LightGBM-based large-size women dress size recommendation method and system

Abstract

The invention relates to the technical field of e-commerce, and discloses a LightGBM-based large-size women's dress size recommendation method and a LightGBM-based large-size women's dress size recommendation system. The method comprises the steps of preprocessing measurement data and determining a first confidence coefficient, generating a figure feature vector and determining a second confidence coefficient, calling a size prediction model and outputting a prediction probability, taking the maximum value of the prediction probability as a third confidence coefficient and outputting the third confidence coefficient, and determining a fourth confidence coefficient and an enhanced confidence coefficient. The system corresponds to the method. According to the method and the device, the problems of single dimension, static matching and the like in the prior art are solved, the method and the device are adaptive to complex body types of large-size users, trial and error cost is reduced, shopping experience is optimized, and a more intelligent, accurate and dynamic solution is provided for large-size women dress size recommendation.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of electronic commerce, and particularly relates to a large-size female garment size recommendation method and system based on LightGBM. BACKGROUND

[0002] Under the development trend of e-commerce clothing online shopping, large-size female garment users are faced with the problem of complex size selection. Large-size garment size is divided into many types, and there are large standard differences between brands and styles, and the body shape characteristics of large-size users are diverse, including various body shapes, part proportion differences and special cases. However, the existing size recommendation technology has the following defects:

[0003] Only based on single dimension such as height and weight, the complexity of the body shape cannot be reflected;

[0004] Static rule matching is adopted, and personalized learning ability is lacked;

[0005] The problem of non-uniform brand standards is not effectively solved;

[0006] There is no feedback mechanism, and it is difficult to continuously optimize;

[0007] There is no effective strategy for new users or new goods, which results in difficulty in size selection, high trial and error cost and high rate of return and exchange.

[0008] The above defects seriously affect the shopping experience of large-size female garment users, and need to be improved.

[0009] A clothing size recommendation method based on user evaluation and big data analysis is disclosed in Chinese Patent for Invention with the application publication number CN111667330A, and a clothing customization shopping guide method based on non-quantitative body intelligent size matching is disclosed in Chinese Patent for Invention with the application publication number CN120013650A. Both of them propose improvements on size recommendation, but both of them are not good in the explainability of size recommendation and the adaptability to various body shape characteristics.

[0010] In summary, a new large-size female garment size recommendation technical solution is urgently needed. SUMMARY

[0011] The purpose of the application is to provide a large-size female garment size recommendation method and system based on LightGBM, so as to solve the technical problems in the background.

[0012] In order to achieve the above purpose, the application discloses the following technical solutions:

[0013] In the first aspect, the application discloses a large-size female garment size recommendation method based on LightGBM, which comprises the following steps:

[0014] Step one: preprocessing the measurement data obtained after measuring the user's body shape, the measurement data including height, weight, bust, waist and hip circumference; determining a first confidence degree based on the quality of the measurement data, the first confidence degree being used to explain the reliability degree of the recommended size;

[0015] Step two: generating a body shape feature vector for defining the user's body shape condition based on the preprocessed measurement data, determining a second confidence degree for defining the matching condition between the user's body shape and the women's clothing size based on the relationship between the body shape feature vector and the women's clothing size, the second confidence degree being used to explain the reliability degree of the recommended size;

[0016] Step three: calling a size prediction model based on the type of women's clothing to be purchased, the size prediction model being trained based on the body shape feature vector and by a LightGBM algorithm;

[0017] Step four: inputting the body shape feature vector into the size prediction model, outputting the predicted probability corresponding to each women's clothing size, taking the maximum value in the output predicted probability as a third confidence degree, the third confidence degree being used to explain the reliability degree of the recommended size;

[0018] Step five: determining a fourth confidence degree based on historical purchase records, the historical purchase records including the total purchase condition of women's clothing and the purchase condition based on the recommended size, the fourth confidence degree being used to explain the reliability degree of the recommended size;

[0019] Step six: determining the size corresponding to the maximum value in the output predicted probability as the recommended size, and determining an enhanced confidence degree for explaining the reliability degree of the recommended size by statistically analyzing the first confidence degree, the second confidence degree, the third confidence degree and the fourth confidence degree, and outputting the recommended size and the enhanced confidence degree together.

[0020] Preferably, the preprocessing includes body measurement rationality verification, data cleaning, outlier detection and data standardization; wherein the body measurement rationality verification is to determine whether the measurement data is reasonable based on preset body rationality parameters.

[0021] Preferably, the determination process of the first confidence degree includes:

[0022] determining an integrity score based on the ratio of the number of non-empty fields to the total number of fields;

[0023] determining an accuracy score based on the ratio of the number of measurement items passing the body measurement rationality verification to the total number of measurement items;

[0024] determining a consistency score based on the difference obtained by subtracting the ratio of the number of logical conflicts existing in the preprocessed measurement data to the total number of logical relationships from one;

[0025] determine a data quality score based on the integrity score, the accuracy score and the consistency score;

[0026] determine the first confidence level based on the non-empty measurement field number, the total measurement field number and the data quality score.

[0027] Preferably, the generation process of the body feature vector comprises:

[0028] determine a relative proportion feature based on the proportion of bust circumference relative to height, the proportion of waist circumference relative to height and the proportion of hip circumference relative to height;

[0029] determine a body difference feature based on the difference between bust circumference and waist circumference and the difference between hip circumference and waist circumference;

[0030] determine a body type feature based on the bust circumference, the waist circumference, the hip circumference and a preset classification threshold.

