A customized soybean milk formula recommendation method and system using user health data

Abstract

The application relates to the technical field of intelligent health diet, and discloses a customized soybean milk formula recommendation method and system using user health data. The method comprises the following steps: obtaining multi-dimensional original data from a preset formula library; mapping the original data into health-related scores through a nutrition model; performing standardization processing on key attributes to generate a multi-dimensional formula attribute vector; obtaining user health and preference data to construct a user demand vector; determining and normalizing the weight coefficients of each dimension based on the user data; calculating the weighted distance value of the formula attribute vector and the demand vector; and determining the recommended formula by screening and sorting according to the distance value. Through the multi-dimensional weighted matching method combining the user health status and preferences, the application solves the problem that the existing recommendation is one-sided and cannot balance the health goal, taste and production difficulty, and improves the accuracy and practicality of the soybean milk formula recommendation.

Description

Technical Field

[0001] This application relates to the field of intelligent healthy diet technology, and in particular to a method and system for recommending customized soy milk recipes using user health data. Background Technology

[0002] With the popularization of healthy eating concepts and the upgrading of personalized nutritional needs, dietary formula recommendations based on user health data have become a research hotspot in the field of intelligent health management. As a daily nutritious beverage, the health suitability, taste compatibility, and ease of preparation of soy milk formulas have attracted much attention. There is an urgent need for a customized recommendation technology that can accurately balance health goals and personal preferences to meet the personalized dietary needs of different users.

[0003] Current recipe recommendation methods have several shortcomings: First, their limitations are obvious due to their reliance on a single dimension. Some methods focus solely on nutritional dimensions, such as selecting recipes based only on calorie, protein, or dietary fiber content, neglecting users' preferences for sweetness and consistency, as well as the suitability of preparation complexity to their own time and energy. This results in recommended recipes that, while meeting nutritional standards, are difficult for users to accept long-term due to their bland taste and cumbersome preparation. Other methods filter recipes based solely on users' indicated taste preferences, without considering specific health goals such as blood sugar control and weight management. This means that while the recommended recipes may taste appealing, they fail to provide effective health support and cannot meet long-term health management needs. Second, they fail to properly handle the trade-offs between the multiple attributes of a recipe. The health attributes, taste attributes, and preparation complexity of a recipe often have mutual constraints. For example, high-fiber recipes often have a rough texture, low-sugar recipes tend to be bland, and high-protein recipes may increase preparation steps. Existing methods do not have adaptation mechanisms designed for these trade-offs and cannot find a reasonable balance among multiple attributes. Third, the system lacks dynamic adaptation to users' personalized health status. It does not adjust the recommendation logic based on specific health data such as users' blood sugar control target levels, nor does it optimize the weight of production complexity based on differences in user scenarios. As a result, the recommendation results cannot accurately match the needs of users with different health conditions and different usage scenarios. Users need to repeatedly try and fail among a large number of recipes to find a suitable recipe that meets both health requirements and taste and ease of preparation. Overall, the recommendation efficiency and accuracy are low.

[0004] To address the above shortcomings, this application combines user health status data with preference data, and utilizes nutritional models, vector standardization processing, and weighted distance calculation to achieve accurate matching across multiple dimensions, including health, taste, and production complexity. This solves the problems of one-sided recommendations and poor adaptability in existing methods, and improves the accuracy and practicality of soy milk recipe recommendations. Summary of the Invention

[0005] This application provides a method and system for recommending customized soy milk recipes using user health data. It achieves accurate matching of health, taste, and production complexity from multiple dimensions, solving the problems of one-sided recommendations and poor adaptability of existing methods, and improving the accuracy and practicality of soy milk recipe recommendations.

[0006] In a first aspect, this application provides a method for recommending customized soy milk recipes using user health data, the method comprising: S1. Obtain raw data of multiple soy milk recipes from a preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative values ​​of glycemic index, the quantitative values ​​of dietary fiber content, the quantitative values ​​of protein content, taste score, and production complexity score. S2. The raw data is mapped to health-related scores for each soy milk formula using a preset nutritional model. The health-related scores represent the contribution of the corresponding soy milk formula to a specific health goal. S3. Standardize the taste score, production complexity score and health-related score in the multi-dimensional attributes of each soy milk recipe to generate a multi-dimensional recipe attribute vector for each soy milk recipe. S4. Obtain user health status data and user preference data, and construct a user demand vector based on the user health status data and user preference data; S5. Determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on the user health status data and the user preference data, and normalize the weight coefficients to generate standardized weight coefficients. S6. Based on the standardized weight coefficients, calculate the distance between the multi-dimensional recipe attribute vector and the user demand vector, whereby the distance value represents the degree of matching between the soy milk recipe and the user demand. S7. After screening and sorting the soy milk recipes according to the distance value, determine the recommended soy milk recipe that matches the user's needs.

[0007] Secondly, this application provides a customized soy milk recipe recommendation system utilizing user health data, used to implement the aforementioned customized soy milk recipe recommendation method utilizing user health data, the system comprising: The data acquisition module is used to acquire raw data of multiple soy milk recipes from a preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative values ​​of glycemic index, the quantitative values ​​of dietary fiber content, the quantitative values ​​of protein content, taste score, and production complexity score. The scoring mapping module is used to map the raw data into health-related scores for each soy milk formula through a preset nutritional model. The health-related scores represent the contribution of the corresponding soy milk formula to a specific health goal. The vector generation module is used to standardize the taste score, production complexity score and health-related score in the multi-dimensional attributes of each soy milk recipe, and generate a multi-dimensional recipe attribute vector corresponding to each soy milk recipe. The requirement construction module is used to acquire user health status data and user preference data, and construct a user requirement vector based on the user health status data and user preference data. The weight determination module is used to determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on the user health status data and the user preference data, and to normalize the weight coefficients to generate standardized weight coefficients. The distance calculation module is used to calculate the distance value between the multi-dimensional recipe attribute vector and the user demand vector based on the standardized weight coefficients. The distance value represents the degree of matching between the soy milk recipe and the user demand. The recipe recommendation module is used to filter and sort the various soy milk recipes based on the distance value, and determine the recommended soy milk recipe that matches the user's needs.

[0008] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the aforementioned method for recommending customized soy milk recipes using user health data.

