Dance costume recommendation method for guaranteeing cultural authenticity and artistic coherence

Abstract

The invention discloses a dance costume recommendation method for guaranteeing cultural authenticity and artistic continuity, which comprises the following steps: constructing a recommendation model, and predicting quintuple lt by the recommendation model; u, d, c, g, rgt; u represents a specific user, d represents a dance type, c represents a cultural background, g represents an existing costume selection, r represents a recommended costume, and calculation of the total preference score comprises calculation of the total preference score based on a matrix decomposition framework. Calculating a user preference score by fusing the influence of dance types, cultural backgrounds and existing costume selections, predicting the matching degree of the existing costume selections and recommended costumes by considering the constraints of the dance types and the cultural backgrounds, calculating costume matching scores, and calculating recommended costume similarity scores based on user historical selections; and measuring the similarity between the recommended costume and the historical matched costume selected by the existing costume, and calculating a similarity score. The problem of inaccurate dance costume recommendation is solved, and recommendation accuracy and culture and art coherence are ensured.

Description

Technical Field

[0001] The present invention relates to a dance costume recommendation method, and in particular to a dance costume recommendation method that ensures cultural authenticity and artistic consistency. Background Art

[0002] Traditional clothing recommendation systems typically make recommendations based on a user's historical choices or a single feature of clothing (such as color or style), failing to fully consider the complex interrelationships among dance types, cultural backgrounds, and artistic styles. This limitation often results in recommendations that lack cultural authenticity and artistic coherence. For example, when recommending clothing for the Dai Peacock Dance, the system may overlook the cultural context of the Water Splashing Festival and the significant differences in living environments and clothing styles between different branches, resulting in recommended clothing that is inconsistent with the cultural context. Alternatively, when a headdress is already present, the recommended bottoms may be stylistically inconsistent.

[0003] Existing approaches for clothing recommendations primarily rely on collaborative filtering and content recommendation. Collaborative filtering leverages user behavior data to mine preferences, but is ineffective in cold-start scenarios. Content recommendation methods, based on visual or textual descriptions of clothing, struggle to capture the deep relationship between users' personalized needs and cultural context. Furthermore, the integration of multimodal data is underutilized in existing systems, resulting in inaccurate recommendations when matching dance genres and cultural backgrounds. Summary of the Invention

[0004] In view of the above-mentioned defects of the prior art, the present invention provides a dance costume recommendation method that ensures cultural authenticity and artistic consistency, thereby solving the problem of lack of cultural authenticity and artistic consistency in costume recommendations.

[0005] The technical solution of the present invention is as follows: A method for recommending dance costumes that ensures cultural authenticity and artistic consistency, comprising: constructing a recommendation model, and using the recommendation model to predict a five-tuple<u,d,c,g,r> The total preference score of is ranked according to the total preference score, and clothing recommendations are made, where u represents a specific user, d represents the dance type, c represents the cultural background, g represents the existing clothing selection, and r represents the recommended clothing;

[0006] The recommendation model predicts quintuples<u,d,c,g,r> The total preference score consists of the following steps:

[0007] Based on the matrix decomposition framework, the user preference score is calculated by integrating the influence of dance type, cultural background and existing clothing choices.

[0008] Considering the constraints of dance type and cultural background, predict the matching degree between existing costume choices and recommended costumes, and calculate the costume matching score

[0009] Calculate the similarity score of recommended clothing based on user historical selection

[0010] Measure the similarity between the recommended clothing and the historical matching clothing of the existing clothing selections and calculate the similarity score

[0011] Calculating the total preference score Among them, μ, φ, is the weight parameter.

[0012] Furthermore, the user preference score in, and is the normalized latent embedding of users and clothing, and To provide visual and textual representations of recommended clothing items, and is the embedding vector of dance type and cultural background, and Provides visual and textual representations of existing clothing items. and are the user’s visual and textual preference representations, α, β u , β r are global bias, user bias and clothing bias respectively, and η is the weight parameter.