[0031] Preferably, the determination process of the second confidence level comprises:

[0032] A1: determine a single-dimension matching score based on one data in the measurement data and the size range of the corresponding part on a woman's dress;

[0033] A2: weight-sum the single-dimension matching scores in A1 based on the weights corresponding to the parts to obtain a comprehensive matching score, and determine the comprehensive matching score as the second confidence level;

[0034] A3: when the body type feature exists, optimize A2: optimize the weights corresponding to the parts based on the body type feature, weight-sum the single-dimension matching scores in A1 based on the optimized weights to obtain an optimized comprehensive matching score, and determine the optimized comprehensive matching score as the second confidence level; when the body type feature does not exist, A3 is not performed;

[0035] A4: when the elasticity score exists, perform elasticity compensation on the comprehensive matching score in A2 or the optimized comprehensive matching score in A3 to obtain a compensated comprehensive matching score, and determine the compensated comprehensive matching score as the second confidence level; when the elasticity score does not exist, A4 is not performed, and the determination process of the elasticity score is: determine a basic elasticity score and a waist elasticity classification score based on the elasticity parameters of the woman's dress, determine a version adaptability score based on the version of the woman's dress, and determine the elasticity score based on the basic elasticity score, the waist elasticity classification score and the version adaptability score.

[0036] Preferably, the training process of the size prediction model comprises:

[0037] train a plurality of independent LightGBM models based on the woman's dress categories;

[0038] training the independent LightGBM model based on the body feature vector and its historical recommended size, so that the independent LightGBM model outputs the prediction probability of the corresponding size of women's clothing based on the body feature vector;

[0039] The trained independent LightGBM model is used as the size prediction model.

[0040] Preferably, the training process of the size prediction model further includes correcting the weight in the loss function in the LightGBM algorithm based on the number of samples in any size category.

[0041] Preferably, the determination process of the fourth confidence includes:

[0042] Obtaining the total purchase quantity of any type of women's clothing, and counting the purchase quantity of the user based on the recommended size of the women's clothing type;

[0043] Determining the fourth confidence based on the ratio of the purchase quantity to the total purchase quantity.

[0044] Preferably, the determination process of the enhanced confidence is a weighted sum based on the first confidence, the second confidence, the third confidence and the fourth confidence.

[0045] In a second aspect, the present application discloses a large-size women's clothing size recommendation system based on LightGBM, which applies the large-size women's clothing size recommendation method based on LightGBM as described above, and the system includes:

[0046] A data preprocessing module is configured to preprocess measurement data, wherein the measurement data includes the height, weight, bust, waist and hip of a user; a first confidence is determined based on the relationship between the preprocessing process and the data quality to define the quality of the measurement data;

[0047] A body feature analysis module is configured to generate a body feature vector for defining the body condition of a user based on the preprocessed measurement data, and to determine a second confidence for defining the matching condition between the body of a user and the size of women's clothing based on the relationship between the body feature vector and the size of women's clothing;

[0048] A model calling module is configured to call a size prediction model based on the type of women's clothing to be purchased, wherein the size prediction model is trained based on the body feature vector and its historical recommended size in combination with the LightGBM algorithm;

[0049] A model running module is configured to input the body feature vector of step two into the size prediction model of step three, output the prediction probability of each size of women's clothing, and output the maximum value of the prediction probability as a third confidence.

[0050] a historical recommendation analysis module configured to determine a fourth confidence for defining a historical case of size recommendation based on a relationship between historical recommendation size purchase data and actual needs of a user;

[0051] a size recommendation module configured to make a size recommendation based on a size corresponding to the third confidence and to make a statistical analysis based on the first confidence, the second confidence, the third confidence and the fourth confidence to determine an enhanced confidence, the enhanced confidence being used to evaluate a reliable degree of making a size recommendation based on a size corresponding to the first confidence.

[0052] Beneficial effects: the size recommendation method and system for large-size women's clothing based on LightGBM provided by the present application guarantee recommendation accuracy and reliability through multi-dimensional data processing and a confidence system; the measured data is preprocessed and the first confidence is determined to ensure the quality of input data; a body feature vector is generated and the second confidence is determined in combination with matching logic to accurately depict the body shape of a user and the size adaptation degree; a classification model trained based on LightGBM is called to output a prediction probability and a third confidence, and the algorithm advantage is used to improve the prediction effect; the fourth confidence is determined based on historical data to introduce feedback optimization; the enhanced confidence is obtained by weighting multiple confidences to comprehensively evaluate the reliability of recommendation; the problems such as single dimension and static matching of the prior art are solved, the complex body shape of large-size users is adapted, the trial and error cost is reduced, the shopping experience is optimized, and a more intelligent, accurate and dynamic solution is provided for size recommendation of large-size women's clothing. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0054] Figure 1 a flowchart of the size recommendation method for large-size women's clothing based on LightGBM provided by the embodiments of the present application;

[0055] Figure 2 a body type weight matrix of the size recommendation method for large-size women's clothing based on LightGBM provided by the embodiments of the present application;

[0056] Figure 3 a structural block diagram of the size recommendation system for large-size women's clothing based on LightGBM provided by the embodiments of the present application. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of them. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0058] In this document, the term "comprising" is intended to cover non-exclusive inclusion, so that a process, method, article 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 to such a process, method, article or device. Without more limitations, the elements defined by the statement "comprising" do not exclude the presence of additional identical elements in the process, method, article or device including the elements.

[0059] The present application aims at the prominent problem of size selection difficulty of large-size women's clothing users, especially the pain points of various sizes, different brand standards, complex and variable user body shapes, and large local size differences, and proposes a size recommendation method and system capable of comprehensively integrating multi-dimensional data, adapting to multiple body shapes, improving recommendation accuracy and personalized experience, so as to effectively reduce the size selection difficulty and the rate of return and exchange of users, and improve the online shopping satisfaction of large-size users.

[0060] The present application adopts a machine learning algorithm based on LightGBM (Light Gradient Boosting Machine), combines multi-dimensional body measurement data, body type analysis, fabric elasticity evaluation and other feature engineering technologies, and constructs a classified size prediction model. LightGBM is a high-efficiency machine learning framework based on gradient boosting decision tree (GBDT) developed by Microsoft, which specially optimizes the performance bottleneck of traditional GBDT algorithm in terms of memory usage and training speed. In the present size recommendation system, LightGBM is used as the core prediction algorithm to construct a strong classifier by integrating multiple weak learners (decision trees).

