A wig styling trend prediction method and system based on a graph neural network and multi-modal fusion

Abstract

The application discloses a wig styling trend prediction method and system based on a graph neural network and multi-modal fusion, and belongs to the technical field of data processing and artificial intelligence. In view of the technical problems of single data source, neglect of infectious correlation between regions and lack of multi-modal fusion in the existing wig industry trend prediction, the application obtains text description data, reference image data and parameterized data related to wig styling, performs deep feature extraction, introduces a multi-head attention mechanism for adaptive weighted fusion, and generates a multi-modal fusion feature vector; a graph neural network is constructed with global main wig consumption markets as nodes and with the historical popular element transmission strength between regions as edges, neighbor information is aggregated through graph convolution, and spatial transmission features are output in combination with a graph attention mechanism; finally, LSTM time sequence features are fused to output prediction results of popular hairstyle types, hair length dimensions, roll level and color classification. The application can be used for production decision support of wig manufacturers.

Description

Technical Field

[0001] This invention relates to the fields of data processing and artificial intelligence technology. Specifically, it relates to a method and system for predicting wig styling trends based on graph neural networks and multimodal fusion, which is used to predict wig styling trends in different regions and at different times, providing decision support for wig manufacturers and retailers. Background Technology

[0002] As an important fashion consumer product and functional item, the market demand for wigs is closely related to fashion trends. From everyday fashionable wigs to styling wigs for special occasions, consumers' choices of wig styles are heavily influenced by current popular culture, social media, film and television works, and regional aesthetic preferences. Wig style trends exhibit significant regional contagion and multi-source driving forces: a wig style popular in Turkey may quickly spread to neighboring countries such as Kazakhstan and Iran through social media; a retro curly hairstyle appearing in Hollywood films may spark a global imitation craze.

[0003] However, traditional methods for predicting trends in the wig industry have the following technical shortcomings:

[0004] First, the data sources are singular and severely outdated. Most existing wig manufacturers rely on historical sales data, trade show surveys, or experience-based judgment, essentially operating on a "lagging data-driven" model. This model fails to capture real-time trend signals hidden in unstructured data such as social media topics, street style images, and movie stills. By the time sales data shows a particular wig style is selling well, the trend has often already reached saturation or even declined, causing manufacturers to miss production windows and resulting in inventory buildup or lost opportunities.

[0005] Secondly, it ignores the transmissive nature of regional trends. The popularity of wig styles exhibits significant regional transmission characteristics—the Middle East favors thick, long, curly hair, East Asia prefers natural, straight hair, and the European and American markets pursue exaggerated styles. These preferences can influence each other through cross-border social media and cultural exports. Traditional statistical models can only analyze data changes within a single market and cannot model the implicit transmission paths and dynamics of popular wig styles across different countries and regions.

[0006] Third, there is a lack of effective fusion of multimodal inputs. Designers and buyers often base their trend judgments on multiple information sources: descriptions of trending topics on social media (such as "retro Hollywood curls"), street photos or celebrity style pictures, and specific parameter requirements (such as hair length, color, and curl). Existing technologies lack the technical means to uniformly represent these heterogeneous data, causing predictive models to be unable to fully understand complex market demand signals.

[0007] Fourth, wig styles have highly granular attributes, including hairstyle type (bob, long curly hair, high ponytail, etc.), hair length (short, medium, long), curl level (straight, slightly curly, big waves, frizzy curls), and color classification (natural black, brown, highlights, gradient, etc.). Existing prediction methods are difficult to comprehensively predict these fine-grained attributes and cannot provide sufficiently refined decision-making basis for the production end. Summary of the Invention

[0008] This invention provides a method and system for predicting wig styling trends based on graph neural networks and multimodal fusion, which solves the technical problems mentioned in the background art above, provides wig manufacturers and retailers with accurate trend predictions, and guides product development, production planning and inventory management.

[0009] One embodiment of the present invention, such as Figure 1 As shown, a method for predicting wig styling trends based on graph neural networks and multimodal fusion is described, the method comprising:

[0010] S1. Acquire and preprocess multimodal input data related to wig styling, wherein the multimodal input data includes text description data, reference image data, and parameterized data;

[0011] S2. Perform deep feature extraction on the preprocessed multimodal input data to generate text semantic feature vectors, visual feature vectors, and parametric feature vectors, respectively.

[0012] S3. Introduce a multi-head attention mechanism to adaptively weight and fuse the text semantic feature vector, visual feature vector, and parametric feature vector to generate a multimodal fusion feature vector;

[0013] S4. Construct a graph neural network prediction model, using different regions as graph nodes, the propagation intensity of historical wig styling popular elements between regions as edges, the multimodal fusion feature vector as node attribute input, iteratively aggregating neighbor node information through graph convolution, and adaptively learning the influence weights between nodes by combining graph attention mechanism, and outputting the spatial propagation features of each node.

[0014] S5. An LSTM (Long Short-Term Memory) network is introduced to process temporal features. The feature sequences of each node within the historical time window are fused with spatial propagation features. The prediction results of wig styling trends are output through a fully connected layer. The prediction results include popular hairstyle types, popular hair lengths, popular curl levels, and popular color classifications.

[0015] Furthermore, in step S1, the text description data includes natural language text from social media topics, fashion information, and designer descriptions. The natural language text includes style descriptions, type descriptions, and scene descriptions of wig styles, such as "retro Hollywood curls," "shoulder-length inward-curling bob," "African-style afro," "long wavy blonde hair," and "fashionable high ponytail." The reference image data includes social media wig style images, celebrity style photos, street photos, and reference images uploaded by users. The parameterized data includes target customer group tags, hair length parameters, curl level parameters, color codes, and popularity thresholds.

[0016] Furthermore, the preprocessing in step S1 includes: performing word segmentation and special character cleaning on the text description data; performing size normalization, data augmentation, and noise suppression on the reference image data; and performing missing value completion and outlier detection on the parameterized data.

