s}}_{o \leftarrow u}^{P}) \end{array}$

The $\max (\cdot)$ -max method is used to determine the weights, and using the basis vectors ρ and affinity scores ${\bar{s}}_{u \leftarrow o}^{m}$ and ${\bar{s}}_{o \leftarrow u}^{m}$ , the weights of each edge are obtained and assigned to achieve soft pruning of the noisy edges.

2.2.6

Information integration

For the modal eigenvectors of users and garments with the same modality m, an aggregated fusion approach is taken: (21) ${\begin{matrix} u_{f_{m}}^{m} = {\hat{u}}_{f_{m}}^{u o} + \frac{{\hat{u}}_{f_{m}}^{u u}}{‖ {\hat{u}}_{f_{m}}^{u u} ‖} \\ o_{f_{m}^{m}}^{m} = {\hat{o}}_{f_{m}}^{u o} + \frac{{\hat{o}}_{f_{m}}^{o o}}{‖ {\hat{o}}_{f_{m}}^{o o} ‖} + \frac{o_{f_{m}}^{o s}}{‖ o_{f_{m}}^{o s} ‖} \end{matrix}$

For the two different modes I, P of user and clothing, a splicing approach is taken to obtain the eigenvectors of each mode separately: (22) ${\begin{array}{l} {\hat{u}}_{f} = u_{f_{I}}^{l} ‖ u_{f_{P}}^{P} \\ {\hat{o}}_{f} = o_{f_{I}}^{l} ‖ o_{f_{P}}^{P} \end{array}$

The same aggregation approach is taken to integrate the ID embedding vectors of users and garments: (23) ${\begin{matrix} {\hat{u}}_{i d} = {\hat{u}}_{i d}^{u o} + \frac{{\hat{u}}_{i d}^{u u}}{‖ {\hat{u}}_{i d}^{u u} ‖} \\ {\hat{o}}_{i d} = {\hat{o}}_{i d}^{u o} + \frac{{\hat{o}}_{i d}^{o o}}{‖ {\hat{o}}_{i d}^{o o} ‖} + \frac{{\hat{o}}_{i d}^{u o}}{‖ {\hat{o}}_{i d}^{u o} ‖} \end{matrix}$

Finally, the eigenvectors of the different modes are obtained by virtue of the ID embedding vectors: (24) ${\begin{array}{l} u^{*} = {\hat{u}}_{f} ‖ {\hat{u}}_{i d} \\ o^{*} = {\hat{o}}_{f} ‖ {\hat{o}}_{i d} \end{array}$

2.2.7

Model predictions

Implementation of personalized clothing recommendation: this section uses inner product as to predict the likelihood that user u will purchase clothing o. I.e: (25) ${\hat{y}}_{u o} = u^{* T} o^{*}$

Implementation of Clothing Compatibility Modeling: in this paper, it is argued that fashion items should have different importance for this whole set of clothing. It is necessary to distinguish the importance of items in a whole set of clothing through the self-attention mechanism, formally, the attention graph is computed as follows: (26) $A = ρ (W_{1} σ (W_{2} {\hat{o}}_{f m}))$

where ${\hat{o}}_{f_{m}} \in R^{| o | \times d}$ is the embedding matrix of a set of garments, $| o |$ is the number of fashion items within the garments, d is the eigenvector dimension of the bimodal splicing of the fashion items; W₁ and W₂ are two trainable weight matrices. $ρ (\cdot)$ is the $s o f t \max (\cdot)$ function to normalize the attention score of the garment. $σ (\cdot)$ is the activation function $L e a k y R e L U (\cdot)$ . Where modality ${\hat{o}}_{f_{m}}$ is: (27) ${\hat{o}}_{f_{m}} = (α o_{f_{l}}) ‖ (β o_{f_{p}})$

α and β are two hyperparameters used to control the weights to create a score map for each garment as follows: (28) $B = σ (W_{3} σ (W_{4} {\hat{o}}_{f_{m}}))$

where W₃ and W₄ are two trainable weight matrices and garment o has a garment compatibility score: (29) ${\hat{y}}_{o} = \sum_{r = 1}^{R} a_{r}^{T} b_{r}$

where a_r and b_r are rows r, A and C, respectively.