[0061] The first aspect of the present embodiment discloses a large-size women's clothing size recommendation method based on LightGBM as shown in the formula (I): Figure 1 The method comprises the following steps:

[0062] Step 1: Preprocessing the measurement data obtained after measuring the user's body shape, wherein the measurement data includes height, weight, bust, waist and hip; determining a first confidence degree based on the quality of the measurement data, and the first confidence degree is used to explain the reliability of the recommended size;

[0063] Step two: based on the pre-processed measurement data, generate a body feature vector for defining the user's body condition, determine a second confidence for defining the matching condition between the user's body and the dress size based on the relationship between the body feature vector and the dress size, and the second confidence is used to explain the reliability of the recommended size;

[0064] Step three: call the size prediction model based on the type of dress to be purchased, and the size prediction model is trained based on the body feature vector and through the LightGBM algorithm;

[0065] Step four: input the body feature vector into the size prediction model, output the prediction probability corresponding to each dress size, and take the maximum value in the output prediction probability as the third confidence, and the third confidence is used to explain the reliability of the recommended size;

[0066] Step five: determine the fourth confidence based on the historical purchase records, which include the total purchase of dresses and the purchase based on the recommended size, and the fourth confidence is used to explain the reliability of the recommended size;

[0067] Step six: determine the size corresponding to the maximum value in the output prediction probability as the recommended size, and statistically analyze the first confidence, the second confidence, the third confidence and the fourth confidence to determine the enhanced confidence for explaining the reliability of the recommended size, and output the recommended size and the enhanced confidence together.

[0068] Through the above, the effect of cooperative guarantee recommendation is realized based on multiple steps, wherein: the measurement data is pre-processed and the first confidence is determined, the data foundation is consolidated; the body feature vector and the second confidence are generated, the body type and the size are accurately adapted; the model trained by LightGBM is called, and the third confidence is output by the algorithm advantage; the fourth confidence is obtained by combining historical data, and the enhanced confidence is obtained by analyzing multiple confidences, which improves the accuracy and reliability of large-size dress size recommendation and optimizes user experience

[0069] Data preprocessing is a key basic link of size recommendation, which directly affects the training effect and prediction accuracy of the model. Considering the complexity and diversity of the body data of large-size dress users, the present application designs a complete data quality guarantee system.

[0070] Specifically, the preprocessing includes body measurement rationality verification, data cleaning, outlier detection and data standardization; wherein the body measurement rationality verification is to judge whether the measurement data is reasonable based on the pre-set body rationality parameters.

[0071] The body measurement data of large-size female clothing users is usually obtained from multiple channels: user self-measurement, professional measurement, historical purchase records, etc. The data from different sources differ significantly in accuracy, completeness, and consistency. User self-measurement is prone to improper measurement methods and measurement tool errors; professional measurement data are relatively accurate but costly to obtain; historical purchase records can provide a reference but have the problem of inconsistent size standards. Therefore, in the preprocessing, a multi-level verification architecture is adopted, including body measurement rationality verification, data cleaning, outlier detection, and data standardization, to ensure the quality and reliability of the measurement data. The body measurement rationality verification includes basic range verification, BMI rationality verification, body proportion verification, and consistency verification.

[0072] The basic range verification is the first line of defense for data quality control. It filters out obviously abnormal data by using the reasonable value range of the body measurement rationality parameters determined by existing medical and human engineering research. In this embodiment, the reasonable value range of the body measurement rationality parameters is specifically: the maximum and minimum values of height are 220 cm and 120 cm respectively, the maximum and minimum values of weight are 30 kg and 200 kg respectively, the maximum and minimum values of bust are 180 cm and 60 cm respectively, the maximum and minimum values of waist are 150 cm and 50 cm respectively, and the maximum and minimum values of hips are 180 cm and 60 cm respectively.

[0073] Body mass index (BMI) is an important indicator for evaluating the coordination of height and weight. The BMI rationality verification can identify abnormal data with severely unbalanced height and weight. In this embodiment, the body measurement rationality parameter set for BMI is the minimum reasonable value of BMI 13 (corresponding to severe malnutrition) and the maximum reasonable value of BMI 60 (corresponding to extreme obesity).

[0074] The body proportion verification identifies proportionally abnormal data by analyzing the relative relationships between body parts. It is the most complex and important verification link. In this embodiment, the result of body proportion verification is obtained based on the product of the waist-hip ratio, the chest-waist ratio, and the consistency verification result. The reasonable range of the waist-hip ratio is [0.6, 1.2], and the reasonable range of the chest-waist ratio is [0.8, 2.0]. By judging whether it is within the corresponding reasonable range, the waist-hip proportion verification result is assigned a value of 0 or 1, and the chest-waist proportion verification result is assigned a value of 0 or 1. For the consistency verification, the logical consistency within the data is detected by statistical methods to identify possible input errors or improper measurements, including height-weight consistency verification, girth consistency verification, and symmetry verification.

[0075] The height-weight consistency verification is based on a large sample statistical data to establish an expected relationship model of height and weight:

[0076] ExpectedWeight = RegressionCoeff x height ExponentTerm + ErrorTolerance

[0077] wherein, ExpectedWeight is the predicted weight value based on height; RegressionCoeff and ExponentTerm are parameters obtained by regression analysis; ErrorTolerance is the allowed deviation range.

[0078] Circumference consistency, test the rationality between bust, waist, and hips:

[0079] |bust-waist|≤ThresholdBustWaist, |hips-waist|≤ThresholdHipsWaist

[0080] wherein, ThresholdBustWaist is the maximum reasonable range of the difference between bust and waist; ThresholdBustWaist is the maximum reasonable range of the difference between hips and waist.