[0013] Furthermore, the clothing matching score in, Visual and textual representations of existing clothing specifically for matching modeling, is the visual and textual representation of recommended clothing specifically for matching modeling, γ2 and δ are weight parameters, It is a fusion representation of the visual features of existing costumes after they are transferred to the target cultural background style, and then combined with the text features of existing costumes, dance type embedding, and cultural background embedding. The visual features of the recommended costumes are transferred to the target cultural background style and then fused with the text features of the recommended costumes, dance type embedding and cultural background embedding.

[0014] Furthermore, the recommended clothing similarity score in, and The average visual and textual representations selected for the user's history, and is the visual and textual regularized representation of the recommended clothing, and γ3 is the weight parameter.

[0015] Furthermore, the similarity score in, and Average visual and textual representations of historically matched costumes, and is the visual and textual regularized representation of the recommended clothing, and γ4 is the weight parameter.

[0016] Furthermore, the recommendation model adopts the loss function optimization,

[0017] in, is the Bayesian personalized ranking loss function, is a training dataset, which includes user u, dance type d, cultural background c, existing clothing g, and positive sample clothing r + and negative sample clothing r - combination of and are the overall preference scores of positive samples and negative samples respectively, σ is the Sigmoid function, λ is the regularization coefficient, is the recommended model parameter Θ F The Frobenius norm of ;

[0018] is the cultural consistency regularization term, e c is the embedding vector of cultural background c, Embed the cultural characteristics of clothing r, is the square of the Euclidean distance;

[0019] γ1 is the cultural regularization weight.

[0020] Furthermore, the positive sample clothing r + Select from the user's historical purchases or high-rated clothing, negative sample clothing r - Determined by random sampling or clothing that the user explicitly dislikes.

[0021] Furthermore, the visual and textual representations of existing and recommended clothing items when calculating the user preference scores are obtained by extracting visual and textual features of clothing items using a pre-trained deep neural network, performing nonlinear transformations on them using a multi-layer perceptron, and then normalizing them using batch feature scaling.

[0022] Furthermore, the visual and textual representations of existing clothing when calculating the clothing matching score are obtained by extracting visual and textual features of clothing using a pre-trained deep neural network, performing nonlinear transformations on them using a multi-layer perceptron, and then normalizing them using batch feature scaling.

[0023] The visual and textual representations of the recommended clothing when calculating the clothing matching score are obtained by performing nonlinear transformation of the visual and textual features of the clothing extracted by a pre-trained deep neural network through a multi-layer perceptron and then normalizing them through batch feature scaling. Then, the normalized visual representation of the existing clothing and the visual representation of the recommended clothing are used as input sequences to calculate the attention weight of the Transformer, and the normalized visual representation of the recommended clothing is weighted to obtain the final visual representation of the recommended clothing. The normalized textual representation of the existing clothing and the textual representation of the recommended clothing are used as input sequences to calculate the attention weight of the Transformer, and the normalized textual representation of the recommended clothing is weighted to obtain the final textual representation of the recommended clothing.

[0024] Furthermore, the multilayer perceptron performs nonlinear transformation specifically as follows: for the original feature X

[0025]

[0026] Among them, W k and b k are the projection matrix and bias vector of the kth layer respectively; K is the number of MLP layers; σ is the Sigmoid activation function, and the output feature after transformation is X K .

[0027] The advantages of the technical solution provided by the present invention are:

[0028] Existing folk dance costume recommendation systems are often based on single features (such as color, style) or collaborative filtering methods, ignoring the cultural background and artistic style behind the folk dance, resulting in recommended costumes that may not conform to the cultural norms of specific dance scenes. The present invention ensures that the recommended costumes are highly coordinated in terms of visual elements, cultural connotations and artistic style by constructing a multimodal recommendation framework that integrates dance types, cultural backgrounds and existing costume choices. In addition, the system ensures that the recommended costumes are in line with national cultural traditions through cultural consistency regularization, prevents cultural misuse or mismatching, and improves the accuracy and professionalism of folk dance costume recommendations.

