Big data collection and analysis method and system based on clothing design

Abstract

The application relates to the technical field of clothing design and big data analysis, and discloses a big data acquisition and analysis method and system based on clothing design, wherein the big data acquisition and analysis method based on clothing design comprises the following steps: acquiring multi-source heterogeneous data; adopting a data quality evaluation algorithm and a cleaning technology; extracting a clothing image feature model; quantifying design elements of the clothing image features and analyzing element correlation; adopting an LSTM-based time series prediction model and a multi-factor influence model to obtain a popular trend prediction result; adopting a feature engineering method to obtain multi-dimensional user features; performing user preference clustering analysis; optimizing, quality evaluating and screening candidate design schemes; adopting a market performance prediction model based on a random forest to perform prediction; and performing risk evaluation; through application of a time series deep learning model, the application can capture complex change rules of popular trends and realize accurate prediction of future popular trends.

Description

Technical Field

[0001] This invention relates to the field of clothing design and big data analysis technology, and more specifically, to a method and system for big data collection and analysis based on clothing design. Background Technology

[0002] With the rapid development of the global fashion industry and the increasing demand for personalized products and services, traditional clothing design models are facing enormous challenges. Traditional design relies heavily on the designer's subjective experience and limited market research data, resulting in long design cycles, delayed market response, and inaccurate understanding of consumer needs. This is especially true in the fast fashion market, where trends change rapidly, making it difficult for traditional design methods to keep pace, leading to delayed new product launches, severe inventory buildup, and significant design homogenization.

[0003] In recent years, the widespread application of big data technology across various industries has provided new opportunities for the transformation and upgrading of the apparel design industry. By collecting and analyzing massive amounts of consumer behavior data, fashion trend information, and market sales data, more scientific and accurate decision support can be provided for apparel design. However, data in the apparel design field is characterized by multi-source heterogeneity, strong real-time requirements, and significant subjectivity, and existing big data analytics technologies still have many shortcomings when applied to apparel design.

[0004] In existing technologies, some studies have attempted to apply big data analytics to the field of apparel design, but these generally suffer from limitations such as limited data collection scope, weak targeting of analytical methods, low predictive accuracy, and insufficient personalization. Therefore, there is an urgent need to develop a big data collection and analysis method specifically tailored to the characteristics of apparel design, capable of effectively integrating multi-source data, accurately predicting fashion trends, and providing personalized design recommendations, thus offering scientific technical support and decision-making basis for the apparel design industry. Summary of the Invention

[0005] This invention provides a big data collection and analysis method and system based on clothing design, which solves the technical problems of incomplete data collection, low trend prediction accuracy, and insufficient personalization in related technologies.

[0006] This invention provides a big data collection and analysis method based on clothing design, including:

[0007] Acquire multi-source heterogeneous data, including sales data from e-commerce platforms, user-generated content from social media, professional information from fashion websites, and environmental data; apply data quality assessment algorithms and cleaning techniques to the multi-source heterogeneous data to obtain a standardized dataset.

[0008] Feature engineering was used to extract clothing image features and multi-dimensional user features from a standardized dataset; the design elements of the clothing image features were quantified and the correlation between the elements was analyzed to obtain the feature vector of clothing design elements;

[0009] Environmental data is obtained from a standardized dataset. The feature vectors of clothing design elements and environmental data are input into a time series prediction model based on LSTM and a multi-factor influence model to obtain the fashion trend prediction results.

[0010] User preference clustering analysis is performed on multi-dimensional user characteristics to obtain user preference mining results;

[0011] The user preference mining results and popular trend prediction results are input into the generative adversarial network to obtain candidate design solutions. The candidate design solutions are then optimized, their quality is evaluated, and they are screened to obtain personalized design solutions.

[0012] Market forecasting and risk assessment are conducted on personalized design solutions to obtain preliminary market performance assessment results and risk control recommendations.

[0013] In a preferred embodiment, the process of acquiring the multi-source heterogeneous data specifically includes:

[0014] A distributed crawler cluster is built to monitor e-commerce platforms in real time. Web crawler technology is used to obtain product descriptions, prices, sales volume, and user reviews to obtain e-commerce platform data.

[0015] Based on the acquired e-commerce platform data, social media data is obtained by using keyword matching and topic tracking technologies through social media API calls to acquire fashion user-generated content.

[0016] By integrating data from e-commerce platforms and social media with professional information from fashion websites, multi-source heterogeneous data is obtained.

[0017] In a preferred embodiment, the data quality assessment algorithm and cleaning technique specifically include:

[0018] Acquire multi-source heterogeneous data, perform encoding format detection on the text data in the multi-source heterogeneous data, and convert the detected different encoding formats into UTF-8 encoding.

[0019] Perform format validation and size normalization on image data in multi-source heterogeneous data;

[0020] Outliers in multi-source heterogeneous data are detected using rules and statistical methods. Detected price anomalies, rating anomalies, and timestamp anomalies are marked and removed to obtain a standardized dataset.

[0021] In a preferred embodiment, the process of generating the feature vector of the clothing design elements specifically includes:

[0022] Obtain a standardized dataset and use deep learning feature extraction technology to construct a clothing image feature extraction model for the standardized dataset;

[0023] Basic visual features are extracted from clothing images using a pre-trained ResNet network, and a clothing attribute recognition branch network is constructed based on these basic visual features to identify design elements.

[0024] A color feature vector is established, and a style innovation index is defined by calculating the similarity difference between the design style and historical popular styles, thus obtaining the feature vector of clothing design elements.

[0025] In a preferred embodiment, the step of generating the trend prediction result specifically includes:

[0026] Obtain feature vectors of clothing design elements and environmental data, and establish a multi-factor influence model based on the feature vectors of clothing design elements and environmental data, comprehensively considering seasonal factors, economic factors, cultural event factors and social media popularity factors;

[0027] The multi-factor influence model was trained using an LSTM-based time series forecasting model, with the time window length and forecast time length set.

[0028] By combining prediction models with different architectures and using a weighted average method to fuse the prediction results of each model, the trend prediction results are obtained.

[0029] In a preferred embodiment, the process of generating the user preference mining results specifically includes:

[0030] Obtain the trend prediction results, and build a user behavior dataset based on the trend prediction results. The user behavior dataset includes user browsing history, purchase history, rating history and collection history.

[0031] The collaborative filtering algorithm is used to analyze the user behavior dataset and calculate the user similarity matrix and item similarity matrix;

[0032] A user preference model is constructed based on the user similarity matrix and the item similarity matrix. Users are divided into different preference groups through clustering algorithms to obtain user preference mining results.

[0033] In a preferred embodiment, the generation of the personalized design scheme specifically includes:

[0034] Obtain user preference mining results, and construct design constraints based on user preference mining results. Design constraints include cost constraints, material constraints, and process constraints.

[0035] The design constraints are optimized and solved using a genetic algorithm, with the population size, crossover probability, and mutation probability set.

[0036] The market adaptability and user satisfaction of the design scheme are evaluated by the fitness function, and the final design scheme is selected through multiple generations of evolution to obtain a personalized design scheme.