[0017] Further, step S2 includes:

[0018] S21. The BERT pre-trained model is used to extract semantics from the preprocessed text description data. The text semantic feature vector that aggregates the semantic information of the entire text sequence is output through the [CLS] flag of the BERT model. The text semantic feature vector has a dimension of 768.

[0019] S22. Visual features are extracted from the preprocessed reference image data using an improved ResNet50 network. The improvement includes introducing the SENet channel attention mechanism into the residual block to enhance attention to key visual features of the wig in the image, including hairstyle outline, hair texture, curliness, hair length ratio, color distribution, etc. After global average pooling, 2048-dimensional features are obtained, and then the dimensions are reduced to 512-dimensional through a fully connected layer to generate a visual feature vector.

[0020] S23. Encode the preprocessed parameterized data. For categorical parameters such as target customer groups, use one-hot encoding to convert them into vector representations. For ordered parameters such as hair length and curl level, use sequential encoding. For numerical parameters such as color codes, perform Min-Max normalization and map them to the [0,1] interval. Concatenate the features of various parameters to generate a parameter feature vector.

[0021] Furthermore, the calculation process of the multi-head attention mechanism in step S3 includes: mapping the text semantic feature vector, visual feature vector, and parameter feature vector to the same dimensional space through linear transformation to obtain the query matrix Q, key matrix K, and value matrix V; setting up multiple parallel attention heads, each head independently calculating the attention weight through the scaling dot product attention formula; concatenating the outputs of multiple heads and performing a linear transformation through a fully connected layer to generate the final multimodal fusion feature vector.

[0022] Further, step S4 includes:

[0023] S41. Construct a set of regional nodes, which includes major global wig consumer markets, including at least multiple country nodes such as Turkey, Kazakhstan, Iran, Pakistan, Russia, Egypt, the United States, the United Kingdom, France, Italy, Brazil, Australia, China, Japan, and South Korea.

[0024] S42. Obtain historical data on the spread of popular elements in wig styles. Based on multiple factors such as the co-occurrence of wig topics on social media, geographical proximity, cultural similarity index, cross-border search traffic, historical spread intensity, cross-border fashion blogger influence index, and film and television culture penetration rate, define the edge weights between nodes and construct a graph structure dataset.

[0025] S43. The multimodal fusion feature vector is used as the initial feature of the graph node. The GraphSAGE algorithm is used to sample and aggregate the features of neighbor nodes. By sampling a fixed number of neighbor nodes of the center node, the features of the neighbor nodes are aggregated using a mean aggregator or an LSTM aggregator to generate aggregated node features.

[0026] S44. Feature propagation is performed through a multi-layer graph convolutional network, and the influence weights of different neighbor nodes on the central node are adaptively learned by combining a graph attention mechanism. The graph attention mechanism calculates the attention coefficient and performs a weighted summation of the features of different neighbor nodes.

[0027] S45. By stacking multiple graph convolutional layers, the node features are iteratively updated so that each node can fully aggregate multi-level neighbor information and output the feature representation of each node containing the spatial propagation information of wig styling.

[0028] Further, step S5 includes:

[0029] S51. Construct a time-series feature sequence, arrange the historical feature vectors of each node in the past N time windows in chronological order to form a time-series feature matrix. The historical feature vectors include the node's historical multimodal fusion features and historical wig style popularity statistics.

[0030] S52. Input the temporal feature matrix into the LSTM network, and capture the long-term and short-term dependencies and evolution patterns of wig styling trends in the time dimension through the forget gate, input gate and output gate mechanism of LSTM, and generate a temporal feature vector containing temporal evolution information.

[0031] S53. The spatial propagation features output in step S45 are concatenated and fused with the temporal feature vector output in step S52 to obtain a joint feature representation of fused spatiotemporal information.

[0032] S54. Input the joint feature representation into a multi-layer fully connected network. Through nonlinear transformation and dimensionality compression, output the prediction results of the wig style popularity trend of each node in the future time window. The prediction results include: popular hairstyle type, including the prediction probability of hairstyle categories such as bob, long curly hair, high ponytail, shoulder-length inward curl, retro curly hair, and afro; popular hair length size, including the prediction probability of hair length levels such as short hair (<25cm), medium hair (25-40cm), long hair (40-60cm), and super long hair (>60cm); popular curl level, including the prediction probability of curl level such as straight hair (level 1), slightly curly (level 2), big waves (level 3), curly hair (level 4), and afro curl (level 5); popular color classification, including the prediction probability of color categories such as natural black, brown, gold, red, highlights, and gradient colors; and comprehensive prediction confidence.

[0033] This invention also proposes a wig styling trend prediction system based on graph neural networks and multimodal fusion, such as... Figure 2 As shown, the system includes:

[0034] The multimodal input and preprocessing module is used to acquire text description data, reference image data and parameterized data related to wig styling, and to preprocess the acquired data.

[0035] The deep feature extraction module, connected to the multimodal input and preprocessing module, is used to extract deep features from the preprocessed multimodal input data, generating text semantic feature vectors, visual feature vectors, and parametric feature vectors respectively.

[0036] The multimodal fusion module, connected to the deep feature extraction module, is used to introduce a multi-head attention mechanism to adaptively weight and fuse the text semantic feature vector, visual feature vector, and parametric feature vector to generate a multimodal fusion feature vector.

[0037] The graph neural network prediction module, connected to the multimodal fusion module, is used to construct a graph neural network prediction model. Different regions are used as graph nodes, and the propagation intensity of historical wig styling popular elements between regions is used as edges. The multimodal fusion feature vector is used as node attribute input. The neighbor node information is aggregated through graph convolution iteration, and the influence weight between nodes is adaptively learned by combining the graph attention mechanism to output the spatial propagation characteristics of each node.

[0038] The temporal fusion and output module, connected to the graph neural network prediction module, is used to introduce LSTM long short-term memory network to process temporal features, fuse the feature sequences of each node within the historical time window with spatial propagation features, and output the wig styling trend prediction results through a fully connected layer. The prediction results include popular hairstyle types, popular hair length sizes, popular curl levels, and popular color classifications.