[0029] Traditional clothing recommendation systems find it difficult to fully utilize multimodal data such as visual and textual data, resulting in insufficient matching of recommended clothing in terms of artistic style and cultural attributes. The present invention uses a pre-trained deep learning model to extract the visual features (color, pattern, style) and text features (description, cultural background) of clothing, and performs feature transformation and batch feature scaling normalization through a multi-layer perceptron to ensure that information from different modalities can be effectively integrated. At the same time, the system introduces matrix decomposition technology to integrate user preferences, dance types, cultural background and existing clothing choices to achieve accurate clothing matching and personalized recommendations. For example, when a user chooses Mongolian dance and already has traditional riding boots as a matching element, the system will give priority to recommending clothing with the characteristics of the Humming culture, and will not mistakenly recommend clothing with other ethnic styles, making the recommendation results more accurate, coherent and in line with the cultural context.

[0030] Traditional collaborative filtering recommendation methods struggle to provide accurate recommendations in cold-start scenarios and lack consideration for the overall coordination of clothing combinations. This invention uses consistency regularization modeling to ensure that recommended clothing remains consistent with the user's historical choices, dance genres, and overall cultural background, improving the stability and artistic coordination of recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of a flow chart of a dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below with reference to the following examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading this description, various equivalent modifications to this description by those skilled in the art fall within the scope defined by the claims appended to this application.

[0033] The dance costume recommendation method of this embodiment, which ensures cultural authenticity and artistic consistency, includes:

[0034] S1. Use a pre-trained deep learning model to extract the visual and textual features of Chinese folk dance costumes, and perform feature transformation and batch feature scaling normalization through a multi-layer perceptron to ensure the effective fusion of multimodal information.

[0035] The goal of this invention is to recommend clothing or accessories that match the dance type, cultural background, or existing clothing choices provided by the user. For example, if the user enters "Dai Peacock Dance" and "Water Splashing Festival" as conditions, skirts and accessories that are consistent with Dai cultural characteristics should be recommended. To clearly define this problem, the following mathematical set is introduced:

[0036] User Collection: Represents all potential users.

[0037] Dance type collection: Indicates Chinese folk dance types, such as Mongolian dance, Dai dance, etc.

[0038] Cultural Background Collection: Indicates relevant cultural background.

[0039] Existing clothing selection collection: Indicates that the user has selected clothing, such as tops, headwear, etc.

[0040] Recommended clothing collection: Indicates recommended clothing.

[0041] For ease of description, simplified symbols are used: u represents a specific user, d represents the dance type, c represents the cultural background, g represents the existing clothing selection, and r represents the recommended clothing. Can be used with specific and Associate and generate quintuples through interaction<u,d,c,g,r> , represents the user’s preference or selection record for the recommended clothing r under given conditions.<g,r> Use multimodal feature descriptions, including:

[0042] Visual features: capture information such as color, pattern, style, etc., by vector (already have clothing) and (Recommended clothing) indicates.

[0043] Text features: Contains clothing description and cultural background information, represented by vector and Indicates that d v and d w are the dimensions of visual and textual features respectively.

[0044] Dance type d and cultural background c are used as category features and are represented by embedding vectors: and where d e is the embedding dimension.

[0045] This invention aims to build a recommendation model Prediction quintuple<u,d,c,g,r> Preference score:

[0046]

[0047] Where Θ is the set of model parameters, score It represents the possibility of user u choosing r under conditions d, c, and g.

[0048] Fashion recommendations rely on multimodal features, including visual features (color, pattern, style) and textual features (description, cultural context). These features are extracted using a pre-trained ResNet-50 network. To adapt to the recommendation task, the raw features need to be transformed and normalized to generate representations suitable for subsequent modeling.

[0049] Taking the visual features of recommended clothing r as an example, the original feature v r The nonlinear transformation is performed through a multi-layer perceptron. The process is as follows:

[0050]

[0051] Among them, W k and b k are the projection matrix and bias vector of the kth layer respectively; K is the number of MLP layers; σ is the Sigmoid activation function.

[0052] The output after transformation is

[0053] where d ′ v is the transformed dimension. Similarly, the text feature w r Through another MLP module MLP w Processing, generation

[0054] Since the scale difference between visual and text features may affect the modeling effect, batch feature scaling technology is used for normalization before interactive modeling. Taking visual features as an example, given a batch of clothing representation tensors (l is the batch size), the scaling formula along dimension m is:

[0055]

[0056] in

[0057]

[0058] ∈=1e-12 is a small constant to avoid division by zero. After scaling, we get Text features are also processed as This step ensures that multimodal features contribute evenly in modeling.