[0037] In a preferred embodiment, the process of generating the market performance pre-assessment results and risk control recommendations specifically includes:

[0038] Obtain personalized design solutions, and build a market performance prediction model based on random forest based on the personalized design solutions. The market performance prediction model includes sales forecast, user acceptance forecast and market risk assessment.

[0039] Monte Carlo simulation method is used to conduct multi-dimensional risk assessment of market performance prediction model, and the number of simulations and confidence intervals are set.

[0040] A decision support analysis report is generated based on the results of multi-dimensional risk assessment, providing preliminary market performance assessments and risk control recommendations.

[0041] In a preferred embodiment, it further includes:

[0042] A performance monitoring system is established based on real-time feedback data of personalized design schemes. The accuracy of the assessment is evaluated by comparing the real-time feedback data of personalized design schemes with the market performance pre-evaluation results. Online learning and incremental learning technologies are used to update the market performance prediction model.

[0043] In a preferred embodiment, the big data acquisition and analysis system based on clothing design is used to execute the above-described big data acquisition and analysis method based on clothing design, including:

[0044] The data acquisition module is used to acquire multi-source heterogeneous data, including sales data from e-commerce platforms, user-generated content from social media, professional information from fashion websites, and environmental data. Data quality assessment algorithms and cleaning techniques are used to obtain standardized datasets from the multi-source heterogeneous data.

[0045] The feature extraction module is used to extract clothing image features and multi-dimensional user features from a standardized dataset using feature engineering; it quantifies the design elements of clothing image features and analyzes the correlation between these elements to obtain feature vectors of clothing design elements.

[0046] The trend prediction module is used to acquire environmental data from a standardized dataset. It inputs the feature vectors of clothing design elements and environmental data into a time series prediction model based on LSTM and a multi-factor influence model to obtain the trend prediction results.

[0047] The user analysis module is used to perform user preference clustering analysis on multi-dimensional user characteristics to obtain user preference mining results.

[0048] The design generation module is used to input user preference mining results and popular trend prediction results into the generative adversarial network to obtain candidate design schemes, optimize the candidate design schemes, evaluate their quality and screen them to obtain personalized design schemes;

[0049] The market assessment module is used to conduct market forecasting and risk assessment of personalized design solutions, and to obtain pre-assessment results of market performance and risk control suggestions.

[0050] The update and optimization module establishes a performance monitoring system based on real-time feedback data of personalized design schemes. It compares the accuracy of the real-time feedback data of personalized design schemes with the market performance pre-evaluation results and uses online learning and incremental learning technologies to update the market performance prediction model.

[0051] The beneficial effects of this invention are as follows: By establishing a multi-source heterogeneous data acquisition network and a deep learning prediction model, this invention can comprehensively consider multi-dimensional information such as seasonal factors, economic factors, cultural event factors, and social media popularity. Compared with traditional prediction methods, the prediction accuracy in various aspects such as color trends, style trends, and fabric trends is improved. In particular, through the application of time series deep learning models, it can capture the complex changing patterns of fashion trends, achieve accurate prediction of future fashion trends, provide designers with scientific trend guidance, and effectively reduce the market risks of new product development.

[0052] By constructing a refined user profile model and a design scheme generation mechanism based on generative adversarial networks, personalized design suggestions can be provided according to the characteristics and preferences of different user groups. It not only considers the demographic characteristics and historical behavior of users, but also deeply explores the potential preferences and social network influence of users, realizing the transformation from mass-market design to personalized customization. Through multi-objective optimization algorithms, while meeting the personalized needs of users, it can also take into account business constraints such as market adaptability and cost control, thereby improving user satisfaction and the commercial feasibility of design schemes. Attached Figure Description

[0053] Figure 1 This is a flowchart of the big data collection and analysis method based on clothing design in this invention;

[0054] Figure 2 This is a block diagram of the big data collection and analysis system based on clothing design in this invention;

[0055] Figure 3 It is a bar chart comparing the accuracy of trend predictions;

[0056] Figure 4It is a line graph showing the improvement in user satisfaction;

[0057] Figure 5 It is a radar chart comparing improvements in business performance. Detailed Implementation

[0058] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0059] At least one embodiment of the present invention discloses a big data collection and analysis method based on clothing design, such as Figure 1 As shown, it includes:

[0060] Step 1: Obtain multi-source heterogeneous data, including sales data from e-commerce platforms, user-generated content from social media, professional information from fashion websites, and environmental data; apply data quality assessment algorithms and cleaning techniques to the multi-source heterogeneous data to obtain a standardized dataset;

[0061] Based on the diverse data needs of clothing design, a distributed data acquisition architecture is adopted to obtain a standardized clothing design dataset.

[0062] Step 1.1: Establish a multi-channel data collection network;

[0063] Based on web crawling technology and API calls, a multi-threaded concurrent data collection strategy was employed to obtain product sales data from e-commerce platforms. Specifically, a distributed crawler cluster was built to monitor clothing product information in real time on major e-commerce platforms such as Taobao, JD.com, and Amazon, including product descriptions, prices, sales volume, and user reviews, thus obtaining e-commerce platform data. Simultaneously, based on social media APIs, keyword matching and topic tracking technologies were used to obtain fashion-related user-generated content from platforms such as Instagram, Weibo, and Xiaohongshu, thus obtaining social media data. The e-commerce platform data and social media data were then fused with professional information from fashion websites to obtain multi-source heterogeneous data.

[0064] Step 1.2: Implement data quality control and cleaning;

[0065] Based on the collected multi-source heterogeneous data, a data quality assessment algorithm is used to obtain data integrity and accuracy scores. Encoding format detection and character normalization are performed on text data, converting text with different encoding formats to UTF-8. For image data, format verification and size normalization are performed to ensure that the image resolution and aspect ratio meet the requirements of subsequent processing. Then, outliers are detected based on rules and statistical methods, including price anomalies, rating anomalies, timestamp anomalies, etc., and detected outliers are marked or removed. By establishing a data quality assessment index system, including integrity score, accuracy score, and timeliness score, the comprehensive data quality score is calculated according to the following steps:

[0066] Calculate the data integrity score: count the number of missing values ​​in each field of the data set, calculate the proportion of missing fields to the total number of fields, and subtract this proportion from 1 to get the integrity score. The score ranges from 0 to 1, and the closer the value is to 1, the more complete the data is.

[0067] Calculate the data accuracy score: The consistency of the data is checked by cross-validation. The consistency of the same information in different data sources is compared. The proportion of accurate data to the total data is used as the accuracy score. The score ranges from 0 to 1. The closer the value is to 1, the more accurate the data is.

[0068] Calculate the data timeliness score: Based on the interval between the data collection time and the current time, establish a timeliness decay function. The longer the time interval, the higher the score. The score range is from 0 to 1. The closer the value is to 1, the fresher the data is.

[0069] Determine the weighting coefficients: Based on the quality requirements of different application scenarios, assign weighting coefficients to completeness, accuracy, and timeliness respectively. The sum of the three weighting coefficients equals 1, and the weighting allocation can be dynamically adjusted according to actual needs.

[0070] Calculate the overall score: Multiply the completeness score by its weight coefficient, the accuracy score by its weight coefficient, and the timeliness score by its weight coefficient. Then add the three weighted results together to obtain the final overall data quality score, which ranges from 0 to 1.