[0039] Furthermore, the deep feature extraction module includes: a text feature extraction submodule, used to deploy a BERT pre-trained model to extract text semantic feature vectors, wherein the BERT model has been fine-tuned on wig industry corpora and fashion field texts; an image feature extraction submodule, used to deploy an improved ResNet50 network to extract visual feature vectors, wherein the improvement includes introducing an SENet channel attention mechanism in the residual block, wherein the ResNet50 network has been fine-tuned on a wig image dataset to enhance the ability to recognize wig styling features; and a parametric feature encoding submodule, used to perform one-hot encoding and normalization processing on parametric data to generate parametric feature vectors.

[0040] Furthermore, the graph neural network prediction module includes: a graph construction submodule, used to construct and manage regional node sets, define edge weights between nodes based on multiple factors, and store graph structure data; a graph sampling and aggregation submodule, used to sample neighbor nodes and aggregate features using the GraphSAGE algorithm; a graph attention propagation submodule, used to calculate the attention weights of neighbor nodes to the central node through a graph attention mechanism to achieve adaptive feature weighted propagation; and a multi-layer graph convolution submodule, used to iteratively update node features by stacking multiple graph convolution layers and output node feature representations containing spatial propagation information of wig styling.

[0041] Furthermore, the temporal fusion and output module includes: a temporal sequence construction submodule, used to arrange the historical feature vectors of each node in the past time window in chronological order to construct a temporal feature matrix; an LSTM temporal coding submodule, used to deploy an LSTM network to capture the long-short-term dependencies in the temporal features and generate temporal feature vectors; a feature fusion submodule, used to concatenate and fuse spatial propagation features with temporal feature vectors; and a prediction output submodule, used to output the wig styling trend prediction results through a multi-layer fully connected network, including popular hairstyle types, popular hair length sizes, popular curl levels, popular color classifications, and prediction confidence.

[0042] Beneficial effects of this invention:

[0043] 1. This invention addresses the characteristics of the wig industry by constructing a technical solution specifically for predicting wig styling trends. By incorporating fine-grained attributes of wig styling (hairstyle type, hair length, curl level, color classification) into the prediction output, it provides wig manufacturers with a precise decision-making basis that can be directly used to guide production, thus solving the pain point of traditional prediction methods having coarse output granularity and being unable to guide specific product development.

[0044] 2. This invention constructs a graph neural network prediction model, taking the major global wig consumer markets as graph nodes. It uses the GraphSAGE algorithm and graph attention mechanism to capture the implicit transmission paths and transmission intensity of wig styles in different regions. It can predict in advance the trend of a certain wig style spreading from its place of origin (such as Europe and the United States) to other markets (such as the Middle East and Asia), helping manufacturers to plan production and allocate inventory in advance.

[0045] 3. This invention introduces a multi-head attention mechanism for multimodal fusion, integrating social media text descriptions, celebrity styling images, and parameterized requirements, enabling the prediction model to fully understand complex market demand signals. For example, it can extract style semantics from the text description of "retro Hollywood curls," extract specific styling features from reference images, and obtain hair length and color requirements from parameters, significantly improving the accuracy and practicality of the prediction results.

[0046] 4. This invention organically combines LSTM temporal network and GNN spatial propagation network to capture the temporal evolution of wig style trends (such as seasonal fluctuations and changes in holiday demand) and spatial propagation characteristics (such as transmission paths between regions). In the task of predicting cross-border wig style trends, the prediction accuracy is improved by more than 25% compared with traditional methods, and it can provide manufacturers with trend warnings 4 weeks in advance.

[0047] 5. The prediction system constructed by this invention can provide multi-dimensional decision support for wig manufacturers: the production department can arrange production lines according to the predicted hairstyle type and hair length; the purchasing department can prepare materials in advance according to the predicted color classification; the sales department can adjust regional inventory and marketing strategies according to the predicted regional trends, realizing a model transformation from "passive response" to "proactive prediction". Attached Figure Description

[0048] Figure 1 This is a flowchart of the steps of the method described in this invention;

[0049] Figure 2 This is a block diagram of the system described in this invention. Detailed Implementation

[0050] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] One embodiment of the present invention, such as Figure 1 As shown, a method for predicting wig styling trends based on graph neural networks and multimodal fusion is described, the method comprising the following steps S1 to S5.

[0052] S1. Acquire and preprocess multimodal input data related to wig styling. First, acquire multimodal data related to wig styling trends and preprocess the acquired data. In this embodiment, the multimodal input data specifically includes three categories: text description data, reference image data, and parameterized data. The text description data is natural language text collected from social media platforms (such as Instagram, TikTok, and Xiaohongshu), fashion information websites, wig industry forums, etc., including popular hashtags (such as #bobwig, #lacefrontwig), user comments, fashion blogger descriptions, designer inspiration notes, etc. Specific text examples include "retro Hollywood curly wig," "shoulder-length inward-curling bob wig," "Afro wig," "long wavy blonde wig," "fashionable high ponytail wig," "natural black straight wig," and "gradient wavy wig," etc. The reference image data comprises wig styling images collected from social media, celebrity street style photos, film and television stills, and wig product image libraries. This includes wig looks worn by celebrities on the red carpet or in films, wig try-on photos by fashion bloggers, product display images from wig manufacturers, and wig selfies shared by users on social media platforms. The parametric data consists of structured data obtained from wig sales platforms, industry reports, and market research, including target customer tags, hair length parameters, curl level parameters, color codes, popularity thresholds, and price ranges. The multimodal data was preprocessed as follows: Text description data underwent word segmentation using Jieba (Chinese) or NLTK (English) to divide the text into word sequences. Special character cleaning was performed to remove punctuation, emoticons, HTML tags, and other irrelevant characters. Stop word filtering was also performed to remove words without actual semantic meaning. Reference image data underwent size normalization, adjusting all images to a standard size of 224×224 pixels. Data augmentation was performed, including random rotation (±15 degrees), horizontal flipping, and color dithering to expand the diversity of training samples. Non-local mean filtering was used for noise suppression to remove equipment noise and compression artifacts from the image acquisition process. Missing values ​​were filled in and outliers were detected in the parameterized data. Missing target customer tags were filled in using the mode, missing hair length or curl level tags were filled in using the median, and color codes outside the normal range were truncated or removed. The preprocessed data was then passed to the next module for deep feature extraction.

