Customized costume design recommendation system

Abstract

The invention provides a customized costume design recommendation system, and relates to the technical field of customized costume design recommendation. User emotion data and user historical behavior data are acquired in real time through a user data acquisition module; a preference type analysis module is used for extracting invisible emotion features when the user recommends the clothes style from the user emotion data, and preference type analysis is carried out on the clothes personalized demand of the user through an emotion label vector; determining a recommendation identifier list of the user for costume design matching through a style matching analysis module, and further performing style matching analysis on a costume application scene of the user through the recommendation identifier list; and performing clothing recommendation to the user through a clothing recommendation module according to the clothing demand preference after the preference type analysis and the clothing matching scene after the matching analysis. According to the method, accurate preference type analysis and style matching can be performed on the basis of acquiring the user emotion data and the historical behavior data in real time, so that the personalized level of a clothing recommendation system is improved.

Description

Technical Field

[0001] This application relates to the technical field of customized clothing design recommendation, and more specifically, to a customized clothing design recommendation system. Background Art

[0002] By analyzing the personal needs, emotional preferences, historical behavior data, and applicable scenarios of users, a clothing design plan that meets the personalized needs of users is generated and recommended. Based on big data analysis and machine learning technologies, this system can deeply explore the preferences of users in aspects such as styles, colors, and materials, and combine real-time market trends and clothing design trends to provide accurate customized clothing recommendations.

[0003] However, in existing customized clothing design recommendation systems, there is a lack of real-time collection and analysis of users' emotional data and historical behavior data, making it impossible for the system to accurately capture the dynamic preference changes and real needs of users, resulting in insufficient personalization of the recommendation results and difficulty in effectively adapting to the demand changes of users at different times and scenarios, thus affecting users' purchase decisions and satisfaction. Therefore, how to perform accurate preference type analysis and style matching based on real-time acquisition of users' emotional data and historical behavior data to improve the personalization level of clothing recommendation systems is a difficult problem faced by the industry. Summary of the Invention

[0004] This application provides a customized clothing design recommendation system that can perform accurate preference type analysis and style matching based on real-time acquisition of users' emotional data and historical behavior data to improve the personalization level of clothing recommendation systems.

[0005] In a first aspect, this application provides a customized clothing design recommendation system, which includes:

[0006] A user data collection module for real-time acquisition of users' emotional data and users' historical behavior data in clothing designs in Internet clothing advertisements;

[0007] A preference type analysis module for extracting the implicit emotional features of users when recommending clothing styles from the users' emotional data, generating an emotional label vector for users' clothing design recommendations through the implicit emotional features and the implicit preference information of users' clothing needs, and performing preference type analysis on the clothing personalization needs of users by the emotional label vector;

[0008] A style matching analysis module for associating the clothing attribute indexes in the Internet clothing advertisement content library, determining a recommended identification list for users' clothing design matching according to the clothing attribute indexes and the users' historical behavior data, and then performing style matching analysis on the clothing applicable scenarios of users by the recommended identification list;

[0009] A clothing recommendation module for recommending clothing to users according to the clothing demand preferences after preference type analysis and the clothing matching scenarios after matching analysis.

[0010] In this embodiment, extracting the implicit emotional features of the user when recommending clothing styles from the user emotional data specifically includes:

[0011] Determining the recommendation evolution trend of the user when accepting clothing recommendations according to the user emotional data;

[0012] Extracting the implicit recommendation sequence of the user when recommending clothing styles from the recommendation evolution trend;

[0013] Determining the implicit emotional features of the user when recommending clothing styles according to the implicit recommendation sequence.

[0014] In this embodiment, generating the emotional label vector of the user for clothing design recommendations through the implicit emotional features and the implicit preference information of the user's clothing needs specifically includes:

[0015] Performing enhancement processing on the implicit emotional features to obtain the emotional feature tensor of the user;

[0016] Converting the implicit preference information into the preference semantic vector of the user's clothing needs;

[0017] Performing cross-modal fusion on the emotional feature tensor and the preference semantic vector to generate a recommendation fusion matrix;

[0018] Mapping the recommendation fusion matrix into a preset emotional label space to generate the emotional label encoding of the user for clothing design recommendations;

[0019] Determining the emotional label vector of the user for clothing design recommendations according to the emotional label encoding.

[0020] In this embodiment, performing preference type analysis on the user's clothing personalized needs by the emotional label vector specifically includes:

[0021] Mapping the emotional label vector into a high-dimensional feature space to generate the enhanced emotional label attributes;

[0022] Clustering the emotional label attributes to obtain preference type clusters;

[0023] Correcting the preference type clusters, and further performing preference type analysis on the user's clothing personalized needs.

[0024] In this embodiment, determining the recommended identification list of the user for clothing design matching according to the clothing attribute index and the user historical behavior data specifically includes:

[0025] Associate and map the clothing attribute index and the user's historical behavior data to obtain the associated feature information of the user in clothing recommendation;

[0026] Quantify the preference fit degree between the clothing and the user's preference according to the associated feature information;

[0027] Start the screening decision maker to screen according to the preference fit degree to obtain a similar recommendation set for clothing recommendation;

[0028] Determine the recommended identification list of the user's clothing design and matching according to the similar recommendation set.

