An item information pushing method combining item attributes and user preferences

Abstract

The application provides an article information pushing method combining article attributes and user preferences, comprising: obtaining digestive system disease types, digestive tract absorption capacity, digestive fluid secretion conditions and physiological characteristic data related to digestive function of a plurality of users, grouping the users through a clustering algorithm to obtain user groups with different digestive function characteristics; obtaining dosage form, particle size and surface area attribute information of health products, reducing the dimensions of the health products through principal component analysis according to disintegration time, dissolution rate and osmotic pressure characteristics of the dosage form to obtain a key attribute set of health products with different dosage forms; obtaining health product information searched by the user and digestive condition data of the user, and judging whether the user belongs to a group with weak digestive tract absorption capacity or a group with normal digestive function according to a set digestive threshold.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to an item information pushing method combining item attributes and user preferences. BACKGROUND

[0002] There are challenges in matching health product formulations with individualized user digestive functions. Different users have different types of digestive system diseases, digestive tract absorption capacities, and digestive fluid secretion conditions, while health products have different formulations, particle sizes, and surface area attributes. Accurately assessing the characteristics of a user's digestive function and precisely matching them with the characteristics of a health product formulation is an important task. In particular, for groups with weak digestive tract absorption capacity, it is particularly important to choose formulations that are easier to absorb. At the same time, the user's digestive function condition may change over time, and the actual absorption effect of the health product is difficult to accurately assess. This makes it difficult to establish a dynamic and individualized recommendation mechanism. Whether based on static assessment or dynamic tracking, current methods cannot fully consider individual differences, health product characteristics, and dynamic changes in digestive function. These factors together constitute a complex technical problem, namely, how to achieve optimal matching of health product formulations with user digestive functions while considering multiple factors such as individual differences, health product characteristics, and dynamic changes, in order to provide truly personalized and precise health solutions. This means that the convenience and accuracy of collaborative matching in personalized health product formulation recommendation are currently low. SUMMARY

[0003] The present application provides an item information pushing method combining item attributes and user preferences, mainly including:

[0004] Obtain the digestive system disease types, digestive tract absorption capacities, digestive fluid secretion conditions, and physiological characteristic data related to digestive function of multiple users, group the users through a clustering algorithm, and obtain user groups with different digestive function characteristics;

[0005] Obtain the formulation, particle size, and surface area attribute information of health products, reduce the dimensionality of the health products using principal component analysis based on the disintegration time, dissolution, and osmotic pressure characteristics of the formulation, and obtain a key attribute set of health products of different formulations;

[0006] For different digestive function groups, analyze their digestive tract peristalsis frequency, mucosal barrier function, and characteristics of the balance state of the flora, combine the adhesion, release control mechanism, and target delivery capacity attributes of the health product formulation, and construct a correlation matrix of user digestive characteristics and health product formulation adaptability;

[0007] According to the association matrix, the acceptance and absorption efficiency scores of different user groups to health care products of different types are calculated by using a collaborative filtering algorithm to obtain a user-dosage form score matrix, and a digestion threshold is set to determine the groups with weak digestive absorption capacity and the groups with normal digestive function.

[0008] The health care product information searched by the user and the user's digestion data are obtained, and according to the set digestion threshold, it is determined whether the user belongs to the group with weak digestive absorption capacity or the group with normal digestive function.

[0009] If the user belongs to the group with weak digestive absorption capacity, the health care products with high dissolution, small particle size and large surface area such as liquid and soft capsule are preferentially recommended according to the health care product information searched by the user and the user-dosage form score matrix, and if the user belongs to the group with normal digestive function, the health care products of different types suitable for the user's preference are recommended according to the tolerance of the digestive tract to specific ingredients, so as to obtain the final recommendation strategy.

[0010] The feedback of the user to the recommended health care products is continuously tracked, the absorption efficiency evaluation and satisfaction score of the user are obtained, the recommendation strategy is dynamically optimized, the matching degree of the health care product dosage form and the user's digestion characteristics is improved, and the personalized and precise health care scheme is provided for the user.

[0011] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0012] The present application discloses an item information pushing method combining item attributes and user preferences. The method classifies users and health care products by obtaining user digestive system data and health care product dosage form attributes, using clustering and principal component analysis. Then, an association matrix of user digestion characteristics and health care product dosage form adaptability is constructed, a scoring matrix is calculated using a collaborative filtering algorithm, and a digestion threshold is set. According to the health care product information searched by the user and the digestion condition, the user group type is determined and the corresponding recommendation is given. For the group with weak digestive absorption capacity, liquid and soft capsule health care products with high dissolution and small particle size are preferentially recommended; for the group with normal digestive function, the appropriate dosage form is recommended according to the tolerance. The present application further optimizes the recommendation strategy through reinforcement learning algorithm, improves the matching degree, provides personalized and precise health care scheme for the user, and effectively solves the problem of mismatch between health care product dosage form and user digestion characteristics. BRIEF DESCRIPTION OF DRAWINGS

[0013] Fig. 1 The flowchart of the item information pushing method combining item attributes and user preferences of the present application.

[0014] Fig. 2 The schematic diagram of the item information pushing method combining item attributes and user preferences of the present application.

[0015] Fig. 3 Another embodiment of the application is an item information pushing method combining item attributes and user preferences. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the application will be described clearly and in detail below with reference to the drawings in the embodiments of the application. The described embodiments are only some of the embodiments of the application.

[0017] As Figs. 1-3 The item information pushing method combining item attributes and user preferences can specifically include the following steps.

[0018] In step S101, the types of digestive system diseases, the absorption capacity of the digestive tract, the secretion of digestive juice, and the physiological characteristic data related to the digestive function of a plurality of users are acquired, and the users are grouped by using a clustering algorithm to obtain user groups with different digestive function characteristics.

[0019] The types of digestive system diseases, the absorption capacity of the digestive tract, the secretion of digestive juice, and the physiological characteristic data of the users are acquired, the original data are cleaned and preprocessed to obtain standardized user characteristic vectors, the user characteristic vectors are analyzed by using a clustering algorithm, the optimal number of clustering clusters is determined according to a contour coefficient, if the contour coefficient reaches a maximum value, the corresponding number of clusters is determined as the optimal value, the boundary users are divided again in combination with a hierarchical clustering method according to the preliminary grouping result, the similarity of the user characteristics in the group is calculated by using an AGNES algorithm, the cosine similarity is used to measure the similarity, if the similarity between the users is lower than a set threshold, the user is removed, the types of digestive system diseases, the absorption capacity of the digestive tract, the secretion of digestive juice, and the physiological characteristic data of the users in each group are statistically analyzed according to the clustering result label, the proportion of different types of digestive system diseases, the mean and standard deviation of the absorption capacity of the digestive tract, the median and interquartile range of the secretion of digestive juice, and the distribution of the main physiological characteristics in each group are calculated, and based on the statistical results, a user digestive function characteristic discrimination rule set is constructed, and new users are filtered step by step by using the discrimination rule set to determine the groups to which the new users belong.