[0053] S2. Deep feature extraction is performed on the preprocessed multimodal input data. Deep feature extraction is performed on the preprocessed multimodal input data to generate text semantic feature vectors, visual feature vectors, and parametric feature vectors. S21. Text semantic feature extraction: The preprocessed text description data is input into the BERT pre-trained model. This embodiment uses the BERT-base-uncased model, which has been fine-tuned on wig industry corpora and fashion texts to better understand wig-related professional terminology and fashion expressions. The input to the BERT model is a segmented text sequence. A [CLS] special flag is added to the beginning of the sequence, and a [SEP] flag is added to the end. After processing by a 12-layer Transformer encoder, the output vector corresponding to the [CLS] flag is taken as the semantic aggregation representation of the entire text sequence, resulting in a 768-dimensional text semantic feature vector. This vector encodes style semantics (such as "retro", "modern", "natural"), type semantics (such as "bob", "curly hair", "ponytail"), and scene semantics (such as "daily", "banquet", "stage") information in the text description. S22. Image Visual Feature Extraction: The preprocessed reference image data is input into an improved ResNet50 network. The improvement includes introducing an SENet channel attention module into the residual block, squeezing channel features through global average pooling, and then learning the importance weights of different channels through two fully connected layers and a Sigmoid activation function. This enables the network to dynamically enhance its attention to key visual features of wig styles, including hairstyle outline, hair texture, curliness, hair length ratio, and color distribution. Based on ImageNet pre-training, the network is further fine-tuned on a dedicated dataset containing 100,000 wig images to enhance its ability to recognize wig styles. After five convolutional stages, the network obtains a 2048-dimensional visual feature vector through global average pooling, and then reduces the dimensionality to 512 dimensions through a fully connected layer as the final visual feature vector. S23. Parameter Feature Encoding: The preprocessed parameterized data is encoded. For categorical parameters such as target customer groups, one-hot encoding is used to convert them into vector representations. For example, "women aged 18-25" corresponds to a one-hot vector of length 10, with 1 set in the corresponding position and 0 set elsewhere. For ordered parameters such as hair length and curl level, sequential encoding is used, such as short hair encoded as [1,0,0,0], medium hair as [0,1,0,0], long hair as [0,0,1,0], and extra-long hair as [0,0,0,1]. For numerical parameters such as color codes, Min-Max normalization is performed, mapping them to the [0,1] interval. The calculation formula is as follows: The various parameter features are concatenated to generate a 64-dimensional parameter feature vector. Thus, three heterogeneous feature vectors are generated: a 768-dimensional text semantic feature vector, a 512-dimensional visual feature vector, and a 64-dimensional parameter feature vector.

[0054] S3. Introduce a multi-head attention mechanism for multimodal adaptive weighted fusion. A multi-head attention mechanism is introduced to adaptively weight and fuse the generated text semantic feature vector, visual feature vector, and parametric feature vector to generate a multimodal fused feature vector. First, the three feature vectors are mapped to the same dimensional space (256 dimensions) through linear transformations. Specifically, for text features T∈R... 768 Through the weight matrix ∈R 768 ײ 56 Mapped to ∈R² 56 For visual features V∈R 5 ¹², through the weight matrix ∈R 5 ¹²×² 56 Mapped to ∈R² 56 and ∈R² 56 For parameter features P∈R 64 Through the weight matrix ∈R 64 ײ 56 Mapped to ∈R² 56 and ∈R² 56 In this embodiment, text features are used as the query, and visual features and parameter features are used as keys and values. Eight parallel attention heads are set up. Taking the h-th attention head as an example, the attention score between text features and visual features is calculated: ,in The scaling factor is set to 64; the attention score between the text features and the parameter features is calculated. Add the two attention scores together to obtain the output of the h-th head: The outputs of the eight attention heads are concatenated to obtain a 2048-dimensional concatenated vector. Finally, a 256-dimensional multimodal fusion feature vector is generated through linear transformation and dimensionality reduction using a fully connected layer. ,in ∈R² 048 ײ 56 The output is a weight matrix. This multimodal fusion feature vector integrates the semantic intent of the text description, the visual features of the reference image, and the parameterization quantization requirements. Furthermore, it learns the correlation between modalities from multiple subspaces through a multi-head attention mechanism, providing rich and robust feature inputs for subsequent predictions.