2.3

Analysis and design of clothing recommendation system

2.3.1

Demand analysis

1)

System functionality requirement analysis

When users buy clothes online, the recommendation strategy provided by the system changes because of the different demand scenarios of each person. In order to better optimize the user experience, the system needs to implement the following functions, functional requirements analysis example shown in Figure 1.

2)

System non-functional requirements analysis

Performance Requirements: It is necessary to ensure that the system can quickly respond to user needs and enhance the user’s interactive experience. The system needs to be able to automatically select the clothing recommendation function or clothing matching function according to the shopping needs of each user, and display the recommended clothing to the user through the front-end page of the system, so that the user basically does not feel the delay in the use of the process. At the same time, the throughput of the system also needs to be high enough to support a large number of users using the system at the same time.

Security requirements: the system should set different permissions for different roles, thus enhancing the security performance of the system. Ordinary users can only see their own data and recommended interfaces, while the administrator can see the data information of all users, can view the system log file information, through the analysis of log files to observe the current system whether there are attacks, violations and whether the error reporting events, to take appropriate ways to deal with them, and at the same time, can also add other administrator users. Users in the login process, the password should be encrypted storage, rather than plaintext storage, to prevent the password leakage, causing great losses to the user. When the system faces sudden failure problems, it can ensure that the data is not lost and the impact on the users is minimized.

2.3.2

Overall system design

1)

System software hierarchy design

The system architecture is shown in Figure 2.

2)

System Functional Module Design

According to the analysis of system functional requirements, this paper designs five major modules necessary for the system, which are login and registration module, clothing group recommendation module, single clothing recommendation module, clothing matching module, and auxiliary module. The login and registration module includes ordinary user registration, ordinary user login, administrator login, and password modification; the clothing group recommendation module recommends a group of garments to the user by considering the factors of clothing category information, user multi-intentions, and correlation between garments; the single garment recommendation module recommends a single garment to the user based on the graph neural network structure of the global graph and the session graph, and by considering the target focus mechanism; the clothing collocation module recommends a single garment to the user by Considering the visual information, textual information, popularity information, and user preference information of the clothing, it successfully matches the corresponding underwear for the user; the auxiliary module includes the functions of viewing clothing details, viewing and adding to shopping cart, viewing and posting comments, and data management. The functional module design of the system is shown in Figure 3.

3

Results of clothing matching recommendation system under graph neural network

In order to evaluate the effectiveness of the proposed improved GRCN method, this paper conducts comparative experiments on the two tasks of compatibility modeling and personalized set recommendation, respectively, and quantitatively compares this method with some existing methods and case studies. The comparison methods used on the POG dataset are Random, SiameseNet, Bi-LSTM, FHN, NGNN, HGAN, and this paper’s method. In order to compare the performance of the models more intuitively, this paper chooses two models, FHN and NGNN, which are more outstanding in the quantitative comparison, to carry out a further case study to demonstrate the effectiveness of the optimized GRCN method in compatibility modeling and personalized set recommendation. GRCN method in compatibility modeling.

3.1

Accuracy of different methods in the FITB mission

Table 1 shows the experimental results obtained by the GRCN network proposed in this paper with other mainstream methods for the FITB task on the POG dataset. In comparison with several models mentioned above, the GRCN model proposed in this paper achieves the best performance. Compared to the HGAN method, this method not only considers the separate direction of message passing from the lower nodes to the upper nodes to give micro-level information to the embedded representation of the nodes, but also considers the intra-layer message passing and the message passing from the upper to the lower layers to encode the rich meso- and macro-level information into the embedded representation of the nodes, which solves the problem of unidirectional message passing for the node features as well as the graph topology the problem of poor encoding performance.

Table 1.