[0029] In this embodiment, the screening decision maker refers to a system component that automatically selects and screens a clothing recommendation set suitable for the user according to the user's preference fit degree and other relevant features.

[0030] In this embodiment, the style matching analysis of the clothing applicable scenarios of the user by the recommended identification list specifically includes:

[0031] Conduct semantic analysis on the clothing information in the recommended identification list to obtain the implicit semantic coupling between the clothing and various scenarios;

[0032] Match the style characteristics of the clothing with the style requirements of different applicable scenarios according to the implicit semantic coupling to obtain the adaptation degree score of each piece of clothing in each scenario;

[0033] Classify and integrate the clothing according to the adaptation degree score to determine the style matching result of the user in different clothing applicable scenarios.

[0034] In this embodiment, the clothing recommendation to the user according to the clothing demand preference after preference type analysis and the clothing matching scenario after matching analysis specifically includes:

[0035] Deeply integrate the clothing demand preference after preference type analysis with the clothing matching scenario after matching analysis to generate a composite recommendation feature set of user personalization and scenario adaptability;

[0036] Dynamically sort the recommended identification list based on the composite recommendation feature set to obtain a candidate clothing set that meets the user's needs and scenarios;

[0037] Recommend the clothing in the candidate clothing set to the user.

[0038] In this embodiment, the emotion label vector refers to the vectorized representation of the user emotion label encoding.

[0039] In this embodiment, the user emotion data and user historical behavior data of the clothing design in the Internet clothing advertisement are obtained by reading the clothing recommendation storage unit of the Internet clothing advertisement.

[0040] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0041] The user data acquisition module is used to obtain in real time the user's emotional data and historical behavior data of clothing design in Internet clothing advertisements; the preference type analysis module extracts the implicit emotional characteristics of the user when recommending clothing styles from the user's emotional data, generates an emotional label vector for the user's clothing design recommendation through the implicit emotional characteristics and the implicit preference information of the user's clothing needs, and performs preference type analysis on the user's personalized clothing needs through the emotional label vector; the style matching analysis module associates the clothing attribute indexes in the Internet clothing advertisement content library, determines a recommended identification list for the user's clothing design matching according to the clothing attribute indexes and the user's historical behavior data, and then performs style matching analysis on the clothing application scenarios suitable for the user through the recommended identification list; the clothing recommendation module recommends clothing to the user according to the clothing demand preferences after preference type analysis and the clothing matching scenarios after matching analysis.

[0042] Therefore, in this application, personalized recommendations can be made when the system cannot accurately capture the dynamic preference changes and real needs of users; among them, by obtaining the user's emotional data and historical behavior data in Internet clothing advertisements in real time, the dynamic needs and emotional changes of users can be comprehensively captured, ensuring that the recommendation system can quickly respond to the needs of users and improve the accuracy and personalization level of recommendations. By extracting implicit emotional characteristics from the user's emotional data and generating an emotional label vector in combination with the implicit preference information of the user's clothing needs, the potential clothing preferences of users can be deeply explored, ensuring that the recommendations more accurately meet the personalized needs of users and improving the satisfaction and shopping experience of users. By associating clothing attribute indexes and user historical behavior data, accurate matching can be performed between different clothing combinations and user demand scenarios, ensuring that the recommendation of each piece of clothing can fit the user's usage scenario and enhancing the relevance and practicality of clothing recommendations. By combining the clothing demands after preference type analysis and the clothing application scenarios after matching analysis for recommendation, personalized and scenario-matched clothing recommendations can be provided, improving the efficiency of the user's shopping decision-making and increasing the conversion rate and user satisfaction of the recommendation system.

[0043] In summary, the technical solution adopted in this application can perform accurate preference type analysis and style matching on the basis of obtaining the user's emotional data and historical behavior data in real time, so as to improve the personalization level of the clothing recommendation system. Brief Description of the Drawings

[0044] To more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only the embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0045] Figure 1 is a module structure diagram of a customized clothing design recommendation system provided by the present application;

[0046] Figure 2 is a flow schematic diagram of determining an emotional label vector provided by the present application;

[0047] Figure 3 is a flow schematic diagram of determining a recommended identifier list provided by the present application. Detailed implementation manners

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0049] The embodiments of the present application provide a customized clothing design recommendation system. Its core is to obtain the user's emotional data and user's historical behavior data of clothing design in Internet clothing advertisements in real time through the user data collection module; extract the implicit emotional features of the user's clothing style recommendation from the user's emotional data through the preference type analysis module, generate an emotional label vector for the user's clothing design recommendation through the implicit emotional features and the implicit preference information of the user's clothing needs, and perform preference type analysis on the user's clothing personalized needs through the emotional label vector; associate the clothing attribute index in the Internet clothing advertisement content library through the style matching analysis module, determine the recommended identifier list for the user's clothing design matching according to the clothing attribute index and the user's historical behavior data, and then perform style matching analysis on the clothing applicable scenarios of the user through the recommended identifier list; recommend clothing to the user according to the clothing demand preferences after preference type analysis and the clothing matching scenarios after matching analysis. By adopting the above solution, accurate preference type analysis and style matching can be performed on the basis of real-time acquisition of user emotional data and historical behavior data, so as to improve the personalization level of the clothing recommendation system.