[0009] This application proposes a method and system for customized soy milk recipe recommendations based on user health data. It achieves precise matching across multiple dimensions, including health, taste, and preparation complexity, solving the problems of one-sided recommendations and poor adaptability in existing methods, and improving the accuracy and practicality of soy milk recipe recommendations. Compared with existing technologies, the beneficial effects of this application's technical solution are at least as follows: First, by mapping the raw data of soy milk recipes to health-related scores through a preset nutritional model, and combining the blood sugar control target level information in the user's health status data with the dynamic adaptation recommendation logic, the health attributes of the recipes are accurately matched with the user's specific health goals, providing the user with effective healthy diet support. Secondly, by taking into account users' preferences for sweetness and consistency, as well as the complexity of recipe preparation, a reasonable balance is found between health contribution, taste acceptance, and ease of preparation through the construction of multi-dimensional attribute vectors and weighted distance calculation, so as to avoid recommended recipes being abandoned by users due to poor taste or complicated operation. Third, by standardizing the multi-dimensional attributes and health-related scores of the formula, and dynamically adjusting the weight coefficients based on user data, the accuracy of formula recommendations can be improved, the cost of repeated trial and error for users can be reduced, and the acceptance and stickiness of users to the recommended formulas can be increased. Fourth, by combining difficulty threshold filtering with distance value sorting, recipes that match the user's time and energy are prioritized for recommendation, further improving the practicality of the recommended solutions and meeting the personalized dietary needs of users in different scenarios. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a method for recommending customized soy milk recipes using user health data, as described in this application. Figure 2 This is a two-dimensional matching diagram of health and taste in this application; Figure 3 This is a graph showing the comparison of average satisfaction levels in this application; Figure 4 This is a schematic diagram of the structure of a customized soy milk recipe recommendation system that utilizes user health data, as described in this application. Detailed Implementation

[0012] This application provides a method and system for recommending customized soy milk recipes using user health data. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 An embodiment of a method for recommending customized soy milk recipes using user health data, as described in this application, includes: S1. Obtain raw data of multiple soy milk recipes from the preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative value of glycemic index, the quantitative value of dietary fiber content, the quantitative value of protein content, taste score, and production complexity score.

[0014] In one specific embodiment, performing step S1 includes the following steps: Extract multi-dimensional attribute data of each soy milk recipe from the preset recipe library; Perform integrity checks on the multi-dimensional attribute data of each soy milk recipe, check for missing data in each field, and remove soy milk recipes with missing data. For the soy milk recipe after integrity verification, the multi-dimensional attribute data is organized and integrated into a structured raw dataset according to the preset field order and data type rules.

[0015] Specifically, the recipe library is a pre-built database storing soy milk recipe information. Each recipe entry in the library corresponds to a unique recipe identifier, and the recipe identifier forms a mapping relationship with multi-dimensional attribute data. The multi-dimensional attribute data under each recipe entry includes quantitative values ​​of nutrients, glycemic index, dietary fiber content, protein content, taste score, and production complexity score. Among them, the quantitative values ​​of nutrients correspond to the measured values ​​of detectable macronutrients and micronutrients per 100mL of finished soy milk, including the content values ​​of carbohydrates, fats, calcium, phosphorus, and potassium, in g / 100mL or mg / 100mL. The quantitative value of glycemic index corresponds to the GI value of the finished product determined by in vitro digestion experiments or human blood glucose experiments, with a value range of 0-100. The quantitative value of dietary fiber content corresponds to the total content of soluble and insoluble dietary fiber in the finished product, in g / 100mL. The quantitative value of protein content corresponds to the content of complete protein in the finished product, in g / 100mL. The taste score and production complexity score are derived from a comprehensive evaluation system, with values ​​ranging from 1 to 10. The taste score is determined by trained sensory evaluators in conjunction with feedback from ordinary users, covering aspects such as sweetness balance, consistency, smoothness, and residual beany flavor. The final score is formed by weighted aggregation. The production complexity score is evaluated based on the complete preparation process of the recipe, with evaluation dimensions including the number of raw material pretreatment steps, the total time spent on core preparation steps, the type and quantity of specialized equipment required, and the difficulty of raw material pretreatment. It is also quantitatively calculated through comprehensive weighting.

[0016] Using structured query statements, all recipe entries in the recipe library are traversed. Attribute data for each recipe entry across its six core dimensions is extracted row by row according to the recipe identifier's coding order. After extraction, a field non-empty validation algorithm is used to verify the integrity of the multi-dimensional attribute data for each soy milk recipe. The validation targets are the six core dimension fields corresponding to each recipe identifier: quantified values ​​of nutritional components, glycemic index, dietary fiber content, protein content, taste rating, and production complexity rating. The validation rules are defined using the following formula: ,in This indicates a product, where m ranges from 1 to 6, corresponding to the six core dimension fields. The validation result for the m-th dimension field. This indicates that the field has passed validation. This indicates that the field failed the validation. P represents the overall validation result for a single recipe; P=1 indicates that the recipe passed the integrity validation, and P=0 indicates that the recipe failed the integrity validation. The validation logic for a single dimension field is as follows: the field value exists, the value falls within the valid range, and the value type exactly matches the preset type. The value is assigned as 1, and otherwise as 0. The valid range for the quantitative values ​​of nutritional components, dietary fiber content, and protein content is a floating-point number ≥ 0. The valid range for the quantitative value of glycemic index is a floating-point number from 0 to 100. The valid range for taste score and preparation complexity score is an integer from 1 to 10. During the verification process, the verification calculation is performed on each of the six dimension fields corresponding to each recipe identifier to obtain the corresponding P value. The full data corresponding to the recipe with P=0 is removed from the extracted dataset, and only the recipe data with P=1 is retained.

[0017] For soy milk recipes that pass integrity verification, multi-dimensional attribute data is organized according to field order and data type rules. The field order is fixed as follows: unique recipe identifier, quantitative values ​​of nutritional components, quantitative values ​​of glycemic index, quantitative values ​​of dietary fiber content, quantitative values ​​of protein content, taste score, and production complexity score. The data type rules are as follows: unique recipe identifier is a string type, quantitative values ​​of nutritional components, glycemic index, dietary fiber content, and protein content are double-precision floating-point types, and taste score and production complexity score are integer types. During the organization process, for each recipe data that passes verification, the arrangement order of each dimension data is first adjusted according to the field order, and then the format of each dimension data is converted according to the data type rules. During the format conversion process, a mapping relationship is formed between the original data and the converted data of each dimension. The numerical value of the original data does not change during the conversion process; only the storage type and arrangement order of the data are adjusted. For example, the original data for the taste score of a certain recipe is a floating-point number 7.2. It is converted into an integer number 7 according to the rules. The converted value corresponds to the same taste dimension of the same recipe as the original value, without changing the core representation of the taste attribute of the recipe. All recipe data after order adjustment and type conversion are arranged according to the coding order of the recipe's unique identifier and integrated into a structured original dataset. The original dataset is stored in a two-dimensional table structure. Each row of data corresponds to a unique recipe identifier, and each column of data corresponds to a fixed dimension attribute. The column index and dimension attribute form a mapping relationship, and the row and recipe form a mapping relationship.

[0018] S2. The raw data is mapped to health-related scores for each soy milk recipe using a pre-set nutritional model. The health-related scores represent the contribution of the corresponding soy milk recipe to a specific health goal.