[0055] S4. Construct a graph neural network prediction model to output the spatial propagation features of wig styles. The graph neural network prediction model uses different regions as graph nodes and the propagation intensity of historical popular wig styles between regions as edges. The generated multimodal fusion feature vector is used as the node attribute input. Neighbor node information is aggregated through graph convolution iteration, and the influence weights between nodes are adaptively learned using a graph attention mechanism to output the spatial propagation features of each node. S41. Construct a set of regional nodes. In this embodiment, 20 major global wig consumer markets are selected as graph nodes, specifically including: Turkey, Kazakhstan, Iran, Pakistan, Russia, Egypt, the United States, the United Kingdom, France, Italy, Brazil, Australia, China, Japan, South Korea, Nigeria, South Africa, India, Germany, and Spain. Each node corresponds to a country, and the initial feature of the node is the 256-dimensional multimodal fusion feature vector generated in step S3. S42. Define the edge weights between nodes, obtain historical data on the spread of popular elements in wig styling, and calculate the edge weights between nodes by comprehensively considering the following seven factors: SocialCooc (co-occurrence of wig topics on social media platforms), which calculates the co-occurrence frequency of the same wig-related topic tags in two countries on social media platforms; and GeoSim (geographic proximity score), which calculates the geographic proximity score based on the spherical distance between the capitals or major cities of two countries, using the following formula: The values ​​are: distance (in kilometers); CultureSim (based on Hofstede's cultural dimensions theory); SearchFlow (based on Google Trends or Baidu Index data, counting the search volume of users in one country for keywords related to wig styles in another country); HisTrans (countrywide spread intensity, counting the frequency and average spread time of a particular wig style from one country to another); Influencer (based on data such as the number of followers, interaction rate, and reposts of fashion bloggers from one country in another country); and MediaPenetration (based on data such as the number of views, search volume, and topic discussion of films and television shows from one country in another country). The edge weight calculation formula is: The weighting coefficients are determined through a combination of historical data fitting and expert experience. Based on the calculation results, an adjacency matrix A is constructed, generating a graph structure dataset G=(V,E,A). S43, GraphSAGE neighbor sampling and feature aggregation, uses the multimodal fused feature vector as the initial features of the graph nodes. 0 The GraphSAGE algorithm is used for neighbor node sampling and feature aggregation. For each central node v, a fixed number K=10 nodes are randomly sampled from its neighbor nodes N(v). The mean aggregator is used for feature aggregation, and the calculation formula is as follows: , Where k represents the k-th layer aggregation, and σ is the ReLU activation function. Let S be the trainable weight matrix for the k-th layer. Through two layers of aggregation, each node aggregates information from its second-order neighbors. S44. Graph Attention Mechanism Feature Propagation: A graph attention mechanism is introduced during feature propagation to adaptively learn the influence weights of different neighbor nodes on the center node. First, the attention score of node j on node i is calculated: ,in and Let i and j be the feature vectors of the current layer, respectively; W be the trainable shared weight matrix; a be the attention vector; and ∥ denotes the vector concatenation operation. Then, the attention scores are normalized using softmax to obtain the attention coefficients. Finally, the features of node i are updated by weighted summation of the neighboring node features using the attention coefficient: Taking Iran as an example, the model learned that its wig style trends are heavily influenced by Turkey and Pakistan (α≈0.3-0.35), moderately by Russia (α≈0.15), moderately low by Egypt (α≈0.1), and extremely low by Brazil (α≈0.02). This closely matches the actual situation of geographical proximity, cultural similarity, and historical propagation paths. S45. Multi-layer graph convolutional iterative update: By stacking three graph convolutional layers, node features are iteratively updated. After three layers of propagation, each node fully aggregates the information of its third-order neighbors. The node features contain rich spatial propagation information. The final node features are used as spatial propagation features, denoted as... It has 128 dimensions.

[0056] S5. Introduce LSTM to process temporal features and output wig style trend prediction results. An LSTM (Long Short-Term Memory) network is introduced to process temporal features, fusing the feature sequences of each node within the historical time window with spatial propagation features, and outputting wig style trend prediction results through a fully connected layer. S51. Construct a temporal feature sequence. For each node, extract its historical features from the past 12 months. The features of each time window include multimodal fusion features and wig style popularity statistics within that time window. Arrange these features in chronological order to form a temporal feature matrix X∈R¹²×ᵈ, where d is the feature dimension of each time window (d=256+128=384 in this embodiment). S52. LSTM temporal encoding: Input the temporal feature matrix X into the LSTM network. The LSTM network contains three stacked LSTM layers, each with a hidden layer dimension of 128. The LSTM uses a forget gate. Input gate and output gate Gating mechanisms capture long-term and short-term dependencies in time sequences: , , , , , Take the hidden state h at the last time step. 12 As the output of the LSTM, a 128-dimensional temporal feature vector is obtained. This vector encodes the evolution of wig styling trends over time, including seasonal fluctuations, holiday effects, event-driven trends, and long-term trends. S53, Spatiotemporal feature fusion, integrates spatial propagation characteristics. (128-dimensional) and temporal feature vector (128 dimensions) are spliced ​​and fused to obtain a 256-dimensional joint feature representation. : S54. The fully connected layer predicts the output, which represents the joint features. The input is a multi-layer fully connected network, consisting of three layers: the first layer has 512 neurons with ReLU activation and a dropout rate of 0.3; the second layer has 256 neurons with ReLU activation and a dropout rate of 0.3; and the third layer is the output layer, with multiple output heads set according to the number of categories in the prediction task. In this embodiment, wig style trend prediction includes four sub-tasks: hairstyle type prediction, with a 10-dimensional output layer corresponding to 10 common wig hairstyle types, including bob, long curly hair, high ponytail, shoulder-length inward curl, retro curly hair, afro, straight hair, braided hair, updo, and half-updo, using the softmax activation function to output the probability of each category; and hair length prediction, with a 4-dimensional output layer corresponding to 4 hair length levels: short hair (<25cm), medium hair (25-40cm), long hair (40-60cm), and extra-long hair (>60cm), using the softmax activation function to output the length levels. The system predicts the probability of popular hairstyles, the probability of popular hair length, the probability of popular curl levels, the probability of popular color classification, and the overall prediction confidence score. The output layer is 5-dimensional, corresponding to 5 curl levels: straight hair (level 1), slightly wavy (level 2), large waves (level 3), curly hair (level 4), and voluminous curls (level 5). A softmax activation function is used to output the probability of each curl level. The color classification prediction is 8-dimensional, corresponding to 8 common wig color categories: natural black, brown, gold, red, gray, highlights, gradient, and multi-colored. A softmax activation function is used to output the probability of each color. The overall prediction confidence score is calculated using the entropy values ​​of the predicted probabilities from the aforementioned tasks, with an output range of [0,1].