[0020] Specifically, the user's digestive system disease type, digestive tract absorption capacity, digestive juice secretion condition, and physiological characteristic data are obtained, and the original data is cleaned and preprocessed. A missing value proportion threshold of 20% is set, and samples exceeding the threshold are removed. For continuous data, the 3 times standard deviation method is used to detect outliers, and for discrete data, the frequency analysis method is used to detect outliers. The continuous data is normalized by the minimum and maximum method, and the discrete data is converted by one-hot encoding to generate a standardized user feature vector. The user feature vector is analyzed using K-means clustering, and the Silhouette Coefficient is used to determine the optimal number of clusters. By calculating the Silhouette Coefficient under different cluster numbers, the cluster number with the maximum Silhouette Coefficient is selected as the optimal value. The K-means++ method is used to select the initial center point, and the cluster centers are iteratively calculated until the center point position changes less than a preset threshold or the maximum iteration number is reached. According to the Euclidean distance, the users are divided into the nearest cluster, and the preliminary user grouping result is obtained. For the preliminary grouping result, hierarchical clustering is used to perform secondary division on the boundary users. The AGNES (Agglomerative Nesting) algorithm is used to start from a single sample and gradually merge the most similar clusters. The similarity between users is measured using cosine similarity, and a similarity threshold of 75% quantile of the overall similarity distribution is set for similarity filtering to remove users below the threshold. Principal component analysis is performed on the user features of each group to reduce dimensionality, and the first two principal components are selected for visualization to obtain the optimized user groups with different digestive function characteristics. Based on the clustering result label, the user's digestive system disease type distribution, digestive tract absorption capacity level, digestive juice secretion condition, and physiological characteristic data are statistically analyzed. The proportion of different digestive system disease types, the mean and standard deviation of the digestive tract absorption capacity, the median and interquartile range of the digestive juice secretion amount, and the distribution of the main physiological characteristics are calculated. Based on these statistical results, a user digestive function characteristic discrimination rule set is constructed to realize the rapid grouping and feature recognition of new users. The discrimination rule set includes disease type judgment, absorption capacity evaluation, secretion condition classification, and physiological characteristic matching, and the new user's group is determined through step-by-step screening. When obtaining the user's digestive system data, the disease types such as gastric ulcer and colitis, the absorption capacity indicators such as glucose absorption rate, the digestive juice secretion conditions such as gastric acid secretion amount, and the physiological characteristics such as body mass index are collected. The original data is cleaned, and it is found that there are 25% missing values in the gastric acid secretion amount, which exceeds the 20% threshold, so these samples are removed. The 3 times standard deviation method is used to detect outliers, and outliers such as body mass index of 45 are removed. The continuous data such as glucose absorption rate is normalized by the minimum and maximum method to map it to the 0-1 interval. The discrete data such as disease type is one-hot encoded, such as gastric ulcer is encoded as [1, 0, 0] and colitis is encoded as [0, 1, 0].The processed user feature vector is subjected to K-means clustering analysis, and the silhouette coefficient is used to determine the optimal cluster number. The silhouette coefficient is calculated when the cluster number is 2 to 10, and it is found that the silhouette coefficient is maximum (0.68) when the cluster number is 5, and 5 is selected as the optimal cluster number. The K-means++ method is used to select the initial center point, and the cluster centers are calculated iteratively. The maximum number of iterations is set to 100, and the center point position change threshold is set to 0.001. After 87 iterations, the center point position change is less than the threshold, and the iteration is stopped. According to the Euclidean distance, the users are divided into the nearest cluster, and the preliminary grouping result is obtained. According to the preliminary grouping result, the AGNES algorithm is used to perform secondary division on the boundary users. Starting from 5000 single samples, the most similar clusters are gradually merged. The cosine similarity between users is calculated, and the similarity threshold is set to 0.85 (75% quantile of the overall similarity distribution). For user pairs with a similarity less than 0.85, they are divided into different clusters. Principal component analysis is performed on the user features of each group, and the first two principal components are selected for visualization. The first principal component explains 65% of the variance, mainly related to the digestive absorption capacity; the second principal component explains 20% of the variance, mainly related to the secretion of digestive juice. According to the clustering results, statistical analysis is performed on each group of users. For example, in the first group, the proportion of patients with gastric ulcer is 40%, and the proportion of patients with colitis is 35%; the average glucose absorption rate is 0.72±0.15; the median gastric acid secretion is 2.5 mEq / h, and the interquartile range is 1.2 mEq / h; the average body mass index is 23.5±3.2. Based on these statistical results, a user digestive function feature discrimination rule set is constructed. For example, if the glucose absorption rate of a new user is less than 0.6 and the gastric acid secretion is less than 1.5 mEq / h, it is classified as a weak digestive function group.

[0021] In step S102, the dosage form, particle size, and surface area attribute information of the health care product are obtained. Based on the disintegration time, dissolution rate, and osmotic pressure characteristics of the dosage form, principal component analysis is used to reduce the dimension of the health care product, and the key attribute set of different dosage forms of health care products is obtained.

[0022] Health product dosage form information and physical and chemical attribute data are acquired, the physical and chemical attribute data including particle size data, surface area attribute, disintegration time measurement results, dissolution curve data, and osmotic pressure measurement values. Z-score method is used to detect outliers in the physical and chemical attribute data, and if the absolute value of the Z-score of a data point is greater than a preset threshold, the data point is marked as an outlier. For missing values in the physical and chemical attribute data, the mean value of the corresponding attribute is used for filling. The continuous data in the physical and chemical attribute data are standardized to obtain standardized numerical values. The dosage form information is converted by one-hot encoding to obtain encoded dosage form data. According to the standardized numerical values and the encoded dosage form data, a standardized health product feature matrix is generated. The Pearson correlation coefficients between attributes in the standardized health product feature matrix are calculated to obtain a correlation coefficient matrix. According to the correlation coefficient matrix, the strength of the correlation between attributes is determined. If the absolute value of the correlation coefficient is greater than a preset threshold, it is determined to be a significant correlation. From the significantly correlated attribute pairs, one is selected as the input variable of principal component analysis to obtain a principal component analysis input variable set. Principal component analysis method is used to reduce the dimension of the principal component analysis input variable set. The covariance matrix of the principal component analysis input variable set is calculated, and the eigenvalues and eigenvectors of the covariance matrix are solved. The eigenvalues are arranged in descending order, and the contribution rate and cumulative contribution rate of each eigenvalue are calculated. The eigenvectors whose cumulative contribution rate reaches a preset threshold are selected as principal components to obtain the reduced health product feature representation. The physical meaning of each principal component in the reduced health product feature representation is explained, and the key attributes corresponding to each principal component are determined. The distribution of different dosage forms on the principal components is analyzed, and the mean and variance of each dosage form on the main principal components are calculated. According to the mean and variance, attributes with significant group differences are extracted to obtain a key attribute set of different dosage form health products.