[0019] In one specific embodiment, performing step S2 includes the following steps: Input the raw dataset into a pre-built nutrition model; Based on a specific pre-set health goal, a nutritional model is used to assign corresponding weight coefficients to each quantitative value in the original dataset, and a linear weighted summation calculation is performed to output the original health-related score for each soy milk recipe. The original health-related scores were normalized to obtain the health-related scores for each soy milk formula.

[0020] Specifically, the pre-constructed nutritional model is a linear weighted model. The core of the model uses preset weight coefficients to linearly combine multi-dimensional quantitative values, achieving a quantitative assessment of the health contribution of soy milk formulas. The model training process focuses on a specific health goal centered on blood glucose control. 600 sets of soy milk formulas covering different bean ingredients (soybeans, black beans, red beans, etc.) and additives (oats, barley, red dates, etc.) are selected as training samples. Each sample includes complete quantitative values ​​for nutritional components, glycemic index, dietary fiber content, and protein content, as well as a health contribution reference value assessed by nutrition experts based on clinical blood glucose control effects. During training, the mean absolute error is used as the loss function, and stochastic gradient descent is employed to optimize the weight coefficients of each dimension. The learning rate is set to 0.015, the batch size is 20, and the number of iterations is 1200. In each iteration, the weight coefficients are adjusted by calculating the error between the predicted value and the reference value until the loss function converges to below 0.003. After training, the weight coefficients of each dimension are fixed to ensure a stable original health-related score when the same quantitative value is input.

[0021] The structured raw dataset is completely input into the nutrition model in the order of the formula identifiers. During the data input process, each sub-quantified value of the nutrients (carbohydrates, fats, calcium, phosphorus, potassium), glycemic index, dietary fiber content, and protein content corresponds one-to-one with the model's preset input items. The model first assigns sub-weight coefficients to each sub-quantified value of the nutrients. For example, the weight coefficient C1 for carbohydrate content is -0.28, C2 for fat content is -0.12, C3 for calcium content is 0.09, C4 for phosphorus content is 0.06, and C5 for potassium content is 0.08. The model then uses a formula... The model calculates a comprehensive weighted result A1 of the nutritional components, where B1 to B5 are the quantified values ​​of carbohydrates, fats, calcium, phosphorus, and potassium, respectively. The model performs a linear weighted summation on each input quantified value, and the calculation process is carried out using the formula... The formula is implemented as follows: H represents the original health-related score, A1 is the comprehensive weighted result of the quantified values ​​of nutrients, A2 is the quantified value of glycemic index, A3 is the quantified value of dietary fiber content, A4 is the quantified value of protein content, and W1 to W4 are the total weight coefficients corresponding to each dimension. Substituting the quantified values ​​of the four dimensions for each soy milk recipe into the formula yields a unique H value, which is the original health-related score of that recipe. The original health-related scores of different recipes are independent of each other and maintain a strict one-to-one correspondence with the recipe label, ensuring the traceability of the health contribution assessment results for each recipe.

[0022] After the original health-related scores of all soy milk recipes were calculated, they were normalized using a min-max normalization algorithm. After normalization, the range of the original health-related scores was uniformly mapped to the interval [0, 1], thus obtaining the health-related score for each soy milk recipe. During the processing, a unique numerical transformation relationship was formed between the original health-related score and the normalized health-related score for each recipe. The relative order of the scores remained unchanged, and only the numerical scale was unified.

[0023] S3. Standardize the taste score, production complexity score, and health-related score in the multi-dimensional attributes of each soy milk recipe to generate a multi-dimensional recipe attribute vector for each soy milk recipe.

[0024] In one specific embodiment, performing step S3 includes the following steps: Extract the taste score, production complexity score, and health-related score from the multi-dimensional attributes of each soy milk recipe; The extracted health-related scores, taste scores, and production complexity scores are converted using a pre-defined standardization algorithm to generate corresponding standardized attribute values. For each soy milk recipe, the corresponding standardized attribute values ​​are combined in a preset dimensional order to generate a multi-dimensional recipe attribute vector for each soy milk recipe.

[0025] Specifically, when the extraction operation is initiated, the system first calls the structured raw dataset and the health-related score results. A unique mapping relationship is established between the two through the formula identifier. Each formula identifier is precisely associated with its own taste score, production complexity score, and health-related score, ensuring clear attribution of the three types of scores without overlap, mismatch, or data omission. The taste score originates from a pre-set sensory evaluation system, comprehensively assessed by professional tasters combined with feedback from ordinary users, covering multiple dimensions such as sweetness suitability, consistency, smoothness, and residual beany flavor. A final single score result is formed after comprehensive weighting. The production complexity score is based on the complete preparation process of the formula, considering factors such as the number of raw material pretreatment steps, the time spent on core preparation steps, the type of specialized equipment required, and the difficulty of pretreatment such as raw material crushing or soaking. A quantitative score is also formed through comprehensive evaluation. The health-related score is the result mapped and normalized by a nutritional model, directly reflecting the formula's contribution to specific health goals such as blood sugar control. During the extraction process, the system iterates through all recipes that have passed the integrity check in the order of recipe identifiers, extracts the taste score and production complexity score of the corresponding recipe from the structured raw dataset, and then groups and stores the three types of scores according to the recipe identifiers based on the health-related scores corresponding to the same recipe identifier, forming a structured data group of "recipe identifier - three types of scores".

[0026] The standardization stage employs the Z-score algorithm, with the core objective of eliminating scale differences among the three types of scores caused by varying original evaluation systems. Since taste, production complexity, and health-related scores originate from different evaluation logics, their original numerical ranges differ significantly—for example, taste and production complexity scores might be integers from 1 to 10, while health-related scores are decimals from 0 to 1. Directly using these for subsequent matching calculations would lead to an imbalance in the influence of different attributes, failing to objectively reflect the recipe's overall performance in terms of health, taste, and production difficulty. By referencing the overall distribution of similar scores across all recommended soy milk recipes, each score for a single recipe is adjusted accordingly: taking taste scores as an example, the overall distribution characteristics of all recipe taste scores are first statistically analyzed, and then, based on the position of the individual recipe's taste score within this distribution, it is converted into a standardized attribute value within a uniform range; production complexity and health-related scores are processed using the same logic. The adjusted standardized attribute values ​​not only preserve the relative order of merit of each recipe in the same category of ratings, but also ensure that the values ​​of the three types of attributes fall within a unified range, achieving horizontal comparability between different attributes. This solves the technical problem in the background technology that multi-dimensional attributes cannot participate in matching calculations due to different scales, allowing the three types of attributes—health contribution, taste performance, and production difficulty—to be fairly considered under the same standard.