[0081] Symmetry test, for symmetrical body parts (such as left and right bust, left and right waist, etc.), test the measurement consistency.

[0082] It should be noted that the physiological significance of the proportion range in the embodiment is specifically:

[0083] Waist-hip ratio 0.6-0.75: pear-shaped body, hip circumference is significantly larger than waist circumference

[0084] Waist-hip ratio 0.75-0.85: hourglass-shaped body, ideal proportion

[0085] Waist-hip ratio 0.85-1.0: apple-shaped body, relatively large waist circumference

[0086] Waist-hip ratio 1.0-1.2: straight tube-shaped body, small difference between waist and hip

[0087] Bust-waist ratio 0.8-1.0: may exist measurement error

[0088] Bust-waist ratio 1.0-1.3: normal female body proportion

[0089] Bust-waist ratio 1.3-2.0: bust is significantly larger than waist, typical female characteristics

[0090] Based on the preprocessing process, the embodiment evaluates the quality of the test data.

[0091] Specifically, the determination process of the first confidence level includes:

[0092] determine the integrity score based on a ratio of the number of non-empty fields to the total number of fields;

[0093] determine the accuracy score based on a ratio of the number of measurement items passing the body measurement plausibility check to the total number of measurement items;

[0094] determine the consistency score based on a difference obtained by subtracting a ratio of the number of logical conflicts existing in the preprocessed measurement data to the total number of logical relationships from 1;

[0095] determine the data quality score based on the integrity score, the accuracy score and the consistency score;

[0096] determine the first confidence based on the number of non-empty measurement fields, the total number of measurement fields and the data quality score.

[0097] In this embodiment, the data quality score is obtained by weighted sum of the integrity score, the accuracy score and the consistency score, and the sum of the corresponding weights of the three is 1; the total number of measurement fields is usually 5, corresponding to height, weight, bust, waist and hip, and the product of the ratio of the number of non-empty measurement fields to the total number of measurement fields and the data quality score is defined as the first confidence.

[0098] It should be noted that the data cleaning, outlier detection and data standardization in this embodiment can be any of the existing ones, for example, detecting numerical outliers using an improved box plot method, detecting multi-dimensional outliers using Mahalanobis distance, and following the principles of conservatism, traceability and recoverability in data cleaning, using Z-Score standardization, Min-Max standardization and robust standardization, etc.

[0099] Through the above, the following technical effects are achieved: improving the quality of model training data and reducing the influence of noise; enhancing the stability and reliability of model prediction; reducing the recommendation error rate caused by data quality problems and providing high-quality basic data for subsequent feature engineering; and laying a solid data foundation for high-precision prediction of the size recommendation system.

[0100] Converting the test data into data for use by the size prediction model is a key step in realizing large-size women's clothing size recommendation based on LightGBM.

[0101] Specifically, the generation process of the body feature vector includes:

[0102] determining the relative proportion feature based on the proportion of bust relative to height, the proportion of waist relative to height and the proportion of hip relative to height;

[0103] determining the body difference feature based on the difference between bust and waist and the difference between hip and waist;

[0104] determining the body type feature based on the bust, the waist, the hip and a preset classification threshold.

[0105] In this embodiment, the relative proportion feature is used to eliminate the height influence, including:

[0106]

[0107] The body shape difference feature is used to capture the body curve, including:

[0108] D hips-waist = hips cm -waist cm

[0109] D bust-waist = bust cm -waist cm

[0110] wherein R bust-height is the proportion of bust to height, dimensionless, typical range [0.45, 0.85]; R waist-height is the proportion of waist to height, dimensionless, typical range [0.35, 0.75]; R hips-height is the proportion of hips to height, dimensionless, typical range [0.45, 0.85]; bust cm is the measurement of bust, unit: centimeter; waist cm is the measurement of waist, unit: centimeter; hips cm is the measurement of hips, unit: centimeter; height cm is the measurement of height, unit: centimeter; D hips-waist is the difference between hips and waist, unit: centimeter, used to represent the lower body curve; D bust-waist is the difference between bust and waist, unit: centimeter, used to represent the upper body curve.

[0111] By the above, the relative proportion feature can eliminate the influence of height on absolute size, better reflecting the relative features of body shape; the body shape difference feature can capture the concave-convex degree of body curve, which is an important basis for body type classification; these features provide standardized input for subsequent machine learning models, improving the generalization ability of the model.

[0112] For body type features, specifically an 8-class body type algorithm based on multi-dimensional body proportion:

[0113] BodyType = f(R waist-hips , R bust-waist , R hips-bust )

[0114] wherein,

[0115] The mathematical representation of the classification algorithm and the priority description are as follows:

[0116] The priority order in classification is: Prominent Pear, Slight Pear, Prominent Apple, Slight Apple, ldeal Hourglass, Hourglass, Straight / Rectangle, and Athletic. When multiple conditions are met simultaneously, the classification with higher priority is selected.

[0117]

[0118] where BodyType is the body type code, taking values {0, 1, 2, 3, 4, 5, 6, 7}; |bust-hips| is the absolute difference between bust and hips; max(bust, hips) is the maximum of bust and hips; and the coefficients 1.10, 1.05, 1.35, 1.25, 0.03, 0.08, 0.75, 0.80, and 0.85 are classification threshold values determined based on ergonomics research.

[0119] In this embodiment, the body type code correspondence is as follows:

[0120] 0: Prominent Pear

[0121] 1: Slight Pear

[0122] 2: Prominent Apple

[0123] 3: Slight Apple

[0124] 4: ldeal Hourglass

[0125] 5: Hourglass

[0126] 6: Straight / Rectangle

[0127] 7: Athletic

[0128] Each body type has its unique clothing matching characteristics, which directly affect the sizing recommendation strategy:

[0129] 1. Pear-shaped body: The lower body is relatively full, and the fit of the hips and thigh circumference needs to be considered.