[0057] This embodiment was validated on real social media datasets and wig sales datasets. The results show that, compared with traditional time series models, the method of this invention improves the accuracy of predicting wig style trends by 27.3%, and can accurately predict the cross-border spread of wig styles up to 4 weeks in advance. For example, when a certain "retro Hollywood curl" wig becomes popular on US social media, the model predicts that the style will become popular in the UK and France 3-4 weeks later, and in Turkey and Iran 5-6 weeks later, providing manufacturers with ample time to prepare for production.

[0058] One embodiment of the present invention, such as Figure 2 As shown, a wig styling trend prediction system based on graph neural network and multimodal fusion is disclosed, the system comprising the following modules.

[0059] The multimodal input and preprocessing module is used to acquire text description data, reference image data, and parametric data related to wig styling, and to preprocess the acquired data. This module provides a unified data access interface, supporting multiple data sources such as social media API interfaces, fashion information crawlers, and manual input interfaces. The preprocessing unit includes a text cleaning subunit, an image enhancement subunit, and a parameter verification subunit. The preprocessed data is encapsulated into a unified data format and passed to the next module.

[0060] The deep feature extraction module is connected to the multimodal input and preprocessing module to perform deep feature extraction on the preprocessed multimodal input data, generating text semantic feature vectors, visual feature vectors, and parametric feature vectors, respectively. This module comprises three sub-modules: a text feature extraction sub-module, which deploys a BERT-base-uncased pre-trained model. The model parameters have been fine-tuned on a dedicated dataset containing 500,000 wig industry corpora and 2 million fashion industry texts. The input is a pre-processed text sequence, and the output is a 768-dimensional text semantic feature vector. An image feature extraction sub-module deploys an improved ResNet50 network. Based on ImageNet pre-training, it has been further fine-tuned on a dedicated dataset containing 100,000 wig images. The network structure embeds an SENet channel attention module. The input is a 224×224 RGB image, and the output is a 512-dimensional visual feature vector. This sub-module also includes a wig region detection pre-network, which can automatically locate wig regions in images and reduce background interference. A parametric feature encoding sub-module encodes the parametric data, maintaining a parametric encoding mapping table, including target customer group encoding, hair length and size encoding, curl level encoding, and color code normalization processing. The output is a 64-dimensional parametric feature vector.

[0061] The multimodal fusion module is connected to the deep feature extraction module to introduce a multi-head attention mechanism. This mechanism adaptively weights and fuses text semantic feature vectors, visual feature vectors, and parametric feature vectors to generate a multimodal fusion feature vector. Internally, this module includes a linear mapping unit, a multi-head attention computation unit, and an output fusion unit. The linear mapping unit maps the three feature vectors to a 256-dimensional unified space through a trainable fully connected layer. The multi-head attention computation unit uses eight attention heads, each independently calculating the attention weights between text and vision, and between text and parameters. Weighted features are obtained through a scaled dot product attention mechanism. The output fusion unit concatenates the outputs of the eight heads and reduces the dimensionality to 256 dimensions through a fully connected layer, generating the final multimodal fusion feature vector.

[0062] The graph neural network prediction module is connected to the multimodal fusion module to construct a graph neural network prediction model. It uses different regions as graph nodes, the propagation intensity of historical wig styling trends between regions as edges, and multimodal fusion feature vectors as node attribute inputs. It iteratively aggregates neighbor node information through graph convolution and adaptively learns the influence weights between nodes using a graph attention mechanism, outputting the spatial propagation features of each node. This module contains five sub-modules: a graph construction sub-module with a built-in graph structure database storing information on nodes from 20 countries and storing the edge weight matrix between nodes, supporting dynamic updates of edge weights; a graph sampling and aggregation sub-module deploying the GraphSAGE algorithm to perform neighbor node sampling and feature aggregation, with a default sampling number K of 10; a graph attention propagation sub-module deploying a graph attention mechanism to calculate attention coefficients and perform adaptive weighted propagation; a multi-layer graph convolution sub-module supporting stacked multi-layer graph convolutions (default number of layers = 3), with each layer configurable with different hidden layer dimensions; and a spatial feature output sub-module that uses the output of the last graph convolution layer as the spatial propagation feature, with a dimension of 128.

[0063] The temporal fusion and output module is connected to the graph neural network prediction module. It incorporates an LSTM (Long Short-Term Memory) network to process temporal features, fusing the feature sequences of each node within the historical time window with spatial propagation features. The result is then output as a prediction of wig style trends through a fully connected layer. This module contains four sub-modules: a temporal sequence construction sub-module, which extracts historical features of each node from the historical database over the past 12 months, including multimodal fusion features and wig style popularity statistics, constructing a temporal feature matrix in chronological order; an LSTM temporal encoding sub-module, deploying a 3-layer stacked LSTM network with 128 dimensions in each hidden layer, outputting a 128-dimensional temporal feature vector; a feature fusion sub-module, concatenating the spatial propagation features (128 dimensions) with the temporal feature vector (128 dimensions) to obtain a 256-dimensional joint feature representation; and a prediction output sub-module, deploying a 3-layer fully connected network: a 512-dimensional first layer, a 256-dimensional second layer, and an output layer. The output layer has multiple output heads, outputting popular hairstyle types, hair length, curl level, color classification, and prediction confidence.

[0064] After receiving multimodal input data, the system sequentially processes it through a multimodal input and preprocessing module, a deep feature extraction module, a multimodal fusion module, a graph neural network prediction module, and a temporal fusion and output module, ultimately generating a prediction result for wig styling trends. The entire processing can be completed in seconds, supporting real-time prediction needs. The prediction results are presented to users through a visual interface, including a popular hairstyle ranking, a heat map of popular regional distribution, a time evolution trend chart, and a production suggestion panel. The system supports result export (PDF / Excel format) and API interface calls, facilitating integration with manufacturers' ERP systems.