[0027] The vector generation stage treats each recipe as an independent processing unit. Following a pre-defined fixed dimensional order, standardized health-related scores, taste scores, and production complexity scores are sequentially combined to form a multi-dimensional recipe attribute vector. This pre-defined dimensional order is set during system initialization and maintained consistently, for example, fixed as "health-related score dimension - taste score dimension - production complexity score dimension." This ensures that the attribute vector structure of all recipes is completely consistent, with each position in the vector corresponding to a fixed attribute type, avoiding errors in subsequent matching calculations due to disordered dimensional order. Each recipe's attribute vector is strictly associated with its own recipe identifier, forming a traceable "recipe identifier - attribute vector" correspondence, facilitating accurate matching of subsequent calculation results with specific recipes.

[0028] S4. Obtain user health status data and user preference data, and construct a user demand vector based on the user health status data and user preference data.

[0029] In one specific embodiment, performing step S4 includes the following steps: Extract health status-related data from user history, including information on blood glucose control target levels; The user's taste preference data is obtained through a real-time input interface. The taste preference data includes information on the degree of preference for sweetness and thickness. Structure the health status-related data and taste preference data to generate a standardized user characteristic dataset; Based on the standardized user feature dataset, the target demand values ​​corresponding to each dimension of the multi-dimensional recipe attribute vector are determined. The preset standard demand vector is adjusted based on the target demand values ​​to construct the user demand vector.

[0030] Specifically, user history records are stored in a database associated with their personal account, encompassing previously completed health questionnaires, periodic statistical data uploaded from blood glucose monitoring devices, and health assessment conclusions annotated by nutritionists. When extracting health status-related data, the user's unique identifier is used as an index to locate the user's latest health assessment record. The structured fields within are parsed to extract the level information of blood glucose control targets. This level information is a pre-defined discrete classification, specifically divided into three categories: strict blood glucose control, general blood glucose control, and maintaining the status quo. Strict blood glucose control corresponds to users who need to strictly limit blood glucose fluctuations and have extremely high requirements for the glycemic index of the soy milk formula; general blood glucose control corresponds to users who need to moderately control blood glucose and allow the formula to have a moderate level of glycemic index; maintaining the status quo corresponds to users with stable blood glucose levels and no special restrictions on the glycemic index of the formula. During the extraction process, it is ensured that each user corresponds to only one set of blood glucose control target level information, with no duplicate or conflicting data. This step directly addresses the technical problem in the background technology that existing recommendation methods lack dynamic adaptation to users' personalized health status. By accurately extracting users' core health needs data, it lays the foundation for subsequent matching of health dimensions.

[0031] The real-time input interface uses a mobile application interface, offering two input methods: a slider and a level selection button. Users can choose their preferred operation mode and input their sweetness and thickness preferences by dragging the slider or clicking the level button. The system maps user input results to integer scores from 0 to 10, with scores positively correlated with preference levels. A score of 0 represents no preference (e.g., a sweetness score of 0 indicates a complete dislike of sweetness, and a thickness score of 0 indicates a preference for an extremely thin texture); a score of 10 represents a very high preference (e.g., a sweetness score of 10 indicates a strong desire for sweetness, and a thickness score of 10 indicates a preference for a thick, pasty texture). The interface receives and stores user input in real time, ensuring the timeliness of preference data. Each user's sweetness and thickness preferences correspond to independent scores, and the two do not interfere with each other. By directly obtaining users' personalized taste needs, subsequent recommendations can balance health and taste compatibility.

[0032] When structuring the extracted health status-related data and taste preference data, the process first involves encoding and scaling the blood glucose control target level information. Strict blood glucose control, moderate blood glucose control, and maintaining the status quo are encoded as 3, 2, and 1 respectively, and then adjusted to the 0-1 range using a unified rule. This transforms different levels of health needs into quantifiable, uniformly scaled data. Similarly, the original scores for sweetness and consistency preferences are adjusted to the 0-1 range using the corresponding rules to ensure consistency between the numerical ranges of health data and preference data. After processing, a standardized user feature dataset is generated. This dataset uses a unique user identifier as an index, and each record contains values ​​for three dimensions: normalized blood glucose control target value, normalized sweetness preference value, and normalized consistency preference value. Each standardized value in each dimension forms a unique mapping relationship with the original data, while preserving the relative magnitudes of the original data. This processing method solves the problem in the background technology where multi-dimensional data have inconsistent scales and cannot collaboratively participate in matching calculations, allowing health needs and taste preference data to be processed under the same standard.

[0033] The preset standard demand vector is a three-dimensional vector, with the dimension order consistent with the multi-dimensional recipe attribute vector: health-related dimension, taste dimension, and production complexity dimension. The initial baseline value for each dimension is set to 0.5, representing a medium level of demand target. When determining the demand target values ​​for each dimension based on the standardized user feature dataset, the demand target value for the health-related dimension is directly related to the normalized value of the blood sugar control target. The higher the normalized value (i.e., the stricter the blood sugar control requirement), the higher the demand target value for the health-related dimension, ensuring that the health attributes of the recipe are accurately matched with the user's blood sugar control needs. The demand target value for the taste dimension is determined by a weighted sum of the normalized values ​​for sweetness preference and thickness preference. For example, the demand target value for the taste dimension = normalized value for sweetness preference × 0.6 + normalized value for thickness preference × 0.4, forming a target value that can comprehensively reflect the user's taste preferences. The demand target value for the production complexity dimension temporarily adopts the initial baseline value of 0.5 from the standard demand vector and will be further optimized in the future based on user scenario tags. After determining the target values ​​for each dimension, the corresponding dimension values ​​in the standard demand vector are replaced with the calculated target values ​​to generate a user demand vector. This vector is perfectly aligned with the dimensions of the multi-dimensional recipe attribute vector, and the value of each dimension directly reflects the intensity of the user's demand for that attribute. This solves the technical problem in the background technology of not being able to find a reasonable balance between multiple attributes such as health, taste, and production complexity. For example, a user's health assessment record shows that their blood sugar control target is strict sugar control, which corresponds to a normalized value of 1 after structured processing. The sweetness preference score input through the real-time input interface is 8, and the thickness preference score is 6, with normalized values ​​of 0.8 and 0.6 respectively after processing. The target value for health-related dimensions is set to 0.9 based on the normalized value of 1, the target value for taste dimension is determined to be 0.72 by combining the two preferences, and the production complexity dimension remains at 0.5. The final constructed user demand vector is [0.9, 0.72, 0.5]. This vector clearly presents the user's core needs for high health adaptability, a sweet and thick taste, and moderate production difficulty, ensuring that subsequent recipe recommendations can accurately respond to these personalized demands.

[0034] S5. Determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on user health status data and user preference data, and normalize the weight coefficients to generate standardized weight coefficients.