[0130] 2. Apple-shaped body: The upper body is relatively full, and the fit of the bust and shoulder width needs to be considered.

[0131] 3. Hourglass-shaped body: The upper and lower bodies are coordinated, and the waist is obvious, which is the closest to the ideal proportion.

[0132] 4. Straight body shape: relatively flat, with little difference between parts

[0133] 5. Athletic body shape: well-developed muscles, tight body, need to consider the stretchability of the fabric

[0134] Through the above, the body feature vector generated by the relative proportion feature, the body difference feature and the body type feature provides accurate body data for size recommendation.

[0135] To improve the explainability of the recommended size, the application analyzes the matching of the body and the size based on the relationship between the body and the size.

[0136] Specifically, the determination process of the second confidence includes:

[0137] A1: determining a single-dimensional matching score based on one data in the measurement data and the size range of the corresponding part on women's clothing;

[0138] In this embodiment, the single-dimensional matching score adopts the following matching score function for the user measurement value and the product size:

[0139]

[0140] Wherein, u is the measurement value of the user at a certain body part, in centimeters; [min, max] is the size range of the product, corresponding to the minimum suitable value (in centimeters) of the part in the product size table, and the maximum suitable value (in centimeters) of the part in the product size table; Score(u, min, max) is the single-dimensional matching score, taking value [0, 1]; is the center value of the size range; width = max-min is the width of the size range, indicating the deviation amount of the size range; |u-center| is the absolute deviation of the user measurement value from the center value; deviation = max(0, u-max, min-u) is the deviation amount of the user measurement value from the size range; The coefficients 0.3 and 0.7 are the score attenuation parameters determined based on experience; The score logic is: when the user measurement value is within the size range, the score is between 0.7-1.0, the closer to the center value, the higher the score; When the user measurement value is out of the size range, the score is between 0-0.7, the greater the deviation, the lower the score.

[0141] A2: weighting and summing the single-dimensional matching scores in A1 based on the weights corresponding to the parts to obtain a comprehensive matching score, and determining the comprehensive matching score as the second confidence;

[0142] In this embodiment, the comprehensive matching score includes:

[0143] OverallScore = weight bust x Scorebust + weight waist x Score waist + weight hips x Score hips

[0144] wherein, OverallScore is the comprehensive matching score, taking value [0, 1]; weight bust , weight waist and weight hips are weight coefficients of bust, waist and hips respectively; Score bust is the single-dimension matching score of bust; Score waist is the single-dimension matching score of waist; Score hips is the single-dimension matching score of hips; weight bust + weight waist + weight hips = 1, and weight i ≥ 0 for i∈{bust, waist, hips}; the default weight setting is: weight bust = 0.8, weight waist = 0.15 and weight thips = 0.05.

[0145] A3: When there is a body type feature, optimizing A2: optimizing the weight corresponding to the part based on the body type feature, weighting and summing the single-dimension matching scores in A1 based on the optimized weight, obtaining an optimized comprehensive matching score, and determining the optimized comprehensive matching score as the second confidence; when there is no body type feature, A3 is not executed;

[0146] In this embodiment, the weight of each part is adjusted according to different body types:

[0147]

[0148] wherein, OptimizedScore is the comprehensive score optimized according to the body type; weight bodytype,part is the weight of a specific body part under a specific body type; part∈{bust, waist, hips} represents three parts of bust, waist and hips; bodytpe∈{0, 1, 2, 3, 4, 5, 6, 7} represents 8 body types; Score part is the single-dimension matching score of the corresponding body part.

[0149] A4: When there is a stretch score, the comprehensive matching score in A2 or the optimized comprehensive matching score in A3 is compensated for elasticity to obtain a compensated comprehensive matching score, and the compensated comprehensive matching score is determined as the second confidence; when there is no stretch score, A4 is not performed, and the process of determining the stretch score is as follows: a basic stretch score and a waist elasticity classification score are determined based on the stretch parameters of the dress, a pattern adaptability score is determined based on the pattern of the dress, and the stretch score is determined based on the basic stretch score, the waist elasticity classification score and the pattern adaptability score.

[0150] The embodiment adopts a stretch score function (0-10 point system):

[0151] StretchScore=min(10,S basic +S waist +S fit )

[0152] Wherein, StretchScore is the stretch score, the value range is [0, 10], the higher the score, the better the elasticity of the dress; S basic is the basic stretch score, the value is {0, 3}; S waist is the waist elasticity classification score, the value is {0, 1, 2, 3, 4}; S fit is the pattern adaptability score, the value is {0, 1, 2, 3}; min(10, ·) ensures that the total score does not exceed 10 points.

[0153] Basic stretch score:

[0154]

[0155] Wherein, is an indicator function, which is 1 when the garment fabric has elasticity, and 0 otherwise, and the coefficient 3 represents the weight value of the basic elasticity.

[0156] Waist elasticity classification score:

[0157]

[0158] Wherein, S waist is the waist elasticity classification score, the value is {0, 1, 2, 3, 4}; level 0: completely inelastic, such as pure cotton, linen, etc.; level 1: slightly elastic (1-3% elastic fiber), such as fabric containing a small amount of spandex; level 2: regular elasticity (3-8% elastic fiber), such as common stretch cotton; level 3: high elasticity (8-15% elastic fiber), such as sports fabric; level 4: super high elasticity (> 15% elastic fiber), such as yoga suit fabric.

[0159] Pattern adaptability score:

[0160]

[0161] wherein FitType is the garment fit type, which affects the score of body adaptability; tight fit has the most stringent requirement on body shape, and the lowest score; loose fit has the strongest inclusiveness on body shape, and the highest score.