[0050] To better understand the above technical solutions, the following will elaborate on the above technical solutions in combination with the accompanying drawings of the specification and specific implementation manners. Refer to Figure 1As shown in the figure, this is a module structure diagram of a customized clothing design recommendation system according to this embodiment of the present application. The recommendation system includes: a user data collection module 100, a preference type analysis module 200, a style matching analysis module 300, and a clothing recommendation module 400, which are described as follows:

[0051] The user data collection module 100 is used to obtain the user's emotional data and user's historical behavior data of clothing design in Internet clothing advertisements in real time.

[0052] Specifically, first, obtain Internet clothing advertisement data through web crawlers (Scrapy, Selenium) or APIs, including behavioral data such as advertisement content, user comments, likes, clicks, and stay time. Use log analysis (such as ELK Stack) to parse the interaction records between users and advertisements, and regard the interaction records between users and advertisements as the user's historical behavior data. Then, use an NLP sentiment analysis model (such as BERT, LSTM, TextCNN) to perform sentiment polarity classification on text data such as comments and bullet screens, output sentiment scores, and then combine a computer vision model (such as ResNet, EfficientNet) to parse the user's expressions and body languages to enhance the accuracy of the emotional data, that is, obtain the user's emotional data. Finally, record the user's behavior trajectory through Clickstream analysis (such as Google Analytics, Kafka + Flink), including browsing, clicking, collecting, purchasing, etc., and store it in the clothing recommendation storage unit of the Internet clothing advertisement. The user's emotional data and user's historical behavior data of clothing design in the Internet clothing advertisement can be obtained by reading the clothing recommendation storage unit of the Internet clothing advertisement.

[0053] It should be noted that in this application, the user's emotional data refers to the data of the emotional tendency shown by the user during the advertisement interaction; the user's historical behavior data refers to the operation data of the user's past browsing, clicking, collecting, and purchasing.

[0054] The preference type analysis module 200 is used to extract the implicit emotional features of the user for clothing style recommendation from the user's emotional data, generate an emotional label vector for the user's clothing design recommendation through the implicit emotional features and the implicit preference information of the user's clothing needs, and perform preference type analysis on the user's clothing personalized needs through the emotional label vector.

[0055] In this embodiment, the implicit emotional features of the user for clothing style recommendation can be extracted from the user's emotional data in the following specific way, that is:

[0056] Determine the recommendation evolution trend when the user accepts clothing recommendations according to the user's emotional data;

[0057] Extract the implicit recommendation sequence when the user makes a clothing style recommendation from the described recommendation evolution trend;

[0058] Determine the implicit emotional characteristics when the user makes a clothing style recommendation according to the implicit recommendation sequence.

[0059] When specifically implemented, first, collect the emotional data of the user in clothing recommendations, including behavior data such as browsing duration, clicks, collections, purchases, and comment emotions. Combine natural language processing (NLP) to analyze the comments or feedback of the user on the recommended content, identify the emotional tendency, sort the user's historical recommendation data by time, construct a time series data set, and record the recommendation acceptance of the user at different time points. Use a moving average to calculate the short-term change trend of the user acceptance, and at the same time use a time series prediction algorithm (such as LSTM or Transformer) to predict the long-term trend, and identify the rising, falling, or stable state of the recommendation acceptance to form a recommendation evolution trend curve, and use the recommendation evolution trend curve as the recommendation evolution trend. Then, in the recommendation evolution trend, use a sequence pattern mining algorithm (such as PrefixSpan) to identify the acceptance patterns of the user in different time periods, for example: "click after the first recommendation - browse multiple times - finally purchase", "click but no interaction - jump out multiple times in a short time - no longer visit", this pattern represents the implicit acceptance sequence of the user for the recommendation. Use sequence embedding technology (such as BERT, Word2Vec) to convert the recommendation sequence into a numerical representation, and use the numerically represented recommendation sequence as the implicit recommendation sequence. Finally, input the user's implicit recommendation sequence into a deep learning model (such as Bi-LSTM+Attention) to automatically extract the long-term emotional preference characteristics of the user. Identify the critical moments of recommendation acceptance through the self-attention mechanism (Self-Attention), such as the time points when the user's emotions fluctuate strongly. Combine a clustering algorithm (such as K-Means) to classify the user's implicit emotional characteristics, such as "prefer simple style", "like retro style", "wait and see attitude towards trendy style", etc., and use the classified user's implicit emotional characteristics as the implicit emotional characteristics when the user makes a clothing style recommendation.

[0060] It should be noted that in this application, the recommendation evolution trend refers to the change pattern of the user's acceptance of the recommended content in the time dimension; the implicit recommendation sequence refers to a series of behavior trajectories of the user in the recommendation system; the implicit emotional characteristics refer to the emotional preference characteristics obtained based on the evolution of the user's recommendation behavior.