[0023] Specifically, information on the health product dosage form, particle size data, surface area attributes, disintegration time measurements, dissolution curve data, and osmotic pressure measurements were obtained. Outliers were detected using the Z-score method, with a threshold of ±3. Data points outside this range were marked as outliers. Missing values ​​were filled using the attribute mean. Continuous data such as particle size, surface area, disintegration time, dissolution rate, and osmotic pressure were normalized using a Min-Max normalization process, mapping values ​​to the [0, 1] interval. Discrete data such as dosage form were one-hot encoded to generate a standardized health product feature matrix. Based on the standardized feature matrix, the Pearson correlation coefficient matrix was calculated between each attribute to determine the strength of the correlation between the attributes. A correlation coefficient greater than 0.7 was considered significant. For significantly correlated attribute pairs, such as particle size and surface area, or disintegration time and dissolution rate, one of the two was selected as the input variable for principal component analysis to reduce redundant information. The correlation analysis results determined the input variable set for the principal component analysis, which included key characteristics such as dosage form, particle size, disintegration time, and osmotic pressure. Principal component analysis (PCA) was used to reduce the dimensionality of the input variable set. The covariance matrix was calculated, and the eigenvalues ​​and eigenvectors were found. The eigenvalues ​​were sorted in descending order, and the contribution rate and cumulative contribution rate of each eigenvalue were calculated. A cumulative contribution rate threshold of 85% was set, and the eigenvectors corresponding to the first few eigenvalues ​​whose cumulative contribution rates reached this threshold were selected as principal components. The original data were projected onto these principal components to obtain a reduced-dimensional representation of the health product. The physical meaning of each principal component in the reduced-dimensional representation was explained. Key attributes were identified based on the composition of the principal components. For example, the first principal component might primarily reflect the physical properties of the health product, while the second principal component might primarily reflect its chemical properties. Combined with the original dosage form information, the distribution of different dosage forms on each principal component was analyzed. The mean and variance of each dosage form on the principal components were calculated, and attributes with significant differences between groups were extracted. By determining the key attribute sets for different dosage forms of health products, the corresponding relationship between dosage form and key attributes was determined. For example, tablets might focus more on disintegration time, while capsules might focus more on dissolution rate. When acquiring health product data, we collected information on 100 different health products, including tablets, capsules, and granules. The particle size range was 50-500 μm, and the surface area distribution was 0.5-5 m 2 / g, disintegration time in 5-30 minutes, dissolution curve data points for 10 time points within 0-120 minutes, osmotic pressure measurement in the range of 200-800 mOsm / kg. Use the Z-score method to detect outliers, set the threshold to ±3, find 5 abnormal data points, such as the abnormal value of particle size 800 μm. Fill in the attribute mean value for 3% missing values. The continuous data is processed by Min-Max standardization, such as mapping the particle size 50-500 μm to the interval 0-1. The dosage form is one-hot encoded, such as tablet encoding [1, 0, 0], capsule encoding [0, 1, 0], generating a 100x8 standardized feature matrix. Calculate the Pearson correlation coefficient matrix between attributes, find that the correlation coefficient between particle size and surface area is -0.85, and the correlation coefficient between disintegration time and dissolution is -0.78, both exceeding the significant correlation threshold of 0.7. Select particle size and disintegration time as input variables for principal component analysis, remove surface area and dissolution, and finally determine dosage form, particle size, disintegration time, and osmotic pressure as the input variable set for principal component analysis. Principal component analysis is performed on the input variable set, and four eigenvalues are calculated as 2.5, 0.8, 0.5, and 0.2. Calculate the cumulative contribution rate, find that the cumulative contribution rate of the first two eigenvalues is 82.5%, plus the third eigenvalue reaches 95%, exceeding the threshold of 85%. Select the characteristic vectors corresponding to the first three eigenvalues as the principal components. Project the original data onto these three principal components to obtain a 100x3 reduced dimension feature representation. Analyze the physical meaning of the principal components, find that the first principal component is mainly composed of particle size and disintegration time, reflecting the physical properties of health products; the second principal component is mainly composed of osmotic pressure, reflecting the chemical properties; the third principal component is mainly composed of dosage form information. Calculate the distribution of different dosage forms in each principal component, such as the mean of tablet in the first principal component is 0.7, and the variance is 0.1, while the mean of capsule is 0.4, and the variance is 0.08. Extract the attributes that are significantly different, determine the key attributes of tablets as disintegration time, the key attributes of capsules as dissolution, and the key attributes of granules as particle size and osmotic pressure, forming the correspondence between health product dosage form and key attributes.

[0024] Step S103, for different digestive function groups, analyze their digestive tract peristalsis frequency, mucosal barrier function, and characteristics of flora balance state, combined with the adhesion, release control mechanism, and target delivery ability attributes of health product dosage form, construct the correlation matrix of user digestion characteristics and health product dosage form adaptability.

[0025] The data of gastrointestinal motility frequency, mucosal barrier function index, and bacterial flora balance state parameter of different digestive function groups are acquired, the data of gastrointestinal motility frequency is processed by Z-score standardization method, the mucosal barrier function index is normalized, and the bacterial flora balance state parameter is logarithmically transformed and then standardized. According to the data processed by standardization, a preliminary mapping relationship between the user's digestive characteristics and the attributes of the health product dosage form is constructed, the correlation rules of gastrointestinal motility frequency and dosage form release control mechanism, mucosal barrier function and dosage form adhesion, and bacterial flora balance state and targeted delivery capacity are set. The adhesion experiment data, release control mechanism test results, and targeted delivery capacity evaluation index of the health product dosage form are processed by dimension reduction using principal component analysis, and the principal components with a cumulative contribution rate reaching a preset threshold are selected as feature vectors. The feature vectors are clustered by K-means clustering algorithm to obtain the dosage form attribute clustering results. According to the dosage form attribute clustering results, the Apriori algorithm is used for association rule mining, the user's digestive characteristics are taken as antecedents, the dosage form attribute clustering results are taken as consequents, the support threshold, confidence threshold, and minimum lift are set, the association rule set meeting the threshold conditions is mined, and a preliminary association matrix is generated. In view of the preliminary association matrix, the correlation strength is corrected by using a fuzzy comprehensive evaluation method in combination with the preset digestive characteristics and dosage form adaptability rules in the knowledge base, the adaptability, safety, and effectiveness are selected as evaluation indexes, the weights are set by using the analytic hierarchy process, the corrected correlation degree is calculated, and a final user's digestive characteristics and health product dosage form adaptability association matrix is constructed.

[0026] Specifically, the data of gastrointestinal motility frequency, mucosal barrier function indicators, and bacterial community balance state parameters of different digestive function groups are obtained. The Z-score standardization method is used to process the gastrointestinal motility frequency, the Min-Max standardization is used to normalize the mucosal barrier function indicators, and the logarithmic transformation is used to standardize the bacterial community balance state parameters. The discrete data such as digestive function group type is converted into one-hot encoding to generate a standardized user digestive feature matrix. Based on the historical data, a preliminary mapping relationship between the user digestive features and the health product dosage form attributes is constructed. The association rules between the gastrointestinal motility frequency and the release control mechanism of the dosage form, the mucosal barrier function and the adhesion of the dosage form, and the bacterial community balance state and the targeted delivery capability are set. These preliminary rules are stored in the knowledge base for subsequent construction and optimization of the association matrix. According to the adhesion experiment data, release control mechanism test results, and targeted delivery capability evaluation indicators of the health product dosage form, principal component analysis is used to reduce the dimension of the dosage form attributes. The first few principal components with a cumulative contribution rate of 85% are selected as the new feature vectors. K-means clustering is used to cluster the reduced dosage form attributes, and the number of clusters is set to 5. Iterative calculation is performed until the cluster center is stable or the maximum iteration number 100 is reached. The dosage form attribute clustering results are obtained, and the Euclidean distance between different dosage forms is calculated as the similarity index. Apriori algorithm is used for association rule mining, with user digestive features as antecedents and dosage form attribute clustering results as consequents. The support threshold is set to 0.1, the confidence threshold is set to 0.7, and the minimum lift is set to 1.1. The association rule set that meets the threshold conditions is mined, and a preliminary association matrix is generated. For the preliminary association matrix, the fuzzy comprehensive evaluation method is used to modify the association strength in combination with the pre-set digestive feature and dosage form adaptability rules in the knowledge base. Adaptability, safety, and effectiveness are selected as evaluation indicators, and the weights are set using the analytic hierarchy process. The modified association degree is calculated, and the final user digestive feature and health product dosage form adaptability association matrix is constructed. The data of 1000 users with different digestive function groups are obtained, including gastrointestinal motility frequency ranging from 3 to 12 times per minute, mucosal barrier function indicators ranging from 0 to 100 points, and bacterial community balance state parameters with Shannon diversity index ranging from 1.5 to 4.5. Z-score standardization is applied to the motility frequency to convert the original value to a standard normal distribution; Min-Max standardization is used for mucosal barrier function indicators to map the numerical value to the 0-1 interval; and logarithmic transformation is used for bacterial community balance state parameters after standardization. The digestive function group types, such as normal, weak, and strong, are one-hot encoded to generate a 1000x6 standardized user digestive feature matrix. Based on the historical data, a preliminary mapping relationship is constructed, such as setting the motility frequency of 5-8 times per minute to be related to the sustained-release dosage form, the mucosal barrier function indicators greater than 80 points to be related to the mucosal adhesion dosage form, and the bacterial community diversity index greater than 3.5 to be related to the enteric dosage form.The attribute data of 50 health care product formulations were collected, including adhesion 0-100%, release control time 0-24 hours, and target delivery capacity 0-10 points. Principal component analysis was used for dimension reduction, and the first two principal components with a cumulative contribution rate of 87% were selected as new features. K-means clustering was applied, and the number of clusters was set to 5. After 76 iterations, the cluster centers were stable, and the formulation attribute clustering results were obtained. Association rule mining was performed using the Apriori algorithm, with a support of 0.1, a confidence of 0.7, and a minimum lift of 1.1. Fifty association rules that met the conditions were mined, such as "IF peristalsis frequency = low AND mucosal function = medium THEN formulation = sustained release", with a support of 0.15 and a confidence of 0.82. Combined with the preset rules, the fuzzy comprehensive evaluation method was used to correct the association strength. The weights of adaptability, safety, and effectiveness were set to 0.5, 0.3, and 0.2, respectively. The corrected association degree was calculated, and the final association degree of the above rule was 0.78. A 1000x50 user digestion feature and health care product formulation adaptability association matrix was constructed, with matrix element values ranging from 0 to 1, representing the degree of adaptability.