[0071] Step 1.3: Perform data standardization and format conversion;

[0072] Based on the cleaned data, data standardization techniques are employed to obtain a structured dataset with a unified format. Standardization specifications for apparel design data are established, including color coding standards, style classification systems, and fabric labeling standards, mapping data from different sources to a unified data model. Through data format conversion and field mapping, seamless integration of information from different data sources is ensured.

[0073] Furthermore, incremental data synchronization technology can be used to replace batch data processing. Based on message queues and streaming processing frameworks, real-time data stream processing technology is employed to achieve near real-time data update capabilities. This alternative embodiment can improve the real-time performance of data acquisition, better capture rapidly changing fashion trends, and is particularly suitable for fast fashion brands that need to respond quickly to market changes.

[0074] Step 2: Use feature engineering to extract clothing image features and multi-dimensional user features from the standardized dataset; quantify the design elements of the clothing image features and analyze the correlation of the elements to obtain the feature vector of clothing design elements;

[0075] Based on a standardized clothing design dataset, deep learning feature extraction technology is used to obtain feature vectors of clothing design elements.

[0076] Step 2.1: Construct a clothing image feature extraction model;

[0077] Based on a convolutional neural network architecture, a multi-layer feature extraction strategy is employed to obtain multi-dimensional feature representations of clothing images. Basic visual features of the images are extracted using a pre-trained ResNet network, and then a dedicated clothing attribute recognition branch network is constructed to identify design elements such as color, style, pattern, and fabric. The specific steps for establishing the color feature vector are as follows:

[0078] Image color space conversion: Convert the input RGB image to the HSV color space to better separate color and brightness information.

[0079] Color component extraction: Extracting the values ​​of each color channel from the converted image, including information in three dimensions: hue, saturation, and brightness.

[0080] Color intensity calculation: Calculate the intensity value of each color component in the entire image. The intensity value reflects the importance and proportion of the color in the image.

[0081] Feature vector construction: Arrange the intensity values ​​of all color components in order to form a multi-dimensional vector. Each element of the vector represents the intensity of a color component, with a value range from 0 to 255.

[0082] Vector normalization: Normalize the constructed color feature vectors to ensure that the color features of different images are comparable.

[0083] Step 2.2: Establish a quantitative evaluation system for design elements;

[0084] Based on the extracted visual features, feature fusion and dimensionality reduction techniques are used to obtain a standardized quantitative representation of the design elements. A style innovation index is defined, and the degree of innovation is measured by calculating the similarity difference between the design style and historically popular styles through the following steps:

[0085] Historical style database construction: Collect and organize the feature vectors of all popular styles in history, and establish a complete historical style database as a comparison benchmark.

[0086] Feature vector extraction: Extract features from the current design style to be evaluated to obtain its corresponding multi-dimensional feature vector representation.

[0087] Similarity calculation: The cosine similarity is calculated between the feature vector of the current design and the feature vector of each design in the historical design library to obtain a series of similarity scores.

[0088] Maximum similarity determination: Find the maximum value among all calculated similarity scores. This maximum value represents the highest degree of similarity between the current design and historical styles.

[0089] Innovation Calculation: Subtract the maximum similarity value obtained in step four from 1 to obtain the final style innovation index. The value ranges from 0 to 1, and the larger the value, the higher the innovation.

[0090] The cosine similarity calculation function is implemented as follows: The two input feature vectors A and B are normalized, then the dot product of the two vectors is calculated, and finally divided by the product of their respective magnitudes to obtain a similarity score ranging from -1 to 1. Specifically, for vectors A and B, the similarity is calculated as the dot product of vectors A and B divided by the product of the magnitudes of vectors A and B. This function effectively measures the degree of directional consistency between two design feature vectors and is unaffected by the vector magnitudes.

[0091] Step 2.3: Construct a correlation analysis model for design elements;

[0092] Based on the quantified design element characteristics, an association rule mining algorithm is used to obtain the association relationships between different design elements. By calculating the support and confidence of element combinations, high-frequency design element combination patterns are identified, providing a knowledge base for subsequent design scheme generation. The specific steps for establishing the element association strength matrix are as follows:

[0093] Frequency of element occurrence: Traverse all design samples, count the total number of times each design element appears individually, and create an element frequency statistics table.

[0094] Statistical co-occurrence frequency of elements: Count the number of times any two design elements appear simultaneously in the same design sample, and establish an element co-occurrence frequency matrix.

[0095] Calculate the association strength: For any two elements, divide their co-occurrence frequency by the product of their individual occurrence frequencies to obtain the normalized association strength value.

[0096] Construct the association matrix: Arrange the association strength values ​​of all feature pairs according to the row and column correspondence to form a complete feature association strength matrix. The value of each element in the matrix ranges from 0 to 1.

[0097] Interpretation of correlation strength: The closer the value in the matrix is ​​to 1, the stronger the correlation between the two corresponding elements, and the more likely they are to appear simultaneously in the design; the closer the value is to 0, the weaker the correlation.

[0098] Furthermore, attention-enhanced feature extraction methods can be used to replace traditional convolutional neural networks. Based on the visual Transformer architecture, a self-attention mechanism is employed to obtain more accurate local design detail features. This alternative embodiment can better capture subtle feature differences in clothing design, improve the accuracy of design element recognition, and is particularly suitable for high-end fashion brands with high requirements for design details.

[0099] Step 3: Obtain environmental data from the standardized dataset, input the feature vectors of clothing design elements and environmental data into the LSTM-based time series prediction model and multi-factor influence model to obtain the fashion trend prediction results;

[0100] Based on the feature vectors of design elements and environmental data, a time series prediction model based on Long Short-Term Memory (LSTM) is used to obtain the prediction results of future trends.

[0101] Step 3.1: Construct a multi-factor trend influence model;

[0102] Based on historical trend data in a standardized dataset, a multivariate time series analysis method was used to obtain the weight distribution of trend influencing factors.

[0103] Before building a multi-factor model, data preprocessing of various influencing factors is required to ensure the model's effectiveness:

[0104] The steps for handling seasonal factors are as follows: collect periodic time feature data such as months and quarters, then convert these time features into sine and cosine encoding forms, and eliminate the influence of linear trends in the time series through trigonometric function transformation to ensure that seasonal changes can be correctly identified and processed by the model.

[0105] The steps for processing economic factors are as follows: collect raw data of various economic indicators such as GDP growth rate, consumer price index, and unemployment rate; calculate the mean and standard deviation of each economic indicator in historical data; then perform Z-score standardization on each data point by subtracting the mean and dividing by the standard deviation to ensure that different economic indicators have the same numerical range and distribution characteristics.

[0106] The steps for processing cultural event factors are as follows: identify and collect information on categorized events such as holidays and cultural activities, convert these categorized events into one-hot encoded vector representations, evaluate the impact intensity of each event, and perform a linear transformation on the impact intensity values ​​through the minimum and maximum values ​​to map them into a standardized range of 0-1.

[0107] The steps for processing social media popularity factors are as follows: collect raw data of social media-related indicators such as search popularity, number of topic mentions, and number of reposts; perform logarithmic transformation on these data to reduce the skewness of the data distribution; then calculate the minimum and maximum values ​​of the transformed data; and map the data to the 0-1 interval using the minimum and maximum normalization methods.