The accuracy of different methods in the FITB task

Method	Accuracy rate(%)
Random	0.248
SiameseNet	0.4994
Bi-LSTM	0.6579
FHN	0.7707
NGNN	0.8167
HGAN	0.8783
GRCN	0.9189

3.2

Analysis of experimental results of personalized set recommendation

For the personalized set recommendation task, this section takes four widely used evaluation metrics of recommender systems: HR@N, NDCG@N, Recall@N, Precision@N, GRCN with the results of quantitative comparison of the existing methods on the POG dataset using the comparison methods FPITF, FHN, MF, VBPR and the method proposed in Chapter 3, HGAN, respectively. The analysis of the comparative experimental results of these methods fully demonstrates the superiority of the improved GRCN method proposed in this paper in the task of personalized set recommendation.

Table 2 shows the experimental results obtained from the improved GRCN network proposed in this paper and other mainstream methods for the personalized set recommendation task on the POG dataset. The results show that the experimental results corresponding to the four parameters “HR@10, NDCG@10, Recall@10, and Precision@10” of the GRCN network proposed in this paper are, respectively, 0.2426, 0.1129, 0.0568, and 0.0464, which are overall more significant than other network algorithms. It can be seen that the embedding representation of the nodes empowers the information at the micro level and it takes into account the intra-layer messaging as well as the messaging from the upper to the lower layers. Encoding rich meso- and macro-level information into the embedded representations of nodes solves the problem of poor performance of unidirectional message passing in encoding node features as well as the topology of the graph. Meanwhile, by stacking multi-layer graph convolutional neural networks, the embedded representations of nodes can encode the features of higher-order neighboring nodes to aggregate richer semantic information in the higher-order connectivity, and to solve the problem of low efficiency of message passing between distant nodes and loss of global information. The experimental results prove that the multimodal personalized clothing recommendation system based on graph neural network is reasonable and effective.

Table 2.

The personalized suit recommends the task quantitative contrast

Method	HR@10	NDCG@10	Recall@10	Precision@10
FPITF	0.0476	0.0444	0.0265	0.0118
FHN	0.0868	0.0724	0.0089	0.0013
MF	0.1798	0.071	0.0401	0.027
VBPR	0.1926	0.0702	0.0395	0.0223
NGCF	0.2472	0.0934	0.0358	0.0149
HGAN	0.2345	0.1266	0.0628	0.0146
GRCN	0.2426	0.1129	0.0568	0.0464

3.3

Ablation experiments

In order to further validate the contribution of each component of the optimized GRCN method proposed in this paper, a series of ablation experiment scenarios are designed in this section on the POG dataset for analyzing the impact of each module in the optimized GRCN on the experimental results, and the individual ablation experiments are discussed in detail below. For the compatibility modeling task, Table 3 demonstrates the quantitative analysis of the results of the ablation experiments, setting up the following three variants for the ablation experiments: 1)

Removal of the uniform category graph, which disregards category information (w/o Category) when aggregating messages from neighboring clothing item nodes, and considers only the visual information of clothing items.

2)

Remove intra-layer as well as message passing from lower layer nodes to upper layer nodes (w/o Bi-direction).

3)

Calculate the compatibility score without the attention mechanism (w/o Attention).

Table 3.

Compatibility modeling ablation experiment quantitative comparison table

Method	Accuracy rate(%)
w/o Category	0.8324
w/o Bi-direction	0.9056
w/o Attention	0.8857
GRCN	0.9502

The quantitative comparison table results of the compatibility modeling ablation experiments are shown in 3. 1)

After removing the uniform category graph, the model lacks the a priori knowledge related to category information when updating and optimizing the embedding representation of clothing item nodes, however, there is a strong correlation between the compatibility of clothing items and categories. The experimental results also show that the lack of category information has a large impact on the model performance, demonstrating the importance of assigning category information to clothing item nodes.

2)

For the compatibility modeling task, the optimized GRCN and HGAN basically adopt the same model structure after removing the intra-layer and the message passing from the lower nodes to the upper nodes, and from the experimental results, it is found that the Accuracy index of the GRCN is higher than the w/o Bi-direction in the FITB task, which indicates that the features of the neighboring nodes in the convergence layer to the target node The embedded representation of the target node gives meso-level information to the target node, while propagating the embedded representation of the messages from the parent node in the upper layer to the child nodes in the lower layer provides macro-level information to the nodes in the lower layer, and the clothing single-item node has a stronger expressive ability after the bidirectional graph convolution.