[0027] In step S104, according to the association matrix, the key attribute set of different formulations of health care products is obtained, and the collaborative filtering algorithm is used to calculate the acceptance and absorption efficiency scores of different user groups for various formulations of health care products, to obtain a user and formulation scoring matrix, and a digestion threshold is set to determine the groups with weak digestive absorption capacity and the groups with normal digestive function.

[0028] The user digestion feature and health care product formulation adaptability association matrix is obtained, the key attribute set of different formulations of health care products is combined, and a user and formulation feature matrix is constructed. The key attributes of the formulations are used to optimize the collaborative filtering algorithm, and a formulation attribute matrix is constructed according to the release control mechanism, adhesion, and target delivery capacity of the formulations. The user and formulation acceptance and absorption efficiency scores are combined to obtain a user and formulation scoring matrix. A digestion threshold is set, and the initial threshold is determined from the scoring distribution of the historical data. Five-fold cross-validation is used to try different thresholds on the training set and evaluate the classification effect on the validation set, and the threshold with the highest F1 score is selected as the final digestion threshold. If the comprehensive score of a user for a formulation is lower than the digestion threshold, the user is determined to belong to the group with weak digestive absorption capacity. According to the digestion threshold, all users are classified to obtain the user digestion capacity classification results.

[0029] Specifically, the user digestion feature and health product dosage form adaptability correlation matrix is obtained, and the user-dosage form feature matrix is constructed by combining the key attribute set of different dosage form health products. The discrete features such as user digestive function type and dosage form category are processed by One-Hot encoding, and the continuous features such as digestive tract peristalsis frequency and mucosal barrier function index are processed by Min-Max normalization. The user digestion features and dosage form characteristics are converted into numerical feature vectors. For the key attributes of the dosage form, such as release control mechanism, adhesion, and targeted delivery capability, expert scoring method is used to quantify them into numerical values between 0 and 10, which are integrated into the feature vector to obtain the input data set for subsequent calculation. According to the user and dosage form feature matrix, the matrix decomposition collaborative filtering algorithm is used to consider the user's acceptance and absorption efficiency of the dosage form. The user hidden vector U and the dosage form hidden vector V are initialized, and the dimension of the hidden vector is set to 20. The stochastic gradient descent method is used to optimize the objective function, and the objective function includes the prediction error term, the regularization term and the dosage form attribute similarity term. In each iteration, the prediction score is calculated according to the current U and V, and the error is obtained by comparing it with the actual score, and then U and V are updated. The number of iterations is set to 100 or the convergence condition is reached to stop. The product of U and V is used to predict the comprehensive score of the user for the untried dosage form. The key attributes of the dosage form such as release control mechanism, adhesion and targeted delivery capability are used to optimize the collaborative filtering algorithm. The key attributes of the dosage form such as release control mechanism, adhesion and targeted delivery capability are used as additional features to construct the dosage form attribute matrix F. In the matrix decomposition model, the attribute term is introduced, that is, R≈U·V^T+U·F, where R is the user-dosage form score matrix, and V^T is the transpose of the dosage form hidden vector V. The user hidden vector U, the dosage form hidden vector V and the attribute weight matrix are learned by minimizing the new objective function. This method combines collaborative filtering and content-based recommendation, and improves the prediction accuracy of new dosage forms. The user-dosage form score matrix is generated by combining the user-dosage form acceptance score and the absorption efficiency score. The initial threshold is determined by using the score distribution of the historical data. Five-fold cross-validation is used to try different thresholds on the training set and evaluate the classification effect on the validation set, and the threshold with the highest F1 score is selected as the final digestion threshold. F1 score is the average of precision and recall. Precision is the proportion of samples that are actually positive in all samples predicted as positive, and recall is the proportion of samples that are correctly predicted as positive in all samples that are actually positive. If the user's comprehensive score of the dosage form is lower than the threshold, the user is determined to belong to the group with weak absorption capacity of the digestive tract, otherwise the user is determined to belong to the group with normal digestive function. According to the threshold, all users are classified to obtain the user digestion capacity classification result. The 1000 user digestion features and the adaptability correlation matrix of 50 health product dosage forms are obtained, and the 1000*70 user-dosage form feature matrix is constructed by combining the 20 dosage form key attribute set.The user's digestive function type is normal, weak, and strong, which is One-Hot coded, and the three types are converted into [1, 0, 0], [0, 1, 0], and [0, 0, 1] forms. The digestive tract peristalsis frequency, ranging from 3 to 12 times per minute, is Min-Max normalized and mapped to the 0-1 interval. For key attributes of dosage forms, such as release control mechanism 0-10 points, adhesion 0-100%, and targeted delivery ability 0-10 points, expert scoring method is used for quantification and integrated into the feature vector. The collaborative filtering algorithm of matrix decomposition is used to initialize the 1000x20 user hidden vector U and the 50x20 dosage form hidden vector V, and the hidden vector dimension is set to 20. The stochastic gradient descent method is used to optimize the objective function, the learning rate is set to 0.01, and the regularization parameter is set to 0.1. Iteration 100 times, in each iteration, the mean square error of the predicted score and the actual score is calculated, and when the error changes less than 0.0001, it is stopped in advance. Through the product of U and V, the comprehensive score of the user to the untried dosage form is predicted, and a 1000x50 prediction score matrix is obtained. The 20 key attributes of the dosage form are constructed into a 50x20 dosage form attribute matrix F, and the attribute term is introduced into the matrix decomposition model, i.e. R≈U·V^T+U·F, where R is a 1000x50 user-dosage form score matrix. The new objective function is minimized, and the user hidden vector U, the dosage form hidden vector V, and the 20x20 attribute weight matrix are learned. Based on the score distribution of the historical data, the initial digestion threshold is set to 6.5 points, and the full score is 10 points. Five-fold cross-validation is used, and different thresholds are tried in the range of 5.5-7.5 points with a step of 0.1, and the highest F1 score of 6.8 points is selected as the final digestion threshold. 1000 users are classified, and the users with a comprehensive score lower than 6.8 points are determined as the weak absorption group of the digestive tract, and the users with a comprehensive score higher than 6.8 points are determined as the normal group of digestive function, and finally the classification results of 620 normal groups and 380 weak absorption groups of the digestive tract are obtained.

[0030] In step S105, the user searches for health product information, and the user's digestive condition data is obtained, and according to the set digestion threshold, it is judged whether the user belongs to the weak absorption group of the digestive tract or the normal group of digestive function.