[0108] Through the above systematic preprocessing steps, we can ensure that influence factors of different types and dimensions can be reasonably weighted and effectively represented in the multi-factor model.

[0109] A multi-factor model for trend prediction is established, comprehensively considering seasonal factors, economic factors, cultural event factors, and social media popularity factors. The calculation steps for the trend prediction value are as follows:

[0110] Factor weight determination: Based on historical data analysis and expert experience, weight coefficients are assigned to seasonal factors, economic factors, cultural event factors, and social media popularity factors to ensure that the sum of the four weight coefficients equals 1.

[0111] Seasonal Factor Calculation: Based on the current time point, calculate the impact of seasonal factors on fashion trends, taking into account the cyclical influence of spring, summer, autumn and winter seasons on clothing fashion trends.

[0112] Economic Factors Calculation: Collect current economic indicator data, including GDP growth rate, consumer price index, unemployment rate, etc., and calculate the impact of the economic environment on consumers' purchasing power and willingness to consume fashion.

[0113] Cultural event factor calculation: Identify important cultural events, holidays, social hotspots, etc. in the current period, assess the driving effect of these events on clothing fashion trends, and calculate the impact value of cultural event factors.

[0114] Social media popularity calculation: Statistics on fashion-related search popularity, topic mentions, and reposts on social media platforms are collected to calculate the impact value of social media popularity factors.

[0115] Weighted summation calculation: Multiply the influence value of each factor by its corresponding weight coefficient, and then sum all the weighted results to obtain the basic prediction value.

[0116] Error term handling: Considering the uncertainty in the prediction process, a random error term is added to the basic predicted value. This error term follows a normal distribution and reflects the random fluctuations in the prediction.

[0117] Final forecast: The base forecast is added to the error term to obtain the final trend forecast index.

[0118] Step 3.2: Establish a time series prediction model based on LSTM;

[0119] Based on LSTM, a multi-layer recursive structure is used to obtain long-term dependency patterns in time series data. A deep network architecture consisting of an input layer, multiple LSTM hidden layers, and an output layer is constructed, and the model parameters are trained using the backpropagation algorithm. A time window length of 30 days is set to predict the trend changes over the next 7 days.

[0120] Step 3.3: Optimize the accuracy of trend forecasting;

[0121] Based on error analysis between the predicted results and actual trends, model ensemble techniques are employed to achieve more stable and accurate predictive performance. Multiple predictive models with different architectures, including LSTM, GRU, and Transformer, are combined using a weighted average to fuse their predictions. The mean absolute percentage error (MAPE) metric is used to evaluate predictive accuracy.

[0122] Furthermore, reinforcement learning-optimized prediction models can replace traditional supervised learning methods. Based on the Actor-Critic algorithm framework, an environment-feedback-driven learning strategy is adopted to obtain a predictive model that can adaptively adjust. This alternative embodiment uses market feedback as a reward signal, enabling continuous optimization of the prediction strategy and improving the responsiveness to sudden trends.

[0123] Step 4: Perform user preference clustering analysis on multi-dimensional user characteristics to obtain user preference mining results;

[0124] Based on user behavior data and consumption records, cluster analysis and deep learning techniques are used to obtain a refined user preference model and personalized profile.

[0125] Step 4.1: Construct a multi-dimensional user feature system;

[0126] Based on user purchase history, browsing behavior, social media activity, and other data, feature engineering methods are used to obtain user preference mining results.

[0127] Standardization preprocessing for different types of raw data:

[0128] The steps for processing demographic data are as follows: Categorical variables (such as gender: male / female, geographic location: city name) are converted into one-hot encoded vectors to ensure that each category corresponds to a binary feature dimension; For continuous variables (such as age: 18-65 years old), minimum and maximum normalization are performed through the following steps: Determine the minimum and maximum values ​​of the feature in the dataset, and then for each data point, subtract the minimum value from the original value of the data point, and then divide by the difference between the maximum and minimum values, finally mapping the result to the 0-1 interval.

[0129] The steps for processing user behavior data are as follows: Standardize indicators of different dimensions, including browsing frequency (number of times / day), dwell time (seconds), click behavior (number of times / session), etc. The Z-score standardization method is used for processing: calculate the mean and standard deviation of each indicator in historical data, and then for each data point, subtract the mean from the original value of the data point and divide by the standard deviation to ensure that different behavior indicators have the same numerical range and distribution characteristics.

[0130] The processing steps for preference data are as follows: the user's historical ratings (out of 1 to 5) are normalized to the 0-1 range through a linear transformation. Specifically, the rating is subtracted by 1 and then divided by 4. For missing ratings, the average user rating or the average item rating is used to fill in the missing ratings to ensure data integrity.

[0131] The steps for processing social features are as follows: logarithmic transformation is performed on indicators such as influence index and social activity to eliminate the influence of skewed data distribution, and then the transformed data is mapped to the standardized interval through minimum and maximum normalization methods.

[0132] By statistically analyzing user browsing frequency, dwell time, and click behavior within a time window, and calculating statistical characteristics of the behavioral sequences such as mean, variance, and trend, we can determine user preferences. For preference data, a preference rating matrix is ​​constructed based on users' historical ratings and feedback, and latent preference factors are extracted using matrix factorization techniques. Social features are analyzed by examining users' social network data, including their following relationships, interaction behaviors, and content sharing, to calculate their influence index and social activity. The steps for constructing user feature vectors are as follows:

[0133] Demographic feature extraction: Collect basic information such as users' age, gender, and geographical location to form demographic feature sub-vectors.

[0134] Behavioral feature calculation: Statistics on user browsing frequency, dwell time, click behavior, and other indicators are collected to calculate the statistical features of the behavioral sequence and form behavioral feature sub-vectors.

[0135] Preference feature construction: Based on users' historical ratings and feedback data, potential preference factors are extracted through matrix factorization techniques to form preference feature sub-vectors.

[0136] Social Feature Analysis: Analyze users' social network data such as follower relationships, interaction behaviors, and content sharing to calculate influence index and social activity, forming social feature sub-vectors.

[0137] Feature vector combination: Connect the demographic feature sub-vectors, behavioral feature sub-vectors, preference feature sub-vectors, and social feature sub-vectors in sequence to form a complete user feature vector.

[0138] Step 4.2: Perform user preference cluster analysis;

[0139] Based on multi-dimensional user characteristics, the K-means clustering algorithm is used to obtain subcategories of the user group. Users with similar preferences are grouped into the same category by calculating the Euclidean distance between user feature vectors. The optimal number of clusters is optimized using the elbow rule and silhouette coefficient. The user similarity calculation process is as follows:

[0140] Vector dot product calculation: Calculate the dot product of two user feature vectors, which is to multiply the feature values ​​of corresponding dimensions and then sum them.

[0141] Vector norm calculation: Calculate the Euclidean norm of the two user feature vectors respectively, which is the square root of the sum of the squares of the eigenvalues ​​of each dimension.

[0142] Cosine similarity calculation: Divide the dot product of the vectors by the product of the norms of the two vectors to obtain the cosine similarity value.

[0143] Similarity Interpretation: Cosine similarity ranges from -1 to 1. The closer the value is to 1, the more similar the user preferences are; the closer the value is to -1, the greater the difference in user preferences; and the closer the value is to 0, the less significant the correlation between user preferences.