[0061] Preferably, in this embodiment, generate an emotional label vector for the user's clothing design recommendation through the implicit emotional characteristics and the implicit preference information of the user's clothing needs, refer to Figure 2 As shown, this figure is a schematic flow chart for determining the emotional label vector in some embodiments of this application. The determination of the emotional label vector in this embodiment can be implemented by the following steps:

[0062] In step S21, the invisible emotional features are enhanced to obtain the emotional feature tensor of the user;

[0063] In step S22, the invisible preference information is converted into a preference semantic vector for the user's clothing needs;

[0064] In step S23, the emotional feature tensor and the preference semantic vector are cross - modality fused to generate a recommendation fusion matrix;

[0065] In step S24, the recommendation fusion matrix is mapped into a preset emotional label space to generate an emotional label encoding for the user's clothing design recommendation;

[0066] In step S25, an emotional label vector for the user's clothing design recommendation is determined according to the emotional label encoding.

[0067] In specific implementation, first, normalization processing (such as Min-Max Scaling, Z-score standardization) is adopted to unify the numerical range of user sentiment features and eliminate the influence of data distribution. Key sentiment features are extracted through principal component analysis (PCA) or autoencoder to improve the discriminability of features. The multi-head attention mechanism (Multi-HeadAttention) is used to calculate the weights of sentiment features, highlighting the feature information of the user's long-term preferences. A time series model (such as Bi-LSTM, Transformer) is adopted to further explore the changing trend of user sentiment features. The enhanced sentiment features are represented by a high-dimensional tensor (Tensor), and each dimension represents a fine-grained sentiment feature, such as "the intensity of interest in the retro style" and "the degree of rejection of the trendy style", etc. That is, the sentiment feature tensor of the user is obtained. Next, the user's search keywords, purchase records, and favorite content are collected to analyze the potential intentions of the user's clothing needs. The user's historical comments and social media posts are extracted to identify the descriptive semantic information of the user's preferences. A word vector model (such as Word2Vec, BERT, FastText) is used to convert the user's preference information into a semantic vector. Semantic expansion is carried out in combination with a knowledge graph (Fashion Ontology) to improve the understanding ability of preference information. The clothing style themes concerned by the user are identified through topic modeling (such as LDA, BERT-Topic) to form the user's personalized demand expression, that is, the preference semantic vector of the user's clothing needs is obtained. Again, the co-attention mechanism (Co-Attention) is used to calculate the correlation between the sentiment features and the preference semantic vector to enhance the feature fusion effect. A multi-modal fusion technology (such as Tensor Fusion Network, TFN) is adopted to map the two data representations to the same vector space and generate a recommendation fusion matrix by combining information interaction. Each element of the recommendation fusion matrix represents the influence weight of a certain sentiment feature on a certain preference semantics, for example, "the degree to which the user's preference for the modern style is affected by the interest in the light luxury style". Then, a Softmax classifier or a deep neural network (DNN) is adopted to convert the fusion matrix into a discrete probability distribution of sentiment labels, such as categories like "happy, like, dislike, indifferent", etc. The user's sentiment labels are further segmented through a clustering algorithm (such as K-Means, DBSCAN) to form a more accurate user interest classification, generating the sentiment label encoding for the user's clothing design recommendation. Finally, the probability distribution of the user's sentiment labels is converted into a vector representation, such as One-hot encoding or a continuous vector representation (such as Word2Vec embedding). For example: The probability distribution of sentiment labels is a numerical array representing the belonging degree of the user to different sentiment labels.For example, assume that the system defines the following four types of sentiment tags: Like, Neutral, Dislike, and Surprise. For a certain user, the sentiment probability distribution calculated by the system is: P = (0.6, 0.2, 0.1, 0.1), where P is the sentiment tag probability, Like = 0.6, Neutral = 0.2, Dislike = 0.1, Surprise = 0.1. One-hot encoding is a discrete representation used to identify the sentiment category with the highest probability. Select the sentiment category with the highest probability, represent it as 1, and the other categories as 0. Conversion steps: Identify the category with the highest probability (here it is "Like" with 0.6). Set the position of this category to 1 and the other categories to 0. The conversion result is (1, 0, 0, 0), that is, the sentiment tag vector of the user's preference for clothing design recommendations is obtained. Use an embedding model (such as Word2Vec, FastText, BERT) to convert the discrete sentiment categories into continuous vectors. Through the pre-trained sentiment semantic space, ensure that the vectors of similar sentiment categories are close. Pre-train the embedding representation of sentiment tags using Word2Vec / FastText, or directly use pre-trained BERT to extract semantic vectors. For example, each tag may have a predefined vector: Like → (0.9, 0.1, 0.3), Neutral → (0.2, 0.7, 0.1), Dislike → (0.1, 0.2, 0.9), Surprise → (0.5, 0.8, 0.2). Use the sentiment tag probability P as the weight to perform weighted averaging on the embedding vectors, and the final continuous vector representation is obtained after calculation. For example: (0.62, 0.25, 0.34), that is, the sentiment tag vector of the user's preference for clothing design recommendations is obtained. Then perform normalization processing to ensure the comparability of the sentiment tag vectors of different users.