[0031] The health product information searched by the user is acquired, a TF-IDF algorithm is used to extract feature words from the health product information, the TF-IDF values of the feature words are calculated, and the first preset feature words with the highest TF-IDF values are selected as key features; according to the key features, the appearance frequencies of past user search words are counted in combination with the historical search behaviors of the user, the first preset words with the highest appearance frequencies are selected to form a user health product interest vector; user digestion data is acquired, the digestion enzyme activity, intestinal permeability and gastrointestinal motility indexes in the digestion data are processed by Z-score standardization, and a user digestion feature vector is generated; according to the user digestion feature vector, a user digestion capacity classification model is constructed by using a decision tree algorithm, and the decision tree algorithm uses information gain ratio as a feature selection and splitting standard; if the user digestion feature vector meets the condition of weak digestive tract absorption capacity, a health suggestion is generated from a preset suggestion library, and if the user digestion feature vector does not meet the condition of weak digestive tract absorption capacity, a suggestion for maintaining intestinal health is provided.

[0032] Specifically, the health product information searched by the user and the user's digestion data are obtained, the TF-IDF algorithm is used to extract the health product related feature words from the search information, the TF-IDF value of each word is calculated, and the top 10 words with the highest weight are selected as the key features. Combined with the user's historical search behavior, the word frequency statistical method is used to construct the user's health product interest vector, the frequency of the user's search words in the past 30 days is counted, and the top 20 words with the highest frequency are selected to form the interest vector. At the same time, the user's digestion data is standardized, the digestion enzyme activity, intestinal permeability, gastrointestinal motility and other indicators are standardized by Z-score, and the user's digestion feature vector is generated. According to the pre-set digestion threshold, the C4.5 decision tree algorithm is used to construct the user's digestion capacity classification model. The information gain ratio is used as the feature selection and splitting standard. Starting from the root node, the information gain ratio of each feature is calculated, and the feature with the highest gain ratio is selected as the splitting node. The minimum sample size is set to 50, and the maximum tree depth is set to 5. The decision tree is constructed by recursive method. The pruning technique is used to avoid overfitting, the pessimistic pruning method is used, and the pruning effect of each non-leaf node is evaluated from bottom to top. When the error rate after pruning does not exceed the error rate before pruning plus a standard deviation, the pruning operation is performed. Finally, the decision rule set for judging the user's digestion capacity is obtained. The constructed decision rule set is used to classify and judge the user's digestion feature vector. The threshold values of multiple digestion indicators are set, including the digestion enzyme activity being less than 50 U / L, the intestinal permeability index being higher than 0.05, and the gastrointestinal motility index being less than 70 points. Starting from the root node of the decision tree, the user's feature value is judged level by level until the leaf node is reached. If the conditions of weak absorption capacity of the digestive tract are met, that is, at least two indicators exceed the threshold value, the user is marked as a weak absorption group, otherwise, the user is marked as a normal digestion function group, and the user's digestion capacity classification result is obtained. Based on the user's digestion capacity classification result, the corresponding health suggestions and dietary guidance are generated. For the group with weak absorption capacity of the digestive tract, suitable contents are selected from the pre-set suggestion library, including suggestions such as increasing the intake of easily digestible food, adjusting the dietary structure, and controlling the eating speed. For the group with normal digestion function, general suggestions for maintaining intestinal health are provided. At the same time, according to the user's health product interest vector, the health product use suggestions related to the user's interest are selected from the suggestion library, such as selecting suitable dosage forms and reasonably arranging the taking time, to form an individualized health guidance scheme. For a user searching for "probiotic capsules" information, the TF-IDF algorithm is used to extract keywords such as "probiotics", "intestinal tract", and "digestion" with TF-IDF values of 0.8, 0.6, and 0.5 respectively, and these are selected as the feature words. By analyzing the user's search history in the past 30 days, it is found that "probiotics" appears 50 times, "vitamins" appear 30 times, and "protein powder" appears 20 times, and the interest vector is constructed.At the same time, user digestive data was collected, including digestive enzyme activity (45 U / L), intestinal permeability index (0.06), and gastrointestinal motility index (65 points). After Z-score standardization, these values ​​were -1.2, 1.5, and -0.8, respectively. A classification model was constructed using the C4.5 decision tree algorithm, using the standardized digestive metrics as input. The information gain ratio threshold was set at 0.1, the minimum sample size was set at 50, and the maximum tree depth was set at 5. After training, the decision rule was derived: if digestive enzyme activity < -1.0 and intestinal permeability > 1.0, the user was classified as a poor absorber. This rule was applied to the user, and the criteria were met, so the user was labeled as a poor absorber. Based on the classification results, relevant suggestions were selected from the suggestion library, such as "Choose a probiotic powder that is easily digestible and absorbable, taken 2-3 times daily" and "Increase intake of foods rich in dietary fiber to improve the intestinal environment." Based on the user's interest in probiotics, the additional suggestion "Choose a combination probiotic preparation containing lactic acid bacteria and bifidobacteria, which can help improve intestinal flora balance" was added.

[0033] Step S106: If the user belongs to a group with weak digestive tract absorption capacity, then based on the health product information searched by the user and the user and dosage form rating matrix, liquid and soft capsule health products with high solubility, small particle size, and large surface area are recommended first; if the user belongs to a group with normal digestive function, then based on the tolerance of the digestive tract to specific ingredients, various dosage forms of health products suitable for the user's preferences are recommended to obtain the final recommendation strategy.

[0034] Obtain the user's digestive tract absorption capacity classification results, health product information, and user and dosage form rating matrix, where the health product information is obtained by user search; based on the health product information, use the TF-IDF algorithm to extract health product keywords and attributes, and construct a health product feature vector containing dissolution, particle size, and surface area attribute values; judge the user's digestive tract absorption capacity classification results. If the user belongs to a group with weak digestive tract absorption capacity, set dissolution, particle size, and surface area thresholds to screen out qualified liquid and soft capsule health products; calculate the similarity between users based on the user and dosage form rating matrix, select the user with the highest similarity to construct a neighbor set, combine the rating data of neighbor users and the current user's historical rating, and use the weighted average method to predict the user's preference score for each dosage form; obtain the user's digestive tract tolerance data for specific ingredients, and use a weighted scoring method to rank the candidate health products, where the weighted scoring method includes product attribute matching, user preference score, and tolerance factor, calculate the comprehensive score of each candidate health product, and select the health product with the highest comprehensive score as the final recommendation list.

[0035] Specifically, the user's digestive tract absorption capacity classification result, the searched health product information, and the user-dosage form scoring matrix are obtained, the TF-IDF algorithm is used to extract the health product keywords and attributes from the search information to construct a health product feature vector, including dissolution, particle size, surface area and other attribute values, and the user's scoring data for different dosage forms is extracted from the user-dosage form scoring matrix. A user-component tolerance matrix is established, and the tolerance data is quantified as a score of 0-1 for subsequent recommendation process. According to the user's digestive tract absorption capacity classification result, conditional judgment is performed. If the user belongs to the weak digestive tract absorption capacity group, specific threshold values including dissolution greater than 80%, particle size less than 100 microns, and surface area greater than 1 square meter per gram are set. Liquid and soft capsule health products meeting these conditions are screened to form a preliminary recommendation list. For such users, the physical and chemical properties of the product are given priority to ensure better absorption. If the user belongs to the normal digestive function group, a user-based collaborative filtering algorithm is used. The similarity between users is calculated based on the user-dosage form scoring matrix, and the cosine similarity formula is used for calculation. The top 10 users with the highest similarity are selected to construct a neighbor set. The neighbor user's scoring data and the current user's historical scoring are combined, and the weighted average method is used to predict the user's preference score for each type of dosage form. Combined with the user's digestive tract tolerance data for specific components, a weighted scoring method is used to sort the candidate health products. The weight distribution is set as product attribute matching degree weight 0.4, user preference score weight 0.4, and tolerance factor weight 0.2. The comprehensive score of each candidate health product is calculated, considering product characteristics, user preferences, and tolerance. The top 5 health products with the highest comprehensive score are selected as the final recommendation list, ensuring that the recommendation result meets the user's digestive capacity and preferences, and considering the tolerance to specific components, obtaining personalized recommendation strategies for users with different digestive capacities. For a user searching for "probiotic soft capsules" information, the TF-IDF algorithm is used to extract keywords such as "probiotic", "soft capsule", and "intestinal tract" with TF-IDF values of 0.8, 0.7, and 0.5, respectively. A health product feature vector is constructed, including dissolution of 85%, particle size of 80 microns, and surface area of 1.2 square meters per gram. The user's score for soft capsules is extracted from the user-dosage form scoring matrix, which is 4.5 out of 5. A user-component tolerance matrix is established, and the user's tolerance score for probiotics is 0.9. It is determined that the user belongs to the weak digestive tract absorption capacity group, so the physical and chemical properties are given priority. Soft capsule health products meeting the conditions, dissolution > 80%, particle size < 100 microns, and surface area > 1 square meter per gram, are screened to form a preliminary recommendation list, containing 5 products. A weighted scoring method is used for final sorting, with weights set as product attribute matching degree 0.4, user preference score 0.4, and tolerance factor 0.2.The comprehensive score is calculated, such as the score of a probiotic soft capsule is 0.4*0.95 (attribute matching degree) + 0.4*0.9 (user preference) + 0.2*0.9 (tolerance) = 0.92. Finally, the three products with the highest comprehensive score are selected as the recommended results, forming the personalized recommendation strategy for the user with weak digestive capacity.