[0035] In one specific embodiment, performing step S5 includes the following steps: Extract blood glucose control target level information from user health status data and preference information on sweetness and consistency from user preference data; Based on the blood glucose control target level information, determine the weight coefficients of health-related scores in the multi-dimensional formula attribute vector; Based on the preference information for sweetness and thickness, determine the weight coefficients of taste score in the multi-dimensional recipe attribute vector; Retrieve a preset complexity weight configuration library, which includes scene tags and corresponding basic weight values ​​for production complexity. Based on the user's scene tag, the basic weight value of production complexity corresponding to the user's scene tag is obtained from the preset complexity weight configuration library, and the weight coefficient of production complexity score in the multi-dimensional recipe attribute vector is determined. The weighting coefficients of health-related scores, taste scores, and production complexity scores are normalized to generate standardized weighting coefficients.

[0036] Specifically, the extraction operation uses the user's unique identifier as an index to associate the acquired user health status data with preference data. It accurately extracts the level information of blood sugar control targets from the health status data, namely, strict blood sugar control, general blood sugar control, and maintaining the status quo. At the same time, it extracts the preference information on sweetness and consistency from the preference data. During the extraction process, it ensures that each type of information forms a unique mapping with the user's unique identifier, without data confusion or omission, providing accurate personalized data support for the subsequent determination of weight coefficients.

[0037] When determining the weighting coefficients of health-related scores based on blood glucose control target levels, a fixed correspondence between levels and weighting coefficients is established. This correspondence is based on the logical setting of the contribution of soy milk formulas to blood glucose control. Specifically, strict blood glucose control corresponds to a weighting coefficient of 0.6, general blood glucose control corresponds to a weighting coefficient of 0.3, and maintaining the status quo corresponds to a weighting coefficient of 0.1. By assigning differentiated weights to different blood glucose control needs, the influence of health-related scores in the matching calculation is ensured to accurately match the user's health goals. For example, users with strict blood glucose control have the highest health weight, ensuring that recommended formulas prioritize meeting health requirements such as low glycemic index and high dietary fiber, providing effective healthy dietary support for users. When determining the weighting coefficients for taste scores based on preferences for sweetness and thickness, a weighted average algorithm is used to integrate the two preference data points. The original sweetness and thickness preferences are used as input, assigned weights of 0.6 and 0.4 respectively. A comprehensive preference score is then calculated and mapped to the taste score's weighting coefficients. The mapping rule is: a comprehensive preference score of 0-3 corresponds to a weighting coefficient of 0.1, 4-7 to 0.3, and 8-10 to 0.5. For example, if a user's sweetness preference score is 9 and their thickness preference score is 7, their comprehensive preference score would be... The taste score is weighted at 0.5. By quantifying and integrating users' specific needs for sweetness and consistency, the weighting coefficient of the taste dimension can accurately reflect users' taste preferences, ensuring that the recommended formula meets health standards while also meeting user expectations in terms of taste, thus avoiding users abandoning the product due to poor taste.

[0038] The system retrieves a pre-built structured database containing a pre-defined complexity weight configuration library. This library stores the mapping between scene tags and their corresponding basic complexity weight values. Scene tags include commuting scenarios, daily home scenarios, and weekend leisure scenarios, with corresponding basic complexity weight values ​​of 0.1, 0.3, and 0.5, respectively. Commuting scenarios emphasize ease of preparation and have the lowest basic weight value, meaning that complexity has the least impact on the matching calculation, prioritizing recipes with simple steps and short preparation time. Weekend leisure scenarios allow users to have ample time and have the highest basic weight value, allowing the recommendation of recipes with slightly more steps and more complex processes. User scene tags are automatically determined by a pre-defined rule engine. The rules are defined as follows: if the user initiates a request on a weekday between 06:00 and 09:00, it is determined to be a commuting scenario; if the user initiates a request on a Saturday, Sunday, or public holiday between 14:00 and 18:00, it is determined to be a weekend leisure scenario; all other times are defaulted to daily home scenarios. Based on the user's scenario tag, the corresponding basic weight value of production complexity is matched from the configuration library and used as the weight coefficient of production complexity score in the multi-dimensional recipe attribute vector. Through scenario-based weight allocation, the consideration of production complexity is adapted to the user's actual usage scenario, thereby improving the practicality of recommended recipes.

[0039] The weighting coefficients for health-related scores, taste scores, and production complexity scores are normalized using a proportional normalization method. The logic involves calculating the sum of the three weighting coefficients, then dividing each weighting coefficient by the sum to obtain standardized weighting coefficients. This ensures that the sum of the three standardized weighting coefficients is 1, and that the weight ratios for each dimension are reasonable, preventing recommendation bias caused by any dimension being too high or too low. For example, a user's blood sugar control goal is strict blood sugar control (health weighting coefficient 0.6), their overall taste preference score is 8.2 (taste weighting coefficient 0.5), and their scenario tag is commuting (complexity weighting coefficient 0.1). The sum of the three weighting coefficients is 1.2. The standardized weighting coefficient for the health-related score after normalization is... The standardized weighting coefficient for taste scores is: The standardized weighting coefficients for the complexity score are: The sum of the three standardized weight coefficients is 1. By unifying the weight scale, the weights of the three dimensions of health, taste, and production complexity are comparable, ensuring that the influence of each dimension is reasonably allocated according to user needs in the subsequent weighted distance calculation, achieving multi-dimensional accurate matching, and improving the accuracy and practicality of formula recommendations.

[0040] S6. Based on standardized weight coefficients, calculate the distance between the multi-dimensional recipe attribute vector and the user demand vector. The distance value represents the degree of matching between the soy milk recipe and the user demand.

[0041] In one specific embodiment, performing step S6 includes the following steps: Obtain multi-dimensional recipe attribute vectors and user demand vectors; Based on standardized weight coefficients, the corresponding dimensions of the multi-dimensional recipe attribute vector and the user demand vector are weighted respectively. The weighted Euclidean distance algorithm is used to calculate the distance between the weighted multi-dimensional recipe attribute vector and the weighted user demand vector.

[0042] Specifically, when obtaining the multi-dimensional recipe attribute vector and user demand vector, the recipe identifier and the user's unique identifier are used as dual indexes to ensure accurate data correlation. The multi-dimensional recipe attribute vector is a three-dimensional vector generated in step S3, with the dimension order fixed as health-related rating dimension, taste rating dimension, and production complexity rating dimension. Each vector maintains a unique mapping relationship with the corresponding soy milk recipe, and the values ​​of each dimension in the vector are standardized attribute values, all ranging from 0 to 1. The user demand vector is a three-dimensional vector constructed in step S4, with the dimension order completely consistent with the multi-dimensional recipe attribute vector, corresponding to the user's target values ​​for health, taste, and production complexity, also ranging from 0 to 1, ensuring that the two dimensions are aligned and scaled uniformly, providing basic data support for subsequent distance calculations.