3)

Compared with the optimized GRCN, not using the attention mechanism in calculating the compatibility score makes the performance of the model degrade to a certain extent, which indicates that the style of a suit usually depends on some important clothing items it contains, and the compatibility degree of the suit is also determined by the key clothing items, and the experimental results also prove that the attention mechanism can well distinguish which clothing items have a The experimental results also prove that the attention mechanism can distinguish which clothing items are decisive for the compatibility of suits.

The quantitative comparison results of the personalized suit recommendation task ablation experiment are shown in Table 4. From the experimental results, it can be found that the performance of the model is greatly reduced, indicating that the information of individual clothing items has an important role in the overall modeling, and it also shows that the user’s historical interaction data has a crucial impact on modeling the user’s preference aspect. In addition, it is found that the overall performance of the model decreases to a low degree, with all four indicators decreasing to some extent, apparently removing the message passing from the upper node to the lower node, so that the lower node is unable to receive the macro-level information passed by the upper node, and for the embedded representation of the suit node, it is unable to learn the macro-level information from the messages propagated by the upper user node. The importance of macro-level information in the node update and optimization process is illustrated.

Table 4.

The personalized suit recommends the results of the experiment

Method	HR@10	NDCG@10	Recall@10	Precision@10
w/o Bt	0.2146	0.0877	0.0562	0.0314
w/o Intralayer	0.2429	0.1101	0.0535	0.0396
w/o Tb	0.2377	0.1033	0.0524	0.0335
GRCN	0.2627	0.1154	0.0555	0.0363

3.4

Simulated experimental validation

Select 20 experimenters whose target classification is dresses from 100 experimenters, and obtain the user demand model based on the browsing data of the 20 experimenters. The initial clothing screening is carried out sequentially, and 85 target recommended garments are obtained. Calculate the interest degree of each piece of clothing in the target clothing collection based on the user preference model, and the results of the clothing interest degree calculation are shown in Table 5. The top 10 garments in the clothing interest degree table and the 10 randomly selected garments in the hot list are shown to the experimenters in order, and 20 experimenters are allowed to score the 400 garments. The results of clothing interest degree calculation are shown in Table 5. In this experiment, this paper agrees that garments with higher than 7.5 points are the user’s favorite garments. According to the scoring of the 20 experimenters, and then according to the recommendation system recommended by the recommendation system of this research, there are 397 pieces of 400 garments that are scored higher than 7.5 points, with an accuracy rate of 99.25%, while there are 31 pieces of 85 garments that are randomly recommended that are scored higher than 7.5 points, with an accuracy rate of only 36.47%.

Table 5.

Calculation results of clothing interest

Clothing coding	Costume style	Color	Collar type	Profile	Fabric	Interest
1442	1	2	3	2	2	0.7612
1626	2	2	2	1	5	0.7605
1777	1	2	3	1	4	0.7567
1658	3	2	2	3	3	0.7472
1378	1	2	3	0	3	0.7358
1591	2	2	1	1	3	0.6928
1386	1	2	3	4	4	0.6831
1392	2	2	2	4	4	0.6745
1369	2	2	2	1	1	0.6604
2285	3	2	2	3	4	0.6506
1621	2	2	2	2	2	0.6496
1175	1	2	2	1	2	0.6348
1732	1	1	2	0	2	0.6074
2082	3	2	2	3	1	0.5307
1677	0	2	2	3	4	0.5257
1687	1	2	2	1	1	0.5306
1660	3	2	1	3	1	0.5114
1318	1	2	2	2	3	0.5013
1658	4	2	2	2	4	0.4695
1497	2	2	2	1	3	0.4382

3.4.1

Recommendation result score data in different ways

The ratings of the recommendation results in different ways are shown in Fig. 4. It is found that there is a significant difference in the mean value of the ratings of the garments recommended by the recommender system proposed in this paper and the randomly recommended garments. The mean score of the 20 testers for the model recommended clothing was above 8.54, which is much higher than the 7.5 satisfaction score standard set in this paper. And the mean value of the random recommendation score is 5.83, which is 1.67 points less than the standard satisfaction score. This shows that the recommendation effect of the clothing recommendation system model proposed in this paper is much higher than that of random recommendation.

3.4.2

Sorting accuracy

Ranking accuracy is used to measure how well the clothing interest level obtained by the recommendation model matches the rating in the user’s mind, which can be measured by the Spearman rank correlation coefficient. The accuracy of the ranking in the recommender system with respect to the true ranking is shown in Fig. 5. The results show that the mean value of the total Spearman’s coefficient is about 0.9662 and the ranking accuracy is good.

The Spearman rank correlation coefficient (SRCC) is used to calculate the Pearson’s correlation coefficient between the rank order of the recommended items in the recommender system and the true rank order, which is defined as follows: (30) $S R C C = \frac{\sum (r_{1} (i) - u_{1}) (r_{2} (i) - μ_{2})}{\sqrt{\sum {(r_{1} (i) - μ_{1})}^{2} \sqrt{\sum {(r_{2} (i) - μ_{2})}^{2}}}}$

where $r_{1} (i)$ and $r_{2} (i)$ are the ranking in the recommender system and the true ranking, respectively, and μ is the mean value of the ranking. If the ranking of an item in the recommender system is the same as the true ranking, the value of SRCC is 1.

3.4.3

Coverage

Coverage rate is to describe the ability of the recommendation coefficient to discover the long tail of goods, simply defined as the proportion of items that the recommendation system can recommend out of the total set of items, it can also be understood as the recommendation system needs to make as many garments as possible to be recommended, assuming that the user set of the system is U, and the recommendation system recommends a recommendation list of length N to each user $R (u)$ , and the total number of recommendations is m, then the coverage rate of the recommendation system formula is as follows: (31) $C o v e r a g e = \frac{| U_{u \in U} \cdot R (u) |}{m}$

The coverage rate under different N-values is calculated based on the interest ranking of the 20 experimenters’ target recommendation outfits. The coverage of the recommender system under different N values is shown in Fig. 6. The results show that the average coverage is higher than 90%, indicating a good ability to discover the long tail and a very high coverage rate.

4

Conclusion

This paper proposes a multimodal personalized clothing recommendation model based on graph neural network on the basis of Artificial Intelligence Generated Content (AIGC), and designs a clothing recommendation system based on the user’s needs to complete the process of personalized clothing recommendation to the user. The results show that the multimodal personalized clothing recommendation model under graph neural network can significantly improve the accuracy and efficiency of the personalized clothing recommendation system, and provide users with a more personalized and satisfactory clothing purchase experience. The simulation results of prediction accuracy, sorting accuracy and coverage show that the clothing personalized recommendation model used in this paper has good recommendation effect, and the recommendation results of the recommendation system are in line with the reality, which is highly reasonable.

Idioma:: Inglés

Calendario de la edición:: 1 veces al año
Temas de la revista:: Ciencias de la vida, Ciencias de la vida, otros, Matemáticas, Matemáticas aplicadas, Matemáticas generales, Física, Física, otros

RSS Feed de revista

Research on personalized clothing recommendation system based on AIGC

Xi Lu

Publicado en línea: 26 sept 2025

Recibido: 26 ene 2025

Aceptado: 01 may 2025

DOI: https://doi.org/10.2478/amns-2025-1038

Palabras claveArtificial Intelligence Generated Content (AIGC), Multimodal Personalized Clothing Recommendation System, Graph Neural Network, GRCN, Neighborhood Matrix Normalization

© 2025 Xi Lu, published by Sciendo.

This work is licensed under the Creative Commons Attribution 4.0 International License.

Palabras clave
Artificial Intelligence Generated Content (AIGC), Multimodal Personalized Clothing Recommendation System, Graph Neural Network, GRCN, Neighborhood Matrix Normalization