[0036] In step S107, the feedback of the user to the recommended health products is continuously tracked, the absorption efficiency evaluation and the satisfaction score of the user are obtained, the recommendation strategy is dynamically optimized, the matching degree of the health product dosage form and the digestive characteristics of the user is improved, and the personalized and precise health care scheme is provided for the user.

[0037] The feedback data carrying the user identification number is received, the user feedback feature vector is generated according to the feedback data, the user feedback feature vector includes objective physiological indicators and subjective evaluation data; a dynamic optimization model is constructed by using a Q-learning reinforcement learning algorithm, the dynamic optimization model is trained according to the user feedback feature vector; a Q value table is updated, the Q value table is obtained according to the training result of the dynamic optimization model; each time the formula Q(s, a) = Q(s, a) + a[r + g*max(Q(s', a')) - Q(s, a)] is used, wherein Q(s, a) is the Q value of the action a in the state s, a is the action taken in the current state, a is the learning rate, r is the immediate reward, g is the discount factor, s' is the new state reached by the agent after executing the action a, a' is the action that can be taken in the next state s', max(Q(s', a')) is the Q value of the action a' with the highest Q value in the next state s'; it is judged whether the Q value table meets the adjustment condition, including: querying the recommendation strategy database; wherein the recommendation strategy database includes strategy data with a recommendation strategy identification number; it is judged whether there is a recommendation strategy identification number matched with the Q value table in the recommendation strategy database; if there is a recommendation strategy identification number matched with the Q value table in the recommendation strategy database, the existing recommendation strategy is adjusted; a personalized and precise health care scheme is generated according to the adjusted recommendation strategy, the personalized and precise health care scheme includes a recommended health product combination, a dosage and a time arrangement; push information carrying the personalized and precise health care scheme is sent, the push information is sent by the system to the user terminal through a mobile application.

[0038] Specifically, the feedback data of users on the recommended health products is obtained, including absorption efficiency evaluation and satisfaction score. The physiological indicators that can be obtained, such as heart rate, blood pressure and sleep quality of users are monitored in real time by using a smart bracelet, and the subjective scores filled in by users through a mobile application every day, such as digestive condition and energy level, are combined. These data are integrated into a user feedback feature vector, including two parts of objective physiological indicators and subjective evaluation. A weekly feedback reminder is set to ensure that the user data is continuously obtained. According to the user feedback feature vector, a dynamic optimization model is constructed by using a Q-learning reinforcement learning algorithm. The state space is defined as the user's digestive features, such as the digestive capacity index, the intestinal health score and the current recommended health product combination. The action space includes adjusting the dosage form, adjusting the dosage, replacing the health product category and the like. The reward function is set as the weighted sum of the absorption efficiency score with a weight of 0.6 and the satisfaction score with a weight of 0.4. The Q value table is initialized, and a random value is given to each state-action pair. The Q value is updated by iteration, and the formula

[0039] Q(s,a) = Q(s,a) + a[r + g*max(Q(s',a')) - Q(s,a)], where Q(s,a) is the Q-value of taking action a in state s, i.e., the expected long-term reward that can be obtained by taking a certain action in the current state, s is the current state, referring to the position or condition of the agent in the environment, a is the action taken in the current state, a is the learning rate, determining the extent to which new information affects old Q-values, the value range of a is usually (0,1], if a = 1, then the new information will completely replace the old information; if a is close to 0, the impact of new information is small, r is the immediate reward, the reward signal received by the agent from the environment after executing action a, this reward can be positive or negative, depending on the goodness of the action, g is the discount factor, used to measure the importance of future rewards, the value range of g is [0,1], if g tends to 0, the agent will only consider the immediate reward; if g is close to 1, the agent will consider future rewards more, s' is the next state, i.e., the new state reached by the agent after executing action a, a' is the action that can be taken in the next state s', max(Q(s',a')) is the Q-value of action a' with the highest Q-value among all possible actions in the next state s'. Adjust the existing recommendation strategy using the updated Q-value table. For different combinations of user digestion characteristics and health product dosage forms, calculate the matching score. Use cosine similarity to calculate the similarity between the user feature vector and the health product feature vector to obtain the matching score. Combine Q-value and matching score, and use weighted summation method to consider absorption efficiency, satisfaction and dosage form characteristics, and generate an optimized recommendation list. The matching score weight is set to 0.3, and the Q-value weight is set to 0.7. Based on the optimized recommendation list, combined with user historical data and current digestion status, generate a personalized and precise health care plan. Select the action with the highest Q-value in the current state from the Q-value table to determine the recommended health product combination, dosage and time arrangement. Set the feedback monitoring period to one week, and update the user state and Q-value table every week. Push the personalized plan to the user through the mobile application, and remind the user to feedback in time. Continuously collect user data, retrain the Q-learning model regularly, such as every month, to realize dynamic optimization of the recommendation strategy. For a user taking probiotic soft capsules, the smart bracelet collects data such as average heart rate 70 times / min, blood pressure 120 / 80 mmHg, and sleep quality 7 hours / night. The user fills in the digestion status, 8 / 10, and energy level, 7 / 10, through the mobile application every day. These data are integrated into the user feedback feature vector [70,120,80,7,8,7]. Define the state space as the user's digestion capacity index 0.8 and the currently recommended probiotic soft capsules, 1 billion live bacteria per particle. The action space includes adjusting the dosage ±20% and replacing the dosage form, tablet / powder. The reward function is set to 0.6x8 + 0.4x7 = 7.6.The initial Q-value table is randomly assigned between -1 and 1. After 100 iterations of updating, the learning rate a = 0.1, and the discount factor g = 0.9, the Q-value table converges. The cosine similarity between the user feature vector [0.8, 10] and the health product feature vector [0.9, 12] is 0.98. The optimal recommendation is to increase the dosage by 20% by combining the Q-value, which has a weight of 0.7, and the matching degree score, which has a weight of 0.3. The personalized plan is generated as probiotic soft capsules, 1.2 billion live bacteria per capsule, twice a day. Set a feedback reminder every week, continuously collect data, and update the Q-learning model every month to achieve dynamic optimization.

[0040] [0.9,12] is 0.98. The optimal recommendation is to increase the dosage by 20% by combining the Q-value, which has a weight of 0.7, and the matching degree score, which has a weight of 0.3. The personalized plan is generated as probiotic soft capsules, 1.2 billion live bacteria per capsule, twice a day. Set a feedback reminder every week, continuously collect data, and update the Q-learning model every month to achieve dynamic optimization.