[0068] It should be noted that in this application, the sentiment feature tensor refers to the high-dimensional features of the user's sentiment preferences when receiving clothing recommendations; the preference semantic vector refers to the features that reflect the implicit preferences converted from the user's historical behavior data; the recommendation fusion matrix refers to the high-dimensional data structure that fuses the user's sentiment features and preference semantic vectors; the sentiment tag encoding refers to the numerical representation after the user's sentiment features are mapped to the tag space; the sentiment tag vector refers to the vectorized representation of the user's sentiment tag encoding; the sentiment tag space refers to the range of the user's sentiment preferences reflected by different sentiment tags.

[0069] In this embodiment, the preference type analysis of the user's clothing personalization needs by the sentiment tag vector can be specifically implemented by the following steps, that is:

[0070] Map the emotional label vector into a high-dimensional feature space to generate enhanced emotional label attributes with features;

[0071] Cluster the emotional label attributes to obtain preference type clusters;

[0072] Revise the preference type clusters, and then conduct preference type analysis on the personalized clothing needs of users.

[0073] Specifically, when implementing, first, use a deep neural network (DNN) or a support vector machine (SVM) to map the emotional label vector from the original space to a high-dimensional feature space to capture more complex user emotional features. During the feature mapping process, introduce a non-linear transformation through a non-linear activation function (such as ReLU, sigmoid) to enhance the expression ability of the emotional label vector. Use the self-attention mechanism (Self-Attention) to emphasize important features in the emotional label vector and strengthen the part of the emotional label that has an important impact on user demand prediction. The enhanced emotional label attributes with features are obtained by performing linear or non-linear transformations on the label attributes in the high-dimensional space, such as the emotional features after weight strengthening, and the weights can be obtained from historical expert experience or experiments. Then, use K-means clustering or DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to cluster the emotional label attributes in the high-dimensional feature space. K-means clustering: Preset the number of clusters in advance and divide the emotional label attributes into multiple clusters based on the Euclidean distance. The clustering result represents different user preference types. DBSCAN clustering: There is no need to preset the number of clusters, and it automatically divides the emotional label attributes into dense regions based on density to identify different user preference type clusters. During the clustering process, principal component analysis (PCA) can be used for dimensionality reduction to reduce noise in the high-dimensional space and improve the clustering quality. The clusters after clustering represent different types of clothing demand preferences. For example, one cluster may represent "users who prefer a simple style", while another cluster represents "users who prefer complex designs". Take the clusters after clustering as the preference type clusters. Finally, dynamically revise the clustering result according to user feedback, new behavior data, or business requirements. The clustering clusters can be continuously optimized through an online learning algorithm (such as online K-means) to ensure that the latest behaviors of users are reflected in a timely manner. Adopt a weighted clustering method, consider the weights of different emotional label attributes, and adjust the boundaries of the preference type clusters. According to the revised preference type clusters, analyze the personalized clothing needs of users. For example, if a preference type cluster represents users who "like a casual style", the system can recommend casual-style clothing based on this preference type cluster. For each preference type cluster, use typical feature analysis (such as feature mean or optimal center point) to summarize the clothing preference features of users in this category.

[0074] It should be noted that in this application, the high-dimensional feature space represents a vector space that captures complex patterns and associations through multi-dimensional data mapping. The emotional label attribute refers to the user's emotional preference after mapping the emotional label vector to the high-dimensional space; the preference type cluster refers to a feature set that divides users with similar emotional label attributes into different groups to reflect different clothing demand preference types; the personalized clothing demand is represented as a personalized preference vector of the user for clothing style, design, and function.

[0075] The style matching analysis module 300 is used to associate the clothing attribute index in the Internet clothing advertisement content library, determine a recommended identification list of the user's clothing design and matching according to the clothing attribute index and the user's historical behavior data, and then perform style matching analysis on the clothing applicable scenarios of the user by the recommended identification list.

[0076] In this embodiment, associating the clothing attribute index in the Internet clothing advertisement content library can be implemented in the following manner: First, collect multi-modal data such as text descriptions, pictures, and videos of clothing advertisements from the Internet clothing advertisement content library. These advertisements usually contain attribute information such as design style, material, color, size, seasonality, and applicable scenarios of the clothing. Then, use natural language processing (NLP) techniques (such as named entity recognition, topic modeling, keyword extraction) for the text description to extract key attributes related to clothing (such as "summer", "loose", "casual"). Use a convolutional neural network (CNN) to extract visual features from the picture data, such as the color tone, style, and pattern of the clothing. Use a video analysis model (such as RNN, 3D-CNN) for the video data to extract dynamic information and the matching effect of the clothing. Finally, through multi-modal information fusion, map the extracted text, image, video and other features to a unified clothing attribute index. This can be achieved in the following way: Assign a certain weight to each clothing attribute (such as color, style, applicable scenario), and the weight can be obtained from expert experience, and combine the weighted average method or a multi-modal neural network for fusion to generate a comprehensive attribute index of the clothing, and use the comprehensive attribute index as the clothing attribute index.

[0077] It should be noted that in this application, the clothing attribute index represents the intensity of various attributes of the clothing in the advertisement.