[0144] Step 4.3: Establish a personalized preference prediction model;

[0145] Based on user clustering results and historical preference data, a collaborative filtering algorithm is used to predict user preferences for different design elements. A user item rating matrix is ​​constructed, and matrix factorization techniques are used to mine potential user preference factors. A user preference rating prediction model is established to generate personalized design recommendations for each user.

[0146] Furthermore, a user preference modeling method based on graph neural networks can be used to replace the traditional collaborative filtering algorithm. Based on user social relationship graphs and product association graphs, a graph convolutional network is employed to obtain richer user and product representations. This alternative embodiment can fully utilize user social network information and product attribute relationships, improving the accuracy of preference prediction and the diversity of recommendations.

[0147] Step 5: Input the user preference mining results and popular trend prediction results into the generative adversarial network to obtain candidate design schemes. Optimize, evaluate and screen the candidate design schemes to obtain personalized design schemes.

[0148] Based on the results of trend prediction and user preference mining, generative adversarial networks and multi-objective optimization algorithms are used to obtain personalized design schemes.

[0149] Step 5.1: Construct a generative adversarial network for clothing design;

[0150] Based on the GAN (Generative Adversarial Network) architecture, an adversarial training mechanism between the generator and discriminator is employed to obtain a model capable of generating innovative design solutions. The generator network takes user preference features and popular trend information as input and outputs candidate design solutions. The discriminator network is responsible for evaluating the quality and realism of the generated designs. Through an iterative optimization process, the generator is able to produce design solutions that are both in line with user preferences and innovative.

[0151] Step 5.2: Implement multi-objective design optimization;

[0152] Based on the generated candidate design schemes, a multi-objective genetic algorithm is used to obtain the Pareto optimal solution set.

[0153] Before performing multi-objective optimization, the input parameters of each objective function need to be preprocessed to ensure the rationality of the calculation:

[0154] The preprocessing steps for market adaptability scoring are as follows: It is necessary to normalize indicators of different dimensions, such as trend matching degree (0-1 range), historical sales data (sales volume value), and market capacity (market size value), and use the minimum and maximum normalization methods to unify them into the 0-1 range. Specifically, the minimum and maximum values ​​of each indicator are found, and then the original value is subtracted from the minimum value and divided by the difference between the maximum and minimum values.

[0155] The preprocessing steps for user satisfaction ratings are as follows: the ratings (out of 1 to 5) of different user groups need to be mapped to the 0-1 range through linear transformation, the preference weights (0-1 range) are kept within their original numerical range, the categorical features in the demographic features (a mixture of categorical and numerical features) are converted into one-hot codes, and the numerical features are standardized to ensure that different types of features can be reasonably integrated.

[0156] For the preprocessing steps of innovation scoring: the design similarity (cosine similarity: -1 to 1) needs to be mapped to the 0-1 interval through linear transformation. Specifically, the similarity value is added by 1 and then divided by 2. The novelty index of design elements is processed to the 0-1 interval through the minimum and maximum normalization methods.

[0157] For the preprocessing steps of production costs, it is necessary to standardize the cost components such as raw material costs (currency unit), labor costs (currency unit), and equipment costs (currency unit), and normalize them by using the cost percentage method, that is, dividing each cost component by the total cost, to ensure the numerical stability and comparability of the cost function.

[0158] Establish a multi-objective optimization function that simultaneously considers multiple objectives such as market adaptability, user satisfaction, production cost, and innovation. The steps for performing multi-objective optimization are as follows:

[0159] Objective function definition: Four optimization objectives are determined, namely, maximizing market adaptability, maximizing user satisfaction, maximizing innovation, and minimizing production costs.

[0160] The calculation of the objective function specifically includes:

[0161] Market adaptability score calculation: Based on the trend prediction results, the matching degree between the design scheme and the current trend is calculated. Combined with historical sales data analysis of the market performance of similar designs, a comprehensive market adaptability score is calculated by weighted summation. The weight coefficients are determined based on factors such as trend matching degree, historical performance similarity and market capacity. Finally, it is mapped to the 0 to 1 range through the sigmoid function.

[0162] User satisfaction score calculation: Based on the user preference model, the degree of matching between the design scheme and the preferences of the target user group is calculated. The collaborative filtering algorithm is used to predict the user's rating of the design scheme. Personalized adjustments are made in combination with user profile data. The weighted average method is used to calculate the comprehensive satisfaction score of different user groups.

[0163] Innovation score calculation: The similarity between the design scheme and all designs in the historical design library is calculated, and the minimum similarity value is taken as the basic innovation score. Then, it is adjusted by combining factors such as the novelty of design elements and the uniqueness of the combination, and finally the innovation score in the range of 0 to 1 is obtained.

[0164] Production cost calculation: Based on parameters such as fabric selection, process complexity, and production batch size in the design scheme, the cost estimation model is used to calculate the raw material cost, labor cost, and equipment cost, and the total production cost is obtained by linear weighted summation.

[0165] Multi-objective vector construction: Market adaptability score, user satisfaction score, and innovation score are used as maximization objectives, and the negative value of production cost is used as the maximization objective (to minimize cost), thus constructing a multi-objective function vector.

[0166] Pareto optimal solution search: A multi-objective genetic algorithm is used to search for the Pareto optimal solution set through operations such as population initialization, selection, crossover, and mutation, ensuring balanced optimization among the objectives.

[0167] Solution set evaluation and selection: Evaluate the Pareto optimal solution set and select the most suitable design scheme based on the decision-maker's preferences and actual constraints.

[0168] Step 5.3: Perform a quality assessment and screening of the design schemes;

[0169] Based on the results of multi-objective optimization, the analytic hierarchy process (AHP) and fuzzy comprehensive evaluation were used to obtain a comprehensive score ranking of the design schemes. An evaluation index system for the design schemes was established, including dimensions such as aesthetic score, market potential score, and technical feasibility score. Through a combination of expert review and user surveys, the weight coefficients of each index were determined, and the comprehensive score of the design schemes was calculated.

[0170] Furthermore, reinforcement learning-based design generation methods can be used to replace generative adversarial networks (GANs). Based on the Deep Q-Network (DQN) framework, an interaction mechanism of actions, states, and rewards is employed to obtain a design generation model capable of autonomous learning and improvement. This alternative example, by modeling the design process as a sequential decision problem, allows for better control over the design generation process and results, improving the controllability and quality of the design solutions.

[0171] Step 6: Conduct market forecasting and risk assessment for the personalized design scheme to obtain preliminary market performance assessment results and risk control recommendations;

[0172] Based on personalized design schemes and market environment data, a market performance prediction model based on random forest is used to obtain the pre-assessment results of the design scheme's market performance and risk control suggestions.

[0173] Step 6.1: Establish a market performance forecasting model;

[0174] Based on historical product sales data and design features, a random forest regression algorithm is used to predict the sales potential of the design scheme.

[0175] Personalized design feature extraction: The personalized design scheme generated in step 5 is converted into a quantitative design feature vector, including color preference matching degree for a specific user group (0-1 score), style innovation score (based on the degree of differentiation from historical best-selling products), and richness of personalized elements (the number and complexity of personalized elements such as patterns, decorations, and details).