[0043] When performing weighted processing based on standardized weight coefficients, the standardized weight coefficients are the three-dimensional coefficients generated in step S5, corresponding to the three dimensions of the multi-dimensional recipe attribute vector and the user demand vector, respectively. The sum of the coefficients is 1, and the magnitude of each dimension coefficient reflects the importance of that dimension in the matching calculation. The weighted processing logic is as follows: multiply the value of each dimension in the multi-dimensional recipe attribute vector by the corresponding standardized weight coefficient to obtain the weighted recipe attribute vector; simultaneously, multiply the value of each dimension in the user demand vector by the corresponding standardized weight coefficient to obtain the weighted user demand vector. For example, if a user's standardized weight coefficients are [0.5, 0.417, 0.083], a soy milk recipe's multi-dimensional attribute vector is [0.8, 0.7, 0.3], and the user demand vector is [0.9, 0.72, 0.5], then the values ​​of each dimension of the weighted recipe attribute vector are 0.4, 0.2919, and 0.0249, respectively; and the values ​​of each dimension of the weighted user demand vector are 0.45, 0.3002, and 0.0415, respectively. By assigning higher weight to user-focused dimensions, the distance calculation results are made more aligned with users' personalized needs. For example, the health dimension is given higher weight to users who strictly control their sugar intake, and its influence is amplified in the calculation, ensuring that the recommended formula prioritizes meeting their health goals.

[0044] When calculating the distance value using the weighted Euclidean distance algorithm, the difference between each dimension of the two weighted vectors is quantified to obtain a numerical value reflecting the degree of matching between the recipe and user needs; the smaller the distance value, the higher the degree of matching. First, the difference between the corresponding dimensions of the weighted multi-dimensional recipe attribute vector and the weighted user need vector is calculated, i.e., the weighted recipe value minus the weighted need value for each dimension. Then, the difference for each dimension is squared to eliminate the negative sign effect and amplify significant differences. Next, the squared difference of each dimension is multiplied by the corresponding standardized weight coefficient to further strengthen the impact of high-weight dimension differences on the distance value. Then, the weighted squared differences of all dimensions are summed to obtain the total difference value. Finally, the square root of the total difference value is taken to obtain the final distance value. Continuing the example above, the differences for each dimension are -0.05, -0.0083, and -0.0166; the squared differences are 0.0025, 0.000069, and 0.000276; the weighted squared differences are 0.00125, 0.0000287, and 0.0000229; the total difference is 0.00125 + 0.0000287 + 0.0000229 ≈ 0.0013016; the distance value... By integrating the differences in health, taste, and production complexity into a single distance value using weighted Euclidean distance, a quantitative assessment of multi-dimensional matching is achieved. Simultaneously, the inclusion of weighting coefficients in the algorithm addresses the problem of existing methods failing to consider user needs and prioritizing specific needs, leading to biased recommendation results. This ensures that the distance value not only reflects attribute differences but also the user's level of attention to each dimension, making the assessment of matching more personalized and accurate.

[0045] S7. After filtering and sorting the soy milk recipes based on the distance value, determine the recommended soy milk recipe that matches the user's needs.

[0046] In one specific embodiment, performing step S7 includes the following steps: Obtain the distance value and production complexity score for each soy milk recipe; Soy milk recipes with a complexity score higher than the preset difficulty threshold are removed from the candidate set, generating a preliminary list of matching recipes; Soy milk recipes in the preliminary matching recipe list are sorted in ascending order based on distance value. The soy milk recipe with the smallest distance value is extracted from the ranking results and determined as the recommended soy milk recipe that matches the user's needs.

[0047] Specifically, when obtaining the distance value and production complexity score corresponding to each soy milk recipe, the recipe identifier is used as the unique association index to establish a precise mapping relationship between the two. The distance value is the result calculated by the weighted Euclidean distance algorithm in step S6, directly representing the degree of matching between the recipe and user needs; the smaller the value, the higher the matching degree. The production complexity score uses the original data after integrity verification in step S1, and is comprehensively evaluated based on dimensions such as the number of pretreatment steps, total preparation time, type of special equipment required, and difficulty of raw material pretreatment. The value range is 1-10, with a higher score indicating a more complicated production process and greater operational difficulty. During the acquisition process, the system traverses all recipe data calculated in step S6, matching the corresponding distance value and production complexity score one by one according to the recipe identifier, ensuring that each set of data and recipe has a unique correspondence, with no data mismatch or omission, providing complete and reliable basic data for subsequent screening and sorting.

[0048] When removing soy milk recipes with a complexity score higher than a preset difficulty threshold from the candidate set, the preset difficulty threshold is a general standard set during system initialization, with a value of 6. Users can also personalize this threshold based on their own operational abilities and time constraints, with an adjustment range of 3-8. The filtering logic compares the complexity score of each recipe in the candidate set with the preset threshold one by one. If the score is higher than the threshold, it indicates that the recipe's difficulty exceeds the user's generally acceptable range or their individual tolerance range. Even if it matches the user's needs well, it may be abandoned due to its cumbersome operation, and therefore directly removed from the candidate set. Only recipes with a complexity score lower than or equal to the threshold are retained, and a preliminary matching recipe list is generated after integration. For example, if a user adjusts the difficulty threshold to 5, and a recipe has a complexity score of 7, it is removed because it exceeds the threshold, while recipes with a score of 4 are retained in the preliminary matching recipe list. By filtering recipes with excessively high operational difficulty, the system avoids recommendations being rejected by users due to insufficient practicality, ensuring that the recipes in the preliminary matching list have both a certain degree of matching and are within the user's actual operational capabilities, thus improving the feasibility of the recommended solutions.

[0049] When sorting the soy milk recipes in the initial matching recipe list in ascending order based on distance values, the sorting algorithm employs a stable sorting mechanism. Recipes with smaller distance values ​​are sorted sequentially, while those with larger distance values ​​are sorted later. During the sorting process, if two or more recipes have the same distance value, a secondary sort is performed based on their production complexity score. Recipes with lower production complexity scores are sorted first, ensuring the uniqueness and rationality of the sorting results. For example, if the initial matching recipe list contains three recipes with distance values ​​of 0.036, 0.052, and 0.036, and corresponding production complexity scores of 3, 4, and 2, the sorted order would be: recipe with a distance value of 0.036 and a production complexity score of 2, recipe with a distance value of 0.036 and a production complexity score of 3, and recipe with a distance value of 0.052 and a production complexity score of 4. This clear sorting rule intuitively presents the degree of fit between recipes and needs, allowing users to quickly locate suitable recipes without repeatedly comparing them among a large number of recipes, reducing trial-and-error costs.