[0162] In this embodiment, the body type weight matrix is as shown in Table 1, and the principle is as follows: Figure 2

[0163] Pear shape (0, 1): the highest weight for hip circumference, because the lower body is relatively plump

[0164] Apple shape (2, 3): the highest weight for bust circumference, because the upper body is relatively plump

[0165] Hourglass shape (4, 5): higher weights for bust and waist circumferences, emphasizing the beauty of curves

[0166] Straight and athletic shape (6, 7): balanced distribution of weights for three circumferences

[0167] If the body type is out of the range of 0-7, the default weights are used, i.e., 0.8 for bust circumference, 0.15 for waist circumference, and 0.05 for hip circumference. Further, this embodiment considers the compensation effect of fabric elasticity on size matching, including:

[0168] ElasticCompensatedScore = min(1.0, OverallScore + Bonus elastic )

[0169] wherein ElasticCompensatedScore is the final score after elastic compensation, with a value range of [0, 1]; OverallScore is the aforementioned comprehensive matching score; Bonus elastic is the elastic compensation bonus, with a value of {0.0, 0.1}, including:

[0170]

[0171] WaistElasticLevel is the waist elasticity classification score, with a value of {0, 1, 2, 3, 4}, and when the waist elasticity reaches super-high elasticity (level 4), a reward of 0.1 is given; min(1.0, ·) ensures that the final score does not exceed 1.0.

[0172] This embodiment is based on the existing LightGBM algorithm, and is trained for women's clothing categories respectively. Specifically, the training process of the size prediction model includes:

[0173] ​The plurality of independent LightGBM models are trained based on the female dress categories, wherein the female dress categories include All_Dresses (dresses category), Casual_Tops (casual tops category), Formal_Tops (formal tops category), Outerwear (outerwear category), Fitted_Bottoms (fitted bottoms category), RelaxedBottoms (relaxed bottoms category), Jumpsuits_Rompers (jumpsuits and rompers category), and Swimwear (swimwear category);

[0174] The independent LightGBM model is trained based on the body feature vector and the historical recommended size, so that the independent LightGBM model outputs the prediction probability of the corresponding female dress size based on the body feature vector;

[0175] The trained independent LightGBM model is used as a size prediction model.

[0176] According to the above, the independent LightGBM model is trained according to the female dress categories (8 categories such as dresses and tops), which is suitable for different styles and adapts to different needs. The model is trained based on historical data to accurately output the size prediction probability of each category, improve the category targeting and prediction accuracy of the large-size female dress size recommendation, and optimize the recommendation effect.

[0177] In this embodiment, the original prediction value of each category is converted into a probability distribution:

[0178]

[0179] Further, to avoid overflow of the exponential function, a numerically stable softmax calculation is used:

[0180]

[0181] where P(y=c|x) is the prediction probability of sample x belonging to category c, satisfying F c (x) is the original prediction value (logit value) of sample x in category c; F k (x) is the original prediction value of sample x in category k; C is the total number of categories, which is the number of size categories in this embodiment; c∈{1,2,...,C} is the target category index; k∈{1,2,...,C} is the index of all categories; max k F k (x) is the maximum value of the original prediction value in all categories for numerical stabilization; is an exponential function that converts the logit value to a positive value; x is the feature vector of the input sample.

[0182] Based on the calculation of the above probability distribution, the final predicted category is:

[0183]

[0184] wherein, is the predicted size category.

[0185] In actual women's clothing sales, the categories of size data are unbalanced, which mainly reflects that common sizes such as M and L are relatively more, and large size samples are relatively less. To solve this phenomenon, the embodiment adopts category weight balancing.

[0186] For the size samples of large size women's clothing, the embodiment optimizes the LightGBM algorithm, and specifically, the training process of the size prediction model further includes modifying the weight in the loss function in the LightGBM algorithm based on the number of samples of any size category.

[0187] In the embodiment, the modification of the weight w c is:

[0188]

[0189] wherein, n is the total number of training samples; n c is the number of samples of category c.

[0190] The modified loss function is:

[0191]

[0192] Based on the above, based on the LightGBM algorithm, the computational complexity is significantly reduced, the growth strategy reduces the memory occupation, the corresponding gradient sampling improves the speed while maintaining the model accuracy, and the robustness and scalability are also improved, thereby ensuring the excellent performance of the LightGBM algorithm in large size women's clothing prediction, providing algorithm guarantee for high accuracy and fast response.

[0193] To further improve the explainability of the recommended size, the confidence of the recommended size is adjusted based on historical data, and specifically, the determination process of the fourth confidence includes:

[0194] Obtaining the total purchase quantity of any women's clothing type, and counting the purchase quantity of the user based on the recommended size to purchase the women's clothing type;

[0195] Determining the fourth confidence based on the ratio of the purchase quantity to the total purchase quantity.

[0196] It should be noted that when there is a lack of historical data, the embodiment uses existing similarity techniques such as Euclidean distance similarity or cosine similarity for cold start.

[0197] Based on the first confidence, the second confidence, the third confidence and the fourth confidence, the embodiment optimizes the enhanced confidence. Specifically, the determination process of the enhanced confidence is to perform weighted summation based on the first confidence, the second confidence, the third confidence and the fourth confidence.

[0198] In the embodiment, the weights corresponding to the first confidence, the second confidence, the third confidence and the fourth confidence in the weighted summation of the determination process of the enhanced confidence are a, β, γ and δ, a, β, γ, δ > 0, and satisfy a + β + γ + δ = 1, and in a specific example, the corresponding value conditions are a = 0.6, β = 0.25, γ = 0.1 and δ = 0.05.

[0199] Based on the above, the embodiment achieves the following technical effects:

[0200] 1. Multi-dimensional feature fusion: combining 25+ dimensional features such as body measurement, body type, fabric characteristics, etc., to improve prediction accuracy;

[0201] 2. Classification optimization: training specialized models for different clothing categories to adapt to different pattern requirements;

[0202] 3. Intelligent matching algorithm: size matching score based on mathematical model, quantifying recommendation confidence;

[0203] 4. Elastic compensation mechanism: considering the impact of fabric elasticity on wearing effect, optimizing recommendation results;

[0204] 5. Cold start processing: solving new product recommendation problems through similar category mapping.