[0059] S2. Predict preferences by integrating user preferences, dance types, cultural background, and existing clothing features through matrix decomposition, evaluate clothing matching, and introduce consistency regularization to ensure that the recommended clothing is coordinated in terms of cultural authenticity and artistic coherence.

[0060] User preference modeling is based on a matrix factorization (MF) framework, integrating the influence of dance type, cultural background, and existing clothing choices to predict user preferences for recommended clothing. The user preference scoring function is designed as follows:

[0061]

[0062] in, and is the normalized latent embedding vector of users and clothing, representing the basic preferences of users and clothing; and The visual and textual representations of recommended clothing are dedicated to preference modeling and are processed by MLP transformation and batch feature scaling normalization in S1; and Introducing cultural context information into the embedding vectors of dance types and cultural backgrounds; and The visual and textual representation of existing clothing, after the MLP transformation and batch feature scaling normalization in S1, reflects the impact of existing choices on recommendations; and is the user's preference representation for visual and text features, which is generated by aggregating user historical interaction data after MLP transformation and batch feature scaling normalization in S1; α, β u , β r They are global bias, user bias and clothing bias, which are used to correct individual differences; η∈[0,1] is a dynamic weight parameter, which is adaptively adjusted through the subsequent attention mechanism to balance the importance of visual and textual modalities.

[0063] Embed Table (User Embed E U 、Clothing embedded E R , dance type embedding E D 、Cultural background embedded in E C ) After initialization, it is normalized by L2 norm to ensure computational stability.

[0064] Costume matching modeling aims to predict the matching degree between existing costume g and recommended costume r, taking into account the constraints of dance type d and cultural background c. The scoring function is designed as follows:

[0065]

[0066] in, and It is a visual and textual representation dedicated to matching modeling, which is distinguished from the preference modeling features to avoid information confusion. It undergoes the MLP transformation and batch feature scaling normalization in S1; γ2∈[0,1] is a weight parameter that adjusts the relative contribution of visual and textual modalities; and Ensure that the matching results are consistent with the cultural semantics of the dance type and cultural background, and δ is the weight parameter.

[0067] for and To enhance the robustness of the matching model, a Transformer-based attention mechanism is introduced to dynamically adjust the weights of the features of existing and recommended clothing. The specific steps are:

[0068] Will and As an input sequence, calculate the self-attention weight: in W Q 、W K 、W V is the learnable projection matrix.

[0069] Output attention-weighted features Replace the original Used for subsequent matching score calculation. Text features The same approach applies.

[0070] for and In order to further enhance the cultural authenticity of recommended clothing, a cultural style transfer module is introduced. Based on the pre-trained generative adversarial network (GAN), the visual features of the recommended clothing are transferred to the Transfer the style to the target cultural background c and generate enhanced features The specific steps are:

[0071] Using cultural context embedding As conditional input, drive the generator G:

[0072] Discriminator D evaluation The distribution consistency with the real cultural clothing characteristics, the optimization goal is:

[0073] After the migration With the original features Weighted fusion: Where λ∈[0,1] is the fusion weight.

[0074] Design a dynamic feature fusion layer to combine visual features Text features Dance Type Embedding and cultural context embeddedness Fusion into a unified representation The fusion process is as follows:

[0075] Calculate the attention weight of each modality: ai =softmax(W a tanh(W i f i +b i ))in is the modal feature, W a 、W i 、b i are learnable parameters.

[0076] Weighted fusion: Then input the fully connected layer to further extract high-level features to obtain the final This is achieved by changing the aforementioned part for recommended clothing to the part for existing clothing and performing the same process.

[0077] To ensure that the recommendation results are consistent with the user's historical preferences, user preference consistency regularization is introduced. Based on the user's historical choices (aggregated into a list u by similar dance types or cultural backgrounds r ), calculate the similarity score of the recommended clothing r:

[0078]

[0079] in, and The average visual and textual representations selected for the user's history are generated by weighted averaging of historical interaction data after MLP transformation and batch feature scaling normalization in S1; and It is the regularized representation for clothing recommendations and shares parameters with the preference modeling features; γ3 balances the contributions of visual and textual modalities.