[0050] When extracting the soy milk recipe with the smallest distance value from the sorted results, the system directly reads the first item in the sorted list. This item's recipe achieves an optimal balance in terms of health suitability, taste preference satisfaction, and ease of preparation, and is therefore identified as the recommended soy milk recipe matching the user's needs. If the user enables the "Multiple Recipe Recommendation" function, the system can extract the first N recipes in sorted order (N is the user-preset number, defaulting to 3) and generate a recommendation list for the user to choose from. If this function is not enabled, only the first recipe is extracted as the sole recommendation. During the extraction process, the system simultaneously links the recipe's complete information, including ingredient ratios, preparation steps, nutritional details, and health-related ratings, providing the user with comprehensive reference information. By focusing on the recipe with the smallest distance value, the system ensures the highest degree of matching between the recommended results and the user's needs. Simultaneously, the system supports a multi-recipe recommendation mode, balancing user choice diversity and achieving a balance between personalization and practicality. This allows users to quickly obtain a suitable recipe that meets their health goals, taste preferences, and ease of preparation, improving recommendation efficiency and user experience.

[0051] Please see Figure 2 , Figure 2 The diagram illustrates the two-dimensional matching effect between health and taste, showing the performance of traditional method 1, traditional method 2, and this method in terms of both health goal achievement and taste compatibility. Traditional method 1 is a soy milk recipe recommendation method that matches only the nutritional dimension, while traditional method 2 is a soy milk recipe recommendation method that filters only based on user taste preferences. This demonstrates that traditional method 1, focusing only on the nutritional dimension, achieves a high degree of health goal achievement but has extremely low taste compatibility, failing to meet user taste needs; traditional method 2, focusing only on taste preferences, achieves a high degree of taste compatibility but a low degree of health goal achievement, making it difficult to provide effective health support; while this method achieves a high-dimensional balance in both health goal achievement and taste compatibility, overcoming the limitations of existing methods' single-dimensional matching and achieving precise matching of both health and taste.

[0052] Please see Figure 3 , Figure 3The graph showing the average satisfaction scores and standard deviations for traditional methods 1, 2, and this method is illustrated. The vertical axis represents the average satisfaction score. Traditional method 1 has an average satisfaction score of 2.60, traditional method 2 has an average satisfaction score of 2.85, and this method has an average satisfaction score of 4.24. Traditional methods 1 and 2 have large standard deviations, while this method has a small standard deviation. This demonstrates that traditional recommendation methods that focus solely on nutrition or taste fail to balance user health needs, taste preferences, and ease of preparation. This results in significant differences in user evaluations of recommended recipes, high dispersion in satisfaction data, and a large standard deviation, reflecting poor stability of the recommendation effect. In contrast, this method, through precise matching across multiple dimensions of health, taste, and preparation complexity, can adapt to the needs of users with different health states, taste preferences, and usage scenarios. The satisfaction scores of different users tend to be consistent, with low dispersion and a small standard deviation. This method achieves a significant improvement in average satisfaction while ensuring the stability and universality of the recommendation effect, validating the method's adaptability and reliability across different user groups.

[0053] Please see Figure 4 The following describes a customized soy milk recipe recommendation system based on user health data, as described in an embodiment of this application. This customized soy milk recipe recommendation system based on user health data includes: The data acquisition module is used to acquire raw data of multiple soy milk recipes from a preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative values ​​of glycemic index, the quantitative values ​​of dietary fiber content, the quantitative values ​​of protein content, taste score, and production complexity score. The scoring mapping module is used to map the raw data into health-related scores for each soy milk recipe using a preset nutritional model. The health-related score represents the contribution of the corresponding soy milk recipe to a specific health goal. The vector generation module is used to standardize the taste score, production complexity score and health-related score in the multi-dimensional attributes of each soy milk recipe, and generate a multi-dimensional recipe attribute vector for each soy milk recipe. The requirement building module is used to acquire user health status data and user preference data, and to build user requirement vectors based on the user health status data and user preference data; The weight determination module is used to determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on user health status data and user preference data, and to normalize the weight coefficients to generate standardized weight coefficients. The distance calculation module is used to calculate the distance between the multi-dimensional recipe attribute vector and the user demand vector based on standardized weight coefficients. The distance value represents the degree of matching between the soy milk recipe and the user demand. The recipe recommendation module is used to filter and sort various soy milk recipes based on distance values, and determine the recommended soy milk recipes that match the user's needs.

[0054] Through the collaborative efforts of the aforementioned components, this system constructs a full-chain intelligent adaptation system, from intelligent processing of formula data to precise matching of user needs, and then to personalized formula recommendations. It achieves dynamic weighted matching of soy milk formulas across three dimensions: health attributes, taste preferences, and preparation complexity. This breaks through the limitations of traditional recommendation methods that rely on a single dimension, allowing recommendations to accurately match the health status, taste preferences, and usage scenarios of different users, thus improving the accuracy, practicality, and user suitability of soy milk formula recommendations. Specifically: The data acquisition module, as the core entry point for system data input, provides standardized and structured basic data for soy milk recipes for the entire process. Through integrity verification and structured data organization, it eliminates invalid data, ensuring the accuracy and consistency of the data source for subsequent model calculations and vector generation. It is the fundamental support for the entire recommendation system. The scoring mapping module builds a quantitative bridge between the original recipe data and its health value. Relying on a professional nutritional model, it transforms multi-dimensional nutritional values ​​into intuitive health-related scores, enabling a scientific assessment of the recipe's health contribution and providing a quantitative basis for precise matching of health dimensions. It is the core module for the system to achieve health goal adaptation. The vector generation module eliminates the scale differences between health, taste, and production complexity scores through standardization. It transforms non-homogeneous recipe attributes into multi-dimensional attribute vectors that can be compared horizontally and mathematically calculated. This solves the technical challenge of multi-dimensional attributes not being able to collaboratively participate in matching calculations, and is a key link in achieving multi-dimensional quantitative matching. The demand construction module digitizes and structures user health status and personalized preferences, mining core needs from historical user data and real-time input. It constructs a user demand vector aligned with the recipe attribute vector, making user-specific needs the core guide for recipe matching—a crucial prerequisite for the system's "customized" recommendations. The weight determination module dynamically allocates weights to each dimension based on user health goals, taste preferences, and usage scenarios. Through normalization, it achieves a scientific weight distribution, prioritizing high-priority dimensions in the matching calculation. This personalized and dynamic adaptation of the recommendation logic is a key support for balancing multi-dimensional needs and aligning with core user demands. The distance calculation module, the core computational module for matching recipes with user needs, uses a weighted Euclidean distance algorithm to quantify the matching differences between the recipe attribute vector and the user demand vector. It integrates multi-dimensional attribute differences into a single distance value, enabling a scientific and quantitative assessment of the recipe matching degree and providing precise numerical data for recipe selection and ranking. As the system's output terminal, the recipe recommendation module filters out recipes with insufficient practicality by using a difficulty threshold for preparation, and then sorts them based on distance value to finally determine the optimal matching recipe. This ensures both the matching degree of the recommendation results and the convenience of actual preparation, turning the recommendation results from "data matching" to "practical usability," which is the final step in realizing the value of the recommendation.