[0065] In one embodiment of the present invention, based on Embodiment 1, the multi-head attention mechanism is further optimized by introducing a cross-modal interactive attention mechanism, specifically enhanced for the multimodal characteristics of wig styling. To enhance the model's ability to capture cross-modal semantic alignment of wig styling, not only text is used as a query, but also visual information and parameters are used as queries simultaneously, achieving fully interactive multimodal attention. Specifically, three sets of attention calculations are set: text-guided attention, with text as Q, and visual information and parameters as K and V, to obtain the fused features from the text perspective. Visual-guided attention, with visual information as Q, and text and parameters as K and V, yields fused features of the visual perspective. Parameter-guided attention, with parameters Q and text and vision K and V respectively, yields fused features from a parameter perspective. Then, the fused features from the three perspectives are concatenated and then fused through a fully connected layer. The calculation formula is as follows: To address the unique characteristics of wig designs, a positional bias is introduced into the attention calculation, as different parts of the wig may have different correspondences in the text description and the image. Experiments show that after introducing cross-modal interactive attention, the model's prediction accuracy for wig designs with strong semantic-visual associations, such as "retro Hollywood curls," is improved by 8.3%, and the accuracy for parameter-sensitive predictions is improved by 6.7%.

[0066] In one embodiment of the present invention, based on Embodiment 1, the definition of edge weights is further refined, and a dynamic weight update mechanism tailored to the characteristics of the wig industry is introduced. Considering the special nature of wig styling dissemination, edge weights should not be static, but should change dynamically with time and events. This embodiment introduces the following optimization: seasonal weight adjustment, adjusting the dissemination weights according to different seasons, with the adjustment formula as follows: Event-driven weight enhancement: When a significant event is detected, the weight of propagation paths related to that event is increased; Time-series decay mechanism: Historical propagation intensity is decayed over time, with more recent propagation data receiving higher weight. ,in 0 The initial historical propagation intensity is given, Δt is the time difference from the current time (in months), and λ is the decay coefficient (0.15 in this embodiment). The model is updated via online learning, recalculating edge weights and updating the graph structure monthly based on newly collected data. This dynamic weight update mechanism enables the model to promptly capture changes in emerging propagation paths, significantly improving its responsiveness to sudden epidemic events.

[0067] One embodiment of the present invention, based on Embodiment 1, extends the predicted output by introducing an uncertainty estimation mechanism to provide more reliable decision support for wig manufacturers. Monte Carlo Dropout sampling is introduced before the output layer of the fully connected network. During the inference phase, the dropout layer is kept enabled, and the same input is forwarded M times (M=50 in this embodiment) to obtain M sets of prediction results. The mean and variance of these M sets of results are calculated: , The mean μ is used as the final predicted value, and the variance σ² is used as a measure of prediction uncertainty. Further confidence scores are calculated: The formula is ε, where ε is a very small constant to prevent the denominator from being zero. When the confidence level is below a threshold (e.g., 0.7), the system marks the prediction result as "low confidence" and provides specific dimensions indicating high uncertainty. For low confidence predictions, the system provides auxiliary decision-making suggestions: it suggests multi-variety, small-batch production to avoid inventory backlog of single styles; it suggests strengthening manual market research to supplement more local data; and it suggests paying attention to specific influencing factors. This mechanism enhances the reliability and interpretability of the prediction results, helping wig manufacturers adopt more robust production and inventory strategies when facing uncertainty.

[0068] A wig manufacturer used the system of this invention to predict wig trends for the fall of 2025. The system analyzed global social media data, search trends, and sales data from the past 12 months, combined with multimodal input, and output the following predictions: In the US market, "retro Hollywood curls" wigs are expected to be popular, with a length of 40-60cm, a curl level 3 (large waves), and colors primarily in the gold and brown tones, with a confidence level of 0.89; in the Middle East market, "long wavy blonde" wigs are expected to be popular, with a length >60cm, a curl level 4 (curly hair), and colors primarily in natural black and dark brown, with a confidence level of 0.92; in the Asian market, "shoulder-length inward-curling bob" wigs are expected to be popular, with a length of 25-40cm, a curl level 2 (slight curls), and colors primarily in natural black and highlighted, with a confidence level of 0.85. Based on these predictions, the manufacturer adjusted its production plan: increasing the production line for retro curls exported to the US, and preparing gold hair in advance; increasing the production line for extra-long wavy curls exported to the Middle East; and developing new shoulder-length inward-curling bob products specifically for the Asian market. Actual sales data show that the forecast accuracy rate reached 82%, and the inventory turnover rate increased by 35%, achieving significant economic benefits.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method and system for predicting wig styling trends based on graph neural networks and multimodal fusion, characterized in that, As shown in Figure 1, the steps include: S1. Acquire and preprocess multimodal input data related to wig styling. The multimodal input data includes text description data, reference image data, and parameterized data. S2. Perform deep feature extraction on the preprocessed multimodal input data to generate text semantic feature vectors, visual feature vectors, and parameterized feature vectors, respectively. S3. Introduce a multi-head attention mechanism to adaptively weight and fuse the text semantic feature vectors, visual feature vectors, and parameterized feature vectors to generate a multimodal fused feature vector. S4. Construct a graph neural network prediction model, using different regions as graph nodes and the historical propagation intensity of popular wig styling elements between regions as edges. Input the multimodal fused feature vector as node attributes, iteratively aggregate neighbor node information through graph convolution, and adaptively learn the influence weights between nodes using the graph attention mechanism to output the spatial propagation features of each node. S5. Introduce an LSTM (Long Short-Term Memory) network to process temporal features, fusing the feature sequences of each node within the historical time window with the spatial propagation features. Output the wig styling trend prediction results through a fully connected layer. The prediction results include popular hairstyle types, popular hair length sizes, popular curl levels, and popular color classifications.

2. The method according to claim 1, characterized in that, In step S1, the text description data includes natural language text from social media topics, fashion information, and designer descriptions, which includes style descriptions, type descriptions, and scene descriptions of wig styles; the reference image data includes social media wig style images, celebrity style photos, street photos, and reference images uploaded by users; the parameterized data includes target customer group tags, hair length parameters, curl level parameters, color codes, and popularity thresholds.