[0078] Preferably, in this embodiment, to determine a recommended identification list of the user's clothing design and matching according to the clothing attribute index and the user's historical behavior data, refer to Figure 3 As shown, this figure is a schematic flowchart of determining the recommended identification list in some embodiments of this application. Determining the recommended identification list in this embodiment can be implemented by the following steps:

[0079] In step S31, the clothing attribute index and the user's historical behavior data are associated and mapped to obtain the associated feature information of the user in clothing recommendation;

[0080] In step S32, the preference fit degree between the clothing and the user's preference is quantified according to the associated feature information;

[0081] In step S33, the screening decision maker is activated to perform screening based on the preference fit degree to obtain a similar recommendation set for clothing recommendation;

[0082] In step S34, the recommendation identification list of the user's clothing design and matching is determined according to the similar recommendation set.

[0083] Specifically, first, the clothing attribute index and the user's historical behavior data are collected. The clothing attribute index represents the characteristics of the clothing's design style, color, material, etc.; the user's historical behavior data includes the user's purchase records, browsing behaviors, likes and comments, etc. Use association rule learning or deep learning models (such as neural networks) to match the clothing attribute index with the user's behavior data. For example, if the user often browses casual-style clothing and most of the purchase records are blue T-shirts, then the clothing attributes of blue casual style can be associated with the user's behavior to construct the user's preference feature vector, and the user's preference feature vector is used as the associated feature information of the user in clothing recommendation. Then, based on methods such as Euclidean distance, cosine similarity, or Pearson correlation coefficient, the fit degree between the clothing and the user's preference is quantified. For example, if the user has a higher preference for "casual style" in the historical data and the weight of the casual style in the attribute index of a certain piece of clothing is larger, then the fit degree between this piece of clothing and the user's preference is higher. A weighted clustering algorithm or support vector machine (SVM) can be used to construct a fit degree evaluation model by learning the user's historical behavior. Then, by setting a threshold or sorting mechanism, according to the fit degree between the clothing and the user's preference, the clothing with a high fit degree is selected as the recommended candidate set. For example, the fit degree threshold can be set to 0.8, and all clothing with a fit degree greater than or equal to 0.8 is selected as the similar recommendation set. Use a greedy algorithm or A / B testing to screen the clothing that best meets the user's needs and avoid too many irrelevant recommendation items, that is, obtain the similar recommendation set. Finally, according to the selected similar recommendation set, a unique recommendation identifier is generated for each clothing design and matching. The recommendation identifier can be encoded based on the clothing's category, style, material, etc., to ensure that each recommendation has a unique identifier. According to factors such as the fit degree and the user's interest, the recommendation identifiers are sorted to ensure that the clothing combinations that best meet the user's preferences are displayed first.

[0084] It should be noted that in this application, the associated feature information refers to the feature data describing the relationship between user preferences and clothing features; the preference fit degree refers to the matching degree between the user's needs and the design features of a certain piece of clothing; the similar recommendation set refers to the set of clothing that highly matches the user's needs; the recommendation identifier list refers to the list of unique identification elements for clothing design combinations; the screening decision-maker represents a system component that automatically selects and screens out a set of clothing recommendations suitable for the user based on the user's preference fit degree and other relevant features.

[0085] In this embodiment, the style matching analysis of the user's clothing applicable scenarios by the recommendation identifier list can be specifically carried out in the following manner, that is:

[0086] Perform semantic analysis on the clothing information in the recommendation identifier list to obtain the implicit semantic coupling between the clothing and various scenarios;

[0087] Match the style features of the clothing with the style requirements of different applicable scenarios according to the implicit semantic coupling to obtain the adaptation degree scores of each piece of clothing in each scenario;

[0088] Integrate the clothing into scene categories based on the adaptation degree scores to determine the style matching results of the user in different clothing applicable scenarios.

[0089] In specific implementation, first, natural language processing (NLP) and semantic analysis techniques are used to parse the text of the clothing information in the recommended identifier list. The clothing information may include design descriptions, materials, colors, usage occasions, etc. Key semantic features in the clothing description are extracted through pre-trained language models such as Word2Vec or BERT. For example, by analyzing the vocabulary in the clothing description, such as "summer" and "sports style", it can be recognized that the clothing is suitable for "summer" or "sports" scenarios. The semantic features of each piece of clothing are analyzed in association with the semantic requirements of different scenarios. The semantic relevance between the clothing and the scenarios is measured through similarity calculation (such as cosine similarity) to obtain implicit semantic coupling. Then, the style features of the clothing, such as color tone, style, design elements, etc., are extracted, and the clothing pictures are analyzed through convolutional neural networks (CNNs) or image processing techniques to obtain visual style features. Corresponding style requirements are defined for each application scenario (such as "formal occasion", "casual scenario", "sports scenario"). These requirements can be extracted from the user's historical behavior data through a rule engine or clustering analysis. Through multimodal fusion technology (combining semantic analysis and visual features) and a weighted matching algorithm, the style features of the clothing are matched with the style requirements of different scenarios. According to the matching results, the fitness scores of each piece of clothing in different scenarios are calculated. For example, a piece of "sports style" clothing may have a fitness score of 0.9 in the "sports scenario" and a lower fitness score, perhaps 0.3, in the "formal occasion", that is, the fitness scores of each piece of clothing in each scenario are obtained. Finally, according to the fitness scores, the clothing is classified according to different scenarios. For example, clothing with a fitness score higher than a certain threshold will be classified into the "formal occasion", and clothing with a fitness score lower than the threshold will be classified as "not suitable for this scenario". Through sorting algorithms or clustering algorithms (such as K-means), the clothing is integrated after being classified by scenario to ensure that the recommended list for each scenario contains clothing that meets the requirements of that scenario. According to the classification results, the style matching degrees of each piece of clothing in different scenarios are output. For example, when the user browses the recommended list, they can see "recommendations suitable for the sports scenario" and "recommendations suitable for the formal occasion", that is, the style matching results of the user for different clothing application scenarios are obtained.