[0055] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the method for recommending customized soy milk recipes using user health data.

[0056] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and systems described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0057] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for recommending customized soy milk recipes using user health data, characterized in that, Includes the following steps: S1. Obtain raw data of multiple soy milk recipes from a preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative values ​​of glycemic index, the quantitative values ​​of dietary fiber content, the quantitative values ​​of protein content, taste score, and production complexity score. S2. The raw data is mapped to health-related scores for each soy milk formula using a preset nutritional model. The health-related scores represent the contribution of the corresponding soy milk formula to a specific health goal. S3. Standardize the taste score, production complexity score and health-related score in the multi-dimensional attributes of each soy milk recipe to generate a multi-dimensional recipe attribute vector for each soy milk recipe. S4. Obtain user health status data and user preference data, and construct a user demand vector based on the user health status data and user preference data; S5. Determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on the user health status data and the user preference data, and normalize the weight coefficients to generate standardized weight coefficients. S6. Based on the standardized weight coefficients, calculate the distance between the multi-dimensional recipe attribute vector and the user demand vector, whereby the distance value represents the degree of matching between the soy milk recipe and the user demand. S7. After screening and sorting the soy milk recipes according to the distance value, determine the recommended soy milk recipe that matches the user's needs.

2. The method according to claim 1, characterized in that, Step S1 includes: Extract multi-dimensional attribute data of each soy milk recipe from the preset recipe library; Perform integrity checks on the multi-dimensional attribute data of each soy milk recipe, check for missing data in each field, and remove soy milk recipes with missing data. For the soy milk recipe after integrity verification, the multi-dimensional attribute data is organized and integrated into a structured raw dataset according to the preset field order and data type rules.

3. The method according to claim 2, characterized in that, Step S2 includes: The original dataset is input into a pre-built nutrition model; Based on a preset specific health goal, the nutritional model is used to assign corresponding weight coefficients to each quantitative value in the original dataset, and a linear weighted summation calculation is performed to output the original health-related score corresponding to each soy milk recipe; The original health-related scores were normalized to obtain the health-related scores for each soy milk formula.

4. The method according to claim 1, characterized in that, Step S3 includes: Extract the taste score, production complexity score, and health-related score from the multi-dimensional attributes of each soy milk recipe; The extracted health-related scores, taste scores, and production complexity scores are converted using a pre-defined standardization algorithm to generate corresponding standardized attribute values. For each soy milk recipe, the corresponding standardized attribute values ​​are combined in a preset dimensional order to generate a multi-dimensional recipe attribute vector for each soy milk recipe.

5. The method according to claim 1, characterized in that, Step S4 includes: Extract health status-related data from user history records, including blood glucose control target level information; The user's taste preference data is obtained through a real-time input interface. The taste preference data includes information on the degree of preference for sweetness and thickness. The health status-related data and the taste preference data are structured to generate a standardized user feature dataset; Based on the standardized user feature dataset, the target demand value corresponding to each dimension of the multi-dimensional recipe attribute vector is determined, and the preset standard demand vector is adjusted based on the target demand value to construct the user demand vector.

6. The method according to claim 5, characterized in that, Step S5 includes: Extract blood glucose control target level information from the user's health status data and preference information for sweetness and consistency from the user's preference data; Based on the blood glucose control target level information, determine the weight coefficients of health-related scores in the multi-dimensional formula attribute vector; Based on the preference information for sweetness and thickness, determine the weight coefficients of taste score in the multi-dimensional recipe attribute vector; Obtain a preset complexity weight configuration library, which includes scene tags and corresponding basic weight values ​​for production complexity. Based on the user's scene tag, the basic weight value of production complexity corresponding to the user's scene tag is obtained from the preset complexity weight configuration library, and the weight coefficient of production complexity score in the multi-dimensional recipe attribute vector is determined. The weighting coefficients of health-related scores, taste scores, and production complexity scores are normalized to generate standardized weighting coefficients.

7. The method according to claim 1, characterized in that, Step S6 includes: Obtain the multi-dimensional recipe attribute vector and the user requirement vector; Based on the standardized weight coefficients, the corresponding dimensions of the multi-dimensional recipe attribute vector and the user demand vector are weighted respectively. The weighted Euclidean distance algorithm is used to calculate the distance between the weighted multi-dimensional recipe attribute vector and the weighted user demand vector.

8. The method according to claim 1, characterized in that, Step S7 includes: Obtain the distance value and production complexity score for each soy milk recipe; Soy milk recipes with a complexity score higher than the preset difficulty threshold are removed from the candidate set, generating a preliminary list of matching recipes; Soy milk recipes in the preliminary matching recipe list are sorted in ascending order based on the distance value. The soy milk recipe with the smallest distance value is extracted from the ranking results and determined as the recommended soy milk recipe that matches the user's needs.

9. A customized soy milk recipe recommendation system utilizing user health data, used to implement the customized soy milk recipe recommendation method utilizing user health data as described in any one of claims 1 to 8, characterized in that, include: The data acquisition module is used to acquire raw data of multiple soy milk recipes from a preset recipe library. The raw data represents the multi-dimensional attributes of each soy milk recipe. The multi-dimensional attributes include at least the quantitative values ​​of nutritional components, the quantitative values ​​of glycemic index, the quantitative values ​​of dietary fiber content, the quantitative values ​​of protein content, taste score, and production complexity score. The scoring mapping module is used to map the raw data into health-related scores for each soy milk formula through a preset nutritional model. The health-related scores represent the contribution of the corresponding soy milk formula to a specific health goal. The vector generation module is used to standardize the taste score, production complexity score and health-related score in the multi-dimensional attributes of each soy milk recipe, and generate a multi-dimensional recipe attribute vector corresponding to each soy milk recipe. The requirement construction module is used to acquire user health status data and user preference data, and construct a user requirement vector based on the user health status data and user preference data. The weight determination module is used to determine the weight coefficients of each dimension in the multi-dimensional recipe attribute vector based on the user health status data and the user preference data, and to normalize the weight coefficients to generate standardized weight coefficients. The distance calculation module is used to calculate the distance value between the multi-dimensional recipe attribute vector and the user demand vector based on the standardized weight coefficients. The distance value represents the degree of matching between the soy milk recipe and the user demand. The recipe recommendation module is used to filter and sort the various soy milk recipes based on the distance value, and determine the recommended soy milk recipe that matches the user's needs.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is executed by the processor, it implements a method for recommending customized soy milk recipes using user health data as described in any one of claims 1-8.

Images