[0176] User group suitability assessment: Based on the user preference model constructed in step 4, calculate the matching degree between the personalized design scheme and the target user group's preferences, and generate a user group suitability feature vector, including dimensions such as age group matching degree, spending power matching degree, and style preference matching degree.

[0177] Personalization Quantification: By comparing the degree of difference between personalized design schemes and mainstream market designs, a personalization index is calculated. This index reflects the uniqueness and differentiation level of the design scheme and serves as an important input feature for the predictive model.

[0178] Before building a predictive model, comprehensive data preprocessing of the input features is required:

[0179] The preprocessing steps for the design feature vector are as follows: normalize the color features (RGB values: 0-255) to the 0-1 range by dividing by 255; convert the style features (categorical variables) into one-hot encoded vector representations; and perform Z-score standardization on the fabric features (texture complexity, thickness, etc.), that is, calculate the mean and standard deviation, subtract the mean from each value, and then divide by the standard deviation.

[0180] The preprocessing steps for the market environment feature vector are as follows: seasonal factors are encoded as periodic features; sine and cosine transforms are used to convert time information into continuous numerical features; economic indicators (GDP growth rate, consumer confidence index, etc.) are normalized and mapped to the 0-1 interval using minimum and maximum normalization methods; consumer trend data (search popularity, number of mentions on social media, etc.) are processed through logarithmic transformation to reduce data skewness, and then standardized.

[0181] The preprocessing steps for competitor feature vectors are as follows: competitor prices (currency units) are normalized by price range, i.e., the minimum and maximum price ranges are determined, and then the prices are mapped to the 0-1 range. Competitor style features are converted into similarity scores (0-1 range) with the target design. The similarity scores are calculated by comparing features and similarity. The market share (percentage) is directly divided by 100 to convert it into a value in the 0-1 range.

[0182] Through the above systematic preprocessing steps, it is ensured that features of different types and dimensions can be reasonably weighted and effectively recognized in the random forest model.

[0183] Construct a predictive model incorporating multi-dimensional inputs, including design features, market environment, and competitor information. Predictive metrics include expected sales volume, market share, and profitability. The execution steps of the sales volume prediction model are as follows:

[0184] Feature vector construction: Combine the design feature vector (including design elements such as color, style, and fabric), the market environment feature vector (including factors such as season, economic conditions, and consumption trends), and the competitor feature vector (including information such as competitor prices, styles, and market share) into a complete input feature vector.

[0185] Random Forest Model Construction: Construct multiple decision tree models, each of which is trained based on a randomly sampled training subset and feature subset. The prediction results of multiple decision trees are combined using the bootstrap aggregation method.

[0186] Feature Importance Assessment: Feature selection and importance assessment are performed on the input multi-dimensional feature vector to identify the feature variables that have the greatest impact on sales forecasting.

[0187] Decision tree ensemble prediction: The final prediction value is calculated by integrating the prediction results of multiple decision trees. The mean aggregation strategy is used to reduce the prediction error of individual models and the risk of overfitting.

[0188] Forecast output: Outputs the expected sales forecast, along with the forecast confidence interval and uncertainty assessment.

[0189] Step 6.2: Conduct a multi-dimensional risk assessment;

[0190] Based on the uncertainty analysis of the prediction results, the Monte Carlo simulation method is used to obtain a risk assessment report for the design scheme.

[0191] Personalized risk factor identification: In view of the special characteristics of personalized design solutions, identify risk factors related to the degree of personalization, including market acceptance risk (excessive personalization may lead to niche market), production complexity risk (personalized elements increase production difficulty and cost), and inventory risk (personalized products are more difficult to predict demand).

[0192] User group risk assessment: Based on the characteristics of the target user group of the personalized design plan, assess the risks related to the user group, including the risk of target user group size, the risk of fluctuation in spending power, the risk of changes in preferences, etc., and simulate the impact of changes in different user group parameters on sales performance through Monte Carlo simulation.

[0193] Personalized premium risk analysis: Assess the pricing strategy risk of personalized design solutions, analyze the balance between personalized premium and market acceptance, and determine the optimal pricing range and risk threshold through sensitivity analysis.

[0194] Establish a risk assessment indicator system, including dimensions such as market risk, technological risk, cost risk, and time risk. Through probability distribution simulation and sensitivity analysis, assess the impact of various uncertainties on the probability of project success.

[0195] Step 6.3: Generate decision support reports and recommendations;

[0196] Based on the pre-assessment results and risk analysis, decision tree analysis is used to obtain decision recommendations for different situations.

[0197] Personalized Solution Optimization Suggestions: Based on the forecast results and risk assessment, specific optimization suggestions are provided for personalized design solutions, including suggestions for adjusting personalized elements (such as reducing overly complex personalized elements to control costs), suggestions for segmenting the target user group (such as adjusting the scope of the target user group to expand the market), and suggestions for balancing the degree of personalization (finding the optimal balance between uniqueness and market acceptance).

[0198] Layered decision-making strategy formulation: For design schemes with different levels of personalization, a layered decision-making strategy is formulated, including a small-batch trial production strategy for highly personalized schemes, a phased promotion strategy for moderately personalized schemes, and a large-scale production strategy for low-personalization schemes, providing managers with differentiated investment and production decision-making suggestions.

[0199] Personalized Value Assessment Report: Generates a value assessment report specifically for personalized design solutions, quantifying the intangible benefits brought by personalization, such as premium potential, brand value enhancement, and increased user loyalty, providing decision-makers with a comprehensive return on investment analysis.

[0200] Establish a decision support system to provide managers with comprehensive reports containing information such as expected returns, risk levels, and investment recommendations. Use visual charts to display key indicators and trends to support rapid decision-making.

[0201] Furthermore, a dynamic risk assessment method based on deep reinforcement learning can be used to replace the traditional statistical model. Based on temporal difference learning, a dynamically adjusted risk assessment strategy is employed to obtain a risk control model that can adapt to changes in the market environment. This alternative example, through continuous learning of market feedback, can adjust risk assessment criteria in real time, improving the accuracy and timeliness of risk warnings.

[0202] In one embodiment of the present invention, in order to compare the accuracy of real-time feedback data of personalized design schemes with the market performance pre-evaluation results, the big data collection and analysis method based on clothing design further includes:

[0203] A performance monitoring system is established based on real-time feedback data of personalized design schemes. The accuracy of the assessment is evaluated by comparing the real-time feedback data of personalized design schemes with the market performance pre-evaluation results. Online learning and incremental learning technologies are used to update the market performance prediction model.

[0204] Specifically, it includes the following:

[0205] Based on the actual market performance and user feedback of personalized design solutions, a continuously optimized big data analysis system is obtained by adopting online learning and model update technologies.

[0206] Based on post-launch sales data, user reviews, and market feedback, data stream processing technology is employed to obtain real-time performance feedback information. A multi-channel feedback collection network is constructed, including data sources such as point-of-sale (POS) systems, customer service centers, and social media monitoring. A standardized processing workflow for feedback data is established to ensure the timeliness and accuracy of feedback information.

[0207] Based on the collected feedback data, a model performance monitoring algorithm is used to obtain the accuracy evaluation results of each prediction model.