[0080] This regularization term constrains the recommendation results through historical data, enhancing the consistency of recommended clothing in cultural style and user habits.

[0081] Clothing matching consistency regularization measures the historical matching of recommended clothing r and existing clothing g (list g r ), the score function is:

[0082]

[0083] in, and average visual and textual representations of historically matched costumes; and is the regularized representation of recommended clothing, and γ4 is the weight parameter.

[0084] S3. By integrating clothing matching, user preferences and consistency scores, combining BPR loss and cultural regularization, and using gradient descent to optimize the model, we ensure recommendation accuracy and cultural and artistic coherence.

[0085] The overall preference score is calculated as follows:

[0086]

[0087] in, is the costume matching score, which indicates the coordination between the recommended costume r and the existing costume g under the dance type d and cultural background c. It can be calculated by the similarity of costume features (such as color and style). is the user preference score, which indicates the degree of preference of user u for the recommended clothing r, and is predicted based on the user's historical behavior and preference model. Regularize the score for clothing matching consistency to ensure that the recommended clothing is consistent with existing clothing in cultural context. Regularize the score for user preference consistency to enhance the fit between recommendation results and user cultural preferences. The weight parameters are used to balance the effects of clothing matching, user preference, and consistency regularization.

[0088] To optimize the ranking ability of the recommendation model, the Bayesian Personalized Ranking (BPR) loss function is used to improve the recommendation accuracy by maximizing the score difference between positive and negative samples. The specific design is as follows:

[0089] Loss Function Defined as:

[0090] in, is a training dataset, which includes user u, dance type d, cultural background c, existing clothing g, and positive sample clothing r + and negative sample clothing r - combination. and are the overall preference scores for positive and negative samples, respectively. σ is the Sigmoid function, which maps the score difference to a probability value. λ is the regularization coefficient, which is used to control model complexity and prevent overfitting. is the model parameter Θ f The Frobenius norm of , including all embedding vectors (such as users, clothing features) and weight parameters.

[0091] Positive sample r + It can be selected from the user's historical purchases or high-rated clothing, and the negative sample r - This can be determined through random sampling or clothing that the user clearly dislikes.

[0092] The loss optimizes the relative ranking of positive and negative samples to ensure that the user's preferred clothing ranks high in the recommendation list, while the regularization term improves the generalization ability of the model.

[0093] Cultural consistency regularization term Defined as: where e c is the embedding vector of the cultural background c, which is generated through pre-training of the cultural dataset. Embedding the cultural features of clothing r can be extracted through expert annotation or deep learning models. It is the square of the Euclidean distance, which measures the differences in cultural background and clothing cultural characteristics.

[0094] Combining the above components, the overall loss function Defined as: in For BPR loss, optimize the recommendation ranking. is the cultural consistency regularization to ensure cultural authenticity. γ1 is the cultural regularization weight, which controls the importance of cultural consistency.

[0095] Through the aforementioned loss function, the recommendation model is optimized using gradient descent, and the recommendation model finally predicts the five-tuple<u,d,c,g,r> The total preference score of the user is used to make clothing recommendations based on the ranking of the total preference scores.

Claims

1. A dance costume recommendation method that ensures cultural authenticity and artistic consistency, characterized by: include: Construct a recommendation model, and use the recommendation model to predict the quintuple<u,d,c,g,r> The total preference score of is ranked according to the total preference score, and clothing recommendations are made, where u represents a specific user, d represents the dance type, c represents the cultural background, g represents the existing clothing selection, and r represents the recommended clothing; The recommendation model predicts quintuples<u,d,c,g,r> The total preference scoring consists of the following steps: Based on the matrix decomposition framework, the user preference score is calculated by integrating the influence of dance type, cultural background and existing clothing choices. Considering the constraints of dance type and cultural background, predict the matching degree between existing costume choices and recommended costumes, and calculate the costume matching score Calculate the similarity score of recommended clothing based on user historical selection Measure the similarity between the recommended clothing and the historical matching clothing of the existing clothing selections and calculate the similarity score Calculating the total preference score Among them, μ, φ, is the weight parameter.

2. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 1, characterized in that: The user preference score in, and is the normalized latent embedding of users and clothing, and To provide visual and textual representations of recommended clothing items, and is the embedding vector of dance type and cultural background, and Provides visual and textual representations of existing clothing items. and are the user’s visual and textual preference representations, α, β u , β r are global bias, user bias and clothing bias respectively, and η is the weight parameter.

3. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 1, characterized in that: The clothing matching score The clothing matching score in, Visual and textual representations of existing clothing specifically for matching modeling, is the visual and textual representation of recommended clothing specifically for matching modeling, γ2 and δ are weight parameters, It is a fusion representation of the visual features of existing costumes after they are transferred to the target cultural background style, and then combined with the text features of existing costumes, dance type embedding, and cultural background embedding. The visual features of the recommended costumes are transferred to the target cultural background style, and then fused with the text features of the recommended costumes, dance type embedding, and cultural background embedding. and is the embedding vector of dance type and cultural background, Normalized latent embedding for recommended clothing.

4. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 1, characterized in that: The recommended clothing similarity score in, and the average visual and textual representations selected for the user's history, and is the visual and textual regularized representation of the recommended clothing, and γ3 is the weight parameter.

5. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 1, characterized in that: The similarity score in, and Average visual and textual representations of historically matched costumes, and is the visual and textual regularized representation of the recommended clothing, and γ4 is the weight parameter.

6. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 1, characterized in that: The recommendation model adopts the loss function optimization, in, is the Bayesian personalized ranking loss function, is a training dataset, which includes user u, dance type d, cultural background c, existing clothing g, and positive sample clothing r + and negative sample clothing r - combination of and are the overall preference scores of positive samples and negative samples respectively, σ is the Sigmoid function, λ is the regularization coefficient, is the recommended model parameter Θ F The Frobenius norm of ; is the cultural consistency regularization term, e c is the embedding vector of cultural background c, Embed the cultural characteristics of clothing r, is the square of the Euclidean distance; γ1 is the cultural regularization weight.

7. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 6, characterized in that: The positive sample clothing r + Select from the user's historical purchases or high-rated clothing, negative sample clothing r - Determined by random sampling or clothing that the user explicitly dislikes.

8. The method for recommending dance costumes that ensures cultural authenticity and artistic consistency according to claim 2, wherein: The visual and textual representations of existing and recommended clothing when calculating the user preference score are obtained by extracting visual and textual features of clothing from a pre-trained deep neural network, performing nonlinear transformations on them through a multi-layer perceptron, and then normalizing them through batch feature scaling.

9. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 3, characterized in that: The visual and textual representations of the existing clothing when calculating the clothing matching score are obtained by extracting the visual and textual features of the clothing through a pre-trained deep neural network, performing nonlinear transformations on them through a multi-layer perceptron, and then normalizing them through batch feature scaling; The visual and textual representations of the recommended clothing when calculating the clothing matching score are obtained by performing nonlinear transformation of the visual and textual features of the clothing extracted by a pre-trained deep neural network through a multi-layer perceptron and then normalizing them through batch feature scaling. Then, the normalized visual representation of the existing clothing and the visual representation of the recommended clothing are used as input sequences to calculate the attention weight of the Transformer, and the normalized visual representation of the recommended clothing is weighted to obtain the final visual representation of the recommended clothing. The normalized textual representation of the existing clothing and the textual representation of the recommended clothing are used as input sequences to calculate the attention weight of the Transformer, and the normalized textual representation of the recommended clothing is weighted to obtain the final textual representation of the recommended clothing.

10. The dance costume recommendation method for ensuring cultural authenticity and artistic consistency according to claim 8 or 9, characterized in that: The multi-layer perceptron performs nonlinear transformation specifically as follows: Among them, W k and b k are the projection matrix and bias vector of the kth layer respectively; K is the number of MLP layers; σ is the Sigmoid activation function, and the output feature after transformation is X K .

Images