[0205] It should be noted that other gradient boosting algorithms (such as XGBoost, CatBoost) or deep learning methods (such as neural networks) can be used in the embodiment, but the LightGBM algorithm has the advantages of fast training speed, low memory occupation and high accuracy when processing table data.

[0206] The verification results based on real e-commerce data show that:

[0207] Top-1 accuracy reaches 75-85%; Top-2 accuracy reaches 85-95%; model confidence reaches an average of 0.7+, high-quality matching can reach 0.9+; feature importance body matching features dominate (60%+ weight).

[0208] The second aspect of the embodiment discloses a large-size women's clothing size recommendation system based on LightGBM as shown in Figure 3 The system includes:

[0209] a data preprocessing module configured to preprocess measurement data, the measurement data comprising height, weight, bust, waist and hip of the user, and determine a first confidence degree defining quality of the measurement data based on a relationship between a process and data quality;

[0210] a body feature analysis module configured to generate a body feature vector defining a body condition of the user based on the preprocessed measurement data, and determine a second confidence degree defining a matching condition between the body of the user and a female garment size based on a relationship between the body feature vector and the female garment size;

[0211] a model calling module configured to call a size prediction model based on a type of the female garment to be purchased, the size prediction model being trained based on the body feature vector and historical recommended sizes in combination with a LightGBM algorithm;

[0212] a model running module configured to input the body feature vector of step two into the size prediction model of step three, output a predicted probability corresponding to each female garment size, and output a maximum value of the predicted probability as a third confidence degree;

[0213] a historical recommendation analysis module configured to determine a fourth confidence degree defining a historical condition of size recommendation based on a relationship between historical recommended size purchase data and actual needs of the user;

[0214] a size recommendation module configured to recommend a size based on the size corresponding to the third confidence degree, and statistically analyze the first confidence degree, the second confidence degree, the third confidence degree and the fourth confidence degree to determine an enhanced confidence degree, the enhanced confidence degree being used to evaluate a reliability degree of recommending the size based on the size corresponding to the first confidence degree.

[0215] It should be noted that the LightGBM-based large-size female garment size recommendation system of the present embodiment corresponds to the aforementioned LightGBM-based large-size female garment size recommendation method. Therefore, the contents not specifically described in the LightGBM-based large-size female garment size recommendation system of the present embodiment, such as function definition, working principle and technical effect, can be referred to the records in the aforementioned LightGBM-based large-size female garment size recommendation method, which will not be repeated herein.

[0216] In summary, the large-size women's clothing size recommendation method and system based on LightGBM in the embodiment can guarantee the recommendation accuracy and reliability through multi-dimensional data processing and confidence system; the measured data is preprocessed and the first confidence is determined to ensure the quality of input data; the body feature vector is generated and combined with the matching logic to determine the second confidence to accurately depict the user's body shape and size adaptation; the classification model trained based on LightGBM is called to output the prediction probability and the third confidence, and the algorithm advantage is used to improve the prediction effect; the fourth confidence is determined based on historical data, and feedback optimization is introduced; the enhanced confidence is obtained through multi-confidence weighting to comprehensively evaluate the recommendation reliability; the problems such as single dimension and static matching of the prior art are solved, the complex body shape of large-size users is adapted, the trial and error cost is reduced, the shopping experience is optimized, and a more intelligent, accurate and dynamic solution is provided for large-size women's clothing size recommendation.

[0217] In the embodiments provided in the present application, it should be understood that the embodiments described herein can be realized in hardware, software, firmware, middleware, code, or any appropriate combination thereof. For hardware implementation, the processor can be realized in one or more of the following units: application specific integrated circuit (ASIC), digital signal processor (DSP), digital signal processing device (DSPD), programmable logic device (PLD), field programmable gate array (FPGA), processor, controller, microcontroller, microprocessor, other electronic units designed to implement the functions described herein, or a combination thereof. For software implementation, part or all of the processes of the embodiments can be instructed by a computer program to relevant hardware. In implementation, the above program can be stored in a computer readable storage medium or transmitted as one or more instructions or codes on a computer readable storage medium. The computer readable storage medium includes computer storage medium and communication medium, wherein the communication medium includes any medium that facilitates the transmission of a computer program from one place to another. The storage medium can be any available medium that can be accessed by a computer. The computer readable storage medium can include, but is not limited to, RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program codes in the form of instructions or data structures and can be accessed by a computer.

[0218] Finally, it should be noted that: the above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features, as long as they are within the spirit and principles of the present application. Any modification, equivalent replacement, improvement, etc. made within the scope of the present application shall be included in the protection scope of the present application.

Claims

1. A method for recommending plus-size women's clothing sizes based on LightGBM, characterized in that, The method includes: Step 1: Preprocess the measurement data obtained after measuring the user's body shape, including height, weight, chest circumference, waist circumference, and hip circumference; determine a first confidence level based on the quality of the measurement data, which is used to interpret the reliability of the recommended size; Step 2: Based on the preprocessed measurement data, generate a body feature vector to define the user's body shape. Based on the relationship between the body feature vector and women's clothing size, determine a second confidence level to define the matching between the user's body shape and women's clothing size. The second confidence level is used to explain the reliability of the recommended size. Step 3: Based on the type of women's clothing to be purchased, call the size prediction model, which is trained based on the body feature vector and the LightGBM algorithm; Step 4: Input the body feature vector into the size prediction model, output the prediction probability corresponding to each women's clothing size, and take the maximum value of the output prediction probability as the third confidence level. The third confidence level is used to explain the reliability of the recommended size. Step 5: Determine the fourth confidence level based on historical purchase records, which include total purchases of women's clothing and purchases based on historical recommended sizes. The fourth confidence level is used to explain the reliability of the recommended sizes. Step Six: Determine the size corresponding to the maximum value in the output predicted probabilities as the recommended size. Statistically analyze the first confidence level, the second confidence level, the third confidence level, and the fourth confidence level to determine the enhanced confidence level used to explain the reliability of the recommended size. Output the recommended size and the enhanced confidence level together.