[0208] Establish a performance monitoring system based on personalized design schemes: Build a performance monitoring framework specifically for personalized design schemes, including personalized design scheme identification code management, real-time feedback data collection interface, and pre-evaluation result storage module, to ensure that the predicted results of each personalized design scheme can be accurately matched and compared with the actual performance.

[0209] Conduct comparative analysis of market performance pre-assessment results: Use the market performance pre-assessment results generated in step 6 as a benchmark and conduct a systematic comparison with the real-time feedback data of the personalized design scheme. Specifically, this includes deviation analysis between predicted sales and actual sales, correlation analysis between predicted user acceptance and actual user ratings, and consistency verification between predicted market risks and actual market performance.

[0210] Establish a prediction accuracy evaluation index system: Based on the comparison between the pre-evaluation results and the actual feedback data, calculate the prediction accuracy evaluation index, including the mean absolute percentage error (MAPE) of sales forecast, the Pearson correlation coefficient of user acceptance forecast, the accuracy and recall rate of risk warning, and evaluate the reliability of the market performance prediction model through quantitative indicators.

[0211] By comparing real-time feedback data of personalized design solutions with market performance pre-evaluation results, the performance differences of the prediction model under different levels of personalization, different user groups, and different market environments can be identified, providing precise directions for model optimization.

[0212] Based on the performance diagnostic results, an updated and optimized market performance prediction model is obtained using an online learning algorithm.

[0213] Establish a performance monitoring system based on real-time feedback data: Build a real-time data acquisition and processing framework, including modules for user behavior data stream processing, real-time acquisition of market feedback information, and tracking of design scheme effects, to ensure the timeliness and accuracy of feedback data.

[0214] Implement a pre-assessment result comparison and verification mechanism: compare and analyze the real-time feedback data with the market performance pre-assessment results generated in step 6 to evaluate the accuracy and reliability of the prediction model. Quantify the performance of the prediction model by calculating prediction bias, correlation analysis, and trend consistency testing.

[0215] Online and incremental learning are employed to update the prediction models: Based on comparative validation results, algorithms such as online gradient descent, incremental support vector machines, and streaming learning are used to update the fashion trend prediction model, user preference mining model, and market performance prediction model in real time. An adaptive learning rate adjustment mechanism is implemented to ensure that the models can quickly adapt to market changes.

[0216] Establish model version management and rollback mechanisms: Create a version identifier for each model update, record performance metric changes before and after the update, and quickly roll back to a stable version when the new model's performance degrades, ensuring system reliability and stability. Implement an incremental learning strategy to continuously incorporate new data and feedback information without losing historical knowledge. Verify the effectiveness of model updates through A / B testing to ensure continuous improvement in system performance.

[0217] Furthermore, a distributed model optimization method using federated learning can be used to replace centralized online learning. Based on the federated learning framework, a distributed collaborative training strategy is employed to obtain an optimized model that can utilize data from multiple parties while protecting privacy. This alternative example is particularly suitable for application scenarios that require integrating data from multiple partners but have data privacy constraints, enabling collaborative model optimization while protecting data privacy.

[0218] Big data collection and analysis system based on clothing design, such as Figure 2 As shown, the method for performing the above-mentioned big data collection and analysis based on clothing design includes:

[0219] The data acquisition module is used to acquire multi-source heterogeneous data based on e-commerce platforms and social media; data quality assessment algorithms and cleaning techniques are used to obtain standardized datasets from the multi-source heterogeneous data.

[0220] The feature extraction module is used to extract feature models of clothing images from standardized datasets using deep learning feature extraction technology; quantify the design elements of clothing image features and analyze the correlation between elements to obtain feature vectors of clothing design elements;

[0221] The trend prediction module is used to acquire environmental data from a standardized dataset. It uses an LSTM-based time series prediction model and a multi-factor influence model on the feature vectors of clothing design elements and environmental data to obtain the trend prediction results.

[0222] The user analysis module is used to obtain multi-dimensional user features based on standardized datasets using feature engineering methods; it performs user preference clustering analysis on the multi-dimensional user features, establishes a personalized preference prediction model, and obtains user preference mining results.

[0223] The design generation module is used to input user preference mining results and popular trend prediction results into the generative adversarial network to obtain candidate design schemes, optimize the candidate design schemes, evaluate their quality and screen them to obtain personalized design schemes;

[0224] The market assessment module is used to predict the market performance of personalized design solutions using a random forest-based market performance prediction model; the prediction results are then used to conduct a risk assessment using the Monte Carlo simulation method to obtain the pre-assessment results of market performance and risk control recommendations.

[0225] The update and optimization module is used to establish a performance monitoring system based on real-time feedback data of personalized design schemes, compare the real-time feedback data of personalized design schemes with the market performance pre-evaluation results to assess accuracy, and use online learning and incremental learning technologies to update the market performance prediction model.

[0226] In one embodiment of the present invention, a specific example is provided:

[0227] A 30-day field test was conducted in District B of City A. During the test, a distributed data collection system was deployed, covering 15 major business districts and shopping centers within the area, spanning approximately 50 square kilometers. By setting up intelligent data collection terminals within shopping malls and combining them with a mobile user behavior tracking system, consumer shopping behavior data and preference information were collected in real time. The test system processed over 100,000 data entries daily, covering the entire user behavior process, including browsing, trying on clothes, and purchasing.

[0228] Table 1 shows an example of data collection from e-commerce platforms:

[0229] Table 1: Examples of data collection from e-commerce platforms;

[0230]

[0231] Table 2 shows an example of social media trend data collection:

[0232] Table 2: Examples of Social Media Trend Data Collection;

[0233]

[0234] like Figure 3 As shown, the performance of the method of this invention in various trend prediction tasks is demonstrated. Compared with traditional prediction methods, this invention achieves accuracy improvements of 22%, 26%, 27%, 27%, and 25% in color trend, style trend, fabric trend, pattern trend, and overall trend prediction, respectively. Particularly in fabric trend prediction, the accuracy reaches 89%, superior to the 62% of traditional methods, demonstrating the advantages of multi-dimensional data fusion and deep learning technology.

[0235] like Figure 4 As shown, the improvement in user satisfaction before and after applying the system of this invention is demonstrated. In terms of design innovation, satisfaction increased from 3.2 to 4.3, an increase of 34%; personalization increased from 2.8 to 4.1, an increase of 46%; and popularity compatibility increased from 3.1 to 4.5, an increase of 45%. Overall satisfaction improved from 3.1 to 4.3, indicating that users showed higher approval of the design solution based on big data analysis.

[0236] like Figure 5 As shown, the present invention demonstrates its effectiveness in improving various business metrics. Sales conversion rate increased from 72% to 95%, new product success rate increased from 58% to 85%, inventory turnover rate increased from 65% to 88%, design efficiency increased from 45% to 78%, and market responsiveness increased from 52% to 90%. These improvements validate the commercial value and practicality of the clothing design method based on big data collection and analysis. Through accurate trend prediction and personalized recommendations, the system effectively reduces market risk and improves resource utilization efficiency.