[0090] It should be noted that in this application, implicit semantic coupling refers to the automatic discovery and establishment of potential association relationships between clothing and scenarios by analyzing the descriptions of clothing (such as text, images, etc.) and the requirements of different scenarios without explicit labels or direct associations; the fitness score refers to a numerical value that measures the applicability or matching degree of a piece of clothing in a specific scenario; the style matching result represents the applicable situation and recommended priority of each piece of clothing in different scenarios.

[0091] The clothing recommendation module 400 is used to recommend clothing to users based on the clothing demand preferences analyzed according to the preference types and the clothing matching scenarios analyzed by the matching analysis.

[0092] In this embodiment, the clothing recommendation to users based on the clothing demand preferences analyzed according to the preference types and the clothing matching scenarios analyzed by the matching analysis can be specifically implemented in the following manner, that is:

[0093] Deeply integrate the clothing demand preferences analyzed according to the preference types with the clothing matching scenarios analyzed by the matching analysis to generate a composite recommendation feature set of user personalization and scenario adaptability;

[0094] Dynamically sort the recommendation identifier list based on the composite recommendation feature set to obtain a candidate clothing set that meets the user's needs and scenarios;

[0095] Recommend the clothing in the candidate clothing set to the user.

[0096] When specifically implemented, first, obtain the user's preferences in terms of clothing style, material, color, etc. from the user's historical behavior data and sentiment label analysis. For example, a certain user prefers "casual style" or "simple design". According to the fitness scores of the clothing, match each piece of clothing with multiple scenarios (such as formal, casual, sports, etc.) to obtain the fitness scores of each piece of clothing in different scenarios. Use multimodal learning or feature fusion techniques (such as weighted fusion, feature splicing, etc.) to combine the user's clothing demand preferences with the scenario fitness of the clothing. Through weighting, splicing, or aggregating the two, generate a composite recommendation feature set that reflects the user's personalized needs in a specific scenario and the applicability of the clothing. Then, based on the generated composite recommendation feature set, use a sorting algorithm (such as Gradient Boosting Decision Tree (GBDT) or XGBoost) to sort the recommendation identifier list. The sorting rules combine the user's demand preferences and the clothing scenario fitness, and give priority to recommending clothing with high fitness. For example, the user's demand preferences may tend to be "casual" style, and the fitness score of the clothing is relatively high in the "casual" scenario, and such clothing will be ranked in the front. Multiple dimensions of factors can be considered during the sorting process, including the weights of user preferences, clothing fitness scores, the user's historical behavior, etc., to ensure that the recommendation results are both personalized and meet the scenario requirements. Finally, according to the sorted results, screen out the clothing that meets the user's needs and scenario fitness as the final recommended candidate set. This set contains the clothing combinations that the user is most interested in and can be best used in specific scenarios. Through the user interface (UI), display the final candidate clothing to the user in the form of card display or sliding recommendation, etc., to help the user make a quick choice.

[0097] It should be noted that in this application, the composite recommendation feature set represents the set of the user's preferences for clothing after the fusion of different scenarios; the candidate clothing set represents the set of clothing features that meet the user's needs and scene adaptability during the recommendation process.

[0098] It can be seen that in this application, personalized recommendations can be made when the system cannot accurately capture the dynamic preference changes and real needs of users; among them, by obtaining the user's emotional data and historical behavior data in Internet clothing advertisements in real time, the dynamic needs and emotional changes of users can be comprehensively captured, ensuring that the recommendation system can quickly respond to the needs of users and improve the accuracy and personalization level of recommendations. By extracting implicit emotional features from the user's emotional data and generating an emotional label vector in combination with the implicit preference information of the user's clothing needs, the potential clothing preferences of users can be deeply explored, ensuring that the recommendations more accurately meet the personalized needs of users and improving the user's satisfaction and shopping experience. By associating the clothing attribute index with the user's historical behavior data, accurate matching can be performed between different clothing combinations and user demand scenarios, ensuring that the recommendation of each piece of clothing can fit the user's usage scenario and enhancing the relevance and practicality of clothing recommendations. By combining the clothing needs after preference type analysis and the clothing applicable scenarios after matching analysis for recommendation, personalized and scenario-matched clothing recommendations can be provided, improving the efficiency of the user's shopping decision-making and increasing the conversion rate and user satisfaction of the recommendation system.

[0099] In summary, the technical solution adopted in this application can perform accurate preference type analysis and style matching on the basis of obtaining the user's emotional data and historical behavior data in real time to improve the personalization level of the clothing recommendation system.