2. The plus-size women's clothing size recommendation method based on LightGBM according to claim 1, characterized in that, The preprocessing includes body measurement rationality verification, data cleaning, outlier detection, and data standardization; wherein, the body measurement rationality verification is to determine whether the measurement data is reasonable based on preset body rationality parameters.

3. The plus-size women's clothing size recommendation method based on LightGBM according to claim 2, characterized in that, The process of determining the first confidence level includes: The integrity score is determined based on the ratio of the number of non-empty fields to the total number of fields; An accuracy score is determined based on the ratio of the number of measurements that pass the rationality check of the body measurements to the total number of measurements. The consistency score is determined by the difference between the number of logical conflicts in the preprocessed measurement data and the total number of logical relations. A data quality score is determined based on the integrity score, the accuracy score, and the consistency score; The first confidence level is determined based on the number of non-empty measurement fields, the total number of measurement fields, and the data quality score.

4. The plus-size women's clothing size recommendation method based on LightGBM according to claim 1, characterized in that, The process of generating the body feature vector includes: The relative proportion characteristics are determined based on the ratio of chest circumference to height, the ratio of waist circumference to height, and the ratio of hip circumference to height. Body shape differences are determined based on the difference between bust and waist circumference and the difference between hip and waist circumference. Body type characteristics are determined based on bust, waist, hip circumference and preset classification thresholds.

5. The plus-size women's clothing size recommendation method based on LightGBM according to claim 4, characterized in that, The process of determining the second confidence level includes: A1: A single-dimensional matching score is determined based on one data point from the measurement data and the size range of the corresponding part in women's clothing; A2: The single-dimensional matching scores in A1 are weighted and summed based on the weights corresponding to the parts to obtain a comprehensive matching score, and the comprehensive matching score is determined as the second confidence level; A3: When the body type feature exists, optimize A2: optimize the weights corresponding to the body parts based on the body type feature, and perform a weighted summation of the single-dimensional matching scores in A1 based on the optimized weights to obtain an optimized comprehensive matching score, which is then determined as the second confidence level; when the body type feature does not exist, A3 is not executed. A4: When an elastic score exists, perform elastic compensation on the comprehensive matching score in A2 or the optimized comprehensive matching score in A3 to obtain a compensated comprehensive matching score, and determine the compensated comprehensive matching score as the second confidence level; when the elastic score does not exist, A4 is not executed. The process of determining the elastic score is as follows: determine the basic elastic score and waist elastic grading score based on the elastic parameters of women's clothing, determine the pattern adaptability score based on the pattern of women's clothing, and determine the elastic score based on the basic elastic score, the waist elastic grading score, and the pattern adaptability score.

6. The plus-size women's clothing size recommendation method based on LightGBM according to claim 1, characterized in that, The training process of the size prediction model includes: Multiple independent LightGBM models were obtained by training based on the women's clothing category; The independent LightGBM model is trained based on the body feature vector and its historical recommended sizes, so that the independent LightGBM model outputs the predicted probability of the corresponding women's clothing size based on the body feature vector. The trained, independent LightGBM model is used as the size prediction model.

7. The plus-size women's clothing size recommendation method based on LightGBM according to claim 6, characterized in that, The training process of the size prediction model also includes adjusting the weights in the loss function of the LightGBM algorithm based on the number of samples for any size category.

8. The plus-size women's clothing size recommendation method based on LightGBM according to claim 1, characterized in that, The process of determining the fourth confidence level includes: Get the total purchase quantity of any women's clothing type, and count the purchase quantity of that women's clothing type by users based on the recommended size; The fourth confidence level is determined based on the ratio of the purchase quantity to the total purchase quantity.

9. The plus-size women's clothing size recommendation method based on LightGBM according to claim 1, characterized in that, The process of determining the enhanced confidence level is to perform a weighted summation based on the first confidence level, the second confidence level, the third confidence level, and the fourth confidence level.

10. A plus-size women's clothing size recommendation system based on LightGBM, employing the plus-size women's clothing size recommendation method based on LightGBM as described in any one of claims 1-9, characterized in that, The system includes: A data preprocessing module is used to preprocess measurement data, including the user's height, weight, chest circumference, waist circumference, and hip circumference; and to determine a first confidence level defining the quality of the measurement data based on the relationship between the preprocessing process and data quality. The body feature analysis module is used to generate a body feature vector to define the user's body shape based on the preprocessed measurement data, and to determine a second confidence level to define the matching between the user's body shape and the women's clothing size based on the relationship between the body feature vector and the women's clothing size. The model invocation module is used to invoke the size prediction model based on the type of women's clothing to be purchased. The size prediction model is trained based on body feature vectors and their historical recommended sizes, combined with the LightGBM algorithm. The model running module is used to input the body feature vector from step two into the size prediction model in step three, output the prediction probability corresponding to each women's clothing size, and use the maximum value of the prediction probability as the third confidence level and output it. The historical recommendation analysis module is used to determine the fourth confidence level for defining the historical situation of size recommendations based on the relationship between historical recommended size purchase data and users' actual needs. The size recommendation module recommends sizes based on the sizes corresponding to the third confidence level, and performs statistical analysis based on the first confidence level, the second confidence level, the third confidence level, and the fourth confidence level to determine the enhanced confidence level. The enhanced confidence level is used to evaluate the reliability of size recommendations based on the sizes corresponding to the first confidence level.

Images