[0237] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A big data collection and analysis method based on clothing design, characterized in that, The application relates to a personalized fashion design method based on multi-source heterogeneous data, and belongs to the technical field of fashion design. The application comprises the following steps: acquiring multi-source heterogeneous data, wherein the multi-source heterogeneous data comprises e-commerce platform sales data, social media user-generated content, fashion website professional information and environmental data; adopting a data quality evaluation algorithm and a cleaning technology on the multi-source heterogeneous data to obtain a standardized data set; adopting feature engineering to extract garment image features and multi-dimensional user features from the standardized data set; quantifying design elements of the garment image features and analyzing element correlation to obtain a garment design element feature vector; acquiring environmental data in the standardized data set, inputting the garment design element feature vector and the environmental data into an LSTM-based time series prediction model and a multi-factor influence model to obtain a popular trend prediction result; performing user preference clustering analysis on the multi-dimensional user features to obtain a user preference mining result; inputting the user preference mining result and the popular trend prediction result into a generative adversarial network to obtain a candidate design scheme, and performing optimization, quality evaluation and screening on the candidate design scheme to obtain a personalized design scheme; adopting an adversarial training mechanism of a generator and a discriminator to obtain a model capable of generating innovative design schemes; the generator network takes user preference features and popular trend information as input and outputs a candidate design scheme; the discriminator network is responsible for evaluating the quality and authenticity of the generated design; through an iterative optimization process, the generator can generate design schemes that meet user preferences and have innovation; performing market prediction and risk assessment on the personalized design scheme to obtain market performance pre-evaluation results and risk control suggestions, and the generation process of the market performance pre-evaluation results and the risk control suggestions comprises the following steps: acquiring the personalized design scheme, constructing a market performance prediction model based on a random forest based on the personalized design scheme, and the market performance prediction model comprises sales prediction, user acceptance prediction and market risk assessment; adopting a Monte Carlo simulation method to perform multi-dimensional risk assessment on the market performance prediction model, setting the simulation times and the confidence interval; generating a decision support analysis report based on the multi-dimensional risk assessment result to obtain the market performance pre-evaluation result and the risk control suggestion; 2. The big data collection analysis method based on clothing design according to claim 1, characterized in that, based on real-time feedback data of the personalized design scheme, a performance monitoring system is established, real-time feedback data of the personalized design scheme is compared with the market performance pre-evaluation result to evaluate the accuracy, online learning and incremental learning technology is adopted to update the market performance prediction model; a version identifier is established for each model update, the performance index change before and after the update is recorded, and when the performance of the new model decreases, the stable version can be quickly rolled back. The acquisition process of the multi-source heterogeneous data comprises the following steps: a distributed crawler cluster is constructed to monitor an e-commerce platform in real time, network crawler technology is adopted to acquire commodity description, price, sales and user evaluation data to obtain e-commerce platform data; based on the acquired e-commerce platform data, social media API interface calling is adopted to acquire fashion user-generated content by adopting keyword matching and topic tracking technology to obtain social media data; 3. The big data collection analysis method based on clothing design according to claim 1, characterized in that, the e-commerce platform data and the social media data are fused with the fashion website professional information to obtain multi-source heterogeneous data. The data quality evaluation algorithm and the cleaning technology comprise the following steps: The multi-source heterogeneous data is acquired, the text data in the multi-source heterogeneous data is subjected to encoding format detection, and different encoding formats detected are uniformly converted into UTF-8 encoding; The image data in the multi-source heterogeneous data is subjected to format verification and size normalization processing; Based on rules and statistical methods, abnormal values in the processed multi-source heterogeneous data are detected, price abnormality, score abnormality and timestamp abnormality data detected are marked and removed, and a standardized data set is obtained.

4. The big data collection analysis method based on clothing design according to claim 1, characterized in that, The generation process of the garment design element feature vector specifically includes: The standardized data set is acquired, and a garment image feature extraction model is constructed for the standardized data set by using a deep learning feature extraction technology; Basic visual features of the garment image are extracted through a pre-trained ResNet network, and a garment attribute recognition branch network is constructed based on the basic visual features to recognize design elements; A color feature vector is established, a style innovation degree index is defined by calculating the similarity difference between the design style and the historical popular style, and a garment design element feature vector is obtained.

5. The big data collection analysis method based on clothing design according to claim 1, characterized in that, The generation steps of the popular trend prediction result specifically include: The garment design element feature vector and the environmental data are acquired, a multi-factor influence model is established based on the garment design element feature vector and the environmental data, and seasonal factors, economic factors, cultural event factors and social media heat factors are comprehensively considered; An LSTM-based time series prediction model is used to train the multi-factor influence model, and a time window length and a prediction time length are set; The prediction results of the models are fused by using a weighted average method through combining prediction models with different architectures, and a popular trend prediction result is obtained.

6. The big data collection analysis method based on clothing design according to claim 1, characterized in that, The generation process of the user preference mining result specifically includes: The popular trend prediction result is acquired, a user behavior data set is constructed based on the popular trend prediction result, and the user behavior data set includes user browsing records, purchase records, evaluation records and collection records; A user similarity matrix and an item similarity matrix are calculated by analyzing the user behavior data set using a collaborative filtering algorithm; A user preference model is constructed based on the user similarity matrix and the item similarity matrix, users are divided into different preference groups by using a clustering algorithm, and a user preference mining result is obtained.

7. The big data collection analysis method based on clothing design according to claim 1, characterized in that, The generation of the individualized design scheme specifically includes: The user preference mining result is acquired, design constraint conditions are constructed based on the user preference mining result, and the design constraint conditions include cost constraints, material constraints and process constraints; The design constraint conditions are optimized and solved by using a genetic algorithm, and a population size, a crossover probability and a mutation probability are set; Market adaptability and user satisfaction of the design scheme are evaluated by using a fitness function, and a final design scheme is selected through multi-generation evolution, and an individualized design scheme is obtained.

8. A big data collection and analysis system based on clothing design, characterized by, The method for collecting and analyzing big data based on garment design according to any one of claims 1-7, comprising: A data collection module is used to acquire multi-source heterogeneous data, including e-commerce platform sales data, social media user-generated content, fashion website professional information and environmental data; data quality evaluation algorithm and cleaning technology are used on the multi-source heterogeneous data to obtain a standardized data set; The feature extraction module is configured to extract garment image features and multi-dimensional user features by using feature engineering on the standardized data set; quantize design elements of the garment image features and analyze element correlation to obtain a garment design element feature vector; The trend prediction module is configured to obtain environmental data in the standardized data set, input the garment design element feature vector and the environmental data into an LSTM-based time series prediction model and a multi-factor influence model to obtain a popular trend prediction result; The user analysis module is configured to perform user preference clustering analysis on the multi-dimensional user features to obtain a user preference mining result; The design generation module is configured to input the user preference mining result and the popular trend prediction result into a generative adversarial network to obtain a candidate design scheme, optimize, quality evaluate and screen the candidate design scheme, and obtain a personalized design scheme; The market evaluation module is configured to perform market prediction and risk evaluation on the personalized design scheme to obtain a market performance pre-evaluation result and a risk control suggestion; The update optimization module is configured to establish a performance monitoring system based on real-time feedback data of the personalized design scheme, compare the real-time feedback data of the personalized design scheme with the market performance pre-evaluation result to evaluate accuracy, and update the market performance prediction model by using online learning and incremental learning techniques.

Images