[0100] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in one Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0101] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructing relevant hardware through a program, and this program can be stored in a computer-readable storage medium. The storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disk memories, tape memories, or any other medium that can be used to carry or store data and is computer-readable.

[0102] It should also be noted that the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or further includes elements inherent in such process, method, commodity or device. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, commodity or device including the element.

Claims

1. A custom clothing design recommendation system, characterized in that: The recommendation system comprises: User data collection module, used to obtain user emotion data and user historical behavior data of clothing designs in Internet clothing advertisements in real time; A preference type analysis module is used to extract the invisible emotion features of the user when recommending clothing styles from the user emotion data, generate the emotion label vector of the user's clothing design recommendation through the invisible emotion features and the user's invisible preference information on clothing needs, and perform preference type analysis on the user's personalized clothing needs based on the emotion label vector; A style matching analysis module is used to associate clothing attribute indexes in an Internet clothing advertisement content library, determine a user's recommended identification list for clothing design matching based on the clothing attribute indexes and the user's historical behavior data, and then perform style matching analysis on the user's clothing applicable scenarios based on the recommended identification list; The clothing recommendation module is used to recommend clothing to users based on clothing demand preferences after preference type analysis and clothing matching scenarios after matching analysis.

2. A custom clothing design recommendation system as claimed in claim 1, characterized in that: Extracting the invisible emotional features of the user's clothing style recommendation from the user's emotional data specifically includes: determining, based on the user emotion data, a recommendation evolution trend when the user accepts clothing recommendations; Extracting the invisible recommendation sequence when the user recommends clothing styles from the recommendation evolution trend; The invisible emotional features of the user when recommending clothing styles are determined according to the invisible recommendation sequence.

3. A custom clothing design recommendation system as claimed in claim 1, characterized in that: Generating the user's emotional label vector for clothing design recommendation through the invisible emotional features and the user's invisible preference information for clothing needs specifically includes: Performing enhancement processing on the invisible emotional features to obtain the user's emotional feature tensor; Converting the invisible preference information into a semantic vector of user preference for clothing needs; Cross-modally fusing the sentiment feature tensor and the preference semantic vector to generate a recommendation fusion matrix; Mapping the recommendation fusion matrix into a preset emotional label space to generate an emotional label encoding of the user's recommendation for clothing design; The emotional label vector of the user's recommendation for the clothing design is determined according to the emotional label encoding.

4. A custom clothing design recommendation system as claimed in claim 1, characterized in that: The preference type analysis of the user's personalized clothing needs based on the emotional label vector specifically includes: Mapping the emotion label vector to a high-dimensional feature space to generate an emotion label attribute after feature enhancement; Clustering the sentiment label attributes to obtain preference type clusters; The preference type cluster is modified, and then the preference type analysis is performed on the user's personalized clothing needs.

5. A custom clothing design recommendation system as claimed in claim 1, characterized in that: Determining a user's recommended identification list for clothing design matching according to the clothing attribute index and the user's historical behavior data specifically includes: Associating and mapping the clothing attribute index with the user's historical behavior data to obtain the user's associated feature information in clothing recommendation; quantifying the degree of fit between the clothing and the user's preference according to the associated feature information; Starting the screening decision maker to screen according to the preference fit, and obtaining a similar recommendation set for clothing recommendation; A list of recommended identifications for clothing design combinations by the user is determined according to the similar recommendation set.

6. A custom clothing design recommendation system as claimed in claim 5, characterized in that: The filtering decision maker represents a system component that automatically selects and filters a set of clothing recommendations suitable for a user based on the user's preference fit and other relevant features.

7. A custom clothing design recommendation system as claimed in claim 1, characterized in that: The style matching analysis of the user's clothing applicable scene based on the recommendation identifier list specifically includes: Performing semantic analysis on the clothing information in the recommendation identification list to obtain implicit semantic coupling between clothing and various scenes; According to the implicit semantic coupling, the style features of the clothing are matched with the style requirements of different applicable scenes to obtain the fitness score of each clothing in each scene; The clothing is classified and integrated according to the suitability scores to determine the style matching results of the user in different clothing applicable scenarios.

8. A custom clothing design recommendation system as claimed in claim 1, characterized in that: The clothing recommendation for users based on the clothing demand preferences after the preference type analysis and the clothing matching scenarios after the matching analysis specifically includes: Deeply integrate the clothing demand preferences after preference type analysis with the clothing matching scenarios after matching analysis to generate a composite recommendation feature set of user personalization and scenario adaptability; Dynamically sorting the recommendation identification list based on the composite recommendation feature set to obtain a set of candidate clothing that meets user needs and scenarios; The clothing in the candidate clothing set is recommended to the user.

9. A custom clothing design recommendation system as claimed in claim 1, characterized in that: The emotion label vector refers to the vectorized representation of the user emotion label encoding.

10. A custom clothing design recommendation system as claimed in claim 1, characterized in that: The user emotion data and user historical behavior data of clothing designs in Internet clothing advertisements are obtained by reading a clothing recommendation storage unit of Internet clothing advertisements.

Images