[0041] The above embodiment is only one of the preferred embodiments of the present application and should not be used to limit the protection scope of the present application, but any modification or embellishment made without substantial meaning within the main design idea and spirit of the present application, which still solves the technical problems consistent with the present application, should be included in the protection scope of the present application.

Claims

1. A method for pushing item information based on item attributes and user preferences, characterized in that: The method includes: obtaining digestive system disease types, digestive tract absorption capacity, digestive fluid secretion and physiological characteristic data related to digestive function of multiple users, and grouping the users by clustering algorithm to obtain user groups with different digestive function characteristics; Obtain information on the dosage form, particle size, and surface area of ​​health products. Based on the disintegration time, solubility, and osmotic pressure characteristics of the dosage form, principal component analysis is used to reduce the dimensionality of health products to obtain a set of key attributes for health products of different dosage forms. Analyze the characteristics of digestive tract motility frequency, mucosal barrier function, and bacterial balance for different digestive function groups. Combined with the adhesiveness, release control mechanism, and targeted delivery properties of health product dosage forms, a correlation matrix between user digestive characteristics and health product dosage form compatibility is constructed. Based on the association matrix, the key attribute sets of health products with different dosage forms were obtained. The collaborative filtering algorithm was used to calculate the acceptance and absorption efficiency scores of different user groups for various dosage forms of health products, and the user and dosage form score matrix was obtained. A digestion threshold was set to judge the groups with weak digestive tract absorption capacity and normal digestive function groups, including: obtaining the digestive characteristics of 1,000 users and the adaptability association matrix of 50 health product dosage forms, combining the 20 dosage form key attribute sets, One-Hot encoding the user digestive function types of normal, weak, and strong, and converting the three types into the form of [1,0,0], [0,1,0], and [0,0,1]; Min-Max normalization was performed on the digestive tract peristalsis frequency and mapped to the range of 0-1; the expert scoring method was used to quantify the key attributes of the dosage form and integrate them into the feature vector; the collaborative filtering algorithm of matrix decomposition was used to initialize the 1000×20 user latent vector U and the 50×20 dosage form latent vector V, and the latent vector dimension was set to 20; the stochastic gradient descent method was used to optimize the objective function, and the learning rate was set to 0.

01. The regularization parameter was set to 0.1; the model was iterated 100 times, and the mean square error between the predicted score and the actual score was calculated in each iteration. The model was stopped early when the error change was less than 0.0001. The user's comprehensive score for the untried dosage form was predicted by multiplying U and V to obtain a 1000×50 prediction score matrix. The 20 key attributes of the dosage form were constructed into a 50×20 dosage form attribute matrix F. The attribute terms were introduced into the matrix decomposition model, that is, R≈U·V^T+U·F, where R is a 1000×50 user-dosage form score matrix. , minimize the new objective function, and simultaneously learn the user latent vector U, dosage form latent vector V, and a 20×20 attribute weight matrix; the initial digestion threshold is set to 6.5 points, with a full score of 10 points. Using five-fold cross-validation, different thresholds are tried in the range of 5.5-7.5 points with a step size of 0.1, and the highest F1 score of 6.8 points is selected as the final digestion threshold; 1,000 users are classified, and users with a comprehensive score below 6.8 points are judged to have weak digestive tract absorption capacity, while those with a comprehensive score above 6.8 points are judged to have normal digestive function; Obtain the health product information that the user searches for, as well as the user's digestion data. Based on the set digestion threshold, determine whether the user belongs to the group with weak digestive tract absorption capacity or the group with normal digestive function; If the user has a weak digestive tract absorption capacity, then based on the health product information the user searches for and the user and dosage form rating matrix, priority is given to recommending liquid and soft capsule health products with high solubility, small particle size, and large surface area. If the user has normal digestive function, then based on their digestive tract's tolerance to specific ingredients, various dosage forms of health products that suit their preferences are recommended, resulting in the final recommendation strategy. Continuously track user feedback on recommended health products, obtain user absorption efficiency evaluation and satisfaction scores, dynamically optimize recommendation strategies, improve the matching degree between health product dosage forms and user digestive characteristics, and provide users with personalized and precise health care plans.

2. The method according to claim 1, characterized in that The method comprises obtaining the digestive system disease types, digestive tract absorption capacity, digestive fluid secretion and physiological characteristic data related to digestive function of multiple users, grouping the users by a clustering algorithm, and obtaining user groups with different digestive function characteristics, including: obtaining the user's digestive system disease types, digestive tract absorption capacity and digestive fluid secretion and physiological characteristic data, cleaning and preprocessing the original data to obtain a standardized user feature vector; analyzing the user feature vector by a clustering algorithm, and determining the optimal number of clusters according to the silhouette coefficient; if the silhouette coefficient reaches a maximum value, determining the corresponding number of clusters as the optimal value; based on the preliminary grouping results, performing a secondary division on the boundary users in combination with a hierarchical clustering method; calculating the similarity of user characteristics within the group by an AGNES algorithm, and measuring the similarity by cosine similarity; excluding the user if the similarity between users is lower than a set threshold; performing a statistical analysis on the distribution of digestive system disease types, digestive tract absorption capacity levels, digestive fluid secretion and physiological characteristic data of each group of users according to the clustering result labels; calculating the proportion of different digestive system disease types, the mean and standard deviation of digestive tract absorption capacity, the median and interquartile range of digestive fluid secretion, and the distribution of main physiological characteristics in each group; Based on the statistical results, a set of rules for distinguishing the characteristics of user digestion functions is constructed; new users are screened step by step through the set of rules to determine the group to which the new users belong.

3. The method according to claim 1, characterized in that The method comprises obtaining the dosage form, particle size, and surface area attribute information of the health care product, and performing dimensionality reduction on the health care product according to the disintegration time, dissolution rate, and osmotic pressure characteristics of the dosage form using a principal component analysis method to obtain a set of key attributes of health care products with different dosage forms, including: obtaining dosage form information and physicochemical attribute data of the health care product, the physicochemical attribute data including particle size data, surface area attributes, disintegration time measurement results, dissolution curve data, and osmotic pressure measurement values; detecting outliers in the physicochemical attribute data using a Z-score method, and marking the data point as an outlier if the absolute value of the Z-score of the data point is greater than a preset threshold; filling missing values ​​in the physicochemical attribute data with the mean value of the corresponding attribute; and normalizing continuous data in the physicochemical attribute data to obtain a standardized value. Perform one-hot encoding conversion on the dosage form information to obtain encoded dosage form data; generate a standardized health product feature matrix based on the standardized values ​​and the encoded dosage form data; calculate the Pearson correlation coefficient between each attribute in the standardized health product feature matrix to obtain a correlation coefficient matrix; According to the correlation coefficient matrix, determine the strength of the correlation between attributes; If the absolute value of the correlation coefficient is greater than the preset threshold, it is determined to be significantly correlated; one of the significantly correlated attribute pairs is selected as the input variable of the principal component analysis to obtain the principal component analysis input variable set; the principal component analysis method is used to reduce the dimension of the principal component analysis input variable set; Calculate the covariance matrix of the principal component analysis input variable set and solve the eigenvalues ​​and eigenvectors of the covariance matrix; arrange the eigenvalues ​​in descending order and calculate the contribution rate and cumulative contribution rate of each eigenvalue; select the eigenvectors whose cumulative contribution rate reaches the preset threshold as the principal components to obtain the characteristic representation of health products after dimensionality reduction; explain the physical meaning of each principal component in the characteristic representation of health products after dimensionality reduction and determine the key attributes corresponding to each principal component; analyze the distribution of different dosage forms on the principal components and calculate the mean and variance of each dosage form on the main principal components; based on the mean and variance, extract the attributes with significant differences between groups to obtain the key attribute set of health products with different dosage forms.

4. The method according to claim 1, wherein The method analyzes the characteristics of digestive tract motility frequency, mucosal barrier function, and microbial balance state for different digestive function groups, and constructs a correlation matrix between user digestive characteristics and health product dosage form compatibility in combination with the adhesion, release control mechanism, and targeted delivery ability properties of the health product dosage form, including: obtaining digestive tract motility frequency data, mucosal barrier function indicators, and microbial balance state parameters for different digestive function groups, processing the digestive tract motility frequency data using a Z-score standardization method, normalizing the mucosal barrier function indicators, and performing logarithmic transformation on the microbial balance state parameters before standardization; constructing a preliminary mapping relationship between user digestive characteristics and health product dosage form attributes based on the standardized data, setting association rules between digestive tract motility frequency and dosage form release control mechanism, association rules between mucosal barrier function and dosage form adhesion, and association rules between microbial balance state and targeted delivery ability; and performing dimensionality reduction processing on the adhesion experimental data, release control mechanism test results, and targeted delivery ability evaluation indicators of the health product dosage form using principal component analysis, and selecting the principal component whose cumulative contribution rate reaches a preset threshold as the feature vector; The K-means clustering algorithm was used to cluster the feature vectors to obtain the clustering results of the dosage form attributes. Based on the clustering results of the dosage form attributes, the Apriori algorithm was used to mine association rules. With the user digestion characteristics as the antecedent and the dosage form attribute clustering results as the consequent, support thresholds, confidence thresholds, and minimum lift were set to mine the association rule sets that met the threshold conditions and generate a preliminary association matrix. For the preliminary association matrix, combined with the preset digestive characteristics and dosage form adaptability rules in the knowledge base, the fuzzy comprehensive evaluation method was used to correct the association strength, and adaptability, safety and effectiveness were selected as evaluation indicators. The hierarchical analysis method was used to set the weights, and the corrected association degree was calculated to construct the final user digestive characteristics and health product dosage form adaptability association matrix.

5. The method according to claim 1, wherein The method comprises: obtaining a correlation matrix based on a set of key attributes of health products with different dosage forms, using a collaborative filtering algorithm to calculate the acceptance and absorption efficiency scores of different user groups for various dosage forms of health products, obtaining a user-dosage form score matrix, and setting a digestion threshold for judging groups with weak digestive tract absorption capacity and groups with normal digestive function, including: obtaining a correlation matrix between user digestion characteristics and health product dosage form adaptability, combining a set of key attributes of health products with different dosage forms to construct a user-dosage form characteristic matrix; optimizing the collaborative filtering algorithm using the key attributes of the dosage form, and constructing a dosage form attribute matrix based on the release control mechanism, adhesion, and targeted delivery capability of the dosage form; combining the user dosage form acceptance score and the absorption efficiency score to obtain a user-dosage form score matrix; setting a digestion threshold, and determining an initial threshold from the score distribution of historical data; using five-fold cross validation to try different thresholds on the training set, evaluating the classification effect on the validation set, and selecting the threshold with the highest F1 score as the final digestion threshold; if the user's comprehensive score for the dosage form is lower than the digestion threshold, determining that the user belongs to the group with weak digestive tract absorption capacity; and classifying all users according to the digestion threshold to obtain a user digestion capacity classification result.

6. The method according to claim 1, characterized in that The obtaining of the health product information searched by the user and the user's digestion data, and judging whether the user belongs to a group with weak digestive tract absorption capacity or a group with normal digestive function based on a set digestion threshold, includes: obtaining the health product information searched by the user, extracting feature words from the health product information using a TF-IDF algorithm, calculating the TF-IDF values ​​of the feature words, and selecting the first preset feature words with the highest TF-IDF values ​​as key features; Based on key features and combined with the user's historical search behavior, the frequency of occurrence of past user search terms is counted, and the top preset terms with the highest frequency are selected to form the user's health product interest vector; the user's digestion data is obtained, and the digestive enzyme activity, intestinal permeability and gastrointestinal motility indicators in the digestion data are normalized using Z-score to generate a user digestion feature vector; based on the user's digestion feature vector, a decision tree algorithm is used to construct a user digestion ability classification model, and the decision tree algorithm uses information gain ratio as the feature selection and splitting criterion; if the user's digestion feature vector meets the condition of weak digestive tract absorption capacity, the corresponding content is selected from the preset suggestion library to generate health suggestions. If the user's digestion feature vector does not meet the condition of weak digestive tract absorption capacity, suggestions for maintaining intestinal health are provided.

7. The method according to claim 1, characterized in that If the user belongs to a group with weak digestive tract absorption capacity, then based on the health product information searched by the user and the user and dosage form rating matrix, liquid and soft capsule health products with high solubility, small particle size, and large surface area are preferentially recommended; if the user belongs to a group with normal digestive function, then based on the tolerance of the digestive tract to specific ingredients, various dosage forms of health products suitable for the user's preferences are recommended, and a final recommendation strategy is obtained, including: obtaining the user's digestive tract absorption capacity classification results and health product information and the user and dosage form rating matrix, wherein the health product information is obtained by the user search; using the TF-IDF algorithm to extract health product keywords and attributes based on the health product information, and constructing a health product feature vector containing dissolution, particle size, and surface area attribute values; judging the user's digestive tract absorption capacity classification results, if the user belongs to a group with weak digestive tract absorption capacity, setting dissolution, particle size, and surface area thresholds to screen out liquid and soft capsule health products that meet the conditions; The similarity between users is calculated based on the user and dosage form rating matrix, and the users with the highest similarity are selected to construct a neighbor set. The rating data of neighbor users and the current user's historical rating are combined to use the weighted average method to predict the user's preference score for various dosage forms. The user's digestive tract tolerance data for specific ingredients is obtained, and the candidate health products are ranked using a weighted scoring method. The weighted scoring method includes product attribute matching, user preference score and tolerance factor. The comprehensive score of each candidate health product is calculated, and the health product with the highest comprehensive score is selected as the final recommendation list.

8. The method according to claim 1, characterized in that The method of continuously tracking user feedback on recommended health products, obtaining user absorption efficiency evaluations and satisfaction scores, dynamically optimizing recommendation strategies, improving the matching degree between health product dosage forms and user digestive characteristics, and providing users with personalized and precise health care plans includes: receiving feedback data carrying a user identification number, and generating a user feedback feature vector based on the feedback data, wherein the user feedback feature vector includes objective physiological indicators and subjective evaluation data; The Q-learning reinforcement learning algorithm is used to build a dynamic optimization model, which is trained based on the user feedback feature vector; Update the Q-value table, which is obtained based on the training results of the dynamic optimization model; each update uses the formula Q(s,a)=Q(s,a)+α[r+γ*max(Q(s',a'))-Q(s,a)], where Q(s,a) is the Q-value of taking action a in state s, a is the action taken in the current state, α is the learning rate, r is the immediate reward, γ is the discount factor, s' is the new state reached by the agent after performing action a, a' is the action that can be taken in the next state s', and max(Q(s',a')) is the Q-value of action a' with the highest Q-value in the next state s'; Determining whether the Q-value table meets the adjustment conditions includes: querying a recommendation strategy database; wherein the recommendation strategy database includes strategy data with a recommendation strategy identification number; determining whether there is a recommendation strategy identification number matching the Q-value table in the recommendation strategy database; if there is a recommendation strategy identification number matching the Q-value table in the recommendation strategy database, adjusting the existing recommendation strategy; generating a personalized precision health care plan based on the adjusted recommendation strategy, the personalized precision health care plan including a recommended combination of health products, dosage, and time schedule; sending a push message carrying the personalized precision health care plan, the push message being sent by the system to the user terminal via a mobile application.

Images