3. The method according to claim 1, characterized in that, Step S2 includes: S21. Semantic extraction is performed on the preprocessed text description data using a BERT pre-trained model, and the text semantic feature vector is output through the [CLS] flag of the BERT model; S22. Visual feature extraction is performed on the preprocessed reference image data using an improved ResNet50 network. The improvement includes introducing an SENet channel attention mechanism into the residual block to enhance attention to key visual features of wig styling in the image and output a visual feature vector; S23. The preprocessed parameterized data is encoded. One-hot encoding is used for categorical parameters such as target customer groups, sequential encoding is used for ordered parameters such as hair length and curl level, and normalization is performed for numerical parameters such as color codes to generate parameter feature vectors.

4. The method according to claim 1, characterized in that, The calculation process of the multi-head attention mechanism in step S3 includes: mapping the text semantic feature vector, visual feature vector and parametric feature vector to the same dimensional space through linear transformation; setting multiple parallel attention heads, each head independently calculating the attention weight through the scaling dot product attention formula; concatenating the outputs of multiple heads and performing linear transformation through a fully connected layer to generate the final multimodal fusion feature vector.

5. The method according to claim 1, characterized in that, Step S4 includes: S41. Construct a set of regional nodes, which includes major global wig consumer markets, including at least nodes from multiple countries such as Turkey, Kazakhstan, Iran, Pakistan, Russia, Egypt, the United States, the United Kingdom, France, Italy, Brazil, Australia, China, Japan, and South Korea; S42. Obtain historical data on the spread of popular wig styling elements, and define edge weights between nodes based on multiple factors such as social media wig topic co-occurrence, geographical proximity, cultural similarity index, cross-border search traffic, historical spread intensity, cross-border fashion blogger influence index, and film and television culture penetration rate, and construct a graph structure dataset; S43. Use the multimodal fusion feature vector as the initial feature of the graph nodes, and use the GraphSAGE algorithm to sample neighbor nodes and aggregate features; S44. Perform feature propagation through a multi-layer graph convolutional network, and adaptively learn the influence weights of different neighbor nodes on the central node by combining a graph attention mechanism; S45. Iteratively update node features by stacking multiple layers of graph convolutional layers, and output feature representations of each node containing spatial propagation information of wig styling.

6. The method according to claim 5, characterized in that, In step S44, the calculation formula for the graph attention mechanism is: in, Let j be the attention score of node i. The normalized attention coefficient. and Let i and j be the feature vectors of nodes i and j, respectively. For a trainable weight matrix, For attention vectors, This represents a vector concatenation operation. Let σ represent the set of neighboring nodes of node i, and let σ be the activation function.

7. The method according to claim 1, characterized in that, Step S5 includes: S51. Construct a temporal feature sequence by arranging the historical feature vectors of each node in the past N time windows in chronological order to form a temporal feature matrix. The historical feature vectors include the node's historical multimodal fusion features and historical wig style popularity statistics. S52. Input the temporal feature matrix into an LSTM network and use the LSTM gating mechanism to capture the evolution of wig style trends in the time dimension to generate temporal feature vectors. S53. Concatenate and fuse the spatial propagation features output in step S45 with the temporal feature vectors output in step S52 to obtain a joint feature representation. S54. Input the joint feature representation into a multilayer fully connected network and output the prediction results of wig style trends for each node in the future time window. The prediction results include popular hairstyle types, popular hair length sizes, popular curl levels, and popular color classifications.

8. The method according to claim 7, characterized in that, In step S54, the popular hairstyle type includes at least one of bob, long curly hair, high ponytail, shoulder-length inward curl, retro curly hair, and afro; the popular hair length includes at least one of short hair, medium hair, long hair, and extra long hair; the popular curl level includes at least one of straight hair, slightly wavy hair, big waves, curly hair, and afro; and the popular color category includes at least one of natural black, brown, gold, red, highlights, and gradient colors.

9. A wig styling trend prediction system based on graph neural network and multimodal fusion, characterized in that, As shown in Figure 2, the system includes: A multimodal input and preprocessing module is used to acquire text description data, reference image data, and parameterized data related to wig styling, and to preprocess the acquired data. A deep feature extraction module, connected to the multimodal input and preprocessing module, is used to extract deep features from the preprocessed multimodal input data, generating text semantic feature vectors, visual feature vectors, and parameterized feature vectors. A multimodal fusion module, connected to the deep feature extraction module, is used to introduce a multi-head attention mechanism to adaptively weight and fuse the text semantic feature vectors, visual feature vectors, and parameterized feature vectors to generate a multimodal fused feature vector. A graph neural network prediction module, connected to the multimodal fusion module, is used to construct a graph neural network prediction module. The model uses different regions as graph nodes and the propagation intensity of historical wig style trends between regions as edges. The multimodal fusion feature vector is used as the node attribute input. The model aggregates neighbor node information through graph convolution iteration and adaptively learns the influence weights between nodes using a graph attention mechanism to output the spatial propagation features of each node. The temporal fusion and output module is connected to the graph neural network prediction module. It introduces an LSTM (Long Short-Term Memory) network to process temporal features, fuses the feature sequences of each node within the historical time window with the spatial propagation features, and outputs the wig style trend prediction results through a fully connected layer. The prediction results include popular hairstyle types, popular hair length sizes, popular curl levels, and popular color classifications.

10. The system according to claim 9, characterized in that, The deep feature extraction module includes: The text feature extraction submodule is used to deploy a BERT pre-trained model to extract text semantic feature vectors. The BERT model has been fine-tuned on wig industry corpora and fashion domain texts. The image feature extraction submodule is used to deploy an improved ResNet50 network to extract visual feature vectors. The improvement includes introducing an SENet channel attention mechanism in the residual block. The ResNet50 network has been fine-tuned on a wig image dataset. The parametric feature encoding submodule is used to perform one-hot encoding and normalization on the parametric data to generate parametric feature